dsotofor / Manuscrit

Commit 2a133b0ce0cea4bbc0db077c67eb74dc771ca52b

Authored by dsotofor 2025-07-11 23:12:09 +0200

Exists in main

version all ok, corrections to do in general conclusion...

Showing 9 changed files with 68 additions and 50 deletions Inline Diff

chapters/TS.aux
chapters/TS.tex
main.aux
main.bbl
main.bib
main.blg
main.log
main.pdf
main.synctex.gz

\chapter{Système de Recommandation dans AI-VT}	1	1	\chapter{Système de Recommandation dans AI-VT}
	2	2
\section{Introduction}	3	3	\section{Introduction}
	4	4
%\colorbox{yellow}{Il y a des différences sémantiques à bien différencier pour éviter les confusions : }\\	5	5	%\colorbox{yellow}{Il y a des différences sémantiques à bien différencier pour éviter les confusions : }\\
%\colorbox{yellow}{Il faut harmoniser : à certains moment, on parle de 3 tests, à d'autres de 3 scénarios.}\\	6	6	%\colorbox{yellow}{Il faut harmoniser : à certains moment, on parle de 3 tests, à d'autres de 3 scénarios.}\\
%\colorbox{yellow}{J'ai tout transformé en séances d'entrainement 1, 2, 3, mais je me suis peut-être trompé}\\	7	7	%\colorbox{yellow}{J'ai tout transformé en séances d'entrainement 1, 2, 3, mais je me suis peut-être trompé}\\
%\colorbox{yellow}{S'il s'agit de scénarios, il faut les décrire.}\\	8	8	%\colorbox{yellow}{S'il s'agit de scénarios, il faut les décrire.}\\
%\\	9	9	%\\
%	10	10	%
%\colorbox{yellow}{Idem pour "modèle", "algorithme", "système", "module", "programme"}\\	11	11	%\colorbox{yellow}{Idem pour "modèle", "algorithme", "système", "module", "programme"}\\
%\colorbox{yellow}{Un système est constitué de modules,}\\	12	12	%\colorbox{yellow}{Un système est constitué de modules,}\\
%\colorbox{yellow}{Un module est constitué de programmes, }\\	13	13	%\colorbox{yellow}{Un module est constitué de programmes, }\\
%\colorbox{yellow}{Un programme suit un algorithme,}\\	14	14	%\colorbox{yellow}{Un programme suit un algorithme,}\\
%\colorbox{yellow}{Un algorithme fonctionne selon un modèle}\\	15	15	%\colorbox{yellow}{Un algorithme fonctionne selon un modèle}\\
%\colorbox{yellow}{Il faudrait revoir cette sémantique en particulier dans le chapitre précédent.}	16	16	%\colorbox{yellow}{Il faudrait revoir cette sémantique en particulier dans le chapitre précédent.}
%\\	17	17	%\\
%	18	18	%
%\colorbox{yellow}{Idem pour "métrique", "équation", "mesure", "résultat"}\\	19	19	%\colorbox{yellow}{Idem pour "métrique", "équation", "mesure", "résultat"}\\
%\colorbox{yellow}{Une métrique permet de définir une graduation (comme le mètre ou le litre),}\\	20	20	%\colorbox{yellow}{Une métrique permet de définir une graduation (comme le mètre ou le litre),}\\
%\colorbox{yellow}{Une équation donne la formule mathématique ou la fonction}\\	21	21	%\colorbox{yellow}{Une équation donne la formule mathématique ou la fonction}\\
%\colorbox{yellow}{permettant de calculer un résultat selon une métrique}\\	22	22	%\colorbox{yellow}{permettant de calculer un résultat selon une métrique}\\
%\colorbox{yellow}{Une mesure est la valeur numérique obtenue, résultant de l'application de l'équation}\\	23	23	%\colorbox{yellow}{Une mesure est la valeur numérique obtenue, résultant de l'application de l'équation}\\
%\colorbox{yellow}{Ainsi, on compare les valeurs dans une métrique, valeurs obtenues en appliquant une équation.}\\	24	24	%\colorbox{yellow}{Ainsi, on compare les valeurs dans une métrique, valeurs obtenues en appliquant une équation.}\\
	25	25
L'un des principaux modules d'un EIAH est le système de recommandation visant à trouver les faiblesses et à réviser la séance d'entraînement proposée initialement par celui-ci. Ce type de module permet donc au système de personnaliser les contenus et les exercices en fonction des besoins et des résultats de chacun des apprenants. Certains auteurs n'hésitent pas à considérer que l'efficacité d'un EIAH dans l'acquisition des connaissances et l'adaptation aux différents types d'apprentissage dépend de ce type de module fondé sur la recommandation \cite{Liu2023}.	26	26	L'un des principaux modules d'un EIAH est le système de recommandation visant à trouver les faiblesses et à réviser la séance d'entraînement proposée initialement par celui-ci. Ce type de module permet donc au système de personnaliser les contenus et les exercices en fonction des besoins et des résultats de chacun des apprenants. Certains auteurs n'hésitent pas à considérer que l'efficacité d'un EIAH dans l'acquisition des connaissances et l'adaptation aux différents types d'apprentissage dépend de ce type de module fondé sur la recommandation \cite{Liu2023}.
	27	27
Les systèmes de recommandation dans les environnements d'apprentissage prennent en compte les exigences, les besoins, le profil, les acquis, compétences, les intérêts et l'évolution de l'apprenant pour adapter et recommander des ressources ou des exercices. Dans ces systèmes, l'adaptation peut être de deux types : l'adaptation de la présentation qui montre aux apprenants des ressources d'étude en fonction de leurs faiblesses et/ou l'adaptation du parcours qui change la structure du cours en fonction du niveau et du style d'apprentissage de chaque apprenant \cite{MUANGPRATHUB2020e05227}.	28	28	Les systèmes de recommandation dans les environnements d'apprentissage prennent en compte les exigences, les besoins, le profil, les acquis, compétences, les intérêts et l'évolution de l'apprenant pour adapter et recommander des ressources ou des exercices. Dans ces systèmes, l'adaptation peut être de deux types : l'adaptation de la présentation qui montre aux apprenants des ressources d'étude en fonction de leurs faiblesses et/ou l'adaptation du parcours qui change la structure du cours en fonction du niveau et du style d'apprentissage de chaque apprenant \cite{MUANGPRATHUB2020e05227}.
	29	29
Parmi les algorithmes les plus prometteurs pouvant aider à proposer des recommandations, nous avons identifié l'algorithme d'échantillonnage de Thompson (TS). Il s'agit d'un algorithme probabiliste appartenant à la catégorie des algorithmes d'apprentissage par renforcement. À l'instant $t$, TS choisit l'action $a_t$ d'un ensemble $A$ d'actions possibles, et obtient une récompense pour celle-ci. À $t+1$, une action $a_{t+1}$ est sélectionnée en tenant compte de la récompense précédente. L'objectif consiste à maximiser la récompense. Selon le principe bayésien, cette maximisation itérative est opérée en suivant une distribution de probabilité évoluant à chaque itération. Cette évolution peut être calculée selon la variante de Bernoulli	30	30	Parmi les algorithmes les plus prometteurs pouvant aider à proposer des recommandations, nous avons identifié l'algorithme d'échantillonnage de Thompson (TS). Il s'agit d'un algorithme probabiliste appartenant à la catégorie des algorithmes d'apprentissage par renforcement. À l'instant $t$, TS choisit l'action $a_t$ d'un ensemble $A$ d'actions possibles, et obtient une récompense pour celle-ci. À $t+1$, une action $a_{t+1}$ est sélectionnée en tenant compte de la récompense précédente. L'objectif consiste à maximiser la récompense. Selon le principe bayésien, cette maximisation itérative est opérée en suivant une distribution de probabilité évoluant à chaque itération. Cette évolution peut être calculée selon la variante de Bernoulli
%\colorbox{yellow}{de la loi de bernoulli ?}	31	31	%\colorbox{yellow}{de la loi de bernoulli ?}
où la récompense n'a que deux valeurs possibles 0 ou 1 (échec ou succès), ou selon une distribution \textit{Beta} définie sur l'intervalle $[0, 1]$ et calculée en fonction de deux valeurs $\alpha$ et $\beta$ \cite{9870279}.	32	32	où la récompense n'a que deux valeurs possibles 0 ou 1 (échec ou succès), ou selon une distribution \textit{Beta} définie sur l'intervalle $[0, 1]$ et calculée en fonction de deux valeurs $\alpha$ et $\beta$ \cite{9870279}.
	33	33
Ce chapitre est divisé en trois parties, la première partie présente un algorithme délivrant des recommandations en fonction des résultats produits par l'apprenant en temps réel. Une partie de cette proposition est publiée dans \cite{Soto2}. Cet algorithme permet l'adaptation automatique en temps réel d'une séance prédéterminée dans l'EIAH AI-VT. Nous considérons qu'elle intervient durant la phase de révision du cycle classique du raisonnement à partir de cas (RàPC). L'algorithme proposé est stochastique et il a été testé selon trois scénarios différents. Les résultats montrent de quelle manière AI-VT peut proposer des recommandations pertinentes selon les faiblesses identifiées de l'apprenant.	34	34	Ce chapitre est divisé en trois parties, la première partie présente un algorithme délivrant des recommandations en fonction des résultats produits par l'apprenant en temps réel. Une partie de cette proposition est publiée dans \cite{Soto2}. Cet algorithme permet l'adaptation automatique en temps réel d'une séance prédéterminée dans l'EIAH AI-VT. Nous considérons qu'elle intervient durant la phase de révision du cycle classique du raisonnement à partir de cas (RàPC). L'algorithme proposé est stochastique et il a été testé selon trois scénarios différents. Les résultats montrent de quelle manière AI-VT peut proposer des recommandations pertinentes selon les faiblesses identifiées de l'apprenant.
	35	35
La deuxième partie de ce chapitre montre l'intégration de tous les algorithmes présentés dans les chapitres précédents à AI-VT. L'algorithme intégré est appliqué au système AI-VT sur des données générées et des données réelles. Plusieurs types de test sont exécutés pour montrer que l'algorithme final permet en effet d'améliorer les capacités d'identification et d'adaptation. Les performances de ce nouveau algorithme sont comparées à celles d'autres algorithmes. Enfin, l'évolution de l'acquisition des connaissances induites par ces nouveaux algorithmes de recommandation stochastiques est analysée dans cette deuxième partie du présent chapitre.	36	36	La deuxième partie de ce chapitre montre l'intégration de tous les algorithmes présentés dans les chapitres précédents à AI-VT. L'algorithme intégré est appliqué au système AI-VT sur des données générées et des données réelles. Plusieurs types de test sont exécutés pour montrer que l'algorithme final permet en effet d'améliorer les capacités d'identification et d'adaptation. Les performances de ce nouveau algorithme sont comparées à celles d'autres algorithmes. Enfin, l'évolution de l'acquisition des connaissances induites par ces nouveaux algorithmes de recommandation stochastiques est analysée dans cette deuxième partie du présent chapitre.
	37	37
Pour terminer, dans la troisième partie de ce chapitre, nous présentons une évolution de ce système de recommandation intégrant le processus de Hawkes. L'intérêt de ce dernier réside dans le fait qu'il utilise une courbe d'oubli, nous permettant ainsi de tenir compte du fait que certaines connaissances et certains mécanismes doivent être rappelés aux apprenants. Cette troisième partie intègre une étude des performances du système de recommandation incluant ce processus stochastique de Hawkes.	38	38	Pour terminer, dans la troisième partie de ce chapitre, nous présentons une évolution de ce système de recommandation intégrant le processus de Hawkes. L'intérêt de ce dernier réside dans le fait qu'il utilise une courbe d'oubli, nous permettant ainsi de tenir compte du fait que certaines connaissances et certains mécanismes doivent être rappelés aux apprenants. Cette troisième partie intègre une étude des performances du système de recommandation incluant ce processus stochastique de Hawkes.
	39	39
\section{Système de recommandation stochastique fondé sur l'échantillonnage de Thompson}	40	40	\section{Système de recommandation stochastique fondé sur l'échantillonnage de Thompson}
\sectionmark{Système de recommandation fondé sur TS}	41	41	\sectionmark{Système de recommandation fondé sur TS}
	42	42
\subsection{Algorithme Proposé}	43	43	\subsection{Algorithme Proposé}
	44	44
L'algorithme proposé, en tant que système de recommandation, prend en compte les notes antérieures des apprenants pour estimer leurs connaissances et leur maîtrise des différentes compétences, sous-compétences et niveaux de complexité au sein du système AI-VT. Puis il adapte les séances pour maximiser l'acquisition des connaissances et la maîtrise des différents domaines contenus dans la même compétence définie.	45	45	L'algorithme proposé, en tant que système de recommandation, prend en compte les notes antérieures des apprenants pour estimer leurs connaissances et leur maîtrise des différentes compétences, sous-compétences et niveaux de complexité au sein du système AI-VT. Puis il adapte les séances pour maximiser l'acquisition des connaissances et la maîtrise des différents domaines contenus dans la même compétence définie.
	46	46
La famille de distributions de probabilité Beta est utilisée pour définir dynamiquement le niveau de complexité (équation \ref{eqBeta}) à proposer à l'apprenant. Cet algorithme permet de recommander des niveaux de complexité non contigus et dans lesquels des lacunes ont été détectées. Les paramètres initiaux des distributions de probabilité peuvent forcer le système à recommander des niveaux de complexité contigus (juste inférieur ou supérieur).	47	47	La famille de distributions de probabilité Beta est utilisée pour définir dynamiquement le niveau de complexité (équation \ref{eqBeta}) à proposer à l'apprenant. Cet algorithme permet de recommander des niveaux de complexité non contigus et dans lesquels des lacunes ont été détectées. Les paramètres initiaux des distributions de probabilité peuvent forcer le système à recommander des niveaux de complexité contigus (juste inférieur ou supérieur).
	48	48
\begin{equation}	49	49	\begin{equation}
B(x, \alpha, \beta) =	50	50	B(x, \alpha, \beta) =
\begin{cases}	51	51	\begin{cases}
\frac{x^{\alpha-1}(1-x)^{\beta - 1}}{\int_0^1 u^{\alpha - 1}(1-u)^{\beta - 1}du} & si \; x \in [0, 1] \\	52	52	\frac{x^{\alpha-1}(1-x)^{\beta - 1}}{\int_0^1 u^{\alpha - 1}(1-u)^{\beta - 1}du} & si \; x \in [0, 1] \\
0&sinon	53	53	0&sinon
\end{cases}	54	54	\end{cases}
\label{eqBeta}	55	55	\label{eqBeta}
\end{equation}	56	56	\end{equation}
	57	57
\begin{table}[!ht]	58	58	\begin{table}[!ht]
\centering	59	59	\centering
\begin{tabular}{ccc}	60	60	\begin{tabular}{ccc}
ID&Description&Domaine\\	61	61	ID&Description&Domaine\\
\hline	62	62	\hline
$c_n$&Niveaux de complexité&$\mathbb{N} \; \| \; c_n>0$\\	63	63	$c_n$&Niveaux de complexité&$\mathbb{N} \; \| \; c_n>0$\\
$g_m$&Valeur maximale dans l'échelle des notes& $\mathbb{N} \;\|\; g_m>0$ \\	64	64	$g_m$&Valeur maximale dans l'échelle des notes& $\mathbb{N} \;\|\; g_m>0$ \\
$g_t$&Seuil de notation &$(0, g_m) \in \mathbb{R}$\\	65	65	$g_t$&Seuil de notation &$(0, g_m) \in \mathbb{R}$\\
$s$&Nombre de parcours définis&$\mathbb{N} \; \| \; s>0$\\	66	66	$s$&Nombre de parcours définis&$\mathbb{N} \; \| \; s>0$\\
$s_c$&Parcours courant fixe défini&$[1, s] \in \mathbb{N}$\\	67	67	$s_c$&Parcours courant fixe défini&$[1, s] \in \mathbb{N}$\\
$\Delta s$&Pas pour les paramètres de la distribution bêta dans le parcours $s$ &$(0,1) \in \mathbb{R}$\\	68	68	$\Delta s$&Pas pour les paramètres de la distribution bêta dans le parcours $s$ &$(0,1) \in \mathbb{R}$\\
$t_m$&Valeur maximale du temps de réponse&$\mathbb{R} \; \| \; t_m>0$\\	69	69	$t_m$&Valeur maximale du temps de réponse&$\mathbb{R} \; \| \; t_m>0$\\
$g_{c}$&Note de l'apprenant à une question de complexité $c$&$[0, g_m] \in \mathbb{R}$\\	70	70	$g_{c}$&Note de l'apprenant à une question de complexité $c$&$[0, g_m] \in \mathbb{R}$\\
$ng_c$&Grade de l'apprenant avec pénalisation du temps &$[0, g_m] \in \mathbb{R}$\\	71	71	$ng_c$&Grade de l'apprenant avec pénalisation du temps &$[0, g_m] \in \mathbb{R}$\\
$t_{c}$&Le temps de réponse à une question de complexité $c$&$[0, t_m] \in \mathbb{R}$\\	72	72	$t_{c}$&Le temps de réponse à une question de complexité $c$&$[0, t_m] \in \mathbb{R}$\\
$ncl$&Nouveau niveau de complexité calculé&$\mathbb{N}$\\	73	73	$ncl$&Nouveau niveau de complexité calculé&$\mathbb{N}$\\
$\alpha_{c}$&Valeur de $\alpha$ dans la complexité $c$&$\mathbb{R} \; \| \; \alpha_{c}>0$\\	74	74	$\alpha_{c}$&Valeur de $\alpha$ dans la complexité $c$&$\mathbb{R} \; \| \; \alpha_{c}>0$\\
$\beta_{c}$&Valeur de $\beta$ dans la complexité $c$&$\mathbb{R} \; \| \; \beta_{c}>0$\\	75	75	$\beta_{c}$&Valeur de $\beta$ dans la complexité $c$&$\mathbb{R} \; \| \; \beta_{c}>0$\\
$\Delta \beta$&Pas initial du paramètre bêta&$\mathbb{N} \; \| \; \Delta \beta >0$\\	76	76	$\Delta \beta$&Pas initial du paramètre bêta&$\mathbb{N} \; \| \; \Delta \beta >0$\\
$\lambda$&Poids de la pénalisation temporelle&$(0,1) \in \mathbb{R}$\\	77	77	$\lambda$&Poids de la pénalisation temporelle&$(0,1) \in \mathbb{R}$\\
$G_c$&Ensemble de $d$ notes dans le niveau de complexité $c$&$\mathbb{R}^d \;, d\in \mathbb{N} \; \| \; d>0$\\	78	78	$G_c$&Ensemble de $d$ notes dans le niveau de complexité $c$&$\mathbb{R}^d \;, d\in \mathbb{N} \; \| \; d>0$\\
$x_c$&Notes moyennes normalisées&$[0, 1] \in \mathbb{R}$\\	79	79	$x_c$&Notes moyennes normalisées&$[0, 1] \in \mathbb{R}$\\
$n_c$&Nombre total de questions dans une séance&$\mathbb{N} \; \| \; n_c>0$\\	80	80	$n_c$&Nombre total de questions dans une séance&$\mathbb{N} \; \| \; n_c>0$\\
$ny_c$&Nombre de questions dans le niveau de complexité $c$&$\mathbb{N} \; \| \; 0<ny_c \le n_c$\\	81	81	$ny_c$&Nombre de questions dans le niveau de complexité $c$&$\mathbb{N} \; \| \; 0<ny_c \le n_c$\\
$y_c$&Proportion de questions dans le niveau de complexité $c$&$[0, 1] \in \mathbb{R}$\\	82	82	$y_c$&Proportion de questions dans le niveau de complexité $c$&$[0, 1] \in \mathbb{R}$\\
$r$&Valeur totale de la métrique définie pour l'adaptabilité&$[0, c_n] \in \mathbb{R}$\\	83	83	$r$&Valeur totale de la métrique définie pour l'adaptabilité&$[0, c_n] \in \mathbb{R}$\\
$sc$&Valeur totale de la métrique de similarité cosinus&$[-1, 1] \in \mathbb{R}$\\	84	84	$sc$&Valeur totale de la métrique de similarité cosinus&$[-1, 1] \in \mathbb{R}$\\
\end{tabular}	85	85	\end{tabular}
\caption{Variables et paramètres du système de recommandation proposé}	86	86	\caption{Variables et paramètres du système de recommandation proposé}
\label{tabPar}	87	87	\label{tabPar}
\end{table}	88	88	\end{table}
	89	89
Le tableau \ref{tabPar} présente les variables de l'algorithme de recommandation. Nous avons considéré que les notes $g_c$ obtenues au niveau de complexité $c$ tiennent compte du temps de réponse. C'est la raison pour laquelle, nous avons défini le grade de l'apprenant $ng_c$ au niveau de complexité $c$. Ce grade calculé selon l'équation \ref{eqsGT}, tient compte d'un poids de pénalisation temporelle $\lambda$.	90	90	Le tableau \ref{tabPar} présente les variables de l'algorithme de recommandation. Nous avons considéré que les notes $g_c$ obtenues au niveau de complexité $c$ tiennent compte du temps de réponse. C'est la raison pour laquelle, nous avons défini le grade de l'apprenant $ng_c$ au niveau de complexité $c$. Ce grade calculé selon l'équation \ref{eqsGT}, tient compte d'un poids de pénalisation temporelle $\lambda$.
	91	91
\begin{equation}	92	92	\begin{equation}
ng_c=g_c- \left(g_c * \lambda * \frac{t_c}{t_m} \right)	93	93	ng_c=g_c- \left(g_c * \lambda * \frac{t_c}{t_m} \right)
\label{eqsGT}	94	94	\label{eqsGT}
\end{equation}	95	95	\end{equation}
	96	96
Dans cet algorithme, la variable de seuil de grade $g_t$ détermine la variabilité de la distribution de probabilité pour chaque niveau de complexité. Les niveaux de complexité des exercices proposés à l'apprenant sont calculés par récompense inverse selon les équations \ref{eqgtc} et \ref{eqltc}. Chaque niveau de complexité est associé à une distribution de probabilité Beta avec des valeurs initiales $\alpha$ et $\beta$ prédéfinies.	97	97	Dans cet algorithme, la variable de seuil de grade $g_t$ détermine la variabilité de la distribution de probabilité pour chaque niveau de complexité. Les niveaux de complexité des exercices proposés à l'apprenant sont calculés par récompense inverse selon les équations \ref{eqgtc} et \ref{eqltc}. Chaque niveau de complexité est associé à une distribution de probabilité Beta avec des valeurs initiales $\alpha$ et $\beta$ prédéfinies.
	98	98
\begin{equation}	99	99	\begin{equation}
ng_c \ge g_t \rightarrow	100	100	ng_c \ge g_t \rightarrow
\begin{cases}	101	101	\begin{cases}
\beta_c=\beta_c+\Delta_s\\	102	102	\beta_c=\beta_c+\Delta_s\\
\beta_{c-1}=\beta_{c-1} + \frac{\Delta_s}{2}\\	103	103	\beta_{c-1}=\beta_{c-1} + \frac{\Delta_s}{2}\\
\alpha_{c+1}=\alpha_{c+1} + \frac{\Delta_s}{2}	104	104	\alpha_{c+1}=\alpha_{c+1} + \frac{\Delta_s}{2}
\end{cases}	105	105	\end{cases}
\label{eqgtc}	106	106	\label{eqgtc}
\end{equation}	107	107	\end{equation}
	108	108
\begin{equation}	109	109	\begin{equation}
ng_c < g_t \rightarrow	110	110	ng_c < g_t \rightarrow
\begin{cases}	111	111	\begin{cases}
\alpha_c=\alpha_c+\Delta_s\\	112	112	\alpha_c=\alpha_c+\Delta_s\\
\alpha_{c-1}=\alpha_{c-1} + \frac{\Delta_s}{2}\\	113	113	\alpha_{c-1}=\alpha_{c-1} + \frac{\Delta_s}{2}\\
\beta_{c+1}=\beta_{c+1} + \frac{\Delta_s}{2}	114	114	\beta_{c+1}=\beta_{c+1} + \frac{\Delta_s}{2}
\end{cases}	115	115	\end{cases}
\label{eqltc}	116	116	\label{eqltc}
\end{equation}	117	117	\end{equation}
	118	118
Pour chaque niveau de complexité $c$, $Beta(\alpha_c, \beta_c)$ fournit une distribution de probabilité $\theta_c$ dont nous calculons l'espérance $\mathbb{E}[\theta_c]$. Le nouveau niveau de complexité $ncl$ correspond à l'espérance maximale obtenue.	119	119	Pour chaque niveau de complexité $c$, $Beta(\alpha_c, \beta_c)$ fournit une distribution de probabilité $\theta_c$ dont nous calculons l'espérance $\mathbb{E}[\theta_c]$. Le nouveau niveau de complexité $ncl$ correspond à l'espérance maximale obtenue.
	120	120
Le détail des pas d'exécution de l'algorithme proposé sont dans l'algorithme \ref{alg2}.	121	121	Le détail des pas d'exécution de l'algorithme proposé sont dans l'algorithme \ref{alg2}.
	122	122
\begin{algorithm}	123	123	\begin{algorithm}
\caption{Algorithme de recommandation stochastique}	124	124	\caption{Algorithme de recommandation stochastique}
\begin{algorithmic}	125	125	\begin{algorithmic}
\State Initialisation de la distribution de probabilité	126	126	\State Initialisation de la distribution de probabilité
\For {\textbf{each} question $q$}	127	127	\For {\textbf{each} question $q$}
\State Soit le niveau de complexité $i$	128	128	\State Soit le niveau de complexité $i$
\State $ng_i=g_i- \left(g_i * \lambda * \frac{t_i}{t_m} \right)$ \Comment{eq \ref{eqsGT}}	129	129	\State $ng_i=g_i- \left(g_i * \lambda * \frac{t_i}{t_m} \right)$ \Comment{eq \ref{eqsGT}}
\State Calculs des paramètres $\alpha_i$ et $\beta_i$ \Comment{eq \ref{eqgtc} et eq \ref{eqltc}}	130	130	\State Calculs des paramètres $\alpha_i$ et $\beta_i$ \Comment{eq \ref{eqgtc} et eq \ref{eqltc}}
\State Choisir $\theta_c$ selon la distribution de probabilité Beta \Comment{$\forall c, \theta_c = Beta(\alpha_c, \beta_c)$}	131	131	\State Choisir $\theta_c$ selon la distribution de probabilité Beta \Comment{$\forall c, \theta_c = Beta(\alpha_c, \beta_c)$}
\State $ncl = max(\mathbb{E}[\theta_c]), \forall c$	132	132	\State $ncl = max(\mathbb{E}[\theta_c]), \forall c$
\EndFor	133	133	\EndFor
\end{algorithmic}	134	134	\end{algorithmic}
\label{alg2}	135	135	\label{alg2}
\end{algorithm}	136	136	\end{algorithm}
	137	137
\subsection{Résultats}	138	138	\subsection{Résultats}
	139	139
Le comportement du module de recommandation a été testé avec des données générées contenant les notes et les temps de réponse de mille apprenants pour cinq niveaux de complexité différents. Ces données sont décrites dans le tableau \ref{tabDataSet}. Les notes des apprenants sont générées selon la loi de probabilité logit-normale considérée comme la plus fidèle dans ce contexte par \cite{Arthurs}.	140	140	Le comportement du module de recommandation a été testé avec des données générées contenant les notes et les temps de réponse de mille apprenants pour cinq niveaux de complexité différents. Ces données sont décrites dans le tableau \ref{tabDataSet}. Les notes des apprenants sont générées selon la loi de probabilité logit-normale considérée comme la plus fidèle dans ce contexte par \cite{Arthurs}.
%\colorbox{yellow}{<== OK pour toi, Daniel ?}	141	141	%\colorbox{yellow}{<== OK pour toi, Daniel ?}
	142	142
L'ensemble de données générées résulte d'une simulation des notes obtenues par des apprenants virtuels ayant répondu à quinze questions réparties sur cinq niveaux de complexité. L'ensemble de données simule, via la distribution de probabilité logit-normale, une faiblesse dans chaque niveau de complexité pour 70\% des apprenants sur les dix premières questions. La difficulté de la complexité est quant à elle simulée en réduisant le score moyen et en augmentant la variance. La figure \ref{figData} montre la manière dont sont réparties les notes selon le niveau de complexité.	143	143	L'ensemble de données générées résulte d'une simulation des notes obtenues par des apprenants virtuels ayant répondu à quinze questions réparties sur cinq niveaux de complexité. L'ensemble de données simule, via la distribution de probabilité logit-normale, une faiblesse dans chaque niveau de complexité pour 70\% des apprenants sur les dix premières questions. La difficulté de la complexité est quant à elle simulée en réduisant le score moyen et en augmentant la variance. La figure \ref{figData} montre la manière dont sont réparties les notes selon le niveau de complexité.
	144	144
\begin{figure}	145	145	\begin{figure}
\includegraphics[width=\textwidth]{./Figures/dataset.png}	146	146	\includegraphics[width=\textwidth]{./Figures/dataset.png}
\caption{Répartition des notes générées selon le niveau de complexité.}	147	147	\caption{Répartition des notes générées selon le niveau de complexité.}
\label{figData}	148	148	\label{figData}
\end{figure}	149	149	\end{figure}
	150	150
\begin{table}[!ht]	151	151	\begin{table}[!ht]
\centering	152	152	\centering
\begin{tabular}{ccc}	153	153	\begin{tabular}{ccc}
ID&Description&Domaine\\	154	154	ID&Description&Domaine\\
\hline	155	155	\hline
$q_{c}$&Niveau de complexité de une question $q$&$[0, c_n] \in \mathbb{N}$\\	156	156	$q_{c}$&Niveau de complexité de une question $q$&$[0, c_n] \in \mathbb{N}$\\
$q_{g,c}$&Note obtenue $g$ pour la question $q$ avec complexité $c$ &$[0,g_m] \in \mathbb{R}$\\	157	157	$q_{g,c}$&Note obtenue $g$ pour la question $q$ avec complexité $c$ &$[0,g_m] \in \mathbb{R}$\\
$q_{t,c}$&Temps employé $t$ pour une question $q$ avec complexité $c$&$[0, t_m] \in \mathbb{R}$\\	158	158	$q_{t,c}$&Temps employé $t$ pour une question $q$ avec complexité $c$&$[0, t_m] \in \mathbb{R}$\\
\end{tabular}	159	159	\end{tabular}
\caption{Description des données utilisées pour l'évaluation.}	160	160	\caption{Description des données utilisées pour l'évaluation.}
\label{tabDataSet}	161	161	\label{tabDataSet}
\end{table}	162	162	\end{table}
	163	163
Toutes les valeurs des paramètres pour tester l'algorithme sont dans le tableau \ref{tabgm1}.	164	164	Toutes les valeurs des paramètres pour tester l'algorithme sont dans le tableau \ref{tabgm1}.
	165	165
\begin{table}[!ht]	166	166	\begin{table}[!ht]
\centering	167	167	\centering
\begin{tabular}{c\|cccccccccccccc}	168	168	\begin{tabular}{c\|cccccccccccccc}
ID&$c_n$&$g_m$&$t_m$&$s$&$s_c$&$\lambda$&$g_t$&$\alpha_{x,1}$&$\alpha_{x,y}$&$\beta_{x,1}$&$\Delta \beta_{x,y}$&$\Delta_1$&$\Delta_2$&$\Delta_3$\\	169	169	ID&$c_n$&$g_m$&$t_m$&$s$&$s_c$&$\lambda$&$g_t$&$\alpha_{x,1}$&$\alpha_{x,y}$&$\beta_{x,1}$&$\Delta \beta_{x,y}$&$\Delta_1$&$\Delta_2$&$\Delta_3$\\
\hline	170	170	\hline
Valeur&5&10&120&3&2&0.25&6 & 2 & 1 & 1 & 1 & 0.3 & 0.5 & 0.7\\	171	171	Valeur&5&10&120&3&2&0.25&6 & 2 & 1 & 1 & 1 & 0.3 & 0.5 & 0.7\\
\end{tabular}	172	172	\end{tabular}
\caption{Valeurs des paramètres pour les scénarios évalués}	173	173	\caption{Valeurs des paramètres pour les scénarios évalués}
\label{tabgm1}	174	174	\label{tabgm1}
\end{table}	175	175	\end{table}
	176	176
La figure \ref{figCmp2} permet de comparer les résultats obtenus par le module proposé, un système de recommandation déterministe et le système AI-VT initial lors d'un \textit{démarrage à froid} (c'est-à-dire sans données historiques ni informations préalables sur le profil de l'apprenant). Sur les graphiques de cette figure, les numéros des questions posées sont reportées en abscisse selon l'ordre chronologique d'apparition durant la séance d’entraînement, le niveau de complexité de chaque question posée est représenté par une couleur différente, et le nombre d'apprenants ayant eu des questions de ce niveau de complexité sont reportés en ordonnées. Ainsi, le système AI-VT initial (premier graphique de la figure) et le système de recommandation déterministe (deuxième graphique) ont tous deux proposé trois questions de niveau de complexité 0 (le plus faible) à tous les apprenants au démarrage de la séance d'entrainement. Nous pouvons remarquer que le système initial est resté sur ce niveau de complexité durant toute la séance (pour les 15 questions du test), tandis que le système de recommandation déterministe a progressivement mixé les complexités des questions posées. Le système de recommandation stochastique décrit dans ce chapitre a quant à lui mixé ces niveaux de complexité dès la première question.	177	177	La figure \ref{figCmp2} permet de comparer les résultats obtenus par le module proposé, un système de recommandation déterministe et le système AI-VT initial lors d'un \textit{démarrage à froid} (c'est-à-dire sans données historiques ni informations préalables sur le profil de l'apprenant). Sur les graphiques de cette figure, les numéros des questions posées sont reportées en abscisse selon l'ordre chronologique d'apparition durant la séance d’entraînement, le niveau de complexité de chaque question posée est représenté par une couleur différente, et le nombre d'apprenants ayant eu des questions de ce niveau de complexité sont reportés en ordonnées. Ainsi, le système AI-VT initial (premier graphique de la figure) et le système de recommandation déterministe (deuxième graphique) ont tous deux proposé trois questions de niveau de complexité 0 (le plus faible) à tous les apprenants au démarrage de la séance d'entrainement. Nous pouvons remarquer que le système initial est resté sur ce niveau de complexité durant toute la séance (pour les 15 questions du test), tandis que le système de recommandation déterministe a progressivement mixé les complexités des questions posées. Le système de recommandation stochastique décrit dans ce chapitre a quant à lui mixé ces niveaux de complexité dès la première question.
	178	178
Ainsi, les systèmes de recommandation permettent de proposer une adaptation progressive du niveau de complexité en fonction des notes obtenues. L'algorithme déterministe génère quatre grandes transitions avec un grand nombre d'apprenants dans les questions 5, 6, 8 et 12, toutes entre des niveaux de complexité contigus. La tendance est à la baisse pour les niveaux 0, 1 et 2 après la huitième question et à la hausse pour les niveaux 1 et 3. L'algorithme stochastique commence par proposer tous les niveaux de complexité possibles tout en privilégiant le niveau 0. Avec ce système, les transitions sont constantes mais pour un petit nombre d'apprenants. La tendance après la dixième question est à la baisse pour les niveaux 0 et 4 et à la hausse pour les niveaux 1, 2 et 3.	179	179	Ainsi, les systèmes de recommandation permettent de proposer une adaptation progressive du niveau de complexité en fonction des notes obtenues. L'algorithme déterministe génère quatre grandes transitions avec un grand nombre d'apprenants dans les questions 5, 6, 8 et 12, toutes entre des niveaux de complexité contigus. La tendance est à la baisse pour les niveaux 0, 1 et 2 après la huitième question et à la hausse pour les niveaux 1 et 3. L'algorithme stochastique commence par proposer tous les niveaux de complexité possibles tout en privilégiant le niveau 0. Avec ce système, les transitions sont constantes mais pour un petit nombre d'apprenants. La tendance après la dixième question est à la baisse pour les niveaux 0 et 4 et à la hausse pour les niveaux 1, 2 et 3.
	180	180
\begin{figure}	181	181	\begin{figure}
\includegraphics[width=\textwidth]{./Figures/comp2.png}	182	182	\includegraphics[width=\textwidth]{./Figures/comp2.png}
\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la première séance avec un démarrage à froid (sans données initiales sur les apprenants). Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}	183	183	\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la première séance avec un démarrage à froid (sans données initiales sur les apprenants). Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}
\label{figCmp2}	184	184	\label{figCmp2}
\end{figure}	185	185	\end{figure}
	186	186
Après la génération de la première séance, le système peut continuer avec une deuxième liste d'exercices. Pour cette partie des tests, les trois algorithmes ont été initialisés avec les mêmes données, et des valeurs égales pour tous les apprenants. La figure \ref{figCmp3} permet de voir que la première transition du système initial n'intervient qu'entre deux séances. Les transitions sont très lentes, et tous les apprenants doivent suivre un chemin identique même s'ils obtiennent des notes différentes au cours de celle-ci.	187	187	Après la génération de la première séance, le système peut continuer avec une deuxième liste d'exercices. Pour cette partie des tests, les trois algorithmes ont été initialisés avec les mêmes données, et des valeurs égales pour tous les apprenants. La figure \ref{figCmp3} permet de voir que la première transition du système initial n'intervient qu'entre deux séances. Les transitions sont très lentes, et tous les apprenants doivent suivre un chemin identique même s'ils obtiennent des notes différentes au cours de celle-ci.
	188	188
Pour leur part, les deux autres systèmes de recommandation testés proposent un fonctionnement différent. L'algorithme déterministe présente trois transitions aux questions 3, 5 et 12. Les tendances sont y relativement homogènes et progressives pour le niveau 3, très variables pour le niveau 2 et fortement décroissantes pour le niveau 0. L'algorithme stochastique quant à lui, propose des transitions douces mais il a tendance à toujours privilégier le niveau le plus faible. Nous pouvons observer une prépondérance du niveau 1 avec ce système. Ici, les niveaux 0 et 1 sont décroissants, le niveau 2 est statique et les niveaux 3 et 4 sont ascendants.	189	189	Pour leur part, les deux autres systèmes de recommandation testés proposent un fonctionnement différent. L'algorithme déterministe présente trois transitions aux questions 3, 5 et 12. Les tendances sont y relativement homogènes et progressives pour le niveau 3, très variables pour le niveau 2 et fortement décroissantes pour le niveau 0. L'algorithme stochastique quant à lui, propose des transitions douces mais il a tendance à toujours privilégier le niveau le plus faible. Nous pouvons observer une prépondérance du niveau 1 avec ce système. Ici, les niveaux 0 et 1 sont décroissants, le niveau 2 est statique et les niveaux 3 et 4 sont ascendants.
	190	190
\begin{figure}	191	191	\begin{figure}
\includegraphics[width=\textwidth]{./Figures/comp3.png}	192	192	\includegraphics[width=\textwidth]{./Figures/comp3.png}
\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la deuxième séance. Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}	193	193	\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la deuxième séance. Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}
\label{figCmp3}	194	194	\label{figCmp3}
\end{figure}	195	195	\end{figure}
	196	196
Les questions de la première et la deuxième séance étant de niveaux 0 et 1, le système a proposé des niveaux de complexité 1 ou 2 pour la troisième séance. La figure \ref{figCmp4} montre que le système initial est très lent à passer d'un niveau à l'autre. La figure \ref{figCmp3} permet de voir la première transition du système initial ne réagissant qu'aux notes obtenues dans les séances précédentes et non à celles de la séance en cours. Dans ce cas, l'algorithme de recommandation déterministe adopte la même stratégie et propose un changement brutal à tous les apprenants autour de la cinquième question. L'algorithme stochastique continue avec des changements progressifs tout en privilégiant le niveau 2.	197	197	Les questions de la première et la deuxième séance étant de niveaux 0 et 1, le système a proposé des niveaux de complexité 1 ou 2 pour la troisième séance. La figure \ref{figCmp4} montre que le système initial est très lent à passer d'un niveau à l'autre. La figure \ref{figCmp3} permet de voir la première transition du système initial ne réagissant qu'aux notes obtenues dans les séances précédentes et non à celles de la séance en cours. Dans ce cas, l'algorithme de recommandation déterministe adopte la même stratégie et propose un changement brutal à tous les apprenants autour de la cinquième question. L'algorithme stochastique continue avec des changements progressifs tout en privilégiant le niveau 2.
	198	198
\begin{figure}	199	199	\begin{figure}
\includegraphics[width=\textwidth]{./Figures/comp4.png}	200	200	\includegraphics[width=\textwidth]{./Figures/comp4.png}
\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la troisième séance. Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}	201	201	\caption{Niveaux de complexité des questions posées aux apprenants par les trois systèmes testés lors de la troisième séance. Gauche - RàPC, centre - recommandation déterministe (DM), droite - moyenne de 100 exécutions de recommandation stochastique (SM)}
\label{figCmp4}	202	202	\label{figCmp4}
\end{figure}	203	203	\end{figure}
	204	204
Pour comparer numériquement le système initial, l'algorithme déterministe et l'algorithme de recommandation proposé, un ensemble d'équations a été défini (équation \ref{eqMetric1} et équation \ref{eqMetric2}). Celles-ci permettent de décrire le système de recommandation idéal si l'objectif de l'apprenant est de suivre un apprentissage standard. Une valeur est calculée pour chaque niveau de complexité en fonction de la moyenne des notes et du nombre de questions recommandées dans ce niveau de complexité. L'objectif de cette mesure est d'attribuer un score élevé aux systèmes de recommandation qui proposent plus d'exercices au niveau de complexité où l'apprenant a obtenu une note moyenne plus basse, lui permettant ainsi de renforcer ses connaissances pour ce niveau de complexité. De la même manière, il est attendu que le système de recommandation propose moins d'exercices aux niveaux de complexité pour lesquels les notes moyennes sont élevées, l'étudiant ayant acquis des connaissances suffisantes à ces niveaux de complexité. Les scores faibles sont attribués aux systèmes qui recommandent peu d'exercices à des niveaux de complexité dont les notes moyennes sont faibles et, inversement, s'ils proposent beaucoup d'exercices à des niveaux de complexité dont les notes moyennes sont élevées.	205	205	Pour comparer numériquement le système initial, l'algorithme déterministe et l'algorithme de recommandation proposé, un ensemble d'équations a été défini (équation \ref{eqMetric1} et équation \ref{eqMetric2}). Celles-ci permettent de décrire le système de recommandation idéal si l'objectif de l'apprenant est de suivre un apprentissage standard. Une valeur est calculée pour chaque niveau de complexité en fonction de la moyenne des notes et du nombre de questions recommandées dans ce niveau de complexité. L'objectif de cette mesure est d'attribuer un score élevé aux systèmes de recommandation qui proposent plus d'exercices au niveau de complexité où l'apprenant a obtenu une note moyenne plus basse, lui permettant ainsi de renforcer ses connaissances pour ce niveau de complexité. De la même manière, il est attendu que le système de recommandation propose moins d'exercices aux niveaux de complexité pour lesquels les notes moyennes sont élevées, l'étudiant ayant acquis des connaissances suffisantes à ces niveaux de complexité. Les scores faibles sont attribués aux systèmes qui recommandent peu d'exercices à des niveaux de complexité dont les notes moyennes sont faibles et, inversement, s'ils proposent beaucoup d'exercices à des niveaux de complexité dont les notes moyennes sont élevées.
	206	206
\begin{equation}	207	207	\begin{equation}
%r_c=x+y-2xy	208	208	%r_c=x+y-2xy
%r_c=x^2+y^2-2x^2y^2	209	209	%r_c=x^2+y^2-2x^2y^2
rp_c(x)=e^{-2(x_{0,c}+x_{1,c}-1)^2} ; \{x \in \mathbb{R}^2 \| 0<=x<=1\}	210	210	rp_c(x)=e^{-2(x_{0,c}+x_{1,c}-1)^2} ; \{x \in \mathbb{R}^2 \| 0<=x<=1\}
\label{eqMetric1}	211	211	\label{eqMetric1}
\end{equation}	212	212	\end{equation}
	213	213
\begin{equation}	214	214	\begin{equation}
r=\sum_{c=0}^{c_n-1} rp_c	215	215	r=\sum_{c=0}^{c_n-1} rp_c
\label{eqMetric2}	216	216	\label{eqMetric2}
\end{equation}	217	217	\end{equation}
	218	218
Les propriétés de la métrique sont :	219	219	Les propriétés de la métrique sont :
\begin{itemize}	220	220	\begin{itemize}
\item $\{\forall x \in \mathbb{R}^2 \| 0<=x<=1\}, rp_c(x)>0$	221	221	\item $\{\forall x \in \mathbb{R}^2 \| 0<=x<=1\}, rp_c(x)>0$
\item $max(rp_c(x))=1; \; if \; x_{0,c}+x_{1,c}=1$	222	222	\item $max(rp_c(x))=1; \; if \; x_{0,c}+x_{1,c}=1$
\item $min(rp_c(x))=0.1353; \; if \; \left ( \sum_{i=1}^2 x_{i,c}=0 \; \lor \; \sum_{i=1}^2 x_{i,c} = 2 \right )$\\	223	223	\item $min(rp_c(x))=0.1353; \; if \; \left ( \sum_{i=1}^2 x_{i,c}=0 \; \lor \; \sum_{i=1}^2 x_{i,c} = 2 \right )$\\
\end{itemize}	224	224	\end{itemize}
	225	225
Dans l'équation \ref{eqMetric1}, $x_{0,c}$ est la moyenne normalisée des notes dans le niveau de complexité $c$ (équation \ref{eqXc}), et $x_{1,c}$ est le nombre normalisé de questions auxquelles des réponses ont été données dans le niveau de complexité $c$ (équation \ref{eqYc}). Ainsi, plus la valeur de $r$ est élevée, meilleure est la recommandation.	226	226	Dans l'équation \ref{eqMetric1}, $x_{0,c}$ est la moyenne normalisée des notes dans le niveau de complexité $c$ (équation \ref{eqXc}), et $x_{1,c}$ est le nombre normalisé de questions auxquelles des réponses ont été données dans le niveau de complexité $c$ (équation \ref{eqYc}). Ainsi, plus la valeur de $r$ est élevée, meilleure est la recommandation.
	227	227
\begin{equation}	228	228	\begin{equation}
x_{0,c}=\frac{<g_c>_{G_c}}{g_m}	229	229	x_{0,c}=\frac{<g_c>_{G_c}}{g_m}
\label{eqXc}	230	230	\label{eqXc}
\end{equation}	231	231	\end{equation}
	232	232
\begin{equation}	233	233	\begin{equation}
x_{1,c}=\frac{ny_c}{n_c}	234	234	x_{1,c}=\frac{ny_c}{n_c}
\label{eqYc}	235	235	\label{eqYc}
\end{equation}	236	236	\end{equation}
	237	237
La figure \ref{figMetric} représente la fonction $rp_c(x)$. La valeur maximale de $r$ dans un niveau de complexité spécifique étant égale à $1$, la valeur maximale globale pour les scénarios testés est égale à 5.	238	238	La figure \ref{figMetric} représente la fonction $rp_c(x)$. La valeur maximale de $r$ dans un niveau de complexité spécifique étant égale à $1$, la valeur maximale globale pour les scénarios testés est égale à 5.
	239	239
\begin{figure}	240	240	\begin{figure}
\includegraphics[width=\textwidth]{./Figures/metric.png}	241	241	\includegraphics[width=\textwidth]{./Figures/metric.png}
\caption{Fonction d'évaluation de la qualité de la recommandation pour un parcours standard}	242	242	\caption{Fonction d'évaluation de la qualité de la recommandation pour un parcours standard}
\label{figMetric}	243	243	\label{figMetric}
\end{figure}	244	244	\end{figure}
	245	245
Les résultats des calculs de la métrique $rp_c(x)$ établie pour le système initial et les deux algorithmes dans les trois scénarios testés sont présentés dans le tableau \ref{tabRM}.	246	246	Les résultats des calculs de la métrique $rp_c(x)$ établie pour le système initial et les deux algorithmes dans les trois scénarios testés sont présentés dans le tableau \ref{tabRM}.
	247	247
\begin{table}[!ht]	248	248	\begin{table}[!ht]
\centering	249	249	\centering
\begin{tabular}{cccccccc}	250	250	\begin{tabular}{cccccccc}
&$c_0$&$c_1$&$c_2$&$c_3$&$c_4$&Total ($r$)&Total ($\%$)\\	251	251	&$c_0$&$c_1$&$c_2$&$c_3$&$c_4$&Total ($r$)&Total ($\%$)\\
\hline	252	252	\hline
Test 1\\	253	253	Test 1\\
\hline	254	254	\hline
RàPC&0.5388&-&-&-&-&0.5388&10.776\\	255	255	RàPC&0.5388&-&-&-&-&0.5388&10.776\\
DM&0.8821&0.7282&\textbf{0.9072}&\textbf{0.8759}&-&3.3934&67.868\\	256	256	DM&0.8821&0.7282&\textbf{0.9072}&\textbf{0.8759}&-&3.3934&67.868\\
SM&\textbf{0.9463}&\textbf{0.8790}&0.7782&0.7108&0.6482&\textbf{3.9625}&\textbf{79.25}\\	257	257	SM&\textbf{0.9463}&\textbf{0.8790}&0.7782&0.7108&0.6482&\textbf{3.9625}&\textbf{79.25}\\
\hline	258	258	\hline
Test 2\\	259	259	Test 2\\
\hline	260	260	\hline
RàPC&0.9445&\textbf{0.9991}&-&-&-&1.9436&38.872\\	261	261	RàPC&0.9445&\textbf{0.9991}&-&-&-&1.9436&38.872\\
DM&-&0.9443&\textbf{0.8208}&\textbf{0.9623}&-&2.7274&54.548\\	262	262	DM&-&0.9443&\textbf{0.8208}&\textbf{0.9623}&-&2.7274&54.548\\
SM&\textbf{0.9688}&0.9861&0.8067&0.7161&0.6214&\textbf{4.0991}&\textbf{81.982}\\	263	263	SM&\textbf{0.9688}&0.9861&0.8067&0.7161&0.6214&\textbf{4.0991}&\textbf{81.982}\\
\hline	264	264	\hline
Test3\\	265	265	Test3\\
\hline	266	266	\hline
RàPC&-&0.8559&0.7377&-&-&1.5936&31.872	267	267	RàPC&-&0.8559&0.7377&-&-&1.5936&31.872
\\	268	268	\\
DM&-&-&0.5538&\textbf{0.7980}&-&1.3518&27.036\\	269	269	DM&-&-&0.5538&\textbf{0.7980}&-&1.3518&27.036\\
SM&0.9089&\textbf{0.9072}&\textbf{0.9339}&0.7382&0.6544&\textbf{4.1426}&\textbf{82.852}\\	270	270	SM&0.9089&\textbf{0.9072}&\textbf{0.9339}&0.7382&0.6544&\textbf{4.1426}&\textbf{82.852}\\
\end{tabular}	271	271	\end{tabular}
\caption{Résultats de la métrique $rp_c(x)$ (RàPC - Système sans module de recommandation, DM - Module de recommandation déterministe, SM - Module de recommandation stochastique)}	272	272	\caption{Résultats de la métrique $rp_c(x)$ (RàPC - Système sans module de recommandation, DM - Module de recommandation déterministe, SM - Module de recommandation stochastique)}
\label{tabRM}	273	273	\label{tabRM}
\end{table}	274	274	\end{table}
	275	275
Les équations \ref{eqMetricS1} et \ref{eqMetricS2} permettent de caractériser un apprentissage progressif. Dans ce cas, un score élevé est attribué aux systèmes proposant plus d'exercices dans un niveau de complexité où les notes moyennes sont légèrement insuffisantes (4/10), plus flexibles avec des notes moyennes plus basses, et un petit nombre d'exercices pour des notes moyennes élevées. Les scores faibles sont attribués aux systèmes qui recommandent de nombreuses questions dans un niveau de complexité avec des notes moyennes élevées ou faibles.	276	276	Les équations \ref{eqMetricS1} et \ref{eqMetricS2} permettent de caractériser un apprentissage progressif. Dans ce cas, un score élevé est attribué aux systèmes proposant plus d'exercices dans un niveau de complexité où les notes moyennes sont légèrement insuffisantes (4/10), plus flexibles avec des notes moyennes plus basses, et un petit nombre d'exercices pour des notes moyennes élevées. Les scores faibles sont attribués aux systèmes qui recommandent de nombreuses questions dans un niveau de complexité avec des notes moyennes élevées ou faibles.
	277	277
\begin{equation}	278	278	\begin{equation}
rs_c(x)=e^{-\frac{2}{100}(32x_{0,c}^2-28x_{0,c}+10x_{1,c}-4)^2} ; \{x \in \mathbb{R}^2 \| 0<=x<=1\}	279	279	rs_c(x)=e^{-\frac{2}{100}(32x_{0,c}^2-28x_{0,c}+10x_{1,c}-4)^2} ; \{x \in \mathbb{R}^2 \| 0<=x<=1\}
\label{eqMetricS1}	280	280	\label{eqMetricS1}
\end{equation}	281	281	\end{equation}
	282	282
\begin{equation}	283	283	\begin{equation}
r=\sum_{c=0}^{c_n-1} rs_c	284	284	r=\sum_{c=0}^{c_n-1} rs_c
\label{eqMetricS2}	285	285	\label{eqMetricS2}
\end{equation}	286	286	\end{equation}
	287	287
Les propriétés de la métrique sont :	288	288	Les propriétés de la métrique sont :
\begin{itemize}	289	289	\begin{itemize}
\item $\{\forall x \in \mathbb{R}^2 \| 0<=x<=1\}, rs_c(x)>0$	290	290	\item $\{\forall x \in \mathbb{R}^2 \| 0<=x<=1\}, rs_c(x)>0$
\item $max(rs_c(x))=1; \; if \; 16x_{0,c}^2-14x_{0,c}+5x_{1,c}-2=0$\\	291	291	\item $max(rs_c(x))=1; \; if \; 16x_{0,c}^2-14x_{0,c}+5x_{1,c}-2=0$\\
\end{itemize}	292	292	\end{itemize}
	293	293
La figure \ref{figMetric2} représente la fonction $rs_c(x)$. Comme pour $rp_c$, la valeur maximale de $r$ dans un niveau de complexité spécifique étant égale à $1$, la valeur maximale globale pour les scénarios testés est égale à 5.	294	294	La figure \ref{figMetric2} représente la fonction $rs_c(x)$. Comme pour $rp_c$, la valeur maximale de $r$ dans un niveau de complexité spécifique étant égale à $1$, la valeur maximale globale pour les scénarios testés est égale à 5.
	295	295
Les résultats du calcul des métriques pour le système initial et les deux algorithmes dans les trois scénarios définis sont présentés dans le tableau \ref{tabRM2}.	296	296	Les résultats du calcul des métriques pour le système initial et les deux algorithmes dans les trois scénarios définis sont présentés dans le tableau \ref{tabRM2}.
	297	297
\begin{figure}[!ht]	298	298	\begin{figure}[!ht]
\centering	299	299	\centering
\includegraphics[width=\textwidth]{./Figures/metric2.png}	300	300	\includegraphics[width=\textwidth]{./Figures/metric2.png}
\caption{Fonction d'évaluation de la qualité de la recommandation pour un apprentissage progressif.}	301	301	\caption{Fonction d'évaluation de la qualité de la recommandation pour un apprentissage progressif.}
\label{figMetric2}	302	302	\label{figMetric2}
\end{figure}	303	303	\end{figure}
	304	304
\begin{table}[!ht]	305	305	\begin{table}[!ht]
\centering	306	306	\centering
\begin{tabular}{cccccccc}	307	307	\begin{tabular}{cccccccc}
&$c_0$&$c_1$&$c_2$&$c_3$&$c_4$&Total ($r$)&Total ($\%$)\\	308	308	&$c_0$&$c_1$&$c_2$&$c_3$&$c_4$&Total ($r$)&Total ($\%$)\\
\hline	309	309	\hline
Séance 1\\	310	310	Séance 1\\
\hline	311	311	\hline
RàPC&\textbf{0.9979}&-&-&-&-&0.9979&19.96\\	312	312	RàPC&\textbf{0.9979}&-&-&-&-&0.9979&19.96\\
DM&0.8994&0.1908&\textbf{0.3773}&\textbf{0.2990}&-&1.7665&35.33\\	313	313	DM&0.8994&0.1908&\textbf{0.3773}&\textbf{0.2990}&-&1.7665&35.33\\
SM&0.8447&\textbf{0.3012}&0.2536&0.2030&\textbf{0.1709}&\textbf{1.7734}&\textbf{35.47}\\	314	314	SM&0.8447&\textbf{0.3012}&0.2536&0.2030&\textbf{0.1709}&\textbf{1.7734}&\textbf{35.47}\\
\hline	315	315	\hline
Séance 2\\	316	316	Séance 2\\
\hline	317	317	\hline
RàPC&\textbf{0.4724}&\textbf{0.7125}&-&-&-&1.1849&23.70\\	318	318	RàPC&\textbf{0.4724}&\textbf{0.7125}&-&-&-&1.1849&23.70\\
DM&-&0.6310&\textbf{0.3901}&\textbf{0.4253}&-&1.4464&28.93\\	319	319	DM&-&0.6310&\textbf{0.3901}&\textbf{0.4253}&-&1.4464&28.93\\
SM&0.2697&0.7089&0.2634&0.2026&\textbf{0.1683}&\textbf{1.6129}&\textbf{32.26}\\	320	320	SM&0.2697&0.7089&0.2634&0.2026&\textbf{0.1683}&\textbf{1.6129}&\textbf{32.26}\\
\hline	321	321	\hline
Séance 3\\	322	322	Séance 3\\
\hline	323	323	\hline
RàPC&-&\textbf{0.9179}&0.2692&-&-&1.1871&23.74	324	324	RàPC&-&\textbf{0.9179}&0.2692&-&-&1.1871&23.74
\\	325	325	\\
DM&-&-&0.2236&\textbf{0.9674}&-&1.191&23.82\\	326	326	DM&-&-&0.2236&\textbf{0.9674}&-&1.191&23.82\\
SM&0.1873&0.3038&\textbf{0.6345}&0.2394&\textbf{0.1726}&\textbf{1.5376}&\textbf{30.75}\\	327	327	SM&0.1873&0.3038&\textbf{0.6345}&0.2394&\textbf{0.1726}&\textbf{1.5376}&\textbf{30.75}\\
\end{tabular}	328	328	\end{tabular}
\caption{Évaluation des recommandations proposées selon $rs_c(x)$ par les différents systèmes de recommandation testés : RàPC - Système sans module de recommandation, DM - Algorithme deterministique, SM - Algorithme stochastique}	329	329	\caption{Évaluation des recommandations proposées selon $rs_c(x)$ par les différents systèmes de recommandation testés : RàPC - Système sans module de recommandation, DM - Algorithme deterministique, SM - Algorithme stochastique}
\label{tabRM2}	330	330	\label{tabRM2}
\end{table}	331	331	\end{table}
	332	332
En complément, le tableau \ref{tabCS} présente les similarités entre toutes les recommandations faites aux apprenants par les trois systèmes et les trois séances d'entrainement. Pour ce faire, nous avons choisi d'appliquer l'équation \ref{eqCS} permettant de calculer une similarité cosinus entre deux vecteurs $A$ et $B$.	333	333	En complément, le tableau \ref{tabCS} présente les similarités entre toutes les recommandations faites aux apprenants par les trois systèmes et les trois séances d'entrainement. Pour ce faire, nous avons choisi d'appliquer l'équation \ref{eqCS} permettant de calculer une similarité cosinus entre deux vecteurs $A$ et $B$.
	334	334
\begin{equation}	335	335	\begin{equation}
sc=\frac{\sum_{i=1}^n A_i B_i}{\sqrt{\sum_{i=1}^n A_i^2} \sqrt{\sum_{i=1}^n B_i^2}}	336	336	sc=\frac{\sum_{i=1}^n A_i B_i}{\sqrt{\sum_{i=1}^n A_i^2} \sqrt{\sum_{i=1}^n B_i^2}}
\label{eqCS}	337	337	\label{eqCS}
\end{equation}	338	338	\end{equation}
	339	339
\begin{table}[!ht]	340	340	\begin{table}[!ht]
\centering	341	341	\centering
\begin{tabular}{cccc}	342	342	\begin{tabular}{cccc}
Système de recommandation & Séance 1 & Séance 2 & Séance 3\\	343	343	Système de recommandation & Séance 1 & Séance 2 & Séance 3\\
\hline	344	344	\hline
RàPC&1&1&1\\	345	345	RàPC&1&1&1\\
DM&0.9540&0.9887&0.9989\\	346	346	DM&0.9540&0.9887&0.9989\\
SM&\textbf{0.8124}&\textbf{0.8856}&\textbf{0.9244}\\	347	347	SM&\textbf{0.8124}&\textbf{0.8856}&\textbf{0.9244}\\
\end{tabular}	348	348	\end{tabular}
\caption{Moyenne de la diversité des propositions pour tous les apprenants. Une valeur plus faible représente une plus grande diversité. (RàPC - Système sans module de recommandation, DM - Module deterministe, SM - Module stochastique)}	349	349	\caption{Moyenne de la diversité des propositions pour tous les apprenants. Une valeur plus faible représente une plus grande diversité. (RàPC - Système sans module de recommandation, DM - Module deterministe, SM - Module stochastique)}
\label{tabCS}	350	350	\label{tabCS}
\end{table}	351	351	\end{table}
	352	352
\subsection{Discussion et Conclusion}	353	353	\subsection{Discussion et Conclusion}
Avec la génération d'exercices par le système de RàPC initial, AI-VT propose les mêmes exercices à tous les apprenants, et l'évolution des niveaux de complexité est très lente, un changement toutes les trois ou quatre séances environ. En effet, le système ne prend pas en compte les notes obtenues pendant la séance. Les systèmes intégrant l'un des modules de recommandation testés sont plus dynamiques et les évolutions sont plus rapides. En considérant les notes des apprenants, l'algorithme déterministe suggère des changements de niveaux à un grand nombre d'apprenants de manière soudaine, tandis que l'algorithme stochastique est plus axé sur la personnalisation individuelle et les changements de niveau de complexité sont produits pour un petit nombre d'apprenants. Les deux modules de recommandation proposés ont la capacité de détecter les faiblesses des apprenants et d'adapter la séance à leurs besoins particuliers.	354	354	Avec la génération d'exercices par le système de RàPC initial, AI-VT propose les mêmes exercices à tous les apprenants, et l'évolution des niveaux de complexité est très lente, un changement toutes les trois ou quatre séances environ. En effet, le système ne prend pas en compte les notes obtenues pendant la séance. Les systèmes intégrant l'un des modules de recommandation testés sont plus dynamiques et les évolutions sont plus rapides. En considérant les notes des apprenants, l'algorithme déterministe suggère des changements de niveaux à un grand nombre d'apprenants de manière soudaine, tandis que l'algorithme stochastique est plus axé sur la personnalisation individuelle et les changements de niveau de complexité sont produits pour un petit nombre d'apprenants. Les deux modules de recommandation proposés ont la capacité de détecter les faiblesses des apprenants et d'adapter la séance à leurs besoins particuliers.
	355	355
Les données générées ont permis de simuler diverses situations avec les notes de mille apprenants, permettant ainsi d'évaluer le comportement des systèmes de recommandation avec différentes configurations.	356	356	Les données générées ont permis de simuler diverses situations avec les notes de mille apprenants, permettant ainsi d'évaluer le comportement des systèmes de recommandation avec différentes configurations.
	357	357
Les résultats numériques montrent que les distributions des questions dans une séance par les deux modules de recommandation sont différentes bien que la tendance générale soit similaire. Les modules de recommandation proposés tentent de répartir les questions dans tous les niveaux de complexité définis. Globalement, le module de recommandation stochastique a obtenu un meilleur score. En comparaison du système initial, les modules de recommandation (déterministe et stochastique) proposent 15\% à 68\% d'adaptations de la complexité pour tous les niveaux. Pour cette raison, l'approche stochastique sera préférée à l'approche déterministe dans la suite des travaux de recherche.	358	358	Les résultats numériques montrent que les distributions des questions dans une séance par les deux modules de recommandation sont différentes bien que la tendance générale soit similaire. Les modules de recommandation proposés tentent de répartir les questions dans tous les niveaux de complexité définis. Globalement, le module de recommandation stochastique a obtenu un meilleur score. En comparaison du système initial, les modules de recommandation (déterministe et stochastique) proposent 15\% à 68\% d'adaptations de la complexité pour tous les niveaux. Pour cette raison, l'approche stochastique sera préférée à l'approche déterministe dans la suite des travaux de recherche.
	359	359
Selon la métrique de la similarité cosinus, le module de recommandation stochastique augmente la diversité des propositions par rapport au système initial dans les trois séances d'entrainement testées, ce qui indique qu'en plus d'atteindre l'adaptabilité, des propositions personnalisées sont générées tout en maintenant l'objectif de progression des niveaux de compétence des apprenants. La diversité des propositions est une caractéristique essentielle de l'algorithme de recommandation dans ses deux versions.	360	360	Selon la métrique de la similarité cosinus, le module de recommandation stochastique augmente la diversité des propositions par rapport au système initial dans les trois séances d'entrainement testées, ce qui indique qu'en plus d'atteindre l'adaptabilité, des propositions personnalisées sont générées tout en maintenant l'objectif de progression des niveaux de compétence des apprenants. La diversité des propositions est une caractéristique essentielle de l'algorithme de recommandation dans ses deux versions.
	361	361
Les modules de recommandation sont un élément essentiel pour certains EIAH car ils aident à guider le processus d'apprentissage individuel. Ils permettent également d'identifier les faiblesses et de réorienter le processus complet afin d'améliorer les connaissances et les compétences. Les deux modules de recommandation proposés peuvent détecter en temps réel les faiblesses de l'apprenant et tentent de réorienter la séance vers le niveau de complexité le plus adapté. Même si l'ensemble des données générées est une simulation de temps de réponse et de notes fictives d'apprenants fictifs, les tests démontrent la flexibilité et la robustesse des modules de recommandation proposés : les données relatives aux apprenants présentent en effet une grande diversité et obligent le système à s'adapter à différents types de configuration. Par conséquent, il est possible de conclure que les modules de recommandation proposés ont la capacité de fonctionner dans différentes situations et de proposer des chemins alternatifs et personnalisés pour améliorer le processus d'apprentissage global.	362	362	Les modules de recommandation sont un élément essentiel pour certains EIAH car ils aident à guider le processus d'apprentissage individuel. Ils permettent également d'identifier les faiblesses et de réorienter le processus complet afin d'améliorer les connaissances et les compétences. Les deux modules de recommandation proposés peuvent détecter en temps réel les faiblesses de l'apprenant et tentent de réorienter la séance vers le niveau de complexité le plus adapté. Même si l'ensemble des données générées est une simulation de temps de réponse et de notes fictives d'apprenants fictifs, les tests démontrent la flexibilité et la robustesse des modules de recommandation proposés : les données relatives aux apprenants présentent en effet une grande diversité et obligent le système à s'adapter à différents types de configuration. Par conséquent, il est possible de conclure que les modules de recommandation proposés ont la capacité de fonctionner dans différentes situations et de proposer des chemins alternatifs et personnalisés pour améliorer le processus d'apprentissage global.
	363	363
\section{ESCBR-SMA et échantillonnage de Thompson}	364	364	\section{ESCBR-SMA et échantillonnage de Thompson}
\sectionmark{ESCBR-SMA et TS}	365	365	\sectionmark{ESCBR-SMA et TS}
	366	366
La section précédente a démontré l'intérêt de l'intégration d'un module de recommandation afin de proposer des exercices d'un niveau de difficulté adapté aux besoins de l'apprenant en fonction des difficultés décelées au cours de la séance d'entraînement. Le système AI-VT initial fondé sur le cycle du raisonnement à partir de cas et ne proposant que des adaptations entre deux séances d’entraînement consécutives, a été supplanté par l'intégration de modules de recommandation utilisés durant la phase de révision du cycle classique du RàPC. Les deux modules de recommandation testés dans la section précédente étaient l'un déterministe, l'autre stochastique.	367	367	La section précédente a démontré l'intérêt de l'intégration d'un module de recommandation afin de proposer des exercices d'un niveau de difficulté adapté aux besoins de l'apprenant en fonction des difficultés décelées au cours de la séance d'entraînement. Le système AI-VT initial fondé sur le cycle du raisonnement à partir de cas et ne proposant que des adaptations entre deux séances d’entraînement consécutives, a été supplanté par l'intégration de modules de recommandation utilisés durant la phase de révision du cycle classique du RàPC. Les deux modules de recommandation testés dans la section précédente étaient l'un déterministe, l'autre stochastique.
	368	368
La section précédente a également démontré qu'il était possible et intéressant que les niveaux de complexité des exercices proposés puissent suivre des fonctions permettant de les faire fluctuer de manière progressive au cours de la séance, et ce afin que les apprenants ne soient pas confrontés à des difficultés changeant de manière trop abrupte durant l'entraînement. Cette étude nous amène donc à considérer la résolution de la génération d'une séance d'exercices sous l'angle de la régression. Nous proposons donc dans cette partie de montrer de quelle manière nous avons intégré et vérifié l'intérêt des outils définis dans le chapitre précédent dans l'EIAH AI-VT.	369	369	La section précédente a également démontré qu'il était possible et intéressant que les niveaux de complexité des exercices proposés puissent suivre des fonctions permettant de les faire fluctuer de manière progressive au cours de la séance, et ce afin que les apprenants ne soient pas confrontés à des difficultés changeant de manière trop abrupte durant l'entraînement. Cette étude nous amène donc à considérer la résolution de la génération d'une séance d'exercices sous l'angle de la régression. Nous proposons donc dans cette partie de montrer de quelle manière nous avons intégré et vérifié l'intérêt des outils définis dans le chapitre précédent dans l'EIAH AI-VT.
	370	370
\subsection{Concepts Associés}	371	371	\subsection{Concepts Associés}
	372	372
Cette section présente les concepts, les définitions et les algorithmes nécessaires à la compréhension du module proposé. Le paradigme fondamental utilisé dans ce travail est le raisonnement à partir de cas (RàPC), qui permet d'exploiter les connaissances acquises et l'expérience accumulée pour résoudre un problème spécifique. L'idée principale est de rechercher des situations antérieures similaires et d'utiliser l'expérience acquise pour résoudre de nouveaux problèmes. Le RàPC suit classiquement un cycle de quatre étapes pour améliorer la solution d'inférence \cite{jmse11050890}.	373	373	Cette section présente les concepts, les définitions et les algorithmes nécessaires à la compréhension du module proposé. Le paradigme fondamental utilisé dans ce travail est le raisonnement à partir de cas (RàPC), qui permet d'exploiter les connaissances acquises et l'expérience accumulée pour résoudre un problème spécifique. L'idée principale est de rechercher des situations antérieures similaires et d'utiliser l'expérience acquise pour résoudre de nouveaux problèmes. Le RàPC suit classiquement un cycle de quatre étapes pour améliorer la solution d'inférence \cite{jmse11050890}.
	374	374
L'un des algorithmes les plus couramment utilisés dans les EIAH pour adapter le contenu et estimer la progression du niveau de connaissance des apprenants est le BKT (\textit{Bayesian Knowledge Tracing}) \cite{ZHANG2018189}. Cet algorithme utilise quatre paramètres pour estimer la progression des connaissances. $P(k)$ estime la probabilité de connaissance dans une compétence spécifique. $P(w)$, est la probabilité que l'apprenant démontre ses connaissances. $P(s)$, est la probabilité que l'apprenant fasse une erreur. $P(g)$, est la probabilité que l'apprenant ait deviné une réponse. La valeur estimée de la connaissance est mise à jour selon les équations \ref{eqbkt1}, \ref{eqbkt2} et \ref{eqbkt3}. Si la réponse de l'apprenant est correcte, l'équation \ref{eqbkt1} est utilisée, mais si la réponse est incorrecte, l'équation \ref{eqbkt2} est utilisée.	375	375	L'un des algorithmes les plus couramment utilisés dans les EIAH pour adapter le contenu et estimer la progression du niveau de connaissance des apprenants est le BKT (\textit{Bayesian Knowledge Tracing}) \cite{ZHANG2018189}. Cet algorithme utilise quatre paramètres pour estimer la progression des connaissances. $P(k)$ estime la probabilité de connaissance dans une compétence spécifique. $P(w)$, est la probabilité que l'apprenant démontre ses connaissances. $P(s)$, est la probabilité que l'apprenant fasse une erreur. $P(g)$, est la probabilité que l'apprenant ait deviné une réponse. La valeur estimée de la connaissance est mise à jour selon les équations \ref{eqbkt1}, \ref{eqbkt2} et \ref{eqbkt3}. Si la réponse de l'apprenant est correcte, l'équation \ref{eqbkt1} est utilisée, mais si la réponse est incorrecte, l'équation \ref{eqbkt2} est utilisée.
	376	376
\begin{equation}	377	377	\begin{equation}
P(k_{t-1}\|Correct_t)=\frac{P(k_{t-1})(1-P(s))}{P(k_{t-1})(1-P(s))+(1-P(k_{t-1}))P(g)}	378	378	P(k_{t-1}\|Correct_t)=\frac{P(k_{t-1})(1-P(s))}{P(k_{t-1})(1-P(s))+(1-P(k_{t-1}))P(g)}
\label{eqbkt1}	379	379	\label{eqbkt1}
\end{equation}	380	380	\end{equation}
	381	381
\begin{equation}	382	382	\begin{equation}
P(k_{t-1}\|Incorrect_t)=\frac{P(k_{t-1})P(s)}{P(k_{t-1})(P(s))+(1-P(k_{t-1}))(1-P(g))}	383	383	P(k_{t-1}\|Incorrect_t)=\frac{P(k_{t-1})P(s)}{P(k_{t-1})(P(s))+(1-P(k_{t-1}))(1-P(g))}
\label{eqbkt2}	384	384	\label{eqbkt2}
\end{equation}	385	385	\end{equation}
	386	386
\begin{equation}	387	387	\begin{equation}
P(k_{t})=P(k_{t-1}\|evidence_t)+(1-P(k_{t-1}\|evidence_t))P(w)	388	388	P(k_{t})=P(k_{t-1}\|evidence_t)+(1-P(k_{t-1}\|evidence_t))P(w)
\label{eqbkt3}	389	389	\label{eqbkt3}
\end{equation}	390	390	\end{equation}
	391	391
Le module de recommandation proposé, associé à AI-VT, est fondé sur le paradigme de l'apprentissage par renforcement. L'apprentissage par renforcement est une technique d'apprentissage automatique qui permet, par le biais d'actions et de récompenses, d'améliorer les connaissances du système sur une tâche spécifique \cite{NEURIPS2023_9d8cf124}. Nous nous intéressons ici plus particulièrement à l'échantillonnage de Thompson, qui, par le biais d'une distribution de probabilité initiale (distribution a priori) et d'un ensemble de règles de mise à jour prédéfinies, peut adapter et améliorer les estimations initiales d'un processus \cite{pmlr-v238-ou24a}. La distribution de probabilité initiale est généralement définie comme une distribution spécifique de la famille des distributions Beta (équation \ref{fbeta}) avec des valeurs initiales prédéterminées pour $\alpha$ et $\beta$ \cite{math12111758}, \cite{NGUYEN2024111566}.	392	392	Le module de recommandation proposé, associé à AI-VT, est fondé sur le paradigme de l'apprentissage par renforcement. L'apprentissage par renforcement est une technique d'apprentissage automatique qui permet, par le biais d'actions et de récompenses, d'améliorer les connaissances du système sur une tâche spécifique \cite{NEURIPS2023_9d8cf124}. Nous nous intéressons ici plus particulièrement à l'échantillonnage de Thompson, qui, par le biais d'une distribution de probabilité initiale (distribution a priori) et d'un ensemble de règles de mise à jour prédéfinies, peut adapter et améliorer les estimations initiales d'un processus \cite{pmlr-v238-ou24a}. La distribution de probabilité initiale est généralement définie comme une distribution spécifique de la famille des distributions Beta (équation \ref{fbeta}) avec des valeurs initiales prédéterminées pour $\alpha$ et $\beta$ \cite{math12111758}, \cite{NGUYEN2024111566}.
	393	393
%\begin{equation}	394	394	%\begin{equation}
% Beta(x,\alpha,\beta)=\begin{cases}	395	395	% Beta(x,\alpha,\beta)=\begin{cases}
% \frac{(x^{\alpha -1})(1-x)^{\beta -1}}{\int_0^1(u^{\alpha -1})(1-u)^{\beta -1} du}&x \in [0, 1]\\	396	396	% \frac{(x^{\alpha -1})(1-x)^{\beta -1}}{\int_0^1(u^{\alpha -1})(1-u)^{\beta -1} du}&x \in [0, 1]\\
% 0&otherwise	397	397	% 0&otherwise
% \end{cases}	398	398	% \end{cases}
%\end{equation}	399	399	%\end{equation}
	400	400
\begin{equation}	401	401	\begin{equation}
Beta(\theta \| \alpha, \beta) = \frac{\Gamma(\alpha + \beta)}{\Gamma(\alpha) \Gamma(\beta)}\theta^{\alpha-1}(1-\theta)^{\beta-1}	402	402	Beta(\theta \| \alpha, \beta) = \frac{\Gamma(\alpha + \beta)}{\Gamma(\alpha) \Gamma(\beta)}\theta^{\alpha-1}(1-\theta)^{\beta-1}
\label{fbeta}	403	403	\label{fbeta}
\end{equation}	404	404	\end{equation}
	405	405
En utilisant la définition formelle de la fonction $\Gamma$ (équation \ref{eqGamma1}) et en remplaçant certaines variables, une nouvelle expression de la fonction Beta est obtenue (équation \ref{f2beta}).	406	406	En utilisant la définition formelle de la fonction $\Gamma$ (équation \ref{eqGamma1}) et en remplaçant certaines variables, une nouvelle expression de la fonction Beta est obtenue (équation \ref{f2beta}).
	407	407
\begin{equation}	408	408	\begin{equation}
\Gamma(z)=\int_0^\infty e^{-x} x^{z-1} dx	409	409	\Gamma(z)=\int_0^\infty e^{-x} x^{z-1} dx
\label{eqGamma1}	410	410	\label{eqGamma1}
\end{equation}	411	411	\end{equation}
	412	412
\begin{equation}	413	413	\begin{equation}
Beta(\theta \| \alpha, \beta) = \frac{\int_0^\infty e^{-s} s^{\alpha+\beta-1}ds}{\int_0^\infty e^{-u} u^{\alpha-1}du\int_0^\infty e^{-v} v^{\beta-1}dv}\theta^{\alpha-1}(1-\theta)^{\beta-1}	414	414	Beta(\theta \| \alpha, \beta) = \frac{\int_0^\infty e^{-s} s^{\alpha+\beta-1}ds}{\int_0^\infty e^{-u} u^{\alpha-1}du\int_0^\infty e^{-v} v^{\beta-1}dv}\theta^{\alpha-1}(1-\theta)^{\beta-1}
\label{f2beta}	415	415	\label{f2beta}
\end{equation}	416	416	\end{equation}
	417	417
En exprimant les deux intégrales du dénominateur comme une seule intégrale, l'équation \ref{f3Beta} est obtenue.	418	418	En exprimant les deux intégrales du dénominateur comme une seule intégrale, l'équation \ref{f3Beta} est obtenue.
	419	419
\begin{equation}	420	420	\begin{equation}
\int_{u=0}^{\infty}\int_{v=0}^\infty e^{-u-v} u^{\alpha-1} v^{\beta-1}du dv	421	421	\int_{u=0}^{\infty}\int_{v=0}^\infty e^{-u-v} u^{\alpha-1} v^{\beta-1}du dv
\label{f3Beta}	422	422	\label{f3Beta}
\end{equation}	423	423	\end{equation}
	424	424
$u=st$, $v=s(1-t)$, $s=u+v$ et $t=u/(u+v)$ sont ensuite remplacées par le résultat du Jacobien \ref{eqJac}, menant ainsi à l'expression finale définie par l'équation \ref{f4Beta}.	425	425	$u=st$, $v=s(1-t)$, $s=u+v$ et $t=u/(u+v)$ sont ensuite remplacées par le résultat du Jacobien \ref{eqJac}, menant ainsi à l'expression finale définie par l'équation \ref{f4Beta}.
	426	426
\begin{equation}	427	427	\begin{equation}
\left (	428	428	\left (
\begin{matrix}	429	429	\begin{matrix}
\frac{\partial u}{\partial t} & \frac{\partial u}{\partial s}\\	430	430	\frac{\partial u}{\partial t} & \frac{\partial u}{\partial s}\\
\frac{\partial v}{\partial t} & \frac{\partial v}{\partial s}\\	431	431	\frac{\partial v}{\partial t} & \frac{\partial v}{\partial s}\\
\end{matrix}	432	432	\end{matrix}
\right ) =	433	433	\right ) =
\left (	434	434	\left (
\begin{matrix}	435	435	\begin{matrix}
sdt & tds \\	436	436	sdt & tds \\
-sdt & (1-t)ds\\	437	437	-sdt & (1-t)ds\\
\end{matrix}	438	438	\end{matrix}
\right ) = s \; dtds	439	439	\right ) = s \; dtds
\label{eqJac}	440	440	\label{eqJac}
\end{equation}	441	441	\end{equation}
	442	442
\begin{equation}	443	443	\begin{equation}
\int_{s=0}^\infty \int_{t=0}^1 e^{-s}(st)^{\alpha-1}(s(1-t))^{\beta-1}s \; dsdt	444	444	\int_{s=0}^\infty \int_{t=0}^1 e^{-s}(st)^{\alpha-1}(s(1-t))^{\beta-1}s \; dsdt
\label{f4Beta}	445	445	\label{f4Beta}
\end{equation}	446	446	\end{equation}
	447	447
Viennent ensuite les équations \ref{f5Beta} et \ref{f6Beta} en exprimant les intégrales en fonction des variables de substitution indépendantes $s$ et $t$.	448	448	Viennent ensuite les équations \ref{f5Beta} et \ref{f6Beta} en exprimant les intégrales en fonction des variables de substitution indépendantes $s$ et $t$.
%\colorbox{yellow}{, elles expriment?}	449	449	%\colorbox{yellow}{, elles expriment?}
	450	450
\begin{equation}	451	451	\begin{equation}
\int_{s=0}^\infty e^{-s}s^{\alpha+\beta-1}ds \int_{t=0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt	452	452	\int_{s=0}^\infty e^{-s}s^{\alpha+\beta-1}ds \int_{t=0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt
\label{f5Beta}	453	453	\label{f5Beta}
\end{equation}	454	454	\end{equation}
	455	455
\begin{equation}	456	456	\begin{equation}
Beta(\theta \| \alpha, \beta) = \frac{\int_0^\infty e^{-s} s^{\alpha+\beta-1}ds}{\int_{s=0}^\infty e^{-s}s^{\alpha+\beta-1}ds \int_{t=0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt	457	457	Beta(\theta \| \alpha, \beta) = \frac{\int_0^\infty e^{-s} s^{\alpha+\beta-1}ds}{\int_{s=0}^\infty e^{-s}s^{\alpha+\beta-1}ds \int_{t=0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt
}\theta^{\alpha-1}(1-\theta)^{\beta-1}	458	458	}\theta^{\alpha-1}(1-\theta)^{\beta-1}
\label{f6Beta}	459	459	\label{f6Beta}
\end{equation}	460	460	\end{equation}
	461	461
Finalement, la famille de fonctions de distribution Beta peut être calculée selon l'équation \ref{f7Beta}.	462	462	Finalement, la famille de fonctions de distribution Beta peut être calculée selon l'équation \ref{f7Beta}.
	463	463
\begin{equation}	464	464	\begin{equation}
Beta(\theta \| \alpha, \beta) = \frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{\int_{0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt	465	465	Beta(\theta \| \alpha, \beta) = \frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{\int_{0}^1 t^{\alpha-1}(1-t)^{\beta-1}dt
}	466	466	}
\label{f7Beta}	467	467	\label{f7Beta}
\end{equation}	468	468	\end{equation}
	469	469
L'évolution de l'algorithme de recommandation TS résulte du changement des distributions de probabilité. Il est à noter qu'au moment de quantifier l'évolution, le changement et la variabilité doivent être calculés en fonction du temps. Les distributions de probabilités peuvent être comparées pour déterminer leur degré de similitude.	470	470	L'évolution de l'algorithme de recommandation TS résulte du changement des distributions de probabilité. Il est à noter qu'au moment de quantifier l'évolution, le changement et la variabilité doivent être calculés en fonction du temps. Les distributions de probabilités peuvent être comparées pour déterminer leur degré de similitude.
	471	471
Par ailleurs, l'apprentissage automatique utilise la divergence de Kullback-Liebler, qui décrit l'entropie relative de deux distributions de probabilités. Cette fonction est fondée sur le concept d'entropie et le résultat peut être interprété comme la quantité d'informations nécessaires pour obtenir la distribution de probabilité $q$ à partir de la distribution de probabilité $p$. Bien que largement utilisée, la divergence de Kullback-Liebler (équation \ref{dkl}) présente toutefois l'inconvénient de ne pas être une mesure symétrique car elle ne satisfait pas à l'inégalité triangulaire et n'est pas bornée \cite{Li_2024}. Pour remédier à cette difficulté, il est possible d'utiliser la divergence de Jensen-Shannon.	472	472	Par ailleurs, l'apprentissage automatique utilise la divergence de Kullback-Liebler, qui décrit l'entropie relative de deux distributions de probabilités. Cette fonction est fondée sur le concept d'entropie et le résultat peut être interprété comme la quantité d'informations nécessaires pour obtenir la distribution de probabilité $q$ à partir de la distribution de probabilité $p$. Bien que largement utilisée, la divergence de Kullback-Liebler (équation \ref{dkl}) présente toutefois l'inconvénient de ne pas être une mesure symétrique car elle ne satisfait pas à l'inégalité triangulaire et n'est pas bornée \cite{Li_2024}. Pour remédier à cette difficulté, il est possible d'utiliser la divergence de Jensen-Shannon.
	473	473
\begin{equation}	474	474	\begin{equation}
D_{KL}(p(x),q(x))=\int_{-\infty}^{\infty}p(x) log \left(\frac{p(x)}{q(x)} \right)dx	475	475	D_{KL}(p(x),q(x))=\int_{-\infty}^{\infty}p(x) log \left(\frac{p(x)}{q(x)} \right)dx
\label{dkl}	476	476	\label{dkl}
\end{equation}	477	477	\end{equation}
	478	478
La divergence de Jenser-Shannon est fondée sur la divergence de Kullback-Liebler. Une distribution de probabilité auxiliaire $m$ est créée dont la définition est fondée sur les distributions initiales $p$ et $q$ \cite{Kim2024}. L'équation \ref{djs} montre la définition formelle de la divergence de Jensen-Shannon, où $m(x)$ est une distribution de mélange de probabilités fondée sur $p(x)$ et $q(x)$. Celle-ci est calculée selon l'équation \ref{djs2}. Les distributions de probabilité à comparer doivent être continues et définies dans le même domaine.	479	479	La divergence de Jenser-Shannon est fondée sur la divergence de Kullback-Liebler. Une distribution de probabilité auxiliaire $m$ est créée dont la définition est fondée sur les distributions initiales $p$ et $q$ \cite{Kim2024}. L'équation \ref{djs} montre la définition formelle de la divergence de Jensen-Shannon, où $m(x)$ est une distribution de mélange de probabilités fondée sur $p(x)$ et $q(x)$. Celle-ci est calculée selon l'équation \ref{djs2}. Les distributions de probabilité à comparer doivent être continues et définies dans le même domaine.
	480	480
%Jensen-Shannon Divergence (equations \ref{djs}, \ref{djs2}).\\	481	481	%Jensen-Shannon Divergence (equations \ref{djs}, \ref{djs2}).\\
	482	482
\begin{equation}	483	483	\begin{equation}
D_{JS}(p(x),q(x))=\frac{1}{2}D_{KL}(p(x), m(x))+\frac{1}{2}D_{KL}(q(x), m(x))	484	484	D_{JS}(p(x),q(x))=\frac{1}{2}D_{KL}(p(x), m(x))+\frac{1}{2}D_{KL}(q(x), m(x))
\label{djs}	485	485	\label{djs}
\end{equation}	486	486	\end{equation}
	487	487
\begin{equation}	488	488	\begin{equation}
m(x)=\frac{1}{2}p(x)+\frac{1}{2}q(x)	489	489	m(x)=\frac{1}{2}p(x)+\frac{1}{2}q(x)
\label{djs2}	490	490	\label{djs2}
\end{equation}	491	491	\end{equation}
	492	492
La prédiction utilisée dans le module proposé ici a été présentée dans le chapitre \ref{ChapESCBR}. Il s'agit d'un algorithme d'empilement de raisonnement à partir de cas mettant en œuvre deux niveaux d'intégration. Le module utilise globalement la stratégie d'empilement pour exécuter plusieurs algorithmes afin de rechercher des informations dans un ensemble de données et générer des solutions à différents problèmes génériques. En outre une étape d'évaluation permet de sélectionner la solution la plus optimale pour un problème donné en fonction d'une métrique adaptative définie pour les problèmes de régression.	493	493	La prédiction utilisée dans le module proposé ici a été présentée dans le chapitre \ref{ChapESCBR}. Il s'agit d'un algorithme d'empilement de raisonnement à partir de cas mettant en œuvre deux niveaux d'intégration. Le module utilise globalement la stratégie d'empilement pour exécuter plusieurs algorithmes afin de rechercher des informations dans un ensemble de données et générer des solutions à différents problèmes génériques. En outre une étape d'évaluation permet de sélectionner la solution la plus optimale pour un problème donné en fonction d'une métrique adaptative définie pour les problèmes de régression.
	494	494
\subsection{Algorithme Proposé}	495	495	\subsection{Algorithme Proposé}
\label{Sec:TS-ESCBR-SMA}	496	496	\label{Sec:TS-ESCBR-SMA}
	497	497
Nous proposons ici une intégration de l'algorithme d'adaptation stochastique (fondé sur l'échantillonnage de Thompson) avec ESCBR-SMA. Ainsi, le module de recommandation révise la séance en fonction des notes de l'apprenant et ESCBR-SMA effectue une prédiction pour valider l'adaptation générée.	498	498	Nous proposons ici une intégration de l'algorithme d'adaptation stochastique (fondé sur l'échantillonnage de Thompson) avec ESCBR-SMA. Ainsi, le module de recommandation révise la séance en fonction des notes de l'apprenant et ESCBR-SMA effectue une prédiction pour valider l'adaptation générée.
	499	499
L'idée est d'unifier les deux modules en se fondant à la fois sur des informations locales (recommandation fondée sur l'échantillonnage de Thompson (TS) et les informations propres à l'apprenant), et sur des informations globales (cas similaires de la base de connaissances du système de RàPC en suivant le principe du paradoxe de Stein) car le RàPC combine différents observations pour réaliser une estimation.	500	500	L'idée est d'unifier les deux modules en se fondant à la fois sur des informations locales (recommandation fondée sur l'échantillonnage de Thompson (TS) et les informations propres à l'apprenant), et sur des informations globales (cas similaires de la base de connaissances du système de RàPC en suivant le principe du paradoxe de Stein) car le RàPC combine différents observations pour réaliser une estimation.
	501	501
L'architecture de l'algorithme est présentée sur la figure \ref{fig:Amodel}, où l'on peut voir que les deux algorithmes TS et RàPC sont exécutés en parallèle et indépendamment. Des synchronisations sont faites après obtention des résultats de chaque module. Ces résultats sont unifiés via d'une fonction de pondération. La recommandation finale est calculée selon l'équation \ref{eqMixModels_}. Le \textit{paradoxe de Simpson} est un paradoxe dans lequel un phénomène observé dans plusieurs groupes s'inverse lorsque les groupes sont combinés\cite{10.1145/3578337.3605122}. L'unification d'ensembles de données différents peut atténuer ce paradoxe \cite{lei2024analysis}.	502	502	L'architecture de l'algorithme est présentée sur la figure \ref{fig:Amodel}, où l'on peut voir que les deux algorithmes TS et RàPC sont exécutés en parallèle et indépendamment. Des synchronisations sont faites après obtention des résultats de chaque module. Ces résultats sont unifiés via d'une fonction de pondération. La recommandation finale est calculée selon l'équation \ref{eqMixModels_}. Le \textit{paradoxe de Simpson} est un paradoxe dans lequel un phénomène observé dans plusieurs groupes s'inverse lorsque les groupes sont combinés\cite{10.1145/3578337.3605122}. L'unification d'ensembles de données différents peut atténuer ce paradoxe \cite{lei2024analysis}.
	503	503
\begin{figure}	504	504	\begin{figure}
\centering	505	505	\centering
\includegraphics[width=0.7\linewidth]{Figures/Model.png}	506	506	\includegraphics[width=0.7\linewidth]{Figures/Model.png}
\caption{Schéma de l'architecture de l'algorithme proposé}	507	507	\caption{Schéma de l'architecture de l'algorithme proposé}
\label{fig:Amodel}	508	508	\label{fig:Amodel}
\end{figure}	509	509	\end{figure}
	510	510
La première étape est l'adaptation avec l'échantillonnage de Thompson. Vient ensuite la prédiction via ECBR-SMA. Enfin, le processus se termine par la prise de décision concernant la suite de la séance à délivrer à l'apprenant. Le système de recommandation obtient une valeur de probabilité pour tous les niveaux de complexité et l'ECBR-SMA évalue la proposition avec une prédiction pour chaque niveau de complexité. Le tableau \ref{tabvp} présente les variables et les paramètres du module proposé ainsi que les mesures employées.	511	511	La première étape est l'adaptation avec l'échantillonnage de Thompson. Vient ensuite la prédiction via ECBR-SMA. Enfin, le processus se termine par la prise de décision concernant la suite de la séance à délivrer à l'apprenant. Le système de recommandation obtient une valeur de probabilité pour tous les niveaux de complexité et l'ECBR-SMA évalue la proposition avec une prédiction pour chaque niveau de complexité. Le tableau \ref{tabvp} présente les variables et les paramètres du module proposé ainsi que les mesures employées.
	512	512
\begin{table}[!ht]	513	513	\begin{table}[!ht]
\centering	514	514	\centering
\footnotesize	515	515	\footnotesize
\begin{tabular}{c\|c\|>{\centering\arraybackslash}p{8cm}\|c}	516	516	\begin{tabular}{c\|c\|>{\centering\arraybackslash}p{8cm}\|c}
ID&Type&Description&Domain\\	517	517	ID&Type&Description&Domain\\
\hline	518	518	\hline
$\alpha$&p&Paramètre de la distribution beta&$[1, \infty] \in \mathbb{R}$\\	519	519	$\alpha$&p&Paramètre de la distribution beta&$[1, \infty] \in \mathbb{R}$\\
$\beta$&p&Paramètre de la distribution beta&$[1, \infty] \in \mathbb{R}$\\	520	520	$\beta$&p&Paramètre de la distribution beta&$[1, \infty] \in \mathbb{R}$\\
$t$&p&Numéro de l'itération&$\mathbb{N}$\\	521	521	$t$&p&Numéro de l'itération&$\mathbb{N}$\\
$c$&p&Niveau de complexité&$\mathbb{N}$\\	522	522	$c$&p&Niveau de complexité&$\mathbb{N}$\\
$x_c$&p&Notes moyennes par niveau de complexité $c$&$\mathbb{R}$\\	523	523	$x_c$&p&Notes moyennes par niveau de complexité $c$&$\mathbb{R}$\\
$y_c$&p&Nombre de questions par niveau de complexité $c$&$\mathbb{N}$\\	524	524	$y_c$&p&Nombre de questions par niveau de complexité $c$&$\mathbb{N}$\\
$r$&f&Fonction suivie pour la recommandation&$[0,1] \in \mathbb{R}$\\	525	525	$r$&f&Fonction suivie pour la recommandation&$[0,1] \in \mathbb{R}$\\
$k_{t,c}$&v&Évolution de la connaissance dans le temps $t$ pour le niveau de complexité $c$&$[0,1] \in \mathbb{R}$\\	526	526	$k_{t,c}$&v&Évolution de la connaissance dans le temps $t$ pour le niveau de complexité $c$&$[0,1] \in \mathbb{R}$\\
$vk_{t,c}$&v&Évolution de la connaissance pour chaque niveau de complexité $c$&$\mathbb{R}$\\	527	527	$vk_{t,c}$&v&Évolution de la connaissance pour chaque niveau de complexité $c$&$\mathbb{R}$\\
$TS_c$&v&Récompense d'échantillonnage de Thompson pour un niveau de complexité $c$&$[0,1] \in \mathbb{R}$\\	528	528	$TS_c$&v&Récompense d'échantillonnage de Thompson pour un niveau de complexité $c$&$[0,1] \in \mathbb{R}$\\
$TSN_c$&v&Normalisation de $TS_c$ avec d'autres niveaux de complexité&$[0,1] \in \mathbb{R}$\\	529	529	$TSN_c$&v&Normalisation de $TS_c$ avec d'autres niveaux de complexité&$[0,1] \in \mathbb{R}$\\
$ESCBR_c$&v&Prédiction de la note pour un niveau de complexité $c$&$\mathbb{R}_+$\\	530	530	$ESCBR_c$&v&Prédiction de la note pour un niveau de complexité $c$&$\mathbb{R}_+$\\
$p_c$&f&Fonction de densité de probabilité pour le niveau de complexité $c$&$\mathbb{R}_+$\\	531	531	$p_c$&f&Fonction de densité de probabilité pour le niveau de complexité $c$&$\mathbb{R}_+$\\
$D_{JS}$&f&Divergence de Jensen-Shannon&$[0,1] \in \mathbb{R}$\\	532	532	$D_{JS}$&f&Divergence de Jensen-Shannon&$[0,1] \in \mathbb{R}$\\
	533	533
\end{tabular}	534	534	\end{tabular}
\caption{Paramètres (p), variables (v) et fonctions (f) de l'algorithme proposé et des métriques utilisées}	535	535	\caption{Paramètres (p), variables (v) et fonctions (f) de l'algorithme proposé et des métriques utilisées}
\label{tabvp}	536	536	\label{tabvp}
\end{table}	537	537	\end{table}
	538	538
Pour rappel, le processus de recommandation se fait en trois étapes. Tout d'abord, il est nécessaire d'avoir des valeurs aléatoires pour chaque niveau de complexité $c$ en utilisant les distributions de probabilité générées avec le algorithme TS (équation \ref{IntEq1_}). Une fois que toutes les valeurs de probabilité correspondant à tous les niveaux de complexité ont été obtenues, la normalisation de toutes ces valeurs est calculée selon l'équation \ref{IntEq2_}. Les valeurs de normalisation servent de paramètres de priorité pour les prédictions effectuées par ESCBR-SMA (équation \ref{eqMixModels_}). La recommandation finalement proposée est celle dont la valeur est la plus élevée.	539	539	Pour rappel, le processus de recommandation se fait en trois étapes. Tout d'abord, il est nécessaire d'avoir des valeurs aléatoires pour chaque niveau de complexité $c$ en utilisant les distributions de probabilité générées avec le algorithme TS (équation \ref{IntEq1_}). Une fois que toutes les valeurs de probabilité correspondant à tous les niveaux de complexité ont été obtenues, la normalisation de toutes ces valeurs est calculée selon l'équation \ref{IntEq2_}. Les valeurs de normalisation servent de paramètres de priorité pour les prédictions effectuées par ESCBR-SMA (équation \ref{eqMixModels_}). La recommandation finalement proposée est celle dont la valeur est la plus élevée.
	540	540
\begin{equation}	541	541	\begin{equation}
TS_c=rand(Beta(\alpha_c, \beta_c))	542	542	TS_c=rand(Beta(\alpha_c, \beta_c))
\label{IntEq1_}	543	543	\label{IntEq1_}
\end{equation}	544	544	\end{equation}
	545	545
\begin{equation}	546	546	\begin{equation}
TSN_c=\frac{TS_c}{\sum_{i=0}^4TS_i}	547	547	TSN_c=\frac{TS_c}{\sum_{i=0}^4TS_i}
\label{IntEq2_}	548	548	\label{IntEq2_}
\end{equation}	549	549	\end{equation}
	550	550
\begin{equation}	551	551	\begin{equation}
n_c=argmax_c(TSN_c*ESCBR_c)	552	552	n_c=argmax_c(TSN_c*ESCBR_c)
\label{eqMixModels_}	553	553	\label{eqMixModels_}
\end{equation}	554	554	\end{equation}
	555	555
\subsection{Résultats et Discussion}	556	556	\subsection{Résultats et Discussion}
	557	557
Le principal inconvénient posé par la validation d'un tel système « en situation réelle » est la difficulté à collecter des données et à évaluer des systèmes différents dans des conditions strictement similaires. Cette difficulté est accentuée dans les contextes d'apprentissage autorégulés, puisque les apprenants peuvent quitter la plateforme d'apprentissage à tout moment rendant ainsi les données incomplètes \cite{badier:hal-04092828}.	558	558	Le principal inconvénient posé par la validation d'un tel système « en situation réelle » est la difficulté à collecter des données et à évaluer des systèmes différents dans des conditions strictement similaires. Cette difficulté est accentuée dans les contextes d'apprentissage autorégulés, puisque les apprenants peuvent quitter la plateforme d'apprentissage à tout moment rendant ainsi les données incomplètes \cite{badier:hal-04092828}.
	559	559
Pour cette raison, les différentes approches proposées ont été testées sur des données générées : les notes et les temps de réponse de 1000 apprenants fictifs et cinq questions par niveau de complexité. Les notes des apprenants ont été créées en suivant la loi de distribution \textit{logit-normale} que nous avons jugée proche de la réalité de la progression d'un apprentissage.\colorbox{yellow}{lien vers la base générée? ==> Ref de l'url du git}	560	560	Pour cette raison, les différentes approches proposées ont été testées sur des données générées : les notes et les temps de réponse de 1000 apprenants fictifs et cinq questions par niveau de complexité. Les notes des apprenants ont été créées en suivant la loi de distribution \textit{logit-normale} que nous avons jugée proche de la réalité de la progression d'un apprentissage \cite{Data}.
	561	561
Quatre séries de tests ont été effectuées. La première série a été menée sur le système AI-VT intégrant le système de RàPC pour la régression afin de démontrer la capacité de l'algorithme à prédire les notes à différents niveaux de complexité.	562	562	Quatre séries de tests ont été effectuées. La première série a été menée sur le système AI-VT intégrant le système de RàPC pour la régression afin de démontrer la capacité de l'algorithme à prédire les notes à différents niveaux de complexité.
La deuxième série de tests a évalué la progression des connaissances avec TS afin d'analyser la capacité du module à proposer des recommandations personnalisées. Lors de la troisième série de tests, nous avons comparé les algorithmes de recommandation BKT et TS. Enfin, lors de la quatrième série de tests, nous avons comparé TS seul et TS avec ESCBR-SMA.	563	563	La deuxième série de tests a évalué la progression des connaissances avec TS afin d'analyser la capacité du module à proposer des recommandations personnalisées. Lors de la troisième série de tests, nous avons comparé les algorithmes de recommandation BKT et TS. Enfin, lors de la quatrième série de tests, nous avons comparé TS seul et TS avec ESCBR-SMA.
	564	564
\subsubsection{Régression avec ESCBR-SMA pour l'aide à l'apprentissage humain}	565	565	\subsubsection{Régression avec ESCBR-SMA pour l'aide à l'apprentissage humain}
	566	566
Le SMA que nous avons implémenté utilise un raisonnement bayésien, ce qui permet aux agents d'apprendre des données et d'intéragir au cours de l'exécution et de l'exploration.	567	567	Le SMA que nous avons implémenté utilise un raisonnement bayésien, ce qui permet aux agents d'apprendre des données et d'intéragir au cours de l'exécution et de l'exploration.
	568	568
ESCBR-SMA utilise une fonction noyau pour obtenir la meilleure approximation de la solution du problème cible. Dans notre cas, l'obtention de la meilleure solution est un problème NP-Difficile car la formulation est similaire au problème de Fermat-Weber à $N$ dimensions \cite{doi:10.1137/23M1592420}.	569	569	ESCBR-SMA utilise une fonction noyau pour obtenir la meilleure approximation de la solution du problème cible. Dans notre cas, l'obtention de la meilleure solution est un problème NP-Difficile car la formulation est similaire au problème de Fermat-Weber à $N$ dimensions \cite{doi:10.1137/23M1592420}.
	570	570
Les différents scénarios du tableau \ref{tab:scenarios} ont été considérés dans un premier temps. Dans le scénario $E_1$, il s'agit de prédire la note d'un apprenant au premier niveau de complexité, après 3 questions. Le scénario $E_2$ considère les notes de 8 questions et l'objectif est de prédire la note de la neuvième question dans le même niveau de complexité. Le scénario $E_3$ interpole la neuvième note que l'apprenant obtiendrait si la neuvième question était de niveau de complexité supérieur à celui de la huitième question. Cette interpolation est faite sur la base des notes obtenues aux quatre questions précédentes. Le scénario $E_4$ considère 4 questions et le système doit interpoler 2 notes dans un niveau de complexité supérieur.	571	571	Les différents scénarios du tableau \ref{tab:scenarios} ont été considérés dans un premier temps. Dans le scénario $E_1$, il s'agit de prédire la note d'un apprenant au premier niveau de complexité, après 3 questions. Le scénario $E_2$ considère les notes de 8 questions et l'objectif est de prédire la note de la neuvième question dans le même niveau de complexité. Le scénario $E_3$ interpole la neuvième note que l'apprenant obtiendrait si la neuvième question était de niveau de complexité supérieur à celui de la huitième question. Cette interpolation est faite sur la base des notes obtenues aux quatre questions précédentes. Le scénario $E_4$ considère 4 questions et le système doit interpoler 2 notes dans un niveau de complexité supérieur.
	572	572
\begin{table}[!ht]	573	573	\begin{table}[!ht]
\centering	574	574	\centering
\begin{tabular}{ccc}	575	575	\begin{tabular}{ccc}
Scenario&Caractéristiques du problème&Dimension de la solution\\	576	576	Scenario&Caractéristiques du problème&Dimension de la solution\\
\hline	577	577	\hline
$E_1$ & 5 & 1\\	578	578	$E_1$ & 5 & 1\\
$E_2$ & 15& 1\\	579	579	$E_2$ & 15& 1\\
$E_3$ & 9 & 1\\	580	580	$E_3$ & 9 & 1\\
$E_4$ & 9 & 2\\	581	581	$E_4$ & 9 & 2\\
\end{tabular}	582	582	\end{tabular}
\caption{Description des scénarios}	583	583	\caption{Description des scénarios}
\label{tab:scenarios}	584	584	\label{tab:scenarios}
\end{table}	585	585	\end{table}
	586	586
ESCBR-SMA a été comparé aux neuf outils classiquement utilisés pour résoudre la régression consignés dans le tableau \ref{tabAlgs} et selon l'erreur quadratique moyenne (RMSE - \textit{Root Mean Squared Error}), l'erreur médiane absolue (MedAE - \textit{Median Absolute Error}) et l'erreur moyenne absolue (MAE - \textit{Mean Absolute Error}).	587	587	ESCBR-SMA a été comparé aux neuf outils classiquement utilisés pour résoudre la régression consignés dans le tableau \ref{tabAlgs} et selon l'erreur quadratique moyenne (RMSE - \textit{Root Mean Squared Error}), l'erreur médiane absolue (MedAE - \textit{Median Absolute Error}) et l'erreur moyenne absolue (MAE - \textit{Mean Absolute Error}).
	588	588
\begin{table}[!ht]	589	589	\begin{table}[!ht]
\centering	590	590	\centering
\footnotesize	591	591	\footnotesize
\begin{tabular}{ll\|ll}	592	592	\begin{tabular}{ll\|ll}
ID&Algorithm&ID&Algorithm\\	593	593	ID&Algorithm&ID&Algorithm\\
\hline	594	594	\hline
A1&Linear Regression&A6&Polinomial Regression\\	595	595	A1&Linear Regression&A6&Polinomial Regression\\
A2&K-Nearest Neighbor&A7&Ridge Regression\\	596	596	A2&K-Nearest Neighbor&A7&Ridge Regression\\
A3&Decision Tree&A8&Lasso Regression\\	597	597	A3&Decision Tree&A8&Lasso Regression\\
A4&Random Forest (Ensemble)&A9&Gradient Boosting (Ensemble)\\	598	598	A4&Random Forest (Ensemble)&A9&Gradient Boosting (Ensemble)\\
A5&Multi Layer Perceptron&A10&Proposed Ensemble Stacking RàPC\\	599	599	A5&Multi Layer Perceptron&A10&Proposed Ensemble Stacking RàPC\\
\end{tabular}	600	600	\end{tabular}
\caption{Liste des algorithmes évalués }	601	601	\caption{Liste des algorithmes évalués }
\label{tabAlgs}	602	602	\label{tabAlgs}
\end{table}	603	603	\end{table}
	604	604
Le tableau \ref{tab:results} présente les résultats obtenus par les 10 algorithmes sur les quatre scénarios. Ces résultats montrent qu'ESCBR-SMA (A10) et le \textit{Gradient Boosting} (A9) obtiennent toujours les deux meilleurs résultats. Si l'on considère uniquement la RMSE, ESCBR-SMA occupe toujours la première place sauf pour $E_3$ où il est deuxième. Inversement, en considérant l'erreur médiane absolue ou l'erreur moyenne absolue, A10 se classe juste après A9. ESCBR-SMA et le \textit{Gradient Boosting} sont donc efficaces pour interpoler les notes des apprenants.	605	605	Le tableau \ref{tab:results} présente les résultats obtenus par les 10 algorithmes sur les quatre scénarios. Ces résultats montrent qu'ESCBR-SMA (A10) et le \textit{Gradient Boosting} (A9) obtiennent toujours les deux meilleurs résultats. Si l'on considère uniquement la RMSE, ESCBR-SMA occupe toujours la première place sauf pour $E_3$ où il est deuxième. Inversement, en considérant l'erreur médiane absolue ou l'erreur moyenne absolue, A10 se classe juste après A9. ESCBR-SMA et le \textit{Gradient Boosting} sont donc efficaces pour interpoler les notes des apprenants.
	606	606
\begin{table}[!ht]	607	607	\begin{table}[!ht]
\centering	608	608	\centering
\footnotesize	609	609	\footnotesize
\begin{tabular}{c\|cccccccccc}	610	610	\begin{tabular}{c\|cccccccccc}
&\multicolumn{10}{c}{\textbf{Algorithme}}\\	611	611	&\multicolumn{10}{c}{\textbf{Algorithme}}\\
\hline	612	612	\hline
& A1&A2&A3&A4&A5&A6&A7&A8&A9&A10\\	613	613	& A1&A2&A3&A4&A5&A6&A7&A8&A9&A10\\
\textbf{Scenario (Metrique)}\\	614	614	\textbf{Scenario (Metrique)}\\
\hline	615	615	\hline
$E_1$ (RMSE)&0.625&0.565&0.741&0.56&0.606&0.626&0.626&0.681&0.541&\textbf{0.54}\\	616	616	$E_1$ (RMSE)&0.625&0.565&0.741&0.56&0.606&0.626&0.626&0.681&0.541&\textbf{0.54}\\
$E_1$ (MedAE) & 0.387&0.35&0.46&0.338&0.384&0.387&0.387&0.453&\textbf{0.327}&0.347\\	617	617	$E_1$ (MedAE) & 0.387&0.35&0.46&0.338&0.384&0.387&0.387&0.453&\textbf{0.327}&0.347\\
$E_1$ (MAE) &0.485&0.436&0.572&0.429&0.47&0.485&0.485&0.544&\textbf{0.414}&0.417\\	618	618	$E_1$ (MAE) &0.485&0.436&0.572&0.429&0.47&0.485&0.485&0.544&\textbf{0.414}&0.417\\
\hline	619	619	\hline
$E_2$ (RMSE)& 0.562&0.588&0.78&0.571&0.61&0.562&0.562&0.622&0.557&\textbf{0.556}\\	620	620	$E_2$ (RMSE)& 0.562&0.588&0.78&0.571&0.61&0.562&0.562&0.622&0.557&\textbf{0.556}\\
$E_2$ (MedAE)&0.351&0.357&0.464&0.344&0.398&0.351&0.351&0.415&\textbf{0.334}&0.346\\	621	621	$E_2$ (MedAE)&0.351&0.357&0.464&0.344&0.398&0.351&0.351&0.415&\textbf{0.334}&0.346\\
$E_2$ (MAE)&0.433&0.448&0.591&0.437&0.478&0.433&0.433&0.495&\textbf{0.422}&0.429\\	622	622	$E_2$ (MAE)&0.433&0.448&0.591&0.437&0.478&0.433&0.433&0.495&\textbf{0.422}&0.429\\
\hline	623	623	\hline
$E_3$ (RMSE)&0.591&0.59&0.79&0.57&0.632&0.591&0.591&0.644&\textbf{0.555}&0.558\\	624	624	$E_3$ (RMSE)&0.591&0.59&0.79&0.57&0.632&0.591&0.591&0.644&\textbf{0.555}&0.558\\
$E_3$ (MedAE)&0.367&0.362&0.474&0.358&0.404&0.367&0.367&0.433&\textbf{0.336}&0.349\\	625	625	$E_3$ (MedAE)&0.367&0.362&0.474&0.358&0.404&0.367&0.367&0.433&\textbf{0.336}&0.349\\
$E_3$ (MAE)&0.453&0.45&0.598&0.441&0.49&0.453&0.453&0.512&\textbf{0.427}&0.43\\	626	626	$E_3$ (MAE)&0.453&0.45&0.598&0.441&0.49&0.453&0.453&0.512&\textbf{0.427}&0.43\\
\hline	627	627	\hline
$E_4$ (RMSE)&0.591&0.589&0.785&0.568&0.613&0.591&0.591&0.644&0.554&\textbf{0.549}\\	628	628	$E_4$ (RMSE)&0.591&0.589&0.785&0.568&0.613&0.591&0.591&0.644&0.554&\textbf{0.549}\\
$E_4$ (MedAE)&0.367&0.362&0.465&0.57&0.375&0.367&0.367&0.433&\textbf{0.336}&0.343\\	629	629	$E_4$ (MedAE)&0.367&0.362&0.465&0.57&0.375&0.367&0.367&0.433&\textbf{0.336}&0.343\\
$E_4$ (MAE)&0.453&0.45&0.598&0.438&0.466&0.453&0.453&0.512&0.426&\textbf{0.417}\\	630	630	$E_4$ (MAE)&0.453&0.45&0.598&0.438&0.466&0.453&0.453&0.512&0.426&\textbf{0.417}\\
\end{tabular}	631	631	\end{tabular}
\caption{Erreurs moyennes et médianes des interpolations des 10 algorithmes sélectionnés sur les 4 scénarios considérés et obtenues après 100 exécutions.}	632	632	\caption{Erreurs moyennes et médianes des interpolations des 10 algorithmes sélectionnés sur les 4 scénarios considérés et obtenues après 100 exécutions.}
\label{tab:results}	633	633	\label{tab:results}
\end{table}	634	634	\end{table}
	635	635
\subsubsection{Progression des connaissances}	636	636	\subsubsection{Progression des connaissances}
	637	637
L'algorithme de recommandation TS est fondé sur le paradigme bayésien le rendant ainsi particulièrement adapté au problèmes liés à la limitation de la quantité de données et à une incertitude forte. Afin de quantifier la connaissance et de voir sa progression dans le temps avec TS, la divergence de Jensen-Shannon avec la famille de distribution Beta en $t$ et $t-1$ a été mesurée. L'équation \ref{eqprog1} décrit formellement le calcul à effectuer avec les distributions de probabilité en un temps $t$ pour un niveau de complexité $c$, en utilisant la définition $m$ (équation \ref{eqprog2}).	638	638	L'algorithme de recommandation TS est fondé sur le paradigme bayésien le rendant ainsi particulièrement adapté au problèmes liés à la limitation de la quantité de données et à une incertitude forte. Afin de quantifier la connaissance et de voir sa progression dans le temps avec TS, la divergence de Jensen-Shannon avec la famille de distribution Beta en $t$ et $t-1$ a été mesurée. L'équation \ref{eqprog1} décrit formellement le calcul à effectuer avec les distributions de probabilité en un temps $t$ pour un niveau de complexité $c$, en utilisant la définition $m$ (équation \ref{eqprog2}).
	639	639
%\begin{equation}	640	640	%\begin{equation}
\begin{multline}	641	641	\begin{multline}
k_{t,c}=\frac{1}{2}	642	642	k_{t,c}=\frac{1}{2}
\int_{0}^{1}p_c(\alpha_t,\beta_t,x) log \left(\frac{p_c(\alpha_t,\beta_t,x)}{m(p_c(\alpha_{t-1},\beta_{t-1},x),p_c(\alpha_t,\beta_t,x))} \right)dx	643	643	\int_{0}^{1}p_c(\alpha_t,\beta_t,x) log \left(\frac{p_c(\alpha_t,\beta_t,x)}{m(p_c(\alpha_{t-1},\beta_{t-1},x),p_c(\alpha_t,\beta_t,x))} \right)dx
\\	644	644	\\
+\frac{1}{2}	645	645	+\frac{1}{2}
\int_{0}^{1}p_c(\alpha_{t-1},\beta_{t-1},x) log \left(\frac{p_c(\alpha_{t-1},\beta_{t-1},x)}{m(p_c(\alpha_{t-1},\beta_{t-1},x),p_c(\alpha_t,\beta_t,x))} \right)dx	646	646	\int_{0}^{1}p_c(\alpha_{t-1},\beta_{t-1},x) log \left(\frac{p_c(\alpha_{t-1},\beta_{t-1},x)}{m(p_c(\alpha_{t-1},\beta_{t-1},x),p_c(\alpha_t,\beta_t,x))} \right)dx
\label{eqprog1}	647	647	\label{eqprog1}
\end{multline}	648	648	\end{multline}
%\end{equation}	649	649	%\end{equation}
	650	650
\begin{multline}	651	651	\begin{multline}
m(p(\alpha_{(t-1)},\beta_{(t-1)},x),p(\alpha_{t},\beta_{t},x))=\frac{1}{2} \left( \frac{x^{\alpha_{(t-1)}-1}(1-x)^{\beta_{(t-1)}-1}}{\int_0^1 u^{\alpha_{(t-1)}-1}(1-u^{\beta_{(t-1)}-1})du} \right )\\	652	652	m(p(\alpha_{(t-1)},\beta_{(t-1)},x),p(\alpha_{t},\beta_{t},x))=\frac{1}{2} \left( \frac{x^{\alpha_{(t-1)}-1}(1-x)^{\beta_{(t-1)}-1}}{\int_0^1 u^{\alpha_{(t-1)}-1}(1-u^{\beta_{(t-1)}-1})du} \right )\\
+\frac{1}{2} \left (\frac{x^{\alpha_{t}-1}(1-x)^{\beta_{t}-1}}{\int_0^1 u^{\alpha_{t}-1}(1-u^{\beta_{t}-1})du} \right )	653	653	+\frac{1}{2} \left (\frac{x^{\alpha_{t}-1}(1-x)^{\beta_{t}-1}}{\int_0^1 u^{\alpha_{t}-1}(1-u^{\beta_{t}-1})du} \right )
%\end{equation}	654	654	%\end{equation}
\label{eqprog2}	655	655	\label{eqprog2}
\end{multline}	656	656	\end{multline}
	657	657
La progression du nombre total de connaissances en $t$ est la somme des différences entre $t$ et $t-1$ pour tous les $c$ niveaux de complexité calculés avec la divergence de Jensen-Shannon (équation \ref{eqTEK}). Pour ce faire, nous évaluons la progression de la variabilité données par équation \ref{eqVarP}.	658	658	La progression du nombre total de connaissances en $t$ est la somme des différences entre $t$ et $t-1$ pour tous les $c$ niveaux de complexité calculés avec la divergence de Jensen-Shannon (équation \ref{eqTEK}). Pour ce faire, nous évaluons la progression de la variabilité données par équation \ref{eqVarP}.
	659	659
\begin{equation}	660	660	\begin{equation}
vk_{t,c}=\begin{cases}	661	661	vk_{t,c}=\begin{cases}
D_{JS}(Beta(\alpha_{t,c},\beta_{t,c}), Beta(\alpha_{t+1,c},\beta_{t+1,c})), & \frac{\alpha_{t,c}}{\alpha_{t,c}+\beta_{t,c}} < \frac{\alpha_{t+1,c}}{\alpha_{t+1,c}+\beta_{t+1,c}}\\	662	662	D_{JS}(Beta(\alpha_{t,c},\beta_{t,c}), Beta(\alpha_{t+1,c},\beta_{t+1,c})), & \frac{\alpha_{t,c}}{\alpha_{t,c}+\beta_{t,c}} < \frac{\alpha_{t+1,c}}{\alpha_{t+1,c}+\beta_{t+1,c}}\\
-D_{JS}(Beta(\alpha_{t,c},\beta_{t,c}), Beta(\alpha_{t+1,c},\beta_{t+1,c})),& Otherwise	663	663	-D_{JS}(Beta(\alpha_{t,c},\beta_{t,c}), Beta(\alpha_{t+1,c},\beta_{t+1,c})),& Otherwise
\end{cases}	664	664	\end{cases}
\label{eqVarP}	665	665	\label{eqVarP}
\end{equation}	666	666	\end{equation}
	667	667
\begin{equation}	668	668	\begin{equation}
k_t=\sum_{c=4}^{c=0 \lor k_t \neq 0}	669	669	k_t=\sum_{c=4}^{c=0 \lor k_t \neq 0}
\begin{cases}	670	670	\begin{cases}
\alpha_{c-1} vk_{t,c-1};&vk_{t,c} > 0\\	671	671	\alpha_{c-1} vk_{t,c-1};&vk_{t,c} > 0\\
0;&Otherwise	672	672	0;&Otherwise
\end{cases}	673	673	\end{cases}
\label{eqTEK}	674	674	\label{eqTEK}
\end{equation}	675	675	\end{equation}
	676	676
\begin{figure}[!ht]	677	677	\begin{figure}[!ht]
\centering	678	678	\centering
\includegraphics[width=\textwidth]{Figures/kEvol_TS.jpg}	679	679	\includegraphics[width=\textwidth]{Figures/kEvol_TS.jpg}
\caption{Progression des connaissances avec l'échantillonnage de Thompson selon la divergence de Jensen-Shannon}	680	680	\caption{Progression des connaissances avec l'échantillonnage de Thompson selon la divergence de Jensen-Shannon}
\label{fig:evolution}	681	681	\label{fig:evolution}
\end{figure}	682	682	\end{figure}
	683	683
La figure \ref{fig:evolution} montre la progression cumulée des connaissances sur les quinze questions d'une même séance d'entrainement. L'augmentation de la moyenne du niveau de connaissance entre la première et la dernière question de la même séance montre que tous les apprenants ont statistiquement augmenté leur niveau de connaissance. La variabilité augmente à partir de la première question jusqu'à la question neuf, où le système a acquis plus d'informations sur les apprenants. À ce stade, la variabilité diminue et la moyenne augmente.	684	684	La figure \ref{fig:evolution} montre la progression cumulée des connaissances sur les quinze questions d'une même séance d'entrainement. L'augmentation de la moyenne du niveau de connaissance entre la première et la dernière question de la même séance montre que tous les apprenants ont statistiquement augmenté leur niveau de connaissance. La variabilité augmente à partir de la première question jusqu'à la question neuf, où le système a acquis plus d'informations sur les apprenants. À ce stade, la variabilité diminue et la moyenne augmente.
	685	685
\subsubsection{Système de recommandation avec un jeu de données d'étudiants réels}	686	686	\subsubsection{Système de recommandation avec un jeu de données d'étudiants réels}
	687	687
Le système de recommandation TS a été testé avec un ensemble de données adaptées extraites de données réelles d'interactions d'étudiants avec un environnement d'apprentissage virtuel pour différents cours \cite{Kuzilek2017}. Cet ensemble contient les notes de $23366$ apprenants dans différents cours. Les apprenants ont été évalués selon différentes modalités (partiels, projets, QCM). Cet ensemble de données a pu être intégré au jeu de données d'AI-VT (notes, temps de réponse et 5 niveaux de complexité). Le test a consisté à générer une recommandation pour l'avant dernière question en fonction des notes précédentes. Ce test a été exécuté 100 fois pour chaque apprenant. Les nombres de questions recommandées sont reportés sur la figure \ref{fig:stabilityBP} pour chaque niveau de complexité. Celle-ci montre que malgré la stochasticité, la variance globale dans tous les niveaux de complexité est faible en fonction du nombre total d'apprenants et du nombre total de recommandations, et démontre ainsi la stabilité de l'algorithme.\\	688	688	Le système de recommandation TS a été testé avec un ensemble de données adaptées extraites de données réelles d'interactions d'étudiants avec un environnement d'apprentissage virtuel pour différents cours \cite{Kuzilek2017}. Cet ensemble contient les notes de $23366$ apprenants dans différents cours. Les apprenants ont été évalués selon différentes modalités (partiels, projets, QCM) \cite{Data}. Cet ensemble de données a pu être intégré au jeu de données d'AI-VT (notes, temps de réponse et 5 niveaux de complexité). Le test a consisté à générer une recommandation pour l'avant dernière question en fonction des notes précédentes. Ce test a été exécuté 100 fois pour chaque apprenant. Les nombres de questions recommandées sont reportés sur la figure \ref{fig:stabilityBP} pour chaque niveau de complexité. Celle-ci montre que malgré la stochasticité, la variance globale dans tous les niveaux de complexité est faible en fonction du nombre total d'apprenants et du nombre total de recommandations, et démontre ainsi la stabilité de l'algorithme.\\
	689	689
\begin{figure}[!ht]	690	690	\begin{figure}[!ht]
\centering	691	691	\centering
\includegraphics[width=1\linewidth]{Figures/stabilityBoxplot.png}	692	692	\includegraphics[width=1\linewidth]{Figures/stabilityBoxplot.png}
\caption{Nombre de recommandations par niveau de complexité}	693	693	\caption{Nombre de recommandations par niveau de complexité}
\label{fig:stabilityBP}	694	694	\label{fig:stabilityBP}
\end{figure}	695	695	\end{figure}
	696	696
La précision de la recommandation pour tous les apprenants est évaluée en considérant comme comportement correct deux états : i) l'algorithme recommande un niveau où l'apprenant a une note supérieure ou égal à 6 et ii) l'algorithme recommande un niveau inférieur au niveau réel évalué par l'apprenant. Le premier cas montre que l'algorithme a identifié le moment précis où l'apprenant doit augmenter le niveau de complexité, le second cas permet d'établir que l'algorithme propose de renforcer un niveau de complexité plus faible. Puis la précision est calculée comme le rapport entre le nombre d'états correspondant aux comportements corrects définis et le nombre total de recommandations. La figure \ref{fig:precision} montre les résultats de cette métrique après 100 exécutions.\\	697	697	La précision de la recommandation pour tous les apprenants est évaluée en considérant comme comportement correct deux états : i) l'algorithme recommande un niveau où l'apprenant a une note supérieure ou égal à 6 et ii) l'algorithme recommande un niveau inférieur au niveau réel évalué par l'apprenant. Le premier cas montre que l'algorithme a identifié le moment précis où l'apprenant doit augmenter le niveau de complexité, le second cas permet d'établir que l'algorithme propose de renforcer un niveau de complexité plus faible. Puis la précision est calculée comme le rapport entre le nombre d'états correspondant aux comportements corrects définis et le nombre total de recommandations. La figure \ref{fig:precision} montre les résultats de cette métrique après 100 exécutions.\\
	698	698
\begin{figure}[!ht]	699	699	\begin{figure}[!ht]
\centering	700	700	\centering
\includegraphics[width=1\linewidth]{Figures/precision.png}	701	701	\includegraphics[width=1\linewidth]{Figures/precision.png}
\caption{Précision de la recommandation}	702	702	\caption{Précision de la recommandation}
\label{fig:precision}	703	703	\label{fig:precision}
\end{figure}	704	704	\end{figure}
	705	705
\subsubsection{Comparaison entre TS et BKT}	706	706	\subsubsection{Comparaison entre TS et BKT}
	707	707
La figure \ref{fig:EvGrades} permet de comparer la recommandation fondée sur l'échantillonnage de Thompson et celle fondée sur BKT. Cette figure montre l'évolution des notes des apprenants en fonction du nombre de questions auxquelles ils répondent dans la même séance. Dans ce cas, le TS génère moins de variabilité que BKT, mais les évolutions induites par les deux systèmes restent globalement très similaires.	708	708	La figure \ref{fig:EvGrades} permet de comparer la recommandation fondée sur l'échantillonnage de Thompson et celle fondée sur BKT. Cette figure montre l'évolution des notes des apprenants en fonction du nombre de questions auxquelles ils répondent dans la même séance. Dans ce cas, le TS génère moins de variabilité que BKT, mais les évolutions induites par les deux systèmes restent globalement très similaires.
	709	709
\begin{figure}[!ht]	710	710	\begin{figure}[!ht]
\centering	711	711	\centering
\includegraphics[width=\textwidth]{Figures/GradesEv.jpg}	712	712	\includegraphics[width=\textwidth]{Figures/GradesEv.jpg}
\caption{Comparaison de l'évolution des notes entre les systèmes fondés sur TS et BKT.}	713	713	\caption{Comparaison de l'évolution des notes entre les systèmes fondés sur TS et BKT.}
\label{fig:EvGrades}	714	714	\label{fig:EvGrades}
\end{figure}	715	715	\end{figure}
	716	716
Toutefois, si l'on considère l'évolution du niveau de complexité recommandé (figure \ref{fig:EvCL}), TS fait évoluer le niveau de complexité des apprenants, alors que BKT a tendance à laisser les apprenants au même niveau de complexité. Autrement dit, avec BKT, il est difficile d'apprendre de nouveaux sujets ou des concepts plus complexes au sein du même domaine. En comparant les résultats des deux figures (figures \ref{fig:EvGrades} et \ref{fig:EvCL}), TS permet de faire progresser la moyenne des notes et facilite l'évolution des niveaux de complexité.	717	717	Toutefois, si l'on considère l'évolution du niveau de complexité recommandé (figure \ref{fig:EvCL}), TS fait évoluer le niveau de complexité des apprenants, alors que BKT a tendance à laisser les apprenants au même niveau de complexité. Autrement dit, avec BKT, il est difficile d'apprendre de nouveaux sujets ou des concepts plus complexes au sein du même domaine. En comparant les résultats des deux figures (figures \ref{fig:EvGrades} et \ref{fig:EvCL}), TS permet de faire progresser la moyenne des notes et facilite l'évolution des niveaux de complexité.
	718	718
\begin{figure}[!ht]	719	719	\begin{figure}[!ht]
\centering	720	720	\centering
\includegraphics[width=\textwidth]{Figures/LevelsEv.jpg}	721	721	\includegraphics[width=\textwidth]{Figures/LevelsEv.jpg}
\caption{Comparaison de l'évolution des niveaux entre les systèmes de recommandation fondés sur BKT et TS}	722	722	\caption{Comparaison de l'évolution des niveaux entre les systèmes de recommandation fondés sur BKT et TS}
\label{fig:EvCL}	723	723	\label{fig:EvCL}

@article{ZHANG2021100025,	1	1	@article{ZHANG2021100025,
title = {AI technologies for education: Recent research and future directions},	2	2	title = {AI technologies for education: Recent research and future directions},
journal = {Computers and Education: Artificial Intelligence},	3	3	journal = {Computers and Education: Artificial Intelligence},
volume = {2},	4	4	volume = {2},
pages = {100025},	5	5	pages = {100025},
language = {English},	6	6	language = {English},
year = {2021},	7	7	year = {2021},
issn = {2666-920X},	8	8	issn = {2666-920X},
type = {article},	9	9	type = {article},
doi = {https://doi.org/10.1016/j.caeai.2021.100025},	10	10	doi = {https://doi.org/10.1016/j.caeai.2021.100025},
url = {https://www.sciencedirect.com/science/article/pii/S2666920X21000199},	11	11	url = {https://www.sciencedirect.com/science/article/pii/S2666920X21000199},
author = {Ke Zhang. and Ayse Begum Aslan},	12	12	author = {Ke Zhang. and Ayse Begum Aslan},
address={USA},	13	13	address={USA},
affiliation={Wayne State University; Eastern Michigan University},	14	14	affiliation={Wayne State University; Eastern Michigan University},
keywords = {Artificial intelligence, AI, AI in Education},	15	15	keywords = {Artificial intelligence, AI, AI in Education},
abstract = {From unique educational perspectives, this article reports a comprehensive review of selected empirical studies on artificial intelligence in education (AIEd) published in 1993–2020, as collected in the Web of Sciences database and selected AIEd-specialized journals. A total of 40 empirical studies met all selection criteria, and were fully reviewed using multiple methods, including selected bibliometrics, content analysis and categorical meta-trends analysis. This article reports the current state of AIEd research, highlights selected AIEd technologies and applications, reviews their proven and potential benefits for education, bridges the gaps between AI technological innovations and their educational applications, and generates practical examples and inspirations for both technological experts that create AIEd technologies and educators who spearhead AI innovations in education. It also provides rich discussions on practical implications and future research directions from multiple perspectives. The advancement of AIEd calls for critical initiatives to address AI ethics and privacy concerns, and requires interdisciplinary and transdisciplinary collaborations in large-scaled, longitudinal research and development efforts.}	16	16	abstract = {From unique educational perspectives, this article reports a comprehensive review of selected empirical studies on artificial intelligence in education (AIEd) published in 1993–2020, as collected in the Web of Sciences database and selected AIEd-specialized journals. A total of 40 empirical studies met all selection criteria, and were fully reviewed using multiple methods, including selected bibliometrics, content analysis and categorical meta-trends analysis. This article reports the current state of AIEd research, highlights selected AIEd technologies and applications, reviews their proven and potential benefits for education, bridges the gaps between AI technological innovations and their educational applications, and generates practical examples and inspirations for both technological experts that create AIEd technologies and educators who spearhead AI innovations in education. It also provides rich discussions on practical implications and future research directions from multiple perspectives. The advancement of AIEd calls for critical initiatives to address AI ethics and privacy concerns, and requires interdisciplinary and transdisciplinary collaborations in large-scaled, longitudinal research and development efforts.}
}	17	17	}
	18	18
@article{PETROVIC201617,	19	19	@article{PETROVIC201617,
title = {Knowledge-light adaptation approaches in case-based reasoning for radiotherapy treatment planning},	20	20	title = {Knowledge-light adaptation approaches in case-based reasoning for radiotherapy treatment planning},
journal = {Artificial Intelligence in Medicine},	21	21	journal = {Artificial Intelligence in Medicine},
volume = {68},	22	22	volume = {68},
pages = {17-28},	23	23	pages = {17-28},
year = {2016},	24	24	year = {2016},
language = {English},	25	25	language = {English},
issn = {0933-3657},	26	26	issn = {0933-3657},
type = {article},	27	27	type = {article},
doi = {https://doi.org/10.1016/j.artmed.2016.01.006},	28	28	doi = {https://doi.org/10.1016/j.artmed.2016.01.006},
url = {https://www.sciencedirect.com/science/article/pii/S093336571630015X},	29	29	url = {https://www.sciencedirect.com/science/article/pii/S093336571630015X},
author = {Sanja Petrovic and Gulmira Khussainova and Rupa Jagannathan},	30	30	author = {Sanja Petrovic and Gulmira Khussainova and Rupa Jagannathan},
affiliation={Nottingham University},	31	31	affiliation={Nottingham University},
address={UK},	32	32	address={UK},
keywords = {Case-based reasoning, Adaptation-guided retrieval, Machine-learning tools, Radiotherapy treatment planning},	33	33	keywords = {Case-based reasoning, Adaptation-guided retrieval, Machine-learning tools, Radiotherapy treatment planning},
abstract = {Objective	34	34	abstract = {Objective
Radiotherapy treatment planning aims at delivering a sufficient radiation dose to cancerous tumour cells while sparing healthy organs in the tumour-surrounding area. It is a time-consuming trial-and-error process that requires the expertise of a group of medical experts including oncologists and medical physicists and can take from 2 to 3h to a few days. Our objective is to improve the performance of our previously built case-based reasoning (CBR) system for brain tumour radiotherapy treatment planning. In this system, a treatment plan for a new patient is retrieved from a case base containing patient cases treated in the past and their treatment plans. However, this system does not perform any adaptation, which is needed to account for any difference between the new and retrieved cases. Generally, the adaptation phase is considered to be intrinsically knowledge-intensive and domain-dependent. Therefore, an adaptation often requires a large amount of domain-specific knowledge, which can be difficult to acquire and often is not readily available. In this study, we investigate approaches to adaptation that do not require much domain knowledge, referred to as knowledge-light adaptation.	35	35	Radiotherapy treatment planning aims at delivering a sufficient radiation dose to cancerous tumour cells while sparing healthy organs in the tumour-surrounding area. It is a time-consuming trial-and-error process that requires the expertise of a group of medical experts including oncologists and medical physicists and can take from 2 to 3h to a few days. Our objective is to improve the performance of our previously built case-based reasoning (CBR) system for brain tumour radiotherapy treatment planning. In this system, a treatment plan for a new patient is retrieved from a case base containing patient cases treated in the past and their treatment plans. However, this system does not perform any adaptation, which is needed to account for any difference between the new and retrieved cases. Generally, the adaptation phase is considered to be intrinsically knowledge-intensive and domain-dependent. Therefore, an adaptation often requires a large amount of domain-specific knowledge, which can be difficult to acquire and often is not readily available. In this study, we investigate approaches to adaptation that do not require much domain knowledge, referred to as knowledge-light adaptation.
Methodology	36	36	Methodology
We developed two adaptation approaches: adaptation based on machine-learning tools and adaptation-guided retrieval. They were used to adapt the beam number and beam angles suggested in the retrieved case. Two machine-learning tools, neural networks and naive Bayes classifier, were used in the adaptation to learn how the difference in attribute values between the retrieved and new cases affects the output of these two cases. The adaptation-guided retrieval takes into consideration not only the similarity between the new and retrieved cases, but also how to adapt the retrieved case.	37	37	We developed two adaptation approaches: adaptation based on machine-learning tools and adaptation-guided retrieval. They were used to adapt the beam number and beam angles suggested in the retrieved case. Two machine-learning tools, neural networks and naive Bayes classifier, were used in the adaptation to learn how the difference in attribute values between the retrieved and new cases affects the output of these two cases. The adaptation-guided retrieval takes into consideration not only the similarity between the new and retrieved cases, but also how to adapt the retrieved case.
Results	38	38	Results
The research was carried out in collaboration with medical physicists at the Nottingham University Hospitals NHS Trust, City Hospital Campus, UK. All experiments were performed using real-world brain cancer patient cases treated with three-dimensional (3D)-conformal radiotherapy. Neural networks-based adaptation improved the success rate of the CBR system with no adaptation by 12%. However, naive Bayes classifier did not improve the current retrieval results as it did not consider the interplay among attributes. The adaptation-guided retrieval of the case for beam number improved the success rate of the CBR system by 29%. However, it did not demonstrate good performance for the beam angle adaptation. Its success rate was 29% versus 39% when no adaptation was performed.	39	39	The research was carried out in collaboration with medical physicists at the Nottingham University Hospitals NHS Trust, City Hospital Campus, UK. All experiments were performed using real-world brain cancer patient cases treated with three-dimensional (3D)-conformal radiotherapy. Neural networks-based adaptation improved the success rate of the CBR system with no adaptation by 12%. However, naive Bayes classifier did not improve the current retrieval results as it did not consider the interplay among attributes. The adaptation-guided retrieval of the case for beam number improved the success rate of the CBR system by 29%. However, it did not demonstrate good performance for the beam angle adaptation. Its success rate was 29% versus 39% when no adaptation was performed.
Conclusions	40	40	Conclusions
The obtained empirical results demonstrate that the proposed adaptation methods improve the performance of the existing CBR system in recommending the number of beams to use. However, we also conclude that to be effective, the proposed adaptation of beam angles requires a large number of relevant cases in the case base.}	41	41	The obtained empirical results demonstrate that the proposed adaptation methods improve the performance of the existing CBR system in recommending the number of beams to use. However, we also conclude that to be effective, the proposed adaptation of beam angles requires a large number of relevant cases in the case base.}
}	42	42	}
	43	43
@article{ROLDANREYES20151,	44	44	@article{ROLDANREYES20151,
title = {Improvement of online adaptation knowledge acquisition and reuse in case-based reasoning: Application to process engineering design},	45	45	title = {Improvement of online adaptation knowledge acquisition and reuse in case-based reasoning: Application to process engineering design},
journal = {Engineering Applications of Artificial Intelligence},	46	46	journal = {Engineering Applications of Artificial Intelligence},
volume = {41},	47	47	volume = {41},
pages = {1-16},	48	48	pages = {1-16},
affiliation={Université de Toulouse; Instituto Tecnologico de Orizaba},	49	49	affiliation={Université de Toulouse; Instituto Tecnologico de Orizaba},
country={France},	50	50	country={France},
language = {English},	51	51	language = {English},
year = {2015},	52	52	year = {2015},
type = {article},	53	53	type = {article},
issn = {0952-1976},	54	54	issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2015.01.015},	55	55	doi = {https://doi.org/10.1016/j.engappai.2015.01.015},
url = {https://www.sciencedirect.com/science/article/pii/S0952197615000263},	56	56	url = {https://www.sciencedirect.com/science/article/pii/S0952197615000263},
author = {E. {Roldan Reyes} and S. Negny and G. {Cortes Robles} and J.M. {Le Lann}},	57	57	author = {E. {Roldan Reyes} and S. Negny and G. {Cortes Robles} and J.M. {Le Lann}},
keywords = {Case based reasoning, Constraint satisfaction problems, Interactive adaptation method, Online knowledge acquisition, Failure diagnosis and repair},	58	58	keywords = {Case based reasoning, Constraint satisfaction problems, Interactive adaptation method, Online knowledge acquisition, Failure diagnosis and repair},
abstract = {Despite various publications in the area during the last few years, the adaptation step is still a crucial phase for a relevant and reasonable Case Based Reasoning system. Furthermore, the online acquisition of the new adaptation knowledge is of particular interest as it enables the progressive improvement of the system while reducing the knowledge engineering effort without constraints for the expert. Therefore this paper presents a new interactive method for adaptation knowledge elicitation, acquisition and reuse, thanks to a modification of the traditional CBR cycle. Moreover to improve adaptation knowledge reuse, a test procedure is also implemented to help the user in the adaptation step and its diagnosis during adaptation failure. A study on the quality and usefulness of the new knowledge acquired is also driven. As our Knowledge Based Systems (KBS) is more focused on preliminary design, and more particularly in the field of process engineering, we need to unify in the same method two types of knowledge: contextual and general. To realize this, this article proposes the integration of the Constraint Satisfaction Problem (based on general knowledge) approach into the Case Based Reasoning (based on contextual knowledge) process to improve the case representation and the adaptation of past experiences. To highlight its capability, the proposed approach is illustrated through a case study dedicated to the design of an industrial mixing device.}	59	59	abstract = {Despite various publications in the area during the last few years, the adaptation step is still a crucial phase for a relevant and reasonable Case Based Reasoning system. Furthermore, the online acquisition of the new adaptation knowledge is of particular interest as it enables the progressive improvement of the system while reducing the knowledge engineering effort without constraints for the expert. Therefore this paper presents a new interactive method for adaptation knowledge elicitation, acquisition and reuse, thanks to a modification of the traditional CBR cycle. Moreover to improve adaptation knowledge reuse, a test procedure is also implemented to help the user in the adaptation step and its diagnosis during adaptation failure. A study on the quality and usefulness of the new knowledge acquired is also driven. As our Knowledge Based Systems (KBS) is more focused on preliminary design, and more particularly in the field of process engineering, we need to unify in the same method two types of knowledge: contextual and general. To realize this, this article proposes the integration of the Constraint Satisfaction Problem (based on general knowledge) approach into the Case Based Reasoning (based on contextual knowledge) process to improve the case representation and the adaptation of past experiences. To highlight its capability, the proposed approach is illustrated through a case study dedicated to the design of an industrial mixing device.}
}	60	60	}
	61	61
@article{JUNG20095695,	62	62	@article{JUNG20095695,
title = {Integrating radial basis function networks with case-based reasoning for product design},	63	63	title = {Integrating radial basis function networks with case-based reasoning for product design},
journal = {Expert Systems with Applications},	64	64	journal = {Expert Systems with Applications},
volume = {36},	65	65	volume = {36},
number = {3, Part 1},	66	66	number = {3, Part 1},
language = {English},	67	67	language = {English},
pages = {5695-5701},	68	68	pages = {5695-5701},
year = {2009},	69	69	year = {2009},
type = {article},	70	70	type = {article},
issn = {0957-4174},	71	71	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2008.06.099},	72	72	doi = {https://doi.org/10.1016/j.eswa.2008.06.099},
url = {https://www.sciencedirect.com/science/article/pii/S0957417408003667},	73	73	url = {https://www.sciencedirect.com/science/article/pii/S0957417408003667},
author = {Sabum Jung and Taesoo Lim and Dongsoo Kim},	74	74	author = {Sabum Jung and Taesoo Lim and Dongsoo Kim},
affiliation={LG Production Engineering Research Institute; Sungkyul University; Soongsil University},	75	75	affiliation={LG Production Engineering Research Institute; Sungkyul University; Soongsil University},
keywords = {Case-based reasoning (CBR), Radial basis function network (RBFN), Design expert system, Product design},	76	76	keywords = {Case-based reasoning (CBR), Radial basis function network (RBFN), Design expert system, Product design},
abstract = {This paper presents a case-based design expert system that automatically determines the design values of a product. We focus on the design problem of a shadow mask which is a core component of monitors in the electronics industry. In case-based reasoning (CBR), it is important to retrieve similar cases and adapt them to meet design specifications exactly. Notably, difficulties in automating the adaptation process have prevented designers from being able to use design expert systems easily and efficiently. In this paper, we present a hybrid approach combining CBR and artificial neural networks in order to solve the problems occurring during the adaptation process. We first constructed a radial basis function network (RBFN) composed of representative cases created by K-means clustering. Then, the representative case most similar to the current problem was adjusted using the network. The rationale behind the proposed approach is discussed, and experimental results acquired from real shadow mask design are presented. Using the design expert system, designers can reduce design time and errors and enhance the total quality of design. Furthermore, the expert system facilitates effective sharing of design knowledge among designers.}	77	77	abstract = {This paper presents a case-based design expert system that automatically determines the design values of a product. We focus on the design problem of a shadow mask which is a core component of monitors in the electronics industry. In case-based reasoning (CBR), it is important to retrieve similar cases and adapt them to meet design specifications exactly. Notably, difficulties in automating the adaptation process have prevented designers from being able to use design expert systems easily and efficiently. In this paper, we present a hybrid approach combining CBR and artificial neural networks in order to solve the problems occurring during the adaptation process. We first constructed a radial basis function network (RBFN) composed of representative cases created by K-means clustering. Then, the representative case most similar to the current problem was adjusted using the network. The rationale behind the proposed approach is discussed, and experimental results acquired from real shadow mask design are presented. Using the design expert system, designers can reduce design time and errors and enhance the total quality of design. Furthermore, the expert system facilitates effective sharing of design knowledge among designers.}
}	78	78	}
	79	79
@article{CHIU2023100118,	80	80	@article{CHIU2023100118,
title = {Systematic literature review on opportunities, challenges, and future research recommendations of artificial intelligence in education},	81	81	title = {Systematic literature review on opportunities, challenges, and future research recommendations of artificial intelligence in education},
journal = {Computers and Education: Artificial Intelligence},	82	82	journal = {Computers and Education: Artificial Intelligence},
volume = {4},	83	83	volume = {4},
language = {English},	84	84	language = {English},
type = {article},	85	85	type = {article},
pages = {100118},	86	86	pages = {100118},
year = {2023},	87	87	year = {2023},
issn = {2666-920X},	88	88	issn = {2666-920X},
doi = {https://doi.org/10.1016/j.caeai.2022.100118},	89	89	doi = {https://doi.org/10.1016/j.caeai.2022.100118},
url = {https://www.sciencedirect.com/science/article/pii/S2666920X2200073X},	90	90	url = {https://www.sciencedirect.com/science/article/pii/S2666920X2200073X},
author = {Thomas K.F. Chiu and Qi Xia and Xinyan Zhou and Ching Sing Chai and Miaoting Cheng},	91	91	author = {Thomas K.F. Chiu and Qi Xia and Xinyan Zhou and Ching Sing Chai and Miaoting Cheng},
keywords = {Artificial intelligence, Artificial intelligence in education, Systematic review, Learning, Teaching, Assessment},	92	92	keywords = {Artificial intelligence, Artificial intelligence in education, Systematic review, Learning, Teaching, Assessment},
abstract = {Applications of artificial intelligence in education (AIEd) are emerging and are new to researchers and practitioners alike. Reviews of the relevant literature have not examined how AI technologies have been integrated into each of the four key educational domains of learning, teaching, assessment, and administration. The relationships between the technologies and learning outcomes for students and teachers have also been neglected. This systematic review study aims to understand the opportunities and challenges of AIEd by examining the literature from the last 10 years (2012–2021) using matrix coding and content analysis approaches. The results present the current focus of AIEd research by identifying 13 roles of AI technologies in the key educational domains, 7 learning outcomes of AIEd, and 10 major challenges. The review also provides suggestions for future directions of AIEd research.}	93	93	abstract = {Applications of artificial intelligence in education (AIEd) are emerging and are new to researchers and practitioners alike. Reviews of the relevant literature have not examined how AI technologies have been integrated into each of the four key educational domains of learning, teaching, assessment, and administration. The relationships between the technologies and learning outcomes for students and teachers have also been neglected. This systematic review study aims to understand the opportunities and challenges of AIEd by examining the literature from the last 10 years (2012–2021) using matrix coding and content analysis approaches. The results present the current focus of AIEd research by identifying 13 roles of AI technologies in the key educational domains, 7 learning outcomes of AIEd, and 10 major challenges. The review also provides suggestions for future directions of AIEd research.}
}	94	94	}
	95	95
@article{Robertson2014ARO,	96	96	@article{Robertson2014ARO,
title = {A Review of Real-Time Strategy Game AI},	97	97	title = {A Review of Real-Time Strategy Game AI},
author = {Glen Robertson and Ian D. Watson},	98	98	author = {Glen Robertson and Ian D. Watson},
affiliation = {University of Auckland },	99	99	affiliation = {University of Auckland },
keywords = {Game, IA, Real-time strategy},	100	100	keywords = {Game, IA, Real-time strategy},
type={article},	101	101	type={article},
language={English},	102	102	language={English},
abstract = {This literature review covers AI techniques used for real-time strategy video games, focusing specifically on StarCraft. It finds that the main areas of current academic research are in tactical and strategic decision making, plan recognition, and learning, and it outlines the research contributions in each of these areas. The paper then contrasts the use of game AI in academe and industry, finding the academic research heavily focused on creating game-winning agents, while the industry aims to maximize player enjoyment. It finds that industry adoption of academic research is low because it is either inapplicable or too time-consuming and risky to implement in a new game, which highlights an area for potential investigation: bridging the gap between academe and industry. Finally, the areas of spatial reasoning, multiscale AI, and cooperation are found to require future work, and standardized evaluation methods are proposed to produce comparable results between studies.},	103	103	abstract = {This literature review covers AI techniques used for real-time strategy video games, focusing specifically on StarCraft. It finds that the main areas of current academic research are in tactical and strategic decision making, plan recognition, and learning, and it outlines the research contributions in each of these areas. The paper then contrasts the use of game AI in academe and industry, finding the academic research heavily focused on creating game-winning agents, while the industry aims to maximize player enjoyment. It finds that industry adoption of academic research is low because it is either inapplicable or too time-consuming and risky to implement in a new game, which highlights an area for potential investigation: bridging the gap between academe and industry. Finally, the areas of spatial reasoning, multiscale AI, and cooperation are found to require future work, and standardized evaluation methods are proposed to produce comparable results between studies.},
journal = {AI Mag.},	104	104	journal = {AI Mag.},
year = {2014},	105	105	year = {2014},
volume = {35},	106	106	volume = {35},
pages = {75-104}	107	107	pages = {75-104}
}	108	108	}
	109	109
@Inproceedings{10.1007/978-3-642-15973-2_50,	110	110	@Inproceedings{10.1007/978-3-642-15973-2_50,
author={Butdee, S.	111	111	author={Butdee, S.
and Tichkiewitch, S.},	112	112	and Tichkiewitch, S.},
affiliation={University of Technology North Bangkok; Grenoble Institute of Technology},	113	113	affiliation={University of Technology North Bangkok; Grenoble Institute of Technology},
editor={Bernard, Alain},	114	114	editor={Bernard, Alain},
title={Case-Based Reasoning for Adaptive Aluminum Extrusion Die Design Together with Parameters by Neural Networks},	115	115	title={Case-Based Reasoning for Adaptive Aluminum Extrusion Die Design Together with Parameters by Neural Networks},
keywords={Adaptive die design and parameters, Optimal aluminum extrusion, Case-based reasoning, Neural networks},	116	116	keywords={Adaptive die design and parameters, Optimal aluminum extrusion, Case-based reasoning, Neural networks},
booktitle={Global Product Development},	117	117	booktitle={Global Product Development},
year={2011},	118	118	year={2011},
type = {article; proceedings paper},	119	119	type = {article; proceedings paper},
language = {English},	120	120	language = {English},
publisher = {Springer Berlin Heidelberg},	121	121	publisher = {Springer Berlin Heidelberg},
address = {Berlin, Heidelberg},	122	122	address = {Berlin, Heidelberg},
pages = {491--496},	123	123	pages = {491--496},
abstract = {Nowadays Aluminum extrusion die design is a critical task for improving productivity which involves with quality, time and cost. Case-Based Reasoning (CBR) method has been successfully applied to support the die design process in order to design a new die by tackling previous problems together with their solutions to match with a new similar problem. Such solutions are selected and modified to solve the present problem. However, the applications of the CBR are useful only retrieving previous features whereas the critical parameters are missing. In additions, the experience learning to such parameters are limited. This chapter proposes Artificial Neural Network (ANN) to associate the CBR in order to learning previous parameters and predict to the new die design according to the primitive die modification. The most satisfactory is to accommodate the optimal parameters of extrusion processes.},	124	124	abstract = {Nowadays Aluminum extrusion die design is a critical task for improving productivity which involves with quality, time and cost. Case-Based Reasoning (CBR) method has been successfully applied to support the die design process in order to design a new die by tackling previous problems together with their solutions to match with a new similar problem. Such solutions are selected and modified to solve the present problem. However, the applications of the CBR are useful only retrieving previous features whereas the critical parameters are missing. In additions, the experience learning to such parameters are limited. This chapter proposes Artificial Neural Network (ANN) to associate the CBR in order to learning previous parameters and predict to the new die design according to the primitive die modification. The most satisfactory is to accommodate the optimal parameters of extrusion processes.},
isbn = {978-3-642-15973-2}	125	125	isbn = {978-3-642-15973-2}
}	126	126	}
	127	127
@Inproceedings{10.1007/978-3-319-47096-2_11,	128	128	@Inproceedings{10.1007/978-3-319-47096-2_11,
author={Grace, Kazjon	129	129	author={Grace, Kazjon
and Maher, Mary Lou	130	130	and Maher, Mary Lou
and Wilson, David C.	131	131	and Wilson, David C.
and Najjar, Nadia A.},	132	132	and Najjar, Nadia A.},
affiliation={University of North Carolina at Charlotte},	133	133	affiliation={University of North Carolina at Charlotte},
editor={Goel, Ashok	134	134	editor={Goel, Ashok
and D{\'i}az-Agudo, M Bel{\'e}n	135	135	and D{\'i}az-Agudo, M Bel{\'e}n
and Roth-Berghofer, Thomas},	136	136	and Roth-Berghofer, Thomas},
title={Combining CBR and Deep Learning to Generate Surprising Recipe Designs},	137	137	title={Combining CBR and Deep Learning to Generate Surprising Recipe Designs},
keywords={Case-based reasoning, deep learning, recipe design},	138	138	keywords={Case-based reasoning, deep learning, recipe design},
type = {article; proceedings paper},	139	139	type = {article; proceedings paper},
booktitle={Case-Based Reasoning Research and Development},	140	140	booktitle={Case-Based Reasoning Research and Development},
year={2016},	141	141	year={2016},
publisher={Springer International Publishing},	142	142	publisher={Springer International Publishing},
address={Cham},	143	143	address={Cham},
language = {English},	144	144	language = {English},
pages={154--169},	145	145	pages={154--169},
abstract={This paper presents a dual-cycle CBR model in the domain of recipe generation. The model combines the strengths of deep learning and similarity-based retrieval to generate recipes that are novel and valuable (i.e. they are creative). The first cycle generates abstract descriptions which we call ``design concepts'' by synthesizing expectations from the entire case base, while the second cycle uses those concepts to retrieve and adapt objects. We define these conceptual object representations as an abstraction over complete cases on which expectations can be formed, allowing objects to be evaluated for surprisingness (the peak level of unexpectedness in the object, given the case base) and plausibility (the overall similarity of the object to those in the case base). The paper presents a prototype implementation of the model, and demonstrates its ability to generate objects that are simultaneously plausible and surprising, in addition to fitting a user query. This prototype is then compared to a traditional single-cycle CBR system.},	146	146	abstract={This paper presents a dual-cycle CBR model in the domain of recipe generation. The model combines the strengths of deep learning and similarity-based retrieval to generate recipes that are novel and valuable (i.e. they are creative). The first cycle generates abstract descriptions which we call ``design concepts'' by synthesizing expectations from the entire case base, while the second cycle uses those concepts to retrieve and adapt objects. We define these conceptual object representations as an abstraction over complete cases on which expectations can be formed, allowing objects to be evaluated for surprisingness (the peak level of unexpectedness in the object, given the case base) and plausibility (the overall similarity of the object to those in the case base). The paper presents a prototype implementation of the model, and demonstrates its ability to generate objects that are simultaneously plausible and surprising, in addition to fitting a user query. This prototype is then compared to a traditional single-cycle CBR system.},
isbn={978-3-319-47096-2}	147	147	isbn={978-3-319-47096-2}
}	148	148	}
	149	149
@Inproceedings{10.1007/978-3-319-61030-6_1,	150	150	@Inproceedings{10.1007/978-3-319-61030-6_1,
author={Maher, Mary Lou	151	151	author={Maher, Mary Lou
and Grace, Kazjon},	152	152	and Grace, Kazjon},
editor={Aha, David W.	153	153	editor={Aha, David W.
and Lieber, Jean},	154	154	and Lieber, Jean},
affiliation={University of North Carolina at Charlotte},	155	155	affiliation={University of North Carolina at Charlotte},
title={Encouraging Curiosity in Case-Based Reasoning and Recommender Systems},	156	156	title={Encouraging Curiosity in Case-Based Reasoning and Recommender Systems},
keywords={Curiosity, Case-based reasoning, Recommender systems},	157	157	keywords={Curiosity, Case-based reasoning, Recommender systems},
booktitle={Case-Based Reasoning Research and Development},	158	158	booktitle={Case-Based Reasoning Research and Development},
year={2017},	159	159	year={2017},
publisher={Springer International Publishing},	160	160	publisher={Springer International Publishing},
address={Cham},	161	161	address={Cham},
pages={3--15},	162	162	pages={3--15},
language = {English},	163	163	language = {English},
type = {article; proceedings paper},	164	164	type = {article; proceedings paper},
abstract={A key benefit of case-based reasoning (CBR) and recommender systems is the use of past experience to guide the synthesis or selection of the best solution for a specific context or user. Typically, the solution presented to the user is based on a value system that privileges the closest match in a query and the solution that performs best when evaluated according to predefined requirements. In domains in which creativity is desirable or the user is engaged in a learning activity, there is a benefit to moving beyond the expected or ``best match'' and include results based on computational models of novelty and surprise. In this paper, models of novelty and surprise are integrated with both CBR and Recommender Systems to encourage user curiosity.},	165	165	abstract={A key benefit of case-based reasoning (CBR) and recommender systems is the use of past experience to guide the synthesis or selection of the best solution for a specific context or user. Typically, the solution presented to the user is based on a value system that privileges the closest match in a query and the solution that performs best when evaluated according to predefined requirements. In domains in which creativity is desirable or the user is engaged in a learning activity, there is a benefit to moving beyond the expected or ``best match'' and include results based on computational models of novelty and surprise. In this paper, models of novelty and surprise are integrated with both CBR and Recommender Systems to encourage user curiosity.},
isbn={978-3-319-61030-6}	166	166	isbn={978-3-319-61030-6}
}	167	167	}
	168	168
@Inproceedings{Muller,	169	169	@Inproceedings{Muller,
author = {Müller, G. and Bergmann, R.},	170	170	author = {Müller, G. and Bergmann, R.},
affiliation={University of Trier},	171	171	affiliation={University of Trier},
year = {2015},	172	172	year = {2015},
month = {01},	173	173	month = {01},
language = {English},	174	174	language = {English},
type = {article; proceedings paper},	175	175	type = {article; proceedings paper},
abstract = {This paper presents CookingCAKE,a framework for the adaptation of cooking recipes represented as workflows. CookingCAKE integrates and combines several workflow adaptation approaches applied in process-oriented case based reasoning (POCBR) in a single adaptation framework, thus providing a capable tool for the adaptation of cooking recipes. The available case base of cooking workflows is analyzed to generate adaptation knowledge which is used to adapt a recipe regarding restrictions and resources, which the user may define for the preparation of a dish.},	176	176	abstract = {This paper presents CookingCAKE,a framework for the adaptation of cooking recipes represented as workflows. CookingCAKE integrates and combines several workflow adaptation approaches applied in process-oriented case based reasoning (POCBR) in a single adaptation framework, thus providing a capable tool for the adaptation of cooking recipes. The available case base of cooking workflows is analyzed to generate adaptation knowledge which is used to adapt a recipe regarding restrictions and resources, which the user may define for the preparation of a dish.},
booktitle = {International Conference on Case-Based Reasoning},	177	177	booktitle = {International Conference on Case-Based Reasoning},
title = {CookingCAKE: A Framework for the adaptation of cooking recipes represented as workflows},	178	178	title = {CookingCAKE: A Framework for the adaptation of cooking recipes represented as workflows},
keywords={recipe adaptation, workflow adaptation, workflows, process-oriented, case based reasoning}	179	179	keywords={recipe adaptation, workflow adaptation, workflows, process-oriented, case based reasoning}
}	180	180	}
	181	181
@Inproceedings{10.1007/978-3-319-24586-7_20,	182	182	@Inproceedings{10.1007/978-3-319-24586-7_20,
author={Onta{\~{n}}{\'o}n, S.	183	183	author={Onta{\~{n}}{\'o}n, S.
and Plaza, E.	184	184	and Plaza, E.
and Zhu, J.},	185	185	and Zhu, J.},
editor={H{\"u}llermeier, Eyke	186	186	editor={H{\"u}llermeier, Eyke
and Minor, Mirjam},	187	187	and Minor, Mirjam},
affiliation={Drexel University; Artificial Intelligence Research Institute CSIC},	188	188	affiliation={Drexel University; Artificial Intelligence Research Institute CSIC},
title={Argument-Based Case Revision in CBR for Story Generation},	189	189	title={Argument-Based Case Revision in CBR for Story Generation},
keywords={CBR, Case-based reasoning, Story generation},	190	190	keywords={CBR, Case-based reasoning, Story generation},
booktitle={Case-Based Reasoning Research and Development},	191	191	booktitle={Case-Based Reasoning Research and Development},
year={2015},	192	192	year={2015},
publisher={Springer International Publishing},	193	193	publisher={Springer International Publishing},
address={Cham},	194	194	address={Cham},
language = {English},	195	195	language = {English},
pages={290--305},	196	196	pages={290--305},
type = {article; proceedings paper},	197	197	type = {article; proceedings paper},
abstract={This paper presents a new approach to case revision in case-based reasoning based on the idea of argumentation. Previous work on case reuse has proposed the use of operations such as case amalgamation (or merging), which generate solutions by combining information coming from different cases. Such approaches are often based on exploring the search space of possible combinations looking for a solution that maximizes a certain criteria. We show how Revise can be performed by arguments attacking specific parts of a case produced by Reuse, and how they can guide and prevent repeating pitfalls in future cases. The proposed approach is evaluated in the task of automatic story generation.},	198	198	abstract={This paper presents a new approach to case revision in case-based reasoning based on the idea of argumentation. Previous work on case reuse has proposed the use of operations such as case amalgamation (or merging), which generate solutions by combining information coming from different cases. Such approaches are often based on exploring the search space of possible combinations looking for a solution that maximizes a certain criteria. We show how Revise can be performed by arguments attacking specific parts of a case produced by Reuse, and how they can guide and prevent repeating pitfalls in future cases. The proposed approach is evaluated in the task of automatic story generation.},
isbn={978-3-319-24586-7}	199	199	isbn={978-3-319-24586-7}
}	200	200	}
	201	201
@Inproceedings{10.1007/978-3-030-58342-2_20,	202	202	@Inproceedings{10.1007/978-3-030-58342-2_20,
author={Lepage, Yves	203	203	author={Lepage, Yves
and Lieber, Jean	204	204	and Lieber, Jean
and Mornard, Isabelle	205	205	and Mornard, Isabelle
and Nauer, Emmanuel	206	206	and Nauer, Emmanuel
and Romary, Julien	207	207	and Romary, Julien
and Sies, Reynault},	208	208	and Sies, Reynault},
editor={Watson, Ian	209	209	editor={Watson, Ian
and Weber, Rosina},	210	210	and Weber, Rosina},
title={The French Correction: When Retrieval Is Harder to Specify than Adaptation},	211	211	title={The French Correction: When Retrieval Is Harder to Specify than Adaptation},
affiliation={Waseda University; Université de Lorraine},	212	212	affiliation={Waseda University; Université de Lorraine},
keywords={case-based reasoning, retrieval, analogy, sentence correction},	213	213	keywords={case-based reasoning, retrieval, analogy, sentence correction},
booktitle={Case-Based Reasoning Research and Development},	214	214	booktitle={Case-Based Reasoning Research and Development},
year={2020},	215	215	year={2020},
language = {English},	216	216	language = {English},
type = {article; proceedings paper},	217	217	type = {article; proceedings paper},
publisher={Springer International Publishing},	218	218	publisher={Springer International Publishing},
address={Cham},	219	219	address={Cham},
pages={309--324},	220	220	pages={309--324},
abstract={A common idea in the field of case-based reasoning is that the retrieval step can be specified by the use of some similarity measure: the retrieved cases maximize the similarity to the target problem and, then, the adaptation step has to take into account the mismatches between the retrieved cases and the target problem in order to this latter. The use of this methodological schema for the application described in this paper has proven to be non efficient. Indeed, designing a retrieval procedure without the precise knowledge of the adaptation procedure has not been possible. The domain of this application is the correction of French sentences: a problem is an incorrect sentence and a valid solution is a correction of this problem. Adaptation consists in solving an analogical equation that enables to execute the correction of the retrieved case on the target problem. Thus, retrieval has to ensure that this application is feasible. The first version of such a retrieval procedure is described and evaluated: it is a knowledge-light procedure that does not use linguistic knowledge about French.},	221	221	abstract={A common idea in the field of case-based reasoning is that the retrieval step can be specified by the use of some similarity measure: the retrieved cases maximize the similarity to the target problem and, then, the adaptation step has to take into account the mismatches between the retrieved cases and the target problem in order to this latter. The use of this methodological schema for the application described in this paper has proven to be non efficient. Indeed, designing a retrieval procedure without the precise knowledge of the adaptation procedure has not been possible. The domain of this application is the correction of French sentences: a problem is an incorrect sentence and a valid solution is a correction of this problem. Adaptation consists in solving an analogical equation that enables to execute the correction of the retrieved case on the target problem. Thus, retrieval has to ensure that this application is feasible. The first version of such a retrieval procedure is described and evaluated: it is a knowledge-light procedure that does not use linguistic knowledge about French.},
isbn={978-3-030-58342-2}	222	222	isbn={978-3-030-58342-2}
}	223	223	}
	224	224
@Inproceedings{10.1007/978-3-030-01081-2_25,	225	225	@Inproceedings{10.1007/978-3-030-01081-2_25,
author={Smyth, Barry	226	226	author={Smyth, Barry
and Cunningham, P{\'a}draig},	227	227	and Cunningham, P{\'a}draig},
editor={Cox, Michael T.	228	228	editor={Cox, Michael T.
and Funk, Peter	229	229	and Funk, Peter
and Begum, Shahina},	230	230	and Begum, Shahina},
affiliation={University College Dublin},	231	231	affiliation={University College Dublin},
title={An Analysis of Case Representations for Marathon Race Prediction and Planning},	232	232	title={An Analysis of Case Representations for Marathon Race Prediction and Planning},
keywords={Marathon planning, Case representation, Case-based reasoning},	233	233	keywords={Marathon planning, Case representation, Case-based reasoning},
booktitle={Case-Based Reasoning Research and Development},	234	234	booktitle={Case-Based Reasoning Research and Development},
year={2018},	235	235	year={2018},
language = {English},	236	236	language = {English},
publisher={Springer International Publishing},	237	237	publisher={Springer International Publishing},
address={Cham},	238	238	address={Cham},
pages={369--384},	239	239	pages={369--384},
type = {article; proceedings paper},	240	240	type = {article; proceedings paper},
abstract={We use case-based reasoning to help marathoners achieve a personal best for an upcoming race, by helping them to select an achievable goal-time and a suitable pacing plan. We evaluate several case representations and, using real-world race data, highlight their performance implications. Richer representations do not always deliver better prediction performance, but certain representational configurations do offer very significant practical benefits for runners, when it comes to predicting, and planning for, challenging goal-times during an upcoming race.},	241	241	abstract={We use case-based reasoning to help marathoners achieve a personal best for an upcoming race, by helping them to select an achievable goal-time and a suitable pacing plan. We evaluate several case representations and, using real-world race data, highlight their performance implications. Richer representations do not always deliver better prediction performance, but certain representational configurations do offer very significant practical benefits for runners, when it comes to predicting, and planning for, challenging goal-times during an upcoming race.},
isbn={978-3-030-01081-2}	242	242	isbn={978-3-030-01081-2}
}	243	243	}
	244	244
@Inproceedings{10.1007/978-3-030-58342-2_8,	245	245	@Inproceedings{10.1007/978-3-030-58342-2_8,
author={Smyth, Barry	246	246	author={Smyth, Barry
and Willemsen, Martijn C.},	247	247	and Willemsen, Martijn C.},
editor={Watson, Ian	248	248	editor={Watson, Ian
and Weber, Rosina},	249	249	and Weber, Rosina},
affiliation={University College Dublin; Eindhoven University of Technology},	250	250	affiliation={University College Dublin; Eindhoven University of Technology},
title={Predicting the Personal-Best Times of Speed Skaters Using Case-Based Reasoning},	251	251	title={Predicting the Personal-Best Times of Speed Skaters Using Case-Based Reasoning},
keywords={CBR for health and exercise, speed skating, race-time prediction, case representation},	252	252	keywords={CBR for health and exercise, speed skating, race-time prediction, case representation},
booktitle={Case-Based Reasoning Research and Development},	253	253	booktitle={Case-Based Reasoning Research and Development},
year={2020},	254	254	year={2020},
type = {article; proceedings paper},	255	255	type = {article; proceedings paper},
language = {English},	256	256	language = {English},
publisher={Springer International Publishing},	257	257	publisher={Springer International Publishing},
address={Cham},	258	258	address={Cham},
pages={112--126},	259	259	pages={112--126},
abstract={Speed skating is a form of ice skating in which the skaters race each other over a variety of standardised distances. Races take place on specialised ice-rinks and the type of track and ice conditions can have a significant impact on race-times. As race distances increase, pacing also plays an important role. In this paper we seek to extend recent work on the application of case-based reasoning to marathon-time prediction by predicting race-times for speed skaters. In particular, we propose and evaluate a number of case-based reasoning variants based on different case and feature representations to generate track-specific race predictions. We show it is possible to improve upon state-of-the-art prediction accuracy by harnessing richer case representations using shorter races and track-adjusted finish and lap-times.},	260	260	abstract={Speed skating is a form of ice skating in which the skaters race each other over a variety of standardised distances. Races take place on specialised ice-rinks and the type of track and ice conditions can have a significant impact on race-times. As race distances increase, pacing also plays an important role. In this paper we seek to extend recent work on the application of case-based reasoning to marathon-time prediction by predicting race-times for speed skaters. In particular, we propose and evaluate a number of case-based reasoning variants based on different case and feature representations to generate track-specific race predictions. We show it is possible to improve upon state-of-the-art prediction accuracy by harnessing richer case representations using shorter races and track-adjusted finish and lap-times.},
isbn={978-3-030-58342-2}	261	261	isbn={978-3-030-58342-2}
}	262	262	}
	263	263
@Inproceedings{10.1007/978-3-030-58342-2_5,	264	264	@Inproceedings{10.1007/978-3-030-58342-2_5,
author={Feely, Ciara	265	265	author={Feely, Ciara
and Caulfield, Brian	266	266	and Caulfield, Brian
and Lawlor, Aonghus	267	267	and Lawlor, Aonghus
and Smyth, Barry},	268	268	and Smyth, Barry},
editor={Watson, Ian	269	269	editor={Watson, Ian
and Weber, Rosina},	270	270	and Weber, Rosina},
affiliation={University College Dublin},	271	271	affiliation={University College Dublin},
title={Using Case-Based Reasoning to Predict Marathon Performance and Recommend Tailored Training Plans},	272	272	title={Using Case-Based Reasoning to Predict Marathon Performance and Recommend Tailored Training Plans},
keywords={CBR for health and exercise, marathon running, race-time prediction, plan recommendation},	273	273	keywords={CBR for health and exercise, marathon running, race-time prediction, plan recommendation},
booktitle={Case-Based Reasoning Research and Development},	274	274	booktitle={Case-Based Reasoning Research and Development},
year={2020},	275	275	year={2020},
language = {English},	276	276	language = {English},
publisher={Springer International Publishing},	277	277	publisher={Springer International Publishing},
address={Cham},	278	278	address={Cham},
pages={67--81},	279	279	pages={67--81},
type = {article; proceedings paper},	280	280	type = {article; proceedings paper},
abstract={Training for the marathon, especially a first marathon, is always a challenge. Many runners struggle to find the right balance between their workouts and their recovery, often leading to sub-optimal performance on race-day or even injury during training. We describe and evaluate a novel case-based reasoning system to help marathon runners as they train in two ways. First, it uses a case-base of training/workouts and race histories to predict future marathon times for a target runner, throughout their training program, helping runners to calibrate their progress and, ultimately, plan their race-day pacing. Second, the system recommends tailored training plans to runners, adapted for their current goal-time target, and based on the training plans of similar runners who have achieved this time. We evaluate the system using a dataset of more than 21,000 unique runners and 1.5 million training/workout sessions.},	281	281	abstract={Training for the marathon, especially a first marathon, is always a challenge. Many runners struggle to find the right balance between their workouts and their recovery, often leading to sub-optimal performance on race-day or even injury during training. We describe and evaluate a novel case-based reasoning system to help marathon runners as they train in two ways. First, it uses a case-base of training/workouts and race histories to predict future marathon times for a target runner, throughout their training program, helping runners to calibrate their progress and, ultimately, plan their race-day pacing. Second, the system recommends tailored training plans to runners, adapted for their current goal-time target, and based on the training plans of similar runners who have achieved this time. We evaluate the system using a dataset of more than 21,000 unique runners and 1.5 million training/workout sessions.},
isbn={978-3-030-58342-2}	282	282	isbn={978-3-030-58342-2}
}	283	283	}
	284	284
@article{LALITHA2020583,	285	285	@article{LALITHA2020583,
title = {Personalised Self-Directed Learning Recommendation System},	286	286	title = {Personalised Self-Directed Learning Recommendation System},
journal = {Procedia Computer Science},	287	287	journal = {Procedia Computer Science},
volume = {171},	288	288	volume = {171},
pages = {583-592},	289	289	pages = {583-592},
year = {2020},	290	290	year = {2020},
type = {article},	291	291	type = {article},
language = {English},	292	292	language = {English},
note = {Third International Conference on Computing and Network Communications (CoCoNet'19)},	293	293	note = {Third International Conference on Computing and Network Communications (CoCoNet'19)},
issn = {1877-0509},	294	294	issn = {1877-0509},
doi = {https://doi.org/10.1016/j.procs.2020.04.063},	295	295	doi = {https://doi.org/10.1016/j.procs.2020.04.063},
url = {https://www.sciencedirect.com/science/article/pii/S1877050920310309},	296	296	url = {https://www.sciencedirect.com/science/article/pii/S1877050920310309},
author = {T B Lalitha and P S Sreeja},	297	297	author = {T B Lalitha and P S Sreeja},
affiliation={Hindustan Institute of Technology and Science},	298	298	affiliation={Hindustan Institute of Technology and Science},
keywords = {e-Learning, PSDLR, Recommendation System, SDL, Self-Directed Learning},	299	299	keywords = {e-Learning, PSDLR, Recommendation System, SDL, Self-Directed Learning},
abstract = {Modern educational systems have changed drastically bringing in knowledge anywhere as needed by the learner with the evolution of Internet. Availability of knowledge in public domain, capability of exchanging large amount of information and filtering relevant information quickly has enabled disruption to conventional educational system. Thus, future trends are looking towards E-Learning (Electronic Learning) and M-Learning (Mobile Learning) technologies over the Internet for their vast knowledge acquisition. In this paper, the work gives an elaborate context of learning strategies prevailing and emerging with the classification of e-learning Techniques. It majorly focuses on the features and variety of aspects with the e-learning and the choice of learning method involved and facilitate the adoption of new ways for personalized selection on learning resources for SDL (Self-Directed Learning) from the unstructured, large web-based environment. Thereby, proposes a Personalised Self-Directed Learning Recommendation System (PSDLR) based on the personal specifications of the SDL learner. The result offers insight into the perspectives and challenges of Self-Directed Learning based on cognitive and constructive characteristics which majorly incorporates web-based learning and gives path in finding appropriate solutions using machine learning techniques and ontology for the open problems in the respective fields with personalised recommendations and guidelines for future research.}	300	300	abstract = {Modern educational systems have changed drastically bringing in knowledge anywhere as needed by the learner with the evolution of Internet. Availability of knowledge in public domain, capability of exchanging large amount of information and filtering relevant information quickly has enabled disruption to conventional educational system. Thus, future trends are looking towards E-Learning (Electronic Learning) and M-Learning (Mobile Learning) technologies over the Internet for their vast knowledge acquisition. In this paper, the work gives an elaborate context of learning strategies prevailing and emerging with the classification of e-learning Techniques. It majorly focuses on the features and variety of aspects with the e-learning and the choice of learning method involved and facilitate the adoption of new ways for personalized selection on learning resources for SDL (Self-Directed Learning) from the unstructured, large web-based environment. Thereby, proposes a Personalised Self-Directed Learning Recommendation System (PSDLR) based on the personal specifications of the SDL learner. The result offers insight into the perspectives and challenges of Self-Directed Learning based on cognitive and constructive characteristics which majorly incorporates web-based learning and gives path in finding appropriate solutions using machine learning techniques and ontology for the open problems in the respective fields with personalised recommendations and guidelines for future research.}
}	301	301	}
	302	302
@article{Zhou2021,	303	303	@article{Zhou2021,
author={Zhou, Lina	304	304	author={Zhou, Lina
and Wang, Chunxia},	305	305	and Wang, Chunxia},
affiliation={Baotou Medical College},	306	306	affiliation={Baotou Medical College},
title={Research on Recommendation of Personalized Exercises in English Learning Based on Data Mining},	307	307	title={Research on Recommendation of Personalized Exercises in English Learning Based on Data Mining},
journal={Scientific Programming},	308	308	journal={Scientific Programming},
year={2021},	309	309	year={2021},
month={Dec},	310	310	month={Dec},
type = {article},	311	311	type = {article},
language = {English},	312	312	language = {English},
day={21},	313	313	day={21},
publisher={Hindawi},	314	314	publisher={Hindawi},
keywords={Recommender systems, Learning},	315	315	keywords={Recommender systems, Learning},
volume={2021},	316	316	volume={2021},
pages={5042286},	317	317	pages={5042286},
abstract={Aiming at the problems of traditional method of exercise recommendation precision, recall rate, long recommendation time, and poor recommendation comprehensiveness, this study proposes a personalized exercise recommendation method for English learning based on data mining. Firstly, a personalized recommendation model is designed, based on the model to preprocess the data in the Web access log, and cleaning the noise data to avoid its impact on the accuracy of the recommendation results is focused; secondly, the DINA model to diagnose the degree of mastery of students{\&}{\#}x2019; knowledge points is used and the students{\&}{\#}x2019; browsing patterns through fuzzy similar relationships are clustered; and finally, according to the clustering results, the similarity between students and the similarity between exercises are measured, and the collaborative filtering recommendation of personalized exercises for English learning is realized. The experimental results show that the exercise recommendation precision and recall rate of this method are higher, the recommendation time is shorter, and the recommendation results are comprehensive.},	318	318	abstract={Aiming at the problems of traditional method of exercise recommendation precision, recall rate, long recommendation time, and poor recommendation comprehensiveness, this study proposes a personalized exercise recommendation method for English learning based on data mining. Firstly, a personalized recommendation model is designed, based on the model to preprocess the data in the Web access log, and cleaning the noise data to avoid its impact on the accuracy of the recommendation results is focused; secondly, the DINA model to diagnose the degree of mastery of students{\&}{\#}x2019; knowledge points is used and the students{\&}{\#}x2019; browsing patterns through fuzzy similar relationships are clustered; and finally, according to the clustering results, the similarity between students and the similarity between exercises are measured, and the collaborative filtering recommendation of personalized exercises for English learning is realized. The experimental results show that the exercise recommendation precision and recall rate of this method are higher, the recommendation time is shorter, and the recommendation results are comprehensive.},
issn={1058-9244},	319	319	issn={1058-9244},
doi={10.1155/2021/5042286},	320	320	doi={10.1155/2021/5042286},
url={https://doi.org/10.1155/2021/5042286}	321	321	url={https://doi.org/10.1155/2021/5042286}
}	322	322	}
	323	323
@article{INGKAVARA2022100086,	324	324	@article{INGKAVARA2022100086,
title = {The use of a personalized learning approach to implementing self-regulated online learning},	325	325	title = {The use of a personalized learning approach to implementing self-regulated online learning},
journal = {Computers and Education: Artificial Intelligence},	326	326	journal = {Computers and Education: Artificial Intelligence},
volume = {3},	327	327	volume = {3},
pages = {100086},	328	328	pages = {100086},
type = {article},	329	329	type = {article},
language = {English},	330	330	language = {English},
year = {2022},	331	331	year = {2022},
issn = {2666-920X},	332	332	issn = {2666-920X},
doi = {https://doi.org/10.1016/j.caeai.2022.100086},	333	333	doi = {https://doi.org/10.1016/j.caeai.2022.100086},
url = {https://www.sciencedirect.com/science/article/pii/S2666920X22000418},	334	334	url = {https://www.sciencedirect.com/science/article/pii/S2666920X22000418},
author = {Thanyaluck Ingkavara and Patcharin Panjaburee and Niwat Srisawasdi and Suthiporn Sajjapanroj},	335	335	author = {Thanyaluck Ingkavara and Patcharin Panjaburee and Niwat Srisawasdi and Suthiporn Sajjapanroj},
keywords = {Intelligent tutoring system, Personalization, Adaptive learning, E-learning, TAM, Artificial intelligence},	336	336	keywords = {Intelligent tutoring system, Personalization, Adaptive learning, E-learning, TAM, Artificial intelligence},
abstract = {Nowadays, students are encouraged to learn via online learning systems to promote students' autonomy. Scholars have found that students' self-regulated actions impact their academic success in an online learning environment. However, because traditional online learning systems cannot personalize feedback to the student's personality, most students have less chance to obtain helpful suggestions for enhancing their knowledge linked to their learning problems. This paper incorporated self-regulated online learning in the Physics classroom and used a personalized learning approach to help students receive proper learning paths and material corresponding to their learning preferences. This study conducted a quasi-experimental design using a quantitative approach to evaluate the effectiveness of the proposed learning environment in secondary schools. The experimental group of students participated in self-regulated online learning with a personalized learning approach, while the control group participated in conventional self-regulated online learning. The experimental results showed that the experimental group's post-test and the learning-gain score of the experimental group were significantly higher than those of the control group. Moreover, the results also suggested that the student's perceptions about the usefulness of learning suggestions, ease of use, goal setting, learning environmental structuring, task strategies, time management, self-evaluation, impact on learning, and attitude toward the learning environment are important predictors of behavioral intention to learn with the self-regulated online learning that integrated with the personalized learning approach.}	337	337	abstract = {Nowadays, students are encouraged to learn via online learning systems to promote students' autonomy. Scholars have found that students' self-regulated actions impact their academic success in an online learning environment. However, because traditional online learning systems cannot personalize feedback to the student's personality, most students have less chance to obtain helpful suggestions for enhancing their knowledge linked to their learning problems. This paper incorporated self-regulated online learning in the Physics classroom and used a personalized learning approach to help students receive proper learning paths and material corresponding to their learning preferences. This study conducted a quasi-experimental design using a quantitative approach to evaluate the effectiveness of the proposed learning environment in secondary schools. The experimental group of students participated in self-regulated online learning with a personalized learning approach, while the control group participated in conventional self-regulated online learning. The experimental results showed that the experimental group's post-test and the learning-gain score of the experimental group were significantly higher than those of the control group. Moreover, the results also suggested that the student's perceptions about the usefulness of learning suggestions, ease of use, goal setting, learning environmental structuring, task strategies, time management, self-evaluation, impact on learning, and attitude toward the learning environment are important predictors of behavioral intention to learn with the self-regulated online learning that integrated with the personalized learning approach.}
}	338	338	}
	339	339
@article{HUANG2023104684,	340	340	@article{HUANG2023104684,
title = {Effects of artificial Intelligence–Enabled personalized recommendations on learners’ learning engagement, motivation, and outcomes in a flipped classroom},	341	341	title = {Effects of artificial Intelligence–Enabled personalized recommendations on learners’ learning engagement, motivation, and outcomes in a flipped classroom},
journal = {Computers and Education},	342	342	journal = {Computers and Education},
volume = {194},	343	343	volume = {194},
pages = {104684},	344	344	pages = {104684},
year = {2023},	345	345	year = {2023},
language = {English},	346	346	language = {English},
type = {article},	347	347	type = {article},
issn = {0360-1315},	348	348	issn = {0360-1315},
doi = {https://doi.org/10.1016/j.compedu.2022.104684},	349	349	doi = {https://doi.org/10.1016/j.compedu.2022.104684},
url = {https://www.sciencedirect.com/science/article/pii/S036013152200255X},	350	350	url = {https://www.sciencedirect.com/science/article/pii/S036013152200255X},
author = {Anna Y.Q. Huang and Owen H.T. Lu and Stephen J.H. Yang},	351	351	author = {Anna Y.Q. Huang and Owen H.T. Lu and Stephen J.H. Yang},
keywords = {Data science applications in education, Distance education and online learning, Improving classroom teaching},	352	352	keywords = {Data science applications in education, Distance education and online learning, Improving classroom teaching},
abstract = {The flipped classroom approach is aimed at improving learning outcomes by promoting learning motivation and engagement. Recommendation systems can also be used to improve learning outcomes. With the rapid development of artificial intelligence (AI) technology, various systems have been developed to facilitate student learning. Accordingly, we applied AI-enabled personalized video recommendations to stimulate students' learning motivation and engagement during a systems programming course in a flipped classroom setting. We assigned students to control and experimental groups comprising 59 and 43 college students, respectively. The students in both groups received flipped classroom instruction, but only those in the experimental group received AI-enabled personalized video recommendations. We quantitatively measured students’ engagement based on their learning profiles in a learning management system. The results revealed that the AI-enabled personalized video recommendations could significantly improve the learning performance and engagement of students with a moderate motivation level.}	353	353	abstract = {The flipped classroom approach is aimed at improving learning outcomes by promoting learning motivation and engagement. Recommendation systems can also be used to improve learning outcomes. With the rapid development of artificial intelligence (AI) technology, various systems have been developed to facilitate student learning. Accordingly, we applied AI-enabled personalized video recommendations to stimulate students' learning motivation and engagement during a systems programming course in a flipped classroom setting. We assigned students to control and experimental groups comprising 59 and 43 college students, respectively. The students in both groups received flipped classroom instruction, but only those in the experimental group received AI-enabled personalized video recommendations. We quantitatively measured students’ engagement based on their learning profiles in a learning management system. The results revealed that the AI-enabled personalized video recommendations could significantly improve the learning performance and engagement of students with a moderate motivation level.}
}	354	354	}
	355	355
@article{ZHAO2023118535,	356	356	@article{ZHAO2023118535,
title = {A recommendation system for effective learning strategies: An integrated approach using context-dependent DEA},	357	357	title = {A recommendation system for effective learning strategies: An integrated approach using context-dependent DEA},
journal = {Expert Systems with Applications},	358	358	journal = {Expert Systems with Applications},
volume = {211},	359	359	volume = {211},
pages = {118535},	360	360	pages = {118535},
year = {2023},	361	361	year = {2023},
language = {English},	362	362	language = {English},
type = {article},	363	363	type = {article},
issn = {0957-4174},	364	364	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2022.118535},	365	365	doi = {https://doi.org/10.1016/j.eswa.2022.118535},
url = {https://www.sciencedirect.com/science/article/pii/S0957417422016104},	366	366	url = {https://www.sciencedirect.com/science/article/pii/S0957417422016104},
author = {Lu-Tao Zhao and Dai-Song Wang and Feng-Yun Liang and Jian Chen},	367	367	author = {Lu-Tao Zhao and Dai-Song Wang and Feng-Yun Liang and Jian Chen},
keywords = {Recommendation system, Learning strategies, Context-dependent DEA, Efficiency analysis},	368	368	keywords = {Recommendation system, Learning strategies, Context-dependent DEA, Efficiency analysis},
abstract = {Universities have been focusing on increasing individualized training and providing appropriate education for students. The individual differences and learning needs of college students should be given enough attention. From the perspective of learning efficiency, we establish a clustering hierarchical progressive improvement model (CHPI), which is based on cluster analysis and context-dependent data envelopment analysis (DEA) methods. The CHPI clusters students' ontological features, employs the context-dependent DEA method to stratify students of different classes, and calculates measures, such as obstacles, to determine the reference path for individuals with inefficient learning processes. The learning strategies are determined according to the gap between the inefficient individual to be improved and the individuals on the reference path. By the study of college English courses as an example, it is found that the CHPI can accurately recommend targeted learning strategies to satisfy the individual needs of college students so that the learning of individuals with inefficient learning processes in a certain stage can be effectively improved. In addition, CHPI can provide specific, efficient suggestions to improve learning efficiency comparing to existing recommendation systems, and has great potential in promoting the integration of education-related researches and expert systems.}	369	369	abstract = {Universities have been focusing on increasing individualized training and providing appropriate education for students. The individual differences and learning needs of college students should be given enough attention. From the perspective of learning efficiency, we establish a clustering hierarchical progressive improvement model (CHPI), which is based on cluster analysis and context-dependent data envelopment analysis (DEA) methods. The CHPI clusters students' ontological features, employs the context-dependent DEA method to stratify students of different classes, and calculates measures, such as obstacles, to determine the reference path for individuals with inefficient learning processes. The learning strategies are determined according to the gap between the inefficient individual to be improved and the individuals on the reference path. By the study of college English courses as an example, it is found that the CHPI can accurately recommend targeted learning strategies to satisfy the individual needs of college students so that the learning of individuals with inefficient learning processes in a certain stage can be effectively improved. In addition, CHPI can provide specific, efficient suggestions to improve learning efficiency comparing to existing recommendation systems, and has great potential in promoting the integration of education-related researches and expert systems.}
}	370	370	}
	371	371
@article{SU2022109547,	372	372	@article{SU2022109547,
title = {Graph-based cognitive diagnosis for intelligent tutoring systems},	373	373	title = {Graph-based cognitive diagnosis for intelligent tutoring systems},
journal = {Knowledge-Based Systems},	374	374	journal = {Knowledge-Based Systems},
volume = {253},	375	375	volume = {253},
pages = {109547},	376	376	pages = {109547},
year = {2022},	377	377	year = {2022},
language = {English},	378	378	language = {English},
type = {article},	379	379	type = {article},
issn = {0950-7051},	380	380	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2022.109547},	381	381	doi = {https://doi.org/10.1016/j.knosys.2022.109547},
url = {https://www.sciencedirect.com/science/article/pii/S095070512200778X},	382	382	url = {https://www.sciencedirect.com/science/article/pii/S095070512200778X},
author = {Yu Su and Zeyu Cheng and Jinze Wu and Yanmin Dong and Zhenya Huang and Le Wu and Enhong Chen and Shijin Wang and Fei Xie},	383	383	author = {Yu Su and Zeyu Cheng and Jinze Wu and Yanmin Dong and Zhenya Huang and Le Wu and Enhong Chen and Shijin Wang and Fei Xie},
keywords = {Cognitive diagnosis, Graph neural networks, Interpretable machine learning},	384	384	keywords = {Cognitive diagnosis, Graph neural networks, Interpretable machine learning},
abstract = {For intelligent tutoring systems, Cognitive Diagnosis (CD) is a fundamental task that aims to estimate the mastery degree of a student on each skill according to the exercise record. The CD task is considered rather challenging since we need to model inner-relations and inter-relations among students, skills, and questions to obtain more abundant information. Most existing methods attempt to solve this problem through two-way interactions between students and questions (or between students and skills), ignoring potential high-order relations among entities. Furthermore, how to construct an end-to-end framework that can model the complex interactions among different types of entities at the same time remains unexplored. Therefore, in this paper, we propose a graph-based Cognitive Diagnosis model (GCDM) that directly discovers the interactions among students, skills, and questions through a heterogeneous cognitive graph. Specifically, we design two graph-based layers: a performance-relative propagator and an attentive knowledge aggregator. The former is applied to propagate a student’s cognitive state through different types of graph edges, while the latter selectively gathers messages from neighboring graph nodes. Extensive experimental results on two real-world datasets clearly show the effectiveness and extendibility of our proposed model.}	385	385	abstract = {For intelligent tutoring systems, Cognitive Diagnosis (CD) is a fundamental task that aims to estimate the mastery degree of a student on each skill according to the exercise record. The CD task is considered rather challenging since we need to model inner-relations and inter-relations among students, skills, and questions to obtain more abundant information. Most existing methods attempt to solve this problem through two-way interactions between students and questions (or between students and skills), ignoring potential high-order relations among entities. Furthermore, how to construct an end-to-end framework that can model the complex interactions among different types of entities at the same time remains unexplored. Therefore, in this paper, we propose a graph-based Cognitive Diagnosis model (GCDM) that directly discovers the interactions among students, skills, and questions through a heterogeneous cognitive graph. Specifically, we design two graph-based layers: a performance-relative propagator and an attentive knowledge aggregator. The former is applied to propagate a student’s cognitive state through different types of graph edges, while the latter selectively gathers messages from neighboring graph nodes. Extensive experimental results on two real-world datasets clearly show the effectiveness and extendibility of our proposed model.}
}	386	386	}
	387	387
@article{EZALDEEN2022100700,	388	388	@article{EZALDEEN2022100700,
title = {A hybrid E-learning recommendation integrating adaptive profiling and sentiment analysis},	389	389	title = {A hybrid E-learning recommendation integrating adaptive profiling and sentiment analysis},
journal = {Journal of Web Semantics},	390	390	journal = {Journal of Web Semantics},
volume = {72},	391	391	volume = {72},
pages = {100700},	392	392	pages = {100700},
year = {2022},	393	393	year = {2022},
type = {article},	394	394	type = {article},
language = {English},	395	395	language = {English},
issn = {1570-8268},	396	396	issn = {1570-8268},
doi = {https://doi.org/10.1016/j.websem.2021.100700},	397	397	doi = {https://doi.org/10.1016/j.websem.2021.100700},
url = {https://www.sciencedirect.com/science/article/pii/S1570826821000664},	398	398	url = {https://www.sciencedirect.com/science/article/pii/S1570826821000664},
author = {Hadi Ezaldeen and Rachita Misra and Sukant Kishoro Bisoy and Rawaa Alatrash and Rojalina Priyadarshini},	399	399	author = {Hadi Ezaldeen and Rachita Misra and Sukant Kishoro Bisoy and Rawaa Alatrash and Rojalina Priyadarshini},
keywords = {Hybrid E-learning recommendation, Adaptive profiling, Semantic learner profile, Fine-grained sentiment analysis, Convolutional Neural Network, Word embeddings},	400	400	keywords = {Hybrid E-learning recommendation, Adaptive profiling, Semantic learner profile, Fine-grained sentiment analysis, Convolutional Neural Network, Word embeddings},
abstract = {This research proposes a novel framework named Enhanced e-Learning Hybrid Recommender System (ELHRS) that provides an appropriate e-content with the highest predicted ratings corresponding to the learner’s particular needs. To accomplish this, a new model is developed to deduce the Semantic Learner Profile automatically. It adaptively associates the learning patterns and rules depending on the learner’s behavior and the semantic relations computed in the semantic matrix that mutually links e-learning materials and terms. Here, a semantic-based approach for term expansion is introduced using DBpedia and WordNet ontologies. Further, various sentiment analysis models are proposed and incorporated as a part of the recommender system to predict ratings of e-learning resources from posted text reviews utilizing fine-grained sentiment classification on five discrete classes. Qualitative Natural Language Processing (NLP) methods with tailored-made Convolutional Neural Network (CNN) are developed and evaluated on our customized dataset collected for a specific domain and a public dataset. Two improved language models are introduced depending on Skip-Gram (S-G) and Continuous Bag of Words (CBOW) techniques. In addition, a robust language model based on hybridization of these couple of methods is developed to derive better vocabulary representation, yielding better accuracy 89.1% for the CNN-Three-Channel-Concatenation model. The suggested recommendation methodology depends on the learner’s preferences, other similar learners’ experience and background, deriving their opinions from the reviews towards the best learning resources. This assists the learners in finding the desired e-content at the proper time.}	401	401	abstract = {This research proposes a novel framework named Enhanced e-Learning Hybrid Recommender System (ELHRS) that provides an appropriate e-content with the highest predicted ratings corresponding to the learner’s particular needs. To accomplish this, a new model is developed to deduce the Semantic Learner Profile automatically. It adaptively associates the learning patterns and rules depending on the learner’s behavior and the semantic relations computed in the semantic matrix that mutually links e-learning materials and terms. Here, a semantic-based approach for term expansion is introduced using DBpedia and WordNet ontologies. Further, various sentiment analysis models are proposed and incorporated as a part of the recommender system to predict ratings of e-learning resources from posted text reviews utilizing fine-grained sentiment classification on five discrete classes. Qualitative Natural Language Processing (NLP) methods with tailored-made Convolutional Neural Network (CNN) are developed and evaluated on our customized dataset collected for a specific domain and a public dataset. Two improved language models are introduced depending on Skip-Gram (S-G) and Continuous Bag of Words (CBOW) techniques. In addition, a robust language model based on hybridization of these couple of methods is developed to derive better vocabulary representation, yielding better accuracy 89.1% for the CNN-Three-Channel-Concatenation model. The suggested recommendation methodology depends on the learner’s preferences, other similar learners’ experience and background, deriving their opinions from the reviews towards the best learning resources. This assists the learners in finding the desired e-content at the proper time.}
}	402	402	}
	403	403
@article{MUANGPRATHUB2020e05227,	404	404	@article{MUANGPRATHUB2020e05227,
title = {Learning recommendation with formal concept analysis for intelligent tutoring system},	405	405	title = {Learning recommendation with formal concept analysis for intelligent tutoring system},
journal = {Heliyon},	406	406	journal = {Heliyon},
volume = {6},	407	407	volume = {6},
number = {10},	408	408	number = {10},
pages = {e05227},	409	409	pages = {e05227},
language = {English},	410	410	language = {English},
type = {article},	411	411	type = {article},
year = {2020},	412	412	year = {2020},
issn = {2405-8440},	413	413	issn = {2405-8440},
doi = {https://doi.org/10.1016/j.heliyon.2020.e05227},	414	414	doi = {https://doi.org/10.1016/j.heliyon.2020.e05227},
url = {https://www.sciencedirect.com/science/article/pii/S2405844020320703},	415	415	url = {https://www.sciencedirect.com/science/article/pii/S2405844020320703},
author = {Jirapond Muangprathub and Veera Boonjing and Kosin Chamnongthai},	416	416	author = {Jirapond Muangprathub and Veera Boonjing and Kosin Chamnongthai},
keywords = {Computer Science, Learning recommendation, Formal concept analysis, Intelligent tutoring system, Adaptive learning},	417	417	keywords = {Computer Science, Learning recommendation, Formal concept analysis, Intelligent tutoring system, Adaptive learning},
abstract = {The aim of this research was to develop a learning recommendation component in an intelligent tutoring system (ITS) that dynamically predicts and adapts to a learner's style. In order to develop a proper ITS, we present an improved knowledge base supporting adaptive learning, which can be achieved by a suitable knowledge construction. This process is illustrated by implementing a web-based online tutor system. In addition, our knowledge structure provides adaptive presentation and personalized learning with the proposed adaptive algorithm, to retrieve content according to individual learner characteristics. To demonstrate the proposed adaptive algorithm, pre-test and post-test were used to evaluate suggestion accuracy of the course in a class for adapting to a learner's style. In addition, pre- and post-testing were also used with students in a real teaching/learning environment to evaluate the performance of the proposed model. The results show that the proposed system can be used to help students or learners achieve improved learning.}	418	418	abstract = {The aim of this research was to develop a learning recommendation component in an intelligent tutoring system (ITS) that dynamically predicts and adapts to a learner's style. In order to develop a proper ITS, we present an improved knowledge base supporting adaptive learning, which can be achieved by a suitable knowledge construction. This process is illustrated by implementing a web-based online tutor system. In addition, our knowledge structure provides adaptive presentation and personalized learning with the proposed adaptive algorithm, to retrieve content according to individual learner characteristics. To demonstrate the proposed adaptive algorithm, pre-test and post-test were used to evaluate suggestion accuracy of the course in a class for adapting to a learner's style. In addition, pre- and post-testing were also used with students in a real teaching/learning environment to evaluate the performance of the proposed model. The results show that the proposed system can be used to help students or learners achieve improved learning.}
}	419	419	}
	420	420
@article{min8100434,	421	421	@article{min8100434,
author = {Leikola, Maria and Sauer, Christian and Rintala, Lotta and Aromaa, Jari and Lundström, Mari},	422	422	author = {Leikola, Maria and Sauer, Christian and Rintala, Lotta and Aromaa, Jari and Lundström, Mari},
title = {Assessing the Similarity of Cyanide-Free Gold Leaching Processes: A Case-Based Reasoning Application},	423	423	title = {Assessing the Similarity of Cyanide-Free Gold Leaching Processes: A Case-Based Reasoning Application},
journal = {Minerals},	424	424	journal = {Minerals},
volume = {8},	425	425	volume = {8},
type = {article},	426	426	type = {article},
language = {English},	427	427	language = {English},
year = {2018},	428	428	year = {2018},
number = {10},	429	429	number = {10},
url = {https://www.mdpi.com/2075-163X/8/10/434},	430	430	url = {https://www.mdpi.com/2075-163X/8/10/434},
issn = {2075-163X},	431	431	issn = {2075-163X},
keywords={hydrometallurgy, cyanide-free gold, knowledge modelling, case-based reasoning, information retrieval},	432	432	keywords={hydrometallurgy, cyanide-free gold, knowledge modelling, case-based reasoning, information retrieval},
abstract = {Hydrometallurgical researchers, and other professionals alike, invest significant amounts of time reading scientific articles, technical notes, and other scientific documents, while looking for the most relevant information for their particular research interest. In an attempt to save the researcher’s time, this study presents an information retrieval tool using case-based reasoning. The tool was built for comparing scientific articles concerning cyanide-free leaching of gold ores/concentrates/tailings. Altogether, 50 cases of experiments were gathered in a case base. 15 different attributes related to the treatment of the raw material and the leaching conditions were selected to compare the cases. The attributes were as follows: Pretreatment, Overall method, Complexant source, Oxidant source, Complexant concentration, Oxidant concentration, Temperature, pH, Redox-potential, Pressure, Materials of construction, Extraction, Extraction rate, Reagent consumption, and Solid-liquid ratio. The resulting retrieval tool (LeachSim) was able to rank the scientific articles according to their similarity with the user’s research interest. Such a tool could eventually aid the user in finding the most relevant information, but not replace thorough understanding and human expertise.},	433	433	abstract = {Hydrometallurgical researchers, and other professionals alike, invest significant amounts of time reading scientific articles, technical notes, and other scientific documents, while looking for the most relevant information for their particular research interest. In an attempt to save the researcher’s time, this study presents an information retrieval tool using case-based reasoning. The tool was built for comparing scientific articles concerning cyanide-free leaching of gold ores/concentrates/tailings. Altogether, 50 cases of experiments were gathered in a case base. 15 different attributes related to the treatment of the raw material and the leaching conditions were selected to compare the cases. The attributes were as follows: Pretreatment, Overall method, Complexant source, Oxidant source, Complexant concentration, Oxidant concentration, Temperature, pH, Redox-potential, Pressure, Materials of construction, Extraction, Extraction rate, Reagent consumption, and Solid-liquid ratio. The resulting retrieval tool (LeachSim) was able to rank the scientific articles according to their similarity with the user’s research interest. Such a tool could eventually aid the user in finding the most relevant information, but not replace thorough understanding and human expertise.},
doi = {10.3390/min8100434}	434	434	doi = {10.3390/min8100434}
}	435	435	}
	436	436
@article{10.1145/3459665,	437	437	@article{10.1145/3459665,
author = {Cunningham, P\'{a}draig and Delany, Sarah Jane},	438	438	author = {Cunningham, P\'{a}draig and Delany, Sarah Jane},
title = {K-Nearest Neighbour Classifiers - A Tutorial},	439	439	title = {K-Nearest Neighbour Classifiers - A Tutorial},
year = {2021},	440	440	year = {2021},
issue_date = {July 2022},	441	441	issue_date = {July 2022},
publisher = {Association for Computing Machinery},	442	442	publisher = {Association for Computing Machinery},
address = {New York, NY, USA},	443	443	address = {New York, NY, USA},
type={article},	444	444	type={article},
language={English},	445	445	language={English},
volume = {54},	446	446	volume = {54},
number = {6},	447	447	number = {6},
issn = {0360-0300},	448	448	issn = {0360-0300},
url = {https://doi.org/10.1145/3459665},	449	449	url = {https://doi.org/10.1145/3459665},
doi = {10.1145/3459665},	450	450	doi = {10.1145/3459665},
abstract = {Perhaps the most straightforward classifier in the arsenal or Machine Learning techniques is the Nearest Neighbour Classifier—classification is achieved by identifying the nearest neighbours to a query example and using those neighbours to determine the class of the query. This approach to classification is of particular importance, because issues of poor runtime performance is not such a problem these days with the computational power that is available. This article presents an overview of techniques for Nearest Neighbour classification focusing on: mechanisms for assessing similarity (distance), computational issues in identifying nearest neighbours, and mechanisms for reducing the dimension of the data.This article is the second edition of a paper previously published as a technical report [16]. Sections on similarity measures for time-series, retrieval speedup, and intrinsic dimensionality have been added. An Appendix is included, providing access to Python code for the key methods.},	451	451	abstract = {Perhaps the most straightforward classifier in the arsenal or Machine Learning techniques is the Nearest Neighbour Classifier—classification is achieved by identifying the nearest neighbours to a query example and using those neighbours to determine the class of the query. This approach to classification is of particular importance, because issues of poor runtime performance is not such a problem these days with the computational power that is available. This article presents an overview of techniques for Nearest Neighbour classification focusing on: mechanisms for assessing similarity (distance), computational issues in identifying nearest neighbours, and mechanisms for reducing the dimension of the data.This article is the second edition of a paper previously published as a technical report [16]. Sections on similarity measures for time-series, retrieval speedup, and intrinsic dimensionality have been added. An Appendix is included, providing access to Python code for the key methods.},
journal = {ACM Comput. Surv.},	452	452	journal = {ACM Comput. Surv.},
month = {jul},	453	453	month = {jul},
articleno = {128},	454	454	articleno = {128},
numpages = {25},	455	455	numpages = {25},
keywords = {k-Nearest neighbour classifiers}	456	456	keywords = {k-Nearest neighbour classifiers}
}	457	457	}
	458	458
@article{9072123,	459	459	@article{9072123,
author={Sinaga, Kristina P. and Yang, Miin-Shen},	460	460	author={Sinaga, Kristina P. and Yang, Miin-Shen},
journal={IEEE Access},	461	461	journal={IEEE Access},
type={article},	462	462	type={article},
language={English},	463	463	language={English},
title={Unsupervised K-Means Clustering Algorithm},	464	464	title={Unsupervised K-Means Clustering Algorithm},
year={2020},	465	465	year={2020},
volume={8},	466	466	volume={8},
number={},	467	467	number={},
pages={80716-80727},	468	468	pages={80716-80727},
doi={10.1109/ACCESS.2020.2988796}	469	469	doi={10.1109/ACCESS.2020.2988796}
}	470	470	}
	471	471
@article{WANG2021331,	472	472	@article{WANG2021331,
title = {A new prediction strategy for dynamic multi-objective optimization using Gaussian Mixture Model},	473	473	title = {A new prediction strategy for dynamic multi-objective optimization using Gaussian Mixture Model},
journal = {Information Sciences},	474	474	journal = {Information Sciences},
volume = {580},	475	475	volume = {580},
type = {article},	476	476	type = {article},
language = {English},	477	477	language = {English},
pages = {331-351},	478	478	pages = {331-351},
year = {2021},	479	479	year = {2021},
issn = {0020-0255},	480	480	issn = {0020-0255},
doi = {https://doi.org/10.1016/j.ins.2021.08.065},	481	481	doi = {https://doi.org/10.1016/j.ins.2021.08.065},
url = {https://www.sciencedirect.com/science/article/pii/S0020025521008732},	482	482	url = {https://www.sciencedirect.com/science/article/pii/S0020025521008732},
author = {Feng Wang and Fanshu Liao and Yixuan Li and Hui Wang},	483	483	author = {Feng Wang and Fanshu Liao and Yixuan Li and Hui Wang},
keywords = {Dynamic multi-objective optimization, Gaussian Mixture Model, Change type detection, Resampling},	484	484	keywords = {Dynamic multi-objective optimization, Gaussian Mixture Model, Change type detection, Resampling},
abstract = {Dynamic multi-objective optimization problems (DMOPs), in which the environments change over time, have attracted many researchers’ attention in recent years. Since the Pareto set (PS) or the Pareto front (PF) can change over time, how to track the movement of the PS or PF is a challenging problem in DMOPs. Over the past few years, lots of methods have been proposed, and the prediction based strategy has been considered the most effective way to track the new PS. However, the performance of most existing prediction strategies depends greatly on the quantity and quality of the historical information and will deteriorate due to non-linear changes, leading to poor results. In this paper, we propose a new prediction method, named MOEA/D-GMM, which incorporates the Gaussian Mixture Model (GMM) into the MOEA/D framework for the prediction of the new PS when changes occur. Since GMM is a powerful non-linear model to accurately fit various data distributions, it can effectively generate solutions with better quality according to the distributions. In the proposed algorithm, a change type detection strategy is first designed to estimate an approximate PS according to different change types. Then, GMM is employed to make a more accurate prediction by training it with the approximate PS. To overcome the shortcoming of a lack of training solutions for GMM, the Empirical Cumulative Distribution Function (ECDF) method is used to resample more training solutions before GMM training. Experimental results on various benchmark test problems and a classical real-world problem show that, compared with some state-of-the-art dynamic optimization algorithms, MOEA/D-GMM outperforms others in most cases.}	485	485	abstract = {Dynamic multi-objective optimization problems (DMOPs), in which the environments change over time, have attracted many researchers’ attention in recent years. Since the Pareto set (PS) or the Pareto front (PF) can change over time, how to track the movement of the PS or PF is a challenging problem in DMOPs. Over the past few years, lots of methods have been proposed, and the prediction based strategy has been considered the most effective way to track the new PS. However, the performance of most existing prediction strategies depends greatly on the quantity and quality of the historical information and will deteriorate due to non-linear changes, leading to poor results. In this paper, we propose a new prediction method, named MOEA/D-GMM, which incorporates the Gaussian Mixture Model (GMM) into the MOEA/D framework for the prediction of the new PS when changes occur. Since GMM is a powerful non-linear model to accurately fit various data distributions, it can effectively generate solutions with better quality according to the distributions. In the proposed algorithm, a change type detection strategy is first designed to estimate an approximate PS according to different change types. Then, GMM is employed to make a more accurate prediction by training it with the approximate PS. To overcome the shortcoming of a lack of training solutions for GMM, the Empirical Cumulative Distribution Function (ECDF) method is used to resample more training solutions before GMM training. Experimental results on various benchmark test problems and a classical real-world problem show that, compared with some state-of-the-art dynamic optimization algorithms, MOEA/D-GMM outperforms others in most cases.}
}	486	486	}
	487	487
@article{9627973,	488	488	@article{9627973,
author={Xu, Shengbing and Cai, Wei and Xia, Hongxi and Liu, Bo and Xu, Jie},	489	489	author={Xu, Shengbing and Cai, Wei and Xia, Hongxi and Liu, Bo and Xu, Jie},
journal={IEEE Access},	490	490	journal={IEEE Access},
title={Dynamic Metric Accelerated Method for Fuzzy Clustering},	491	491	title={Dynamic Metric Accelerated Method for Fuzzy Clustering},
year={2021},	492	492	year={2021},
type={article},	493	493	type={article},
language={English},	494	494	language={English},
volume={9},	495	495	volume={9},
number={},	496	496	number={},
pages={166838-166854},	497	497	pages={166838-166854},
doi={10.1109/ACCESS.2021.3131368}	498	498	doi={10.1109/ACCESS.2021.3131368}
}	499	499	}
	500	500
@article{9434422,	501	501	@article{9434422,
author={Gupta, Samarth and Chaudhari, Shreyas and Joshi, Gauri and Yağan, Osman},	502	502	author={Gupta, Samarth and Chaudhari, Shreyas and Joshi, Gauri and Yağan, Osman},
journal={IEEE Transactions on Information Theory},	503	503	journal={IEEE Transactions on Information Theory},
title={Multi-Armed Bandits With Correlated Arms},	504	504	title={Multi-Armed Bandits With Correlated Arms},
year={2021},	505	505	year={2021},
language={English},	506	506	language={English},
type={article},	507	507	type={article},
volume={67},	508	508	volume={67},
number={10},	509	509	number={10},
pages={6711-6732},	510	510	pages={6711-6732},
doi={10.1109/TIT.2021.3081508}	511	511	doi={10.1109/TIT.2021.3081508}
}	512	512	}
	513	513
@Inproceedings{8495930,	514	514	@Inproceedings{8495930,
author={Supic, H.},	515	515	author={Supic, H.},
booktitle={2018 IEEE 27th International Conference on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE)},	516	516	booktitle={2018 IEEE 27th International Conference on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE)},
title={Case-Based Reasoning Model for Personalized Learning Path Recommendation in Example-Based Learning Activities},	517	517	title={Case-Based Reasoning Model for Personalized Learning Path Recommendation in Example-Based Learning Activities},
year={2018},	518	518	year={2018},
type={article},	519	519	type={article},
language={English},	520	520	language={English},
volume={},	521	521	volume={},
number={},	522	522	number={},
pages={175-178},	523	523	pages={175-178},
doi={10.1109/WETICE.2018.00040}	524	524	doi={10.1109/WETICE.2018.00040}
}	525	525	}
	526	526
@Inproceedings{9870279,	527	527	@Inproceedings{9870279,
author={Lin, Baihan},	528	528	author={Lin, Baihan},
booktitle={2022 IEEE Congress on Evolutionary Computation (CEC)},	529	529	booktitle={2022 IEEE Congress on Evolutionary Computation (CEC)},
title={Evolutionary Multi-Armed Bandits with Genetic Thompson Sampling},	530	530	title={Evolutionary Multi-Armed Bandits with Genetic Thompson Sampling},
year={2022},	531	531	year={2022},
type={article},	532	532	type={article},
language={English},	533	533	language={English},
volume={},	534	534	volume={},
number={},	535	535	number={},
pages={1-8},	536	536	pages={1-8},
doi={10.1109/CEC55065.2022.9870279}	537	537	doi={10.1109/CEC55065.2022.9870279}
}	538	538	}
	539	539
@article{Obeid,	540	540	@article{Obeid,
author={Obeid, C. and Lahoud, C. and Khoury, H. E. and Champin, P.},	541	541	author={Obeid, C. and Lahoud, C. and Khoury, H. E. and Champin, P.},
title={A Novel Hybrid Recommender System Approach for Student Academic Advising Named COHRS, Supported by Case-based Reasoning and Ontology},	542	542	title={A Novel Hybrid Recommender System Approach for Student Academic Advising Named COHRS, Supported by Case-based Reasoning and Ontology},
journal={Computer Science and Information Systems},	543	543	journal={Computer Science and Information Systems},
type={article},	544	544	type={article},
language={English},	545	545	language={English},
volume={19},	546	546	volume={19},
number={2},	547	547	number={2},
pages={979–1005},	548	548	pages={979–1005},
year={2022},	549	549	year={2022},
doi={https://doi.org/10.2298/CSIS220215011O}	550	550	doi={https://doi.org/10.2298/CSIS220215011O}
}	551	551	}
	552	552
@book{Nkambou,	553	553	@book{Nkambou,
author = {Nkambou, R. and Bourdeau, J. and Mizoguchi, R.},	554	554	author = {Nkambou, R. and Bourdeau, J. and Mizoguchi, R.},
title = {Advances in Intelligent Tutoring Systems},	555	555	title = {Advances in Intelligent Tutoring Systems},
year = {2010},	556	556	year = {2010},
type = {article},	557	557	type = {article},
language = {English},	558	558	language = {English},
publisher = {Springer Berlin, Heidelberg},	559	559	publisher = {Springer Berlin, Heidelberg},
edition = {1}	560	560	edition = {1}
}	561	561	}
	562	562
@book{hajduk2019cognitive,	563	563	@book{hajduk2019cognitive,
title={Cognitive Multi-agent Systems: Structures, Strategies and Applications to Mobile Robotics and Robosoccer},	564	564	title={Cognitive Multi-agent Systems: Structures, Strategies and Applications to Mobile Robotics and Robosoccer},
author={Hajduk, M. and Sukop, M. and Haun, M.},	565	565	author={Hajduk, M. and Sukop, M. and Haun, M.},
type={book},	566	566	type={book},
language={English},	567	567	language={English},
isbn={9783319936857},	568	568	isbn={9783319936857},
series={Studies in Systems, Decision and Control},	569	569	series={Studies in Systems, Decision and Control},
year={2019},	570	570	year={2019},
publisher={Springer International Publishing}	571	571	publisher={Springer International Publishing}
}	572	572	}
	573	573
@article{RICHTER20093,	574	574	@article{RICHTER20093,
title = {The search for knowledge, contexts, and Case-Based Reasoning},	575	575	title = {The search for knowledge, contexts, and Case-Based Reasoning},
journal = {Engineering Applications of Artificial Intelligence},	576	576	journal = {Engineering Applications of Artificial Intelligence},
language = {English},	577	577	language = {English},
type = {article},	578	578	type = {article},
volume = {22},	579	579	volume = {22},
number = {1},	580	580	number = {1},
pages = {3-9},	581	581	pages = {3-9},
year = {2009},	582	582	year = {2009},
issn = {0952-1976},	583	583	issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2008.04.021},	584	584	doi = {https://doi.org/10.1016/j.engappai.2008.04.021},
url = {https://www.sciencedirect.com/science/article/pii/S095219760800078X},	585	585	url = {https://www.sciencedirect.com/science/article/pii/S095219760800078X},
author = {Michael M. Richter},	586	586	author = {Michael M. Richter},
keywords = {Case-Based Reasoning, Knowledge, Processes, Utility, Context},	587	587	keywords = {Case-Based Reasoning, Knowledge, Processes, Utility, Context},
abstract = {A major goal of this paper is to compare Case-Based Reasoning with other methods searching for knowledge. We consider knowledge as a resource that can be traded. It has no value in itself; the value is measured by the usefulness of applying it in some process. Such a process has info-needs that have to be satisfied. The concept to measure this is the economical term utility. In general, utility depends on the user and its context, i.e., it is subjective. Here, we introduce levels of contexts from general to individual. We illustrate that Case-Based Reasoning on the lower, i.e., more personal levels CBR is quite useful, in particular in comparison with traditional informational retrieval methods.}	588	588	abstract = {A major goal of this paper is to compare Case-Based Reasoning with other methods searching for knowledge. We consider knowledge as a resource that can be traded. It has no value in itself; the value is measured by the usefulness of applying it in some process. Such a process has info-needs that have to be satisfied. The concept to measure this is the economical term utility. In general, utility depends on the user and its context, i.e., it is subjective. Here, we introduce levels of contexts from general to individual. We illustrate that Case-Based Reasoning on the lower, i.e., more personal levels CBR is quite useful, in particular in comparison with traditional informational retrieval methods.}
}	589	589	}
	590	590
@Thesis{Marie,	591	591	@Thesis{Marie,
author={Marie, F.},	592	592	author={Marie, F.},
title={COLISEUM-3D. Une plate-forme innovante pour la segmentation d’images médicales par Raisonnement à Partir de Cas (RàPC) et méthodes d’apprentissage de type Deep Learning},	593	593	title={COLISEUM-3D. Une plate-forme innovante pour la segmentation d’images médicales par Raisonnement à Partir de Cas (RàPC) et méthodes d’apprentissage de type Deep Learning},
type={diplomathesis},	594	594	type={diplomathesis},
language={French},	595	595	language={French},
institution={Université de Franche-Comte},	596	596	institution={Université de Franche-Comte},
year={2019}	597	597	year={2019}
}	598	598	}
	599	599
@book{Hoang,	600	600	@book{Hoang,
title = {La formule du savoir. Une philosophie unifiée du savoir fondée sur le théorème de Bayes},	601	601	title = {La formule du savoir. Une philosophie unifiée du savoir fondée sur le théorème de Bayes},
author = {Hoang, L.N.},	602	602	author = {Hoang, L.N.},
type = {book},	603	603	type = {book},
language = {French},	604	604	language = {French},
isbn = {9782759822607},	605	605	isbn = {9782759822607},
year = {2018},	606	606	year = {2018},
publisher = {EDP Sciences}	607	607	publisher = {EDP Sciences}
}	608	608	}
	609	609
@book{Richter2013,	610	610	@book{Richter2013,
title={Case-Based Reasoning (A Textbook)},	611	611	title={Case-Based Reasoning (A Textbook)},
author={Richter, M. and Weber, R.},	612	612	author={Richter, M. and Weber, R.},
type={book},	613	613	type={book},
language={English},	614	614	language={English},
isbn={9783642401664},	615	615	isbn={9783642401664},
year={2013},	616	616	year={2013},
publisher={Springer-Verlag GmbH}	617	617	publisher={Springer-Verlag GmbH}
}	618	618	}
	619	619
@book{kedia2020hands,	620	620	@book{kedia2020hands,
title={Hands-On Python Natural Language Processing: Explore tools and techniques to analyze and process text with a view to building real-world NLP applications},	621	621	title={Hands-On Python Natural Language Processing: Explore tools and techniques to analyze and process text with a view to building real-world NLP applications},
author={Kedia, A. and Rasu, M.},	622	622	author={Kedia, A. and Rasu, M.},
language={English},	623	623	language={English},
type={book},	624	624	type={book},
isbn={9781838982584},	625	625	isbn={9781838982584},
url={https://books.google.fr/books?id=1AbuDwAAQBAJ},	626	626	url={https://books.google.fr/books?id=1AbuDwAAQBAJ},
year={2020},	627	627	year={2020},
publisher={Packt Publishing}	628	628	publisher={Packt Publishing}
}	629	629	}
	630	630
@book{ghosh2019natural,	631	631	@book{ghosh2019natural,
title={Natural Language Processing Fundamentals: Build intelligent applications that can interpret the human language to deliver impactful results},	632	632	title={Natural Language Processing Fundamentals: Build intelligent applications that can interpret the human language to deliver impactful results},
author={Ghosh, S. and Gunning, D.},	633	633	author={Ghosh, S. and Gunning, D.},
language={English},	634	634	language={English},
type={book},	635	635	type={book},
isbn={9781789955989},	636	636	isbn={9781789955989},
url={https://books.google.fr/books?id=i8-PDwAAQBAJ},	637	637	url={https://books.google.fr/books?id=i8-PDwAAQBAJ},
year={2019},	638	638	year={2019},
publisher={Packt Publishing}	639	639	publisher={Packt Publishing}
}	640	640	}
	641	641
@article{Akerblom,	642	642	@article{Akerblom,
title={Online learning of network bottlenecks via minimax paths},	643	643	title={Online learning of network bottlenecks via minimax paths},
author={kerblom, Niklas and Hoseini, Fazeleh Sadat and Haghir Chehreghani, Morteza},	644	644	author={kerblom, Niklas and Hoseini, Fazeleh Sadat and Haghir Chehreghani, Morteza},
language={English},	645	645	language={English},
type={article},	646	646	type={article},
volume = {122},	647	647	volume = {122},
year = {2023},	648	648	year = {2023},
issn = {1573-0565},	649	649	issn = {1573-0565},
doi = {https://doi.org/10.1007/s10994-022-06270-0},	650	650	doi = {https://doi.org/10.1007/s10994-022-06270-0},
url = {https://doi.org/10.1007/s10994-022-06270-0},	651	651	url = {https://doi.org/10.1007/s10994-022-06270-0},
abstract={In this paper, we study bottleneck identification in networks via extracting minimax paths. Many real-world networks have stochastic weights for which full knowledge is not available in advance. Therefore, we model this task as a combinatorial semi-bandit problem to which we apply a combinatorial version of Thompson Sampling and establish an upper bound on the corresponding Bayesian regret. Due to the computational intractability of the problem, we then devise an alternative problem formulation which approximates the original objective. Finally, we experimentally evaluate the performance of Thompson Sampling with the approximate formulation on real-world directed and undirected networks.}	652	652	abstract={In this paper, we study bottleneck identification in networks via extracting minimax paths. Many real-world networks have stochastic weights for which full knowledge is not available in advance. Therefore, we model this task as a combinatorial semi-bandit problem to which we apply a combinatorial version of Thompson Sampling and establish an upper bound on the corresponding Bayesian regret. Due to the computational intractability of the problem, we then devise an alternative problem formulation which approximates the original objective. Finally, we experimentally evaluate the performance of Thompson Sampling with the approximate formulation on real-world directed and undirected networks.}
}	653	653	}
	654	654
@article{Simen,	655	655	@article{Simen,
title={Dynamic slate recommendation with gated recurrent units and Thompson sampling},	656	656	title={Dynamic slate recommendation with gated recurrent units and Thompson sampling},
author={Eide, Simen and Leslie, David S. and Frigessi, Arnoldo},	657	657	author={Eide, Simen and Leslie, David S. and Frigessi, Arnoldo},
language={English},	658	658	language={English},
type={article},	659	659	type={article},
volume = {36},	660	660	volume = {36},
year = {2022},	661	661	year = {2022},
issn = {1573-756X},	662	662	issn = {1573-756X},
doi = {https://doi.org/10.1007/s10618-022-00849-w},	663	663	doi = {https://doi.org/10.1007/s10618-022-00849-w},
url = {https://doi.org/10.1007/s10618-022-00849-w},	664	664	url = {https://doi.org/10.1007/s10618-022-00849-w},
abstract={We consider the problem of recommending relevant content to users of an internet platform in the form of lists of items, called slates. We introduce a variational Bayesian Recurrent Neural Net recommender system that acts on time series of interactions between the internet platform and the user, and which scales to real world industrial situations. The recommender system is tested both online on real users, and on an offline dataset collected from a Norwegian web-based marketplace, FINN.no, that is made public for research. This is one of the first publicly available datasets which includes all the slates that are presented to users as well as which items (if any) in the slates were clicked on. Such a data set allows us to move beyond the common assumption that implicitly assumes that users are considering all possible items at each interaction. Instead we build our likelihood using the items that are actually in the slate, and evaluate the strengths and weaknesses of both approaches theoretically and in experiments. We also introduce a hierarchical prior for the item parameters based on group memberships. Both item parameters and user preferences are learned probabilistically. Furthermore, we combine our model with bandit strategies to ensure learning, and introduce ‘in-slate Thompson sampling’ which makes use of the slates to maximise explorative opportunities. We show experimentally that explorative recommender strategies perform on par or above their greedy counterparts. Even without making use of exploration to learn more effectively, click rates increase simply because of improved diversity in the recommended slates.}	665	665	abstract={We consider the problem of recommending relevant content to users of an internet platform in the form of lists of items, called slates. We introduce a variational Bayesian Recurrent Neural Net recommender system that acts on time series of interactions between the internet platform and the user, and which scales to real world industrial situations. The recommender system is tested both online on real users, and on an offline dataset collected from a Norwegian web-based marketplace, FINN.no, that is made public for research. This is one of the first publicly available datasets which includes all the slates that are presented to users as well as which items (if any) in the slates were clicked on. Such a data set allows us to move beyond the common assumption that implicitly assumes that users are considering all possible items at each interaction. Instead we build our likelihood using the items that are actually in the slate, and evaluate the strengths and weaknesses of both approaches theoretically and in experiments. We also introduce a hierarchical prior for the item parameters based on group memberships. Both item parameters and user preferences are learned probabilistically. Furthermore, we combine our model with bandit strategies to ensure learning, and introduce ‘in-slate Thompson sampling’ which makes use of the slates to maximise explorative opportunities. We show experimentally that explorative recommender strategies perform on par or above their greedy counterparts. Even without making use of exploration to learn more effectively, click rates increase simply because of improved diversity in the recommended slates.}
}	666	666	}
	667	667
@Inproceedings{Arthurs,	668	668	@Inproceedings{Arthurs,
author={Arthurs, Noah and Stenhaug, Ben and Karayev, Sergey and Piech, Chris},	669	669	author={Arthurs, Noah and Stenhaug, Ben and Karayev, Sergey and Piech, Chris},
booktitle={International Conference on Educational Data Mining (EDM)},	670	670	booktitle={International Conference on Educational Data Mining (EDM)},
title={Grades Are Not Normal: Improving Exam Score Models Using the Logit-Normal Distribution},	671	671	title={Grades Are Not Normal: Improving Exam Score Models Using the Logit-Normal Distribution},
year={2019},	672	672	year={2019},
type={article},	673	673	type={article},
language={English},	674	674	language={English},
volume={},	675	675	volume={},
number={},	676	676	number={},
pages={6},	677	677	pages={6},
url={https://eric.ed.gov/?id=ED599204}	678	678	url={https://eric.ed.gov/?id=ED599204}
}	679	679	}
	680	680
@article{Bahramian,	681	681	@article{Bahramian,
title={A Cold Start Context-Aware Recommender System for Tour Planning Using Artificial Neural Network and Case Based Reasoning},	682	682	title={A Cold Start Context-Aware Recommender System for Tour Planning Using Artificial Neural Network and Case Based Reasoning},
author={Bahramian, Zahra and Ali Abbaspour, Rahim and Claramunt, Christophe},	683	683	author={Bahramian, Zahra and Ali Abbaspour, Rahim and Claramunt, Christophe},
language={English},	684	684	language={English},
type={article},	685	685	type={article},
year = {2017},	686	686	year = {2017},
issn = {1574-017X},	687	687	issn = {1574-017X},
doi = {https://doi.org/10.1155/2017/9364903},	688	688	doi = {https://doi.org/10.1155/2017/9364903},
url = {https://doi.org/10.1155/2017/9364903},	689	689	url = {https://doi.org/10.1155/2017/9364903},
abstract={Nowadays, large amounts of tourism information and services are available over the Web. This makes it difficult for the user to search for some specific information such as selecting a tour in a given city as an ordered set of points of interest. Moreover, the user rarely knows all his needs upfront and his preferences may change during a recommendation process. The user may also have a limited number of initial ratings and most often the recommender system is likely to face the well-known cold start problem. The objective of the research presented in this paper is to introduce a hybrid interactive context-aware tourism recommender system that takes into account user’s feedbacks and additional contextual information. It offers personalized tours to the user based on his preferences thanks to the combination of a case based reasoning framework and an artificial neural network. The proposed method has been tried in the city of Tehran in Iran. The results show that the proposed method outperforms current artificial neural network methods and combinations of case based reasoning with <svg xmlns:xlink="http://www.w3.org/1999/xlink" xmlns="http://www.w3.org/2000/svg" style="vertical-align:-0.2063999pt" id="M1" height="9.49473pt" version="1.1" viewBox="-0.0498162 -9.28833 6.66314 9.49473" width="6.66314pt"><g transform="matrix(.013,0,0,-0.013,0,0)"><path id="g113-108" d="M480 416C480 431 465 448 438 448C388 448 312 383 252 330C217 299 188 273 155 237H153L257 680C262 700 263 712 253 712C240 712 183 684 97 674L92 648L126 647C166 646 172 645 163 606L23 -6L29 -12C51 -5 77 2 107 8C115 62 130 128 142 180C153 193 179 220 204 241C231 170 259 106 288 54C317 0 336 -12 358 -12C381 -12 423 2 477 80L460 100C434 74 408 54 398 54C385 54 374 65 351 107C326 154 282 241 263 299C296 332 351 377 403 377C424 377 436 372 445 368C449 366 456 368 462 375C472 386 480 402 480 416Z"/></g></svg>-nearest neighbor methods in terms of user effort, accuracy, and user satisfaction.}	690	690	abstract={Nowadays, large amounts of tourism information and services are available over the Web. This makes it difficult for the user to search for some specific information such as selecting a tour in a given city as an ordered set of points of interest. Moreover, the user rarely knows all his needs upfront and his preferences may change during a recommendation process. The user may also have a limited number of initial ratings and most often the recommender system is likely to face the well-known cold start problem. The objective of the research presented in this paper is to introduce a hybrid interactive context-aware tourism recommender system that takes into account user’s feedbacks and additional contextual information. It offers personalized tours to the user based on his preferences thanks to the combination of a case based reasoning framework and an artificial neural network. The proposed method has been tried in the city of Tehran in Iran. The results show that the proposed method outperforms current artificial neural network methods and combinations of case based reasoning with <svg xmlns:xlink="http://www.w3.org/1999/xlink" xmlns="http://www.w3.org/2000/svg" style="vertical-align:-0.2063999pt" id="M1" height="9.49473pt" version="1.1" viewBox="-0.0498162 -9.28833 6.66314 9.49473" width="6.66314pt"><g transform="matrix(.013,0,0,-0.013,0,0)"><path id="g113-108" d="M480 416C480 431 465 448 438 448C388 448 312 383 252 330C217 299 188 273 155 237H153L257 680C262 700 263 712 253 712C240 712 183 684 97 674L92 648L126 647C166 646 172 645 163 606L23 -6L29 -12C51 -5 77 2 107 8C115 62 130 128 142 180C153 193 179 220 204 241C231 170 259 106 288 54C317 0 336 -12 358 -12C381 -12 423 2 477 80L460 100C434 74 408 54 398 54C385 54 374 65 351 107C326 154 282 241 263 299C296 332 351 377 403 377C424 377 436 372 445 368C449 366 456 368 462 375C472 386 480 402 480 416Z"/></g></svg>-nearest neighbor methods in terms of user effort, accuracy, and user satisfaction.}
}	691	691	}
	692	692
@Thesis{Daubias2011,	693	693	@Thesis{Daubias2011,
author={Sthéphanie Jean-Daubias},	694	694	author={Sthéphanie Jean-Daubias},
title={Ingénierie des profils d'apprenants},	695	695	title={Ingénierie des profils d'apprenants},
type={diplomathesis},	696	696	type={diplomathesis},
language={French},	697	697	language={French},
institution={Université Claude Bernard Lyon 1},	698	698	institution={Université Claude Bernard Lyon 1},
year={2011}	699	699	year={2011}
}	700	700	}
	701	701
@article{Tapalova,	702	702	@article{Tapalova,
author = {Olga Tapalova and Nadezhda Zhiyenbayeva},	703	703	author = {Olga Tapalova and Nadezhda Zhiyenbayeva},
title ={Artificial Intelligence in Education: AIEd for Personalised Learning Pathways},	704	704	title ={Artificial Intelligence in Education: AIEd for Personalised Learning Pathways},
journal = {Electronic Journal of e-Learning},	705	705	journal = {Electronic Journal of e-Learning},
volume = {},	706	706	volume = {},
number = {},	707	707	number = {},
pages = {15},	708	708	pages = {15},
year = {2022},	709	709	year = {2022},
URL = {https://eric.ed.gov/?q=Artificial+Intelligence+in+Education%3a+AIEd+for+Personalised+Learning+Pathways&id=EJ1373006},	710	710	URL = {https://eric.ed.gov/?q=Artificial+Intelligence+in+Education%3a+AIEd+for+Personalised+Learning+Pathways&id=EJ1373006},
language={English},	711	711	language={English},
type={article},	712	712	type={article},
abstract = {Artificial intelligence is the driving force of change focusing on the needs and demands of the student. The research explores Artificial Intelligence in Education (AIEd) for building personalised learning systems for students. The research investigates and proposes a framework for AIEd: social networking sites and chatbots, expert systems for education, intelligent mentors and agents, machine learning, personalised educational systems and virtual educational environments. These technologies help educators to develop and introduce personalised approaches to master new knowledge and develop professional competencies. The research presents a case study of AIEd implementation in education. The scholars conducted the experiment in educational establishments using artificial intelligence in the curriculum. The scholars surveyed 184 second-year students of the Institute of Pedagogy and Psychology at the Abay Kazakh National Pedagogical University and the Kuban State Technological University to collect the data. The scholars considered the collective group discussions regarding the application of artificial intelligence in education to improve the effectiveness of learning. The research identified key advantages to creating personalised learning pathways such as access to training in 24/7 mode, training in virtual contexts, adaptation of educational content to personal needs of students, real-time and regular feedback, improvements in the educational process and mental stimulations. The proposed education paradigm reflects the increasing role of artificial intelligence in socio-economic life, the social and ethical concerns artificial intelligence may pose to humanity and its role in the digitalisation of education. The current article may be used as a theoretical framework for many educational institutions planning to exploit the capabilities of artificial intelligence in their adaptation to personalized learning.}	713	713	abstract = {Artificial intelligence is the driving force of change focusing on the needs and demands of the student. The research explores Artificial Intelligence in Education (AIEd) for building personalised learning systems for students. The research investigates and proposes a framework for AIEd: social networking sites and chatbots, expert systems for education, intelligent mentors and agents, machine learning, personalised educational systems and virtual educational environments. These technologies help educators to develop and introduce personalised approaches to master new knowledge and develop professional competencies. The research presents a case study of AIEd implementation in education. The scholars conducted the experiment in educational establishments using artificial intelligence in the curriculum. The scholars surveyed 184 second-year students of the Institute of Pedagogy and Psychology at the Abay Kazakh National Pedagogical University and the Kuban State Technological University to collect the data. The scholars considered the collective group discussions regarding the application of artificial intelligence in education to improve the effectiveness of learning. The research identified key advantages to creating personalised learning pathways such as access to training in 24/7 mode, training in virtual contexts, adaptation of educational content to personal needs of students, real-time and regular feedback, improvements in the educational process and mental stimulations. The proposed education paradigm reflects the increasing role of artificial intelligence in socio-economic life, the social and ethical concerns artificial intelligence may pose to humanity and its role in the digitalisation of education. The current article may be used as a theoretical framework for many educational institutions planning to exploit the capabilities of artificial intelligence in their adaptation to personalized learning.}
}	714	714	}
	715	715
@article{Auer,	716	716	@article{Auer,
title = {From monolithic systems to Microservices: An assessment framework},	717	717	title = {From monolithic systems to Microservices: An assessment framework},
journal = {Information and Software Technology},	718	718	journal = {Information and Software Technology},
volume = {137},	719	719	volume = {137},
pages = {106600},	720	720	pages = {106600},
year = {2021},	721	721	year = {2021},
issn = {0950-5849},	722	722	issn = {0950-5849},
doi = {https://doi.org/10.1016/j.infsof.2021.106600},	723	723	doi = {https://doi.org/10.1016/j.infsof.2021.106600},
url = {https://www.sciencedirect.com/science/article/pii/S0950584921000793},	724	724	url = {https://www.sciencedirect.com/science/article/pii/S0950584921000793},
author = {Florian Auer and Valentina Lenarduzzi and Michael Felderer and Davide Taibi},	725	725	author = {Florian Auer and Valentina Lenarduzzi and Michael Felderer and Davide Taibi},
keywords = {Microservices, Cloud migration, Software measurement},	726	726	keywords = {Microservices, Cloud migration, Software measurement},
abstract = {Context:	727	727	abstract = {Context:
Re-architecting monolithic systems with Microservices-based architecture is a common trend. Various companies are migrating to Microservices for different reasons. However, making such an important decision like re-architecting an entire system must be based on real facts and not only on gut feelings.	728	728	Re-architecting monolithic systems with Microservices-based architecture is a common trend. Various companies are migrating to Microservices for different reasons. However, making such an important decision like re-architecting an entire system must be based on real facts and not only on gut feelings.
Objective:	729	729	Objective:
The goal of this work is to propose an evidence-based decision support framework for companies that need to migrate to Microservices, based on the analysis of a set of characteristics and metrics they should collect before re-architecting their monolithic system.	730	730	The goal of this work is to propose an evidence-based decision support framework for companies that need to migrate to Microservices, based on the analysis of a set of characteristics and metrics they should collect before re-architecting their monolithic system.
Method:	731	731	Method:
We conducted a survey done in the form of interviews with professionals to derive the assessment framework based on Grounded Theory.	732	732	We conducted a survey done in the form of interviews with professionals to derive the assessment framework based on Grounded Theory.
Results:	733	733	Results:
We identified a set consisting of information and metrics that companies can use to decide whether to migrate to Microservices or not. The proposed assessment framework, based on the aforementioned metrics, could be useful for companies if they need to migrate to Microservices and do not want to run the risk of failing to consider some important information.}	734	734	We identified a set consisting of information and metrics that companies can use to decide whether to migrate to Microservices or not. The proposed assessment framework, based on the aforementioned metrics, could be useful for companies if they need to migrate to Microservices and do not want to run the risk of failing to consider some important information.}
}	735	735	}
	736	736
@Article{jmse10040464,	737	737	@Article{jmse10040464,
AUTHOR = {Zuluaga, Carlos A. and Aristizábal, Luis M. and Rúa, Santiago and Franco, Diego A. and Osorio, Dorie A. and Vásquez, Rafael E.},	738	738	AUTHOR = {Zuluaga, Carlos A. and Aristizábal, Luis M. and Rúa, Santiago and Franco, Diego A. and Osorio, Dorie A. and Vásquez, Rafael E.},
TITLE = {Development of a Modular Software Architecture for Underwater Vehicles Using Systems Engineering},	739	739	TITLE = {Development of a Modular Software Architecture for Underwater Vehicles Using Systems Engineering},
JOURNAL = {Journal of Marine Science and Engineering},	740	740	JOURNAL = {Journal of Marine Science and Engineering},
VOLUME = {10},	741	741	VOLUME = {10},
YEAR = {2022},	742	742	YEAR = {2022},
NUMBER = {4},	743	743	NUMBER = {4},
ARTICLE-NUMBER = {464},	744	744	ARTICLE-NUMBER = {464},
URL = {https://www.mdpi.com/2077-1312/10/4/464},	745	745	URL = {https://www.mdpi.com/2077-1312/10/4/464},
ISSN = {2077-1312},	746	746	ISSN = {2077-1312},
ABSTRACT = {This paper addresses the development of a modular software architecture for the design/construction/operation of a remotely operated vehicle (ROV), based on systems engineering. First, systems engineering and the Vee model are presented with the objective of defining the interactions of the stakeholders with the software architecture development team and establishing the baselines that must be met in each development phase. In the development stage, the definition of the architecture and its connection with the hardware is presented, taking into account the use of the actor model, which represents the high-level software architecture used to solve concurrency problems. Subsequently, the structure of the classes is defined both at high and low levels in the instruments using the object-oriented programming paradigm. Finally, unit tests are developed for each component in the software architecture, quality assessment tests are implemented for system functions fulfillment, and a field sea trial for testing different modules of the vehicle is described. This approach is well suited for the development of complex systems such as marine vehicles and those systems which require scalability and modularity to add functionalities.},	747	747	ABSTRACT = {This paper addresses the development of a modular software architecture for the design/construction/operation of a remotely operated vehicle (ROV), based on systems engineering. First, systems engineering and the Vee model are presented with the objective of defining the interactions of the stakeholders with the software architecture development team and establishing the baselines that must be met in each development phase. In the development stage, the definition of the architecture and its connection with the hardware is presented, taking into account the use of the actor model, which represents the high-level software architecture used to solve concurrency problems. Subsequently, the structure of the classes is defined both at high and low levels in the instruments using the object-oriented programming paradigm. Finally, unit tests are developed for each component in the software architecture, quality assessment tests are implemented for system functions fulfillment, and a field sea trial for testing different modules of the vehicle is described. This approach is well suited for the development of complex systems such as marine vehicles and those systems which require scalability and modularity to add functionalities.},
DOI = {10.3390/jmse10040464}	748	748	DOI = {10.3390/jmse10040464}
}	749	749	}
	750	750
@article{doi:10.1177/1754337116651013,	751	751	@article{doi:10.1177/1754337116651013,
author = {Julien Henriet and Lang Christophe and Philippe Laurent},	752	752	author = {Julien Henriet and Lang Christophe and Philippe Laurent},
title ={Artificial Intelligence-Virtual Trainer: An educative system based on artificial intelligence and designed to produce varied and consistent training lessons},	753	753	title ={Artificial Intelligence-Virtual Trainer: An educative system based on artificial intelligence and designed to produce varied and consistent training lessons},
journal = {Proceedings of the Institution of Mechanical Engineers, Part P: Journal of Sports Engineering and Technology},	754	754	journal = {Proceedings of the Institution of Mechanical Engineers, Part P: Journal of Sports Engineering and Technology},
volume = {231},	755	755	volume = {231},
number = {2},	756	756	number = {2},
pages = {110-124},	757	757	pages = {110-124},
year = {2017},	758	758	year = {2017},
doi = {10.1177/1754337116651013},	759	759	doi = {10.1177/1754337116651013},
URL = {https://doi.org/10.1177/1754337116651013},	760	760	URL = {https://doi.org/10.1177/1754337116651013},
eprint = {https://doi.org/10.1177/1754337116651013},	761	761	eprint = {https://doi.org/10.1177/1754337116651013},
abstract = { AI-Virtual Trainer is an educative system using Artificial Intelligence to propose varied lessons to trainers. The agents of this multi-agent system apply case-based reasoning to build solutions by analogy. However, as required by the field, Artificial Intelligence-Virtual Trainer never proposes the same lesson twice, whereas the same objective may be set many times consecutively. The adaptation process of Artificial Intelligence-Virtual Trainer delivers an ordered set of exercises adapted to the objectives and sub-objectives chosen by trainers. This process has been enriched by including the notion of distance between exercises: the proposed tasks are not only appropriate but are hierarchically ordered. With this new version of the system, students are guided towards their objectives via an underlying theme. Finally, the agents responsible for the different parts of lessons collaborate with each other according to a dedicated protocol and decision-making policy since no exercise must appear more than once in the same lesson. The results prove that Artificial Intelligence-Virtual Trainer, however perfectible, meets the requirements of this field. }	762	762	abstract = { AI-Virtual Trainer is an educative system using Artificial Intelligence to propose varied lessons to trainers. The agents of this multi-agent system apply case-based reasoning to build solutions by analogy. However, as required by the field, Artificial Intelligence-Virtual Trainer never proposes the same lesson twice, whereas the same objective may be set many times consecutively. The adaptation process of Artificial Intelligence-Virtual Trainer delivers an ordered set of exercises adapted to the objectives and sub-objectives chosen by trainers. This process has been enriched by including the notion of distance between exercises: the proposed tasks are not only appropriate but are hierarchically ordered. With this new version of the system, students are guided towards their objectives via an underlying theme. Finally, the agents responsible for the different parts of lessons collaborate with each other according to a dedicated protocol and decision-making policy since no exercise must appear more than once in the same lesson. The results prove that Artificial Intelligence-Virtual Trainer, however perfectible, meets the requirements of this field. }
}	763	763	}
	764	764
@InProceedings{10.1007/978-3-030-01081-2_9,	765	765	@InProceedings{10.1007/978-3-030-01081-2_9,
author="Henriet, Julien	766	766	author="Henriet, Julien
and Greffier, Fran{\c{c}}oise",	767	767	and Greffier, Fran{\c{c}}oise",
editor="Cox, Michael T.	768	768	editor="Cox, Michael T.
and Funk, Peter	769	769	and Funk, Peter
and Begum, Shahina",	770	770	and Begum, Shahina",
title="AI-VT: An Example of CBR that Generates a Variety of Solutions to the Same Problem",	771	771	title="AI-VT: An Example of CBR that Generates a Variety of Solutions to the Same Problem",
booktitle="Case-Based Reasoning Research and Development",	772	772	booktitle="Case-Based Reasoning Research and Development",
year="2018",	773	773	year="2018",
publisher="Springer International Publishing",	774	774	publisher="Springer International Publishing",
address="Cham",	775	775	address="Cham",
pages="124--139",	776	776	pages="124--139",
abstract="AI-Virtual Trainer (AI-VT) is an intelligent tutoring system based on case-based reasoning. AI-VT has been designed to generate personalised, varied, and consistent training sessions for learners. The AI-VT training sessions propose different exercises in regard to a capacity associated with sub-capacities. For example, in the field of training for algorithms, a capacity could be ``Use a control structure alternative'' and an associated sub-capacity could be ``Write a boolean condition''. AI-VT can elaborate a personalised list of exercises for each learner. One of the main requirements and challenges studied in this work is its ability to propose varied training sessions to the same learner for many weeks, which constitutes the challenge studied in our work. Indeed, if the same set of exercises is proposed time after time to learners, they will stop paying attention and lose motivation. Thus, even if the generation of training sessions is based on analogy and must integrate the repetition of some exercises, it also must introduce some diversity and AI-VT must deal with this diversity. In this paper, we have highlighted the fact that the retaining (or capitalisation) phase of CBR is of the utmost importance for diversity, and we have also highlighted that the equilibrium between repetition and variety depends on the abilities learned. This balance has an important impact on the retaining phase of AI-VT.",	777	777	abstract="AI-Virtual Trainer (AI-VT) is an intelligent tutoring system based on case-based reasoning. AI-VT has been designed to generate personalised, varied, and consistent training sessions for learners. The AI-VT training sessions propose different exercises in regard to a capacity associated with sub-capacities. For example, in the field of training for algorithms, a capacity could be ``Use a control structure alternative'' and an associated sub-capacity could be ``Write a boolean condition''. AI-VT can elaborate a personalised list of exercises for each learner. One of the main requirements and challenges studied in this work is its ability to propose varied training sessions to the same learner for many weeks, which constitutes the challenge studied in our work. Indeed, if the same set of exercises is proposed time after time to learners, they will stop paying attention and lose motivation. Thus, even if the generation of training sessions is based on analogy and must integrate the repetition of some exercises, it also must introduce some diversity and AI-VT must deal with this diversity. In this paper, we have highlighted the fact that the retaining (or capitalisation) phase of CBR is of the utmost importance for diversity, and we have also highlighted that the equilibrium between repetition and variety depends on the abilities learned. This balance has an important impact on the retaining phase of AI-VT.",
isbn="978-3-030-01081-2"	778	778	isbn="978-3-030-01081-2"
}	779	779	}
	780	780
@article{BAKUROV2021100913,	781	781	@article{BAKUROV2021100913,
title = {Genetic programming for stacked generalization},	782	782	title = {Genetic programming for stacked generalization},
journal = {Swarm and Evolutionary Computation},	783	783	journal = {Swarm and Evolutionary Computation},
volume = {65},	784	784	volume = {65},
pages = {100913},	785	785	pages = {100913},
year = {2021},	786	786	year = {2021},
issn = {2210-6502},	787	787	issn = {2210-6502},
doi = {https://doi.org/10.1016/j.swevo.2021.100913},	788	788	doi = {https://doi.org/10.1016/j.swevo.2021.100913},
url = {https://www.sciencedirect.com/science/article/pii/S2210650221000742},	789	789	url = {https://www.sciencedirect.com/science/article/pii/S2210650221000742},
author = {Illya Bakurov and Mauro Castelli and Olivier Gau and Francesco Fontanella and Leonardo Vanneschi},	790	790	author = {Illya Bakurov and Mauro Castelli and Olivier Gau and Francesco Fontanella and Leonardo Vanneschi},
keywords = {Genetic Programming, Stacking, Ensemble Learning, Stacked Generalization},	791	791	keywords = {Genetic Programming, Stacking, Ensemble Learning, Stacked Generalization},
abstract = {In machine learning, ensemble techniques are widely used to improve the performance of both classification and regression systems. They combine the models generated by different learning algorithms, typically trained on different data subsets or with different parameters, to obtain more accurate models. Ensemble strategies range from simple voting rules to more complex and effective stacked approaches. They are based on adopting a meta-learner, i.e. a further learning algorithm, and are trained on the predictions provided by the single algorithms making up the ensemble. The paper aims at exploiting some of the most recent genetic programming advances in the context of stacked generalization. In particular, we investigate how the evolutionary demes despeciation initialization technique, ϵ-lexicase selection, geometric-semantic operators, and semantic stopping criterion, can be effectively used to improve GP-based systems’ performance for stacked generalization (a.k.a. stacking). The experiments, performed on a broad set of synthetic and real-world regression problems, confirm the effectiveness of the proposed approach.}	792	792	abstract = {In machine learning, ensemble techniques are widely used to improve the performance of both classification and regression systems. They combine the models generated by different learning algorithms, typically trained on different data subsets or with different parameters, to obtain more accurate models. Ensemble strategies range from simple voting rules to more complex and effective stacked approaches. They are based on adopting a meta-learner, i.e. a further learning algorithm, and are trained on the predictions provided by the single algorithms making up the ensemble. The paper aims at exploiting some of the most recent genetic programming advances in the context of stacked generalization. In particular, we investigate how the evolutionary demes despeciation initialization technique, ϵ-lexicase selection, geometric-semantic operators, and semantic stopping criterion, can be effectively used to improve GP-based systems’ performance for stacked generalization (a.k.a. stacking). The experiments, performed on a broad set of synthetic and real-world regression problems, confirm the effectiveness of the proposed approach.}
}	793	793	}
	794	794
@article{Liang,	795	795	@article{Liang,
author={Liang Mang and Chang Tianpeng and An Bingxing and Duan Xinghai and Du Lili and Wang Xiaoqiao and Miao Jian and Xu Lingyang and Gao Xue and Zhang Lupei and Li Junya and Gao Huijiang},	796	796	author={Liang Mang and Chang Tianpeng and An Bingxing and Duan Xinghai and Du Lili and Wang Xiaoqiao and Miao Jian and Xu Lingyang and Gao Xue and Zhang Lupei and Li Junya and Gao Huijiang},
Title={A Stacking Ensemble Learning Framework for Genomic Prediction},	797	797	Title={A Stacking Ensemble Learning Framework for Genomic Prediction},
Journal={Frontiers in Genetics},	798	798	Journal={Frontiers in Genetics},
year={2021},	799	799	year={2021},
doi ={10.3389/fgene.2021.600040},	800	800	doi ={10.3389/fgene.2021.600040},
PMID={33747037},	801	801	PMID={33747037},
PMCID={PMC7969712}	802	802	PMCID={PMC7969712}
}	803	803	}
	804	804
@Article{cmc.2023.033417,	805	805	@Article{cmc.2023.033417,
AUTHOR = {Jeonghoon Choi and Dongjun Suh and Marc-Oliver Otto},	806	806	AUTHOR = {Jeonghoon Choi and Dongjun Suh and Marc-Oliver Otto},
TITLE = {Boosted Stacking Ensemble Machine Learning Method for Wafer Map Pattern Classification},	807	807	TITLE = {Boosted Stacking Ensemble Machine Learning Method for Wafer Map Pattern Classification},
JOURNAL = {Computers, Materials \& Continua},	808	808	JOURNAL = {Computers, Materials \& Continua},
VOLUME = {74},	809	809	VOLUME = {74},
YEAR = {2023},	810	810	YEAR = {2023},
NUMBER = {2},	811	811	NUMBER = {2},
PAGES = {2945--2966},	812	812	PAGES = {2945--2966},
URL = {http://www.techscience.com/cmc/v74n2/50296},	813	813	URL = {http://www.techscience.com/cmc/v74n2/50296},
ISSN = {1546-2226},	814	814	ISSN = {1546-2226},
ABSTRACT = {Recently, machine learning-based technologies have been developed to automate the classification of wafer map defect patterns during semiconductor manufacturing. The existing approaches used in the wafer map pattern classification include directly learning the image through a convolution neural network and applying the ensemble method after extracting image features. This study aims to classify wafer map defects more effectively and derive robust algorithms even for datasets with insufficient defect patterns. First, the number of defects during the actual process may be limited. Therefore, insufficient data are generated using convolutional auto-encoder (CAE), and the expanded data are verified using the evaluation technique of structural similarity index measure (SSIM). After extracting handcrafted features, a boosted stacking ensemble model that integrates the four base-level classifiers with the extreme gradient boosting classifier as a meta-level classifier is designed and built for training the model based on the expanded data for final prediction. Since the proposed algorithm shows better performance than those of existing ensemble classifiers even for insufficient defect patterns, the results of this study will contribute to improving the product quality and yield of the actual semiconductor manufacturing process.},	815	815	ABSTRACT = {Recently, machine learning-based technologies have been developed to automate the classification of wafer map defect patterns during semiconductor manufacturing. The existing approaches used in the wafer map pattern classification include directly learning the image through a convolution neural network and applying the ensemble method after extracting image features. This study aims to classify wafer map defects more effectively and derive robust algorithms even for datasets with insufficient defect patterns. First, the number of defects during the actual process may be limited. Therefore, insufficient data are generated using convolutional auto-encoder (CAE), and the expanded data are verified using the evaluation technique of structural similarity index measure (SSIM). After extracting handcrafted features, a boosted stacking ensemble model that integrates the four base-level classifiers with the extreme gradient boosting classifier as a meta-level classifier is designed and built for training the model based on the expanded data for final prediction. Since the proposed algorithm shows better performance than those of existing ensemble classifiers even for insufficient defect patterns, the results of this study will contribute to improving the product quality and yield of the actual semiconductor manufacturing process.},
DOI = {10.32604/cmc.2023.033417}	816	816	DOI = {10.32604/cmc.2023.033417}
}	817	817	}
	818	818
@ARTICLE{10.3389/fgene.2021.600040,	819	819	@ARTICLE{10.3389/fgene.2021.600040,
AUTHOR={Liang, Mang and Chang, Tianpeng and An, Bingxing and Duan, Xinghai and Du, Lili and Wang, Xiaoqiao and Miao, Jian and Xu, Lingyang and Gao, Xue and Zhang, Lupei and Li, Junya and Gao, Huijiang},	820	820	AUTHOR={Liang, Mang and Chang, Tianpeng and An, Bingxing and Duan, Xinghai and Du, Lili and Wang, Xiaoqiao and Miao, Jian and Xu, Lingyang and Gao, Xue and Zhang, Lupei and Li, Junya and Gao, Huijiang},
TITLE={A Stacking Ensemble Learning Framework for Genomic Prediction},	821	821	TITLE={A Stacking Ensemble Learning Framework for Genomic Prediction},
JOURNAL={Frontiers in Genetics},	822	822	JOURNAL={Frontiers in Genetics},
VOLUME={12},	823	823	VOLUME={12},
YEAR={2021},	824	824	YEAR={2021},
URL={https://www.frontiersin.org/articles/10.3389/fgene.2021.600040},	825	825	URL={https://www.frontiersin.org/articles/10.3389/fgene.2021.600040},
DOI={10.3389/fgene.2021.600040},	826	826	DOI={10.3389/fgene.2021.600040},
ISSN={1664-8021},	827	827	ISSN={1664-8021},
ABSTRACT={Machine learning (ML) is perhaps the most useful tool for the interpretation of large genomic datasets. However, the performance of a single machine learning method in genomic selection (GS) is currently unsatisfactory. To improve the genomic predictions, we constructed a stacking ensemble learning framework (SELF), integrating three machine learning methods, to predict genomic estimated breeding values (GEBVs). The present study evaluated the prediction ability of SELF by analyzing three real datasets, with different genetic architecture; comparing the prediction accuracy of SELF, base learners, genomic best linear unbiased prediction (GBLUP) and BayesB. For each trait, SELF performed better than base learners, which included support vector regression (SVR), kernel ridge regression (KRR) and elastic net (ENET). The prediction accuracy of SELF was, on average, 7.70% higher than GBLUP in three datasets. Except for the milk fat percentage (MFP) traits, of the German Holstein dairy cattle dataset, SELF was more robust than BayesB in all remaining traits. Therefore, we believed that SEFL has the potential to be promoted to estimate GEBVs in other animals and plants.}	828	828	ABSTRACT={Machine learning (ML) is perhaps the most useful tool for the interpretation of large genomic datasets. However, the performance of a single machine learning method in genomic selection (GS) is currently unsatisfactory. To improve the genomic predictions, we constructed a stacking ensemble learning framework (SELF), integrating three machine learning methods, to predict genomic estimated breeding values (GEBVs). The present study evaluated the prediction ability of SELF by analyzing three real datasets, with different genetic architecture; comparing the prediction accuracy of SELF, base learners, genomic best linear unbiased prediction (GBLUP) and BayesB. For each trait, SELF performed better than base learners, which included support vector regression (SVR), kernel ridge regression (KRR) and elastic net (ENET). The prediction accuracy of SELF was, on average, 7.70% higher than GBLUP in three datasets. Except for the milk fat percentage (MFP) traits, of the German Holstein dairy cattle dataset, SELF was more robust than BayesB in all remaining traits. Therefore, we believed that SEFL has the potential to be promoted to estimate GEBVs in other animals and plants.}
}	829	829	}
	830	830
@article{DIDDEN2023338,	831	831	@article{DIDDEN2023338,
title = {Decentralized learning multi-agent system for online machine shop scheduling problem},	832	832	title = {Decentralized learning multi-agent system for online machine shop scheduling problem},
journal = {Journal of Manufacturing Systems},	833	833	journal = {Journal of Manufacturing Systems},
volume = {67},	834	834	volume = {67},
pages = {338-360},	835	835	pages = {338-360},
year = {2023},	836	836	year = {2023},
issn = {0278-6125},	837	837	issn = {0278-6125},
doi = {https://doi.org/10.1016/j.jmsy.2023.02.004},	838	838	doi = {https://doi.org/10.1016/j.jmsy.2023.02.004},
url = {https://www.sciencedirect.com/science/article/pii/S0278612523000286},	839	839	url = {https://www.sciencedirect.com/science/article/pii/S0278612523000286},
author = {Jeroen B.H.C. Didden and Quang-Vinh Dang and Ivo J.B.F. Adan},	840	840	author = {Jeroen B.H.C. Didden and Quang-Vinh Dang and Ivo J.B.F. Adan},
keywords = {Multi-agent system, Decentralized systems, Learning algorithm, Industry 4.0, Smart manufacturing},	841	841	keywords = {Multi-agent system, Decentralized systems, Learning algorithm, Industry 4.0, Smart manufacturing},
abstract = {Customer profiles have rapidly changed over the past few years, with products being requested with more customization and with lower demand. In addition to the advances in technologies owing to Industry 4.0, manufacturers explore autonomous and smart factories. This paper proposes a decentralized multi-agent system (MAS), including intelligent agents that can respond to their environment autonomously through learning capabilities, to cope with an online machine shop scheduling problem. In the proposed system, agents participate in auctions to receive jobs to process, learn how to bid for jobs correctly, and decide when to start processing a job. The objective is to minimize the mean weighted tardiness of all jobs. In contrast to the existing literature, the proposed MAS is assessed on its learning capabilities, producing novel insights concerning what is relevant for learning, when re-learning is needed, and system response to dynamic events (such as rush jobs, increase in processing time, and machine unavailability). Computational experiments also reveal the outperformance of the proposed MAS to other multi-agent systems by at least 25% and common dispatching rules in mean weighted tardiness, as well as other performance measures.}	842	842	abstract = {Customer profiles have rapidly changed over the past few years, with products being requested with more customization and with lower demand. In addition to the advances in technologies owing to Industry 4.0, manufacturers explore autonomous and smart factories. This paper proposes a decentralized multi-agent system (MAS), including intelligent agents that can respond to their environment autonomously through learning capabilities, to cope with an online machine shop scheduling problem. In the proposed system, agents participate in auctions to receive jobs to process, learn how to bid for jobs correctly, and decide when to start processing a job. The objective is to minimize the mean weighted tardiness of all jobs. In contrast to the existing literature, the proposed MAS is assessed on its learning capabilities, producing novel insights concerning what is relevant for learning, when re-learning is needed, and system response to dynamic events (such as rush jobs, increase in processing time, and machine unavailability). Computational experiments also reveal the outperformance of the proposed MAS to other multi-agent systems by at least 25% and common dispatching rules in mean weighted tardiness, as well as other performance measures.}
}	843	843	}
	844	844
@article{REZAEI20221,	845	845	@article{REZAEI20221,
title = {A Biased Inferential Naivety learning model for a network of agents},	846	846	title = {A Biased Inferential Naivety learning model for a network of agents},
journal = {Cognitive Systems Research},	847	847	journal = {Cognitive Systems Research},
volume = {76},	848	848	volume = {76},
pages = {1-12},	849	849	pages = {1-12},
year = {2022},	850	850	year = {2022},
issn = {1389-0417},	851	851	issn = {1389-0417},
doi = {https://doi.org/10.1016/j.cogsys.2022.07.001},	852	852	doi = {https://doi.org/10.1016/j.cogsys.2022.07.001},
url = {https://www.sciencedirect.com/science/article/pii/S1389041722000298},	853	853	url = {https://www.sciencedirect.com/science/article/pii/S1389041722000298},
author = {Zeinab Rezaei and Saeed Setayeshi and Ebrahim Mahdipour},	854	854	author = {Zeinab Rezaei and Saeed Setayeshi and Ebrahim Mahdipour},
keywords = {Bayesian decision making, Heuristic method, Inferential naivety assumption, Observational learning, Social learning},	855	855	keywords = {Bayesian decision making, Heuristic method, Inferential naivety assumption, Observational learning, Social learning},
abstract = {We propose a Biased Inferential Naivety social learning model. In this model, a group of agents tries to determine the true state of the world and make the best possible decisions. The agents have limited computational abilities. They receive noisy private signals about the true state and observe the history of their neighbors' decisions. The proposed model is rooted in the Bayesian method but avoids the complexity of fully Bayesian inference. In our model, the role of knowledge obtained from social observations is separated from the knowledge obtained from private observations. Therefore, the Bayesian inferences on social observations are approximated using inferential naivety assumption, while purely Bayesian inferences are made on private observations. The reduction of herd behavior is another innovation of the proposed model. This advantage is achieved by reducing the effect of social observations on agents' beliefs over time. Therefore, all the agents learn the truth, and the correct consensus is achieved effectively. In this model, using two cognitive biases, there is heterogeneity in agents' behaviors. Therefore, the growth of beliefs and the learning speed can be improved in different situations. Several Monte Carlo simulations confirm the features of the proposed model. The conditions under which the proposed model leads to asymptotic learning are proved.}	856	856	abstract = {We propose a Biased Inferential Naivety social learning model. In this model, a group of agents tries to determine the true state of the world and make the best possible decisions. The agents have limited computational abilities. They receive noisy private signals about the true state and observe the history of their neighbors' decisions. The proposed model is rooted in the Bayesian method but avoids the complexity of fully Bayesian inference. In our model, the role of knowledge obtained from social observations is separated from the knowledge obtained from private observations. Therefore, the Bayesian inferences on social observations are approximated using inferential naivety assumption, while purely Bayesian inferences are made on private observations. The reduction of herd behavior is another innovation of the proposed model. This advantage is achieved by reducing the effect of social observations on agents' beliefs over time. Therefore, all the agents learn the truth, and the correct consensus is achieved effectively. In this model, using two cognitive biases, there is heterogeneity in agents' behaviors. Therefore, the growth of beliefs and the learning speed can be improved in different situations. Several Monte Carlo simulations confirm the features of the proposed model. The conditions under which the proposed model leads to asymptotic learning are proved.}
}	857	857	}
	858	858
@article{KAMALI2023110242,	859	859	@article{KAMALI2023110242,
title = {An immune inspired multi-agent system for dynamic multi-objective optimization},	860	860	title = {An immune inspired multi-agent system for dynamic multi-objective optimization},
journal = {Knowledge-Based Systems},	861	861	journal = {Knowledge-Based Systems},
volume = {262},	862	862	volume = {262},
pages = {110242},	863	863	pages = {110242},
year = {2023},	864	864	year = {2023},
issn = {0950-7051},	865	865	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2022.110242},	866	866	doi = {https://doi.org/10.1016/j.knosys.2022.110242},
url = {https://www.sciencedirect.com/science/article/pii/S0950705122013387},	867	867	url = {https://www.sciencedirect.com/science/article/pii/S0950705122013387},
author = {Seyed Ruhollah Kamali and Touraj Banirostam and Homayun Motameni and Mohammad Teshnehlab},	868	868	author = {Seyed Ruhollah Kamali and Touraj Banirostam and Homayun Motameni and Mohammad Teshnehlab},
keywords = {Immune inspired multi-agent system, Dynamic multi-objective optimization, Severe and frequent changes},	869	869	keywords = {Immune inspired multi-agent system, Dynamic multi-objective optimization, Severe and frequent changes},
abstract = {In this research, an immune inspired multi-agent system (IMAS) is proposed to solve optimization problems in dynamic and multi-objective environments. The proposed IMAS uses artificial immune system metaphors to shape the local behaviors of agents to detect environmental changes, generate Pareto optimal solutions, and react to the dynamics of the problem environment. Apart from that, agents enhance their adaptive capacity in dealing with environmental changes to find the global optimum, with a hierarchical structure without any central control. This study used a combination of diversity-, multi-population- and memory-based approaches to perform better in multi-objective environments with severe and frequent changes. The proposed IMAS is compared with six state-of-the-art algorithms on various benchmark problems. The results indicate its superiority in many of the experiments.}	870	870	abstract = {In this research, an immune inspired multi-agent system (IMAS) is proposed to solve optimization problems in dynamic and multi-objective environments. The proposed IMAS uses artificial immune system metaphors to shape the local behaviors of agents to detect environmental changes, generate Pareto optimal solutions, and react to the dynamics of the problem environment. Apart from that, agents enhance their adaptive capacity in dealing with environmental changes to find the global optimum, with a hierarchical structure without any central control. This study used a combination of diversity-, multi-population- and memory-based approaches to perform better in multi-objective environments with severe and frequent changes. The proposed IMAS is compared with six state-of-the-art algorithms on various benchmark problems. The results indicate its superiority in many of the experiments.}
}	871	871	}
	872	872
@article{ZHANG2023110564,	873	873	@article{ZHANG2023110564,
title = {A novel human learning optimization algorithm with Bayesian inference learning},	874	874	title = {A novel human learning optimization algorithm with Bayesian inference learning},
journal = {Knowledge-Based Systems},	875	875	journal = {Knowledge-Based Systems},
volume = {271},	876	876	volume = {271},
pages = {110564},	877	877	pages = {110564},
year = {2023},	878	878	year = {2023},
issn = {0950-7051},	879	879	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2023.110564},	880	880	doi = {https://doi.org/10.1016/j.knosys.2023.110564},
url = {https://www.sciencedirect.com/science/article/pii/S0950705123003143},	881	881	url = {https://www.sciencedirect.com/science/article/pii/S0950705123003143},
author = {Pinggai Zhang and Ling Wang and Zixiang Fei and Lisheng Wei and Minrui Fei and Muhammad Ilyas Menhas},	882	882	author = {Pinggai Zhang and Ling Wang and Zixiang Fei and Lisheng Wei and Minrui Fei and Muhammad Ilyas Menhas},
keywords = {Human learning optimization, Meta-heuristic, Bayesian inference, Bayesian inference learning, Individual learning, Social learning},	883	883	keywords = {Human learning optimization, Meta-heuristic, Bayesian inference, Bayesian inference learning, Individual learning, Social learning},
abstract = {Humans perform Bayesian inference in a wide variety of tasks, which can help people make selection decisions effectively and therefore enhances learning efficiency and accuracy. Inspired by this fact, this paper presents a novel human learning optimization algorithm with Bayesian inference learning (HLOBIL), in which a Bayesian inference learning operator (BILO) is developed to utilize the inference strategy for enhancing learning efficiency. The in-depth analysis shows that the proposed BILO can efficiently improve the exploitation ability of the algorithm as it can achieve the optimal values and retrieve the optimal information with the accumulated search information. Besides, the exploration ability of HLOBIL is also strengthened by the inborn characteristics of Bayesian inference. The experimental results demonstrate that the developed HLOBIL is superior to previous HLO variants and other state-of-art algorithms with its improved exploitation and exploration abilities.}	884	884	abstract = {Humans perform Bayesian inference in a wide variety of tasks, which can help people make selection decisions effectively and therefore enhances learning efficiency and accuracy. Inspired by this fact, this paper presents a novel human learning optimization algorithm with Bayesian inference learning (HLOBIL), in which a Bayesian inference learning operator (BILO) is developed to utilize the inference strategy for enhancing learning efficiency. The in-depth analysis shows that the proposed BILO can efficiently improve the exploitation ability of the algorithm as it can achieve the optimal values and retrieve the optimal information with the accumulated search information. Besides, the exploration ability of HLOBIL is also strengthened by the inborn characteristics of Bayesian inference. The experimental results demonstrate that the developed HLOBIL is superior to previous HLO variants and other state-of-art algorithms with its improved exploitation and exploration abilities.}
}	885	885	}
	886	886
@article{HIPOLITO2023103510,	887	887	@article{HIPOLITO2023103510,
title = {Breaking boundaries: The Bayesian Brain Hypothesis for perception and prediction},	888	888	title = {Breaking boundaries: The Bayesian Brain Hypothesis for perception and prediction},
journal = {Consciousness and Cognition},	889	889	journal = {Consciousness and Cognition},
volume = {111},	890	890	volume = {111},
pages = {103510},	891	891	pages = {103510},
year = {2023},	892	892	year = {2023},
issn = {1053-8100},	893	893	issn = {1053-8100},
doi = {https://doi.org/10.1016/j.concog.2023.103510},	894	894	doi = {https://doi.org/10.1016/j.concog.2023.103510},
url = {https://www.sciencedirect.com/science/article/pii/S1053810023000478},	895	895	url = {https://www.sciencedirect.com/science/article/pii/S1053810023000478},
author = {Inês Hipólito and Michael Kirchhoff},	896	896	author = {Inês Hipólito and Michael Kirchhoff},
keywords = {Bayesian Brain Hypothesis, Modularity of the Mind, Cognitive processes, Informational boundaries},	897	897	keywords = {Bayesian Brain Hypothesis, Modularity of the Mind, Cognitive processes, Informational boundaries},
abstract = {This special issue aims to provide a comprehensive overview of the current state of the Bayesian Brain Hypothesis and its standing across neuroscience, cognitive science and the philosophy of cognitive science. By gathering cutting-edge research from leading experts, this issue seeks to showcase the latest advancements in our understanding of the Bayesian brain, as well as its potential implications for future research in perception, cognition, and motor control. A special focus to achieve this aim is adopted in this special issue, as it seeks to explore the relation between two seemingly incompatible frameworks for the understanding of cognitive structure and function: the Bayesian Brain Hypothesis and the Modularity Theory of the Mind. In assessing the compatibility between these theories, the contributors to this special issue open up new pathways of thinking and advance our understanding of cognitive processes.}	898	898	abstract = {This special issue aims to provide a comprehensive overview of the current state of the Bayesian Brain Hypothesis and its standing across neuroscience, cognitive science and the philosophy of cognitive science. By gathering cutting-edge research from leading experts, this issue seeks to showcase the latest advancements in our understanding of the Bayesian brain, as well as its potential implications for future research in perception, cognition, and motor control. A special focus to achieve this aim is adopted in this special issue, as it seeks to explore the relation between two seemingly incompatible frameworks for the understanding of cognitive structure and function: the Bayesian Brain Hypothesis and the Modularity Theory of the Mind. In assessing the compatibility between these theories, the contributors to this special issue open up new pathways of thinking and advance our understanding of cognitive processes.}
}	899	899	}
	900	900
@article{LI2023424,	901	901	@article{LI2023424,
title = {Multi-agent evolution reinforcement learning method for machining parameters optimization based on bootstrap aggregating graph attention network simulated environment},	902	902	title = {Multi-agent evolution reinforcement learning method for machining parameters optimization based on bootstrap aggregating graph attention network simulated environment},
journal = {Journal of Manufacturing Systems},	903	903	journal = {Journal of Manufacturing Systems},
volume = {67},	904	904	volume = {67},
pages = {424-438},	905	905	pages = {424-438},
year = {2023},	906	906	year = {2023},
issn = {0278-6125},	907	907	issn = {0278-6125},
doi = {https://doi.org/10.1016/j.jmsy.2023.02.015},	908	908	doi = {https://doi.org/10.1016/j.jmsy.2023.02.015},
url = {https://www.sciencedirect.com/science/article/pii/S0278612523000390},	909	909	url = {https://www.sciencedirect.com/science/article/pii/S0278612523000390},
author = {Weiye Li and Songping He and Xinyong Mao and Bin Li and Chaochao Qiu and Jinwen Yu and Fangyu Peng and Xin Tan},	910	910	author = {Weiye Li and Songping He and Xinyong Mao and Bin Li and Chaochao Qiu and Jinwen Yu and Fangyu Peng and Xin Tan},
keywords = {Surface roughness, Cutting efficiency, Machining parameters optimization, Graph attention network, Multi-agent reinforcement learning, Evolutionary learning},	911	911	keywords = {Surface roughness, Cutting efficiency, Machining parameters optimization, Graph attention network, Multi-agent reinforcement learning, Evolutionary learning},
abstract = {Improving machining quality and production efficiency is the focus of the manufacturing industry. How to obtain efficient machining parameters under multiple constraints such as machining quality is a severe challenge for manufacturing industry. In this paper, a multi-agent evolutionary reinforcement learning method (MAERL) is proposed to optimize the machining parameters for high quality and high efficiency machining by combining the graph neural network and reinforcement learning. Firstly, a bootstrap aggregating graph attention network (Bagging-GAT) based roughness estimation method for machined surface is proposed, which combines the structural knowledge between machining parameters and vibration features. Secondly, a mathematical model of machining parameters optimization problem is established, which is formalized into Markov decision process (MDP), and a multi-agent reinforcement learning method is proposed to solve the MDP problem, and evolutionary learning is introduced to improve the stability of multi-agent training. Finally, a series of experiments were carried out on the commutator production line, and the results show that the proposed Bagging-GAT-based method can improve the prediction effect by about 25% in the case of small samples, and the MAERL-based optimization method can better deal with the coupling problem of reward function in the optimization process. Compared with the classical optimization method, the optimization effect is improved by 13% and a lot of optimization time is saved.}	912	912	abstract = {Improving machining quality and production efficiency is the focus of the manufacturing industry. How to obtain efficient machining parameters under multiple constraints such as machining quality is a severe challenge for manufacturing industry. In this paper, a multi-agent evolutionary reinforcement learning method (MAERL) is proposed to optimize the machining parameters for high quality and high efficiency machining by combining the graph neural network and reinforcement learning. Firstly, a bootstrap aggregating graph attention network (Bagging-GAT) based roughness estimation method for machined surface is proposed, which combines the structural knowledge between machining parameters and vibration features. Secondly, a mathematical model of machining parameters optimization problem is established, which is formalized into Markov decision process (MDP), and a multi-agent reinforcement learning method is proposed to solve the MDP problem, and evolutionary learning is introduced to improve the stability of multi-agent training. Finally, a series of experiments were carried out on the commutator production line, and the results show that the proposed Bagging-GAT-based method can improve the prediction effect by about 25% in the case of small samples, and the MAERL-based optimization method can better deal with the coupling problem of reward function in the optimization process. Compared with the classical optimization method, the optimization effect is improved by 13% and a lot of optimization time is saved.}
}	913	913	}
	914	914
@inproceedings{10.1145/3290605.3300912,	915	915	@inproceedings{10.1145/3290605.3300912,
author = {Kim, Yea-Seul and Walls, Logan A. and Krafft, Peter and Hullman, Jessica},	916	916	author = {Kim, Yea-Seul and Walls, Logan A. and Krafft, Peter and Hullman, Jessica},
title = {A Bayesian Cognition Approach to Improve Data Visualization},	917	917	title = {A Bayesian Cognition Approach to Improve Data Visualization},
year = {2019},	918	918	year = {2019},
isbn = {9781450359702},	919	919	isbn = {9781450359702},
publisher = {Association for Computing Machinery},	920	920	publisher = {Association for Computing Machinery},
address = {New York, NY, USA},	921	921	address = {New York, NY, USA},
url = {https://doi.org/10.1145/3290605.3300912},	922	922	url = {https://doi.org/10.1145/3290605.3300912},
doi = {10.1145/3290605.3300912},	923	923	doi = {10.1145/3290605.3300912},
abstract = {People naturally bring their prior beliefs to bear on how they interpret the new information, yet few formal models exist for accounting for the influence of users' prior beliefs in interactions with data presentations like visualizations. We demonstrate a Bayesian cognitive model for understanding how people interpret visualizations in light of prior beliefs and show how this model provides a guide for improving visualization evaluation. In a first study, we show how applying a Bayesian cognition model to a simple visualization scenario indicates that people's judgments are consistent with a hypothesis that they are doing approximate Bayesian inference. In a second study, we evaluate how sensitive our observations of Bayesian behavior are to different techniques for eliciting people subjective distributions, and to different datasets. We find that people don't behave consistently with Bayesian predictions for large sample size datasets, and this difference cannot be explained by elicitation technique. In a final study, we show how normative Bayesian inference can be used as an evaluation framework for visualizations, including of uncertainty.},	924	924	abstract = {People naturally bring their prior beliefs to bear on how they interpret the new information, yet few formal models exist for accounting for the influence of users' prior beliefs in interactions with data presentations like visualizations. We demonstrate a Bayesian cognitive model for understanding how people interpret visualizations in light of prior beliefs and show how this model provides a guide for improving visualization evaluation. In a first study, we show how applying a Bayesian cognition model to a simple visualization scenario indicates that people's judgments are consistent with a hypothesis that they are doing approximate Bayesian inference. In a second study, we evaluate how sensitive our observations of Bayesian behavior are to different techniques for eliciting people subjective distributions, and to different datasets. We find that people don't behave consistently with Bayesian predictions for large sample size datasets, and this difference cannot be explained by elicitation technique. In a final study, we show how normative Bayesian inference can be used as an evaluation framework for visualizations, including of uncertainty.},
booktitle = {Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems},	925	925	booktitle = {Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems},
pages = {1–14},	926	926	pages = {1–14},
numpages = {14},	927	927	numpages = {14},
keywords = {bayesian cognition, uncertainty elicitation, visualization},	928	928	keywords = {bayesian cognition, uncertainty elicitation, visualization},
location = {Glasgow, Scotland Uk},	929	929	location = {Glasgow, Scotland Uk},
series = {CHI '19}	930	930	series = {CHI '19}
}	931	931	}
	932	932
@article{DYER2024104827,	933	933	@article{DYER2024104827,
title = {Black-box Bayesian inference for agent-based models},	934	934	title = {Black-box Bayesian inference for agent-based models},
journal = {Journal of Economic Dynamics and Control},	935	935	journal = {Journal of Economic Dynamics and Control},
volume = {161},	936	936	volume = {161},
pages = {104827},	937	937	pages = {104827},
year = {2024},	938	938	year = {2024},
issn = {0165-1889},	939	939	issn = {0165-1889},
doi = {https://doi.org/10.1016/j.jedc.2024.104827},	940	940	doi = {https://doi.org/10.1016/j.jedc.2024.104827},
url = {https://www.sciencedirect.com/science/article/pii/S0165188924000198},	941	941	url = {https://www.sciencedirect.com/science/article/pii/S0165188924000198},
author = {Joel Dyer and Patrick Cannon and J. Doyne Farmer and Sebastian M. Schmon},	942	942	author = {Joel Dyer and Patrick Cannon and J. Doyne Farmer and Sebastian M. Schmon},
keywords = {Agent-based models, Bayesian inference, Neural networks, Parameter estimation, Simulation-based inference, Time series},	943	943	keywords = {Agent-based models, Bayesian inference, Neural networks, Parameter estimation, Simulation-based inference, Time series},
abstract = {Simulation models, in particular agent-based models, are gaining popularity in economics and the social sciences. The considerable flexibility they offer, as well as their capacity to reproduce a variety of empirically observed behaviours of complex systems, give them broad appeal, and the increasing availability of cheap computing power has made their use feasible. Yet a widespread adoption in real-world modelling and decision-making scenarios has been hindered by the difficulty of performing parameter estimation for such models. In general, simulation models lack a tractable likelihood function, which precludes a straightforward application of standard statistical inference techniques. A number of recent works have sought to address this problem through the application of likelihood-free inference techniques, in which parameter estimates are determined by performing some form of comparison between the observed data and simulation output. However, these approaches are (a) founded on restrictive assumptions, and/or (b) typically require many hundreds of thousands of simulations. These qualities make them unsuitable for large-scale simulations in economics and the social sciences, and can cast doubt on the validity of these inference methods in such scenarios. In this paper, we investigate the efficacy of two classes of simulation-efficient black-box approximate Bayesian inference methods that have recently drawn significant attention within the probabilistic machine learning community: neural posterior estimation and neural density ratio estimation. We present a number of benchmarking experiments in which we demonstrate that neural network-based black-box methods provide state of the art parameter inference for economic simulation models, and crucially are compatible with generic multivariate or even non-Euclidean time-series data. In addition, we suggest appropriate assessment criteria for use in future benchmarking of approximate Bayesian inference procedures for simulation models in economics and the social sciences.}	944	944	abstract = {Simulation models, in particular agent-based models, are gaining popularity in economics and the social sciences. The considerable flexibility they offer, as well as their capacity to reproduce a variety of empirically observed behaviours of complex systems, give them broad appeal, and the increasing availability of cheap computing power has made their use feasible. Yet a widespread adoption in real-world modelling and decision-making scenarios has been hindered by the difficulty of performing parameter estimation for such models. In general, simulation models lack a tractable likelihood function, which precludes a straightforward application of standard statistical inference techniques. A number of recent works have sought to address this problem through the application of likelihood-free inference techniques, in which parameter estimates are determined by performing some form of comparison between the observed data and simulation output. However, these approaches are (a) founded on restrictive assumptions, and/or (b) typically require many hundreds of thousands of simulations. These qualities make them unsuitable for large-scale simulations in economics and the social sciences, and can cast doubt on the validity of these inference methods in such scenarios. In this paper, we investigate the efficacy of two classes of simulation-efficient black-box approximate Bayesian inference methods that have recently drawn significant attention within the probabilistic machine learning community: neural posterior estimation and neural density ratio estimation. We present a number of benchmarking experiments in which we demonstrate that neural network-based black-box methods provide state of the art parameter inference for economic simulation models, and crucially are compatible with generic multivariate or even non-Euclidean time-series data. In addition, we suggest appropriate assessment criteria for use in future benchmarking of approximate Bayesian inference procedures for simulation models in economics and the social sciences.}
}	945	945	}
	946	946
@Article{Nikpour2021,	947	947	@Article{Nikpour2021,
author={Nikpour, Hoda	948	948	author={Nikpour, Hoda
and Aamodt, Agnar},	949	949	and Aamodt, Agnar},
title={Inference and reasoning in a Bayesian knowledge-intensive CBR system},	950	950	title={Inference and reasoning in a Bayesian knowledge-intensive CBR system},
journal={Progress in Artificial Intelligence},	951	951	journal={Progress in Artificial Intelligence},
year={2021},	952	952	year={2021},
month={Mar},	953	953	month={Mar},
day={01},	954	954	day={01},
volume={10},	955	955	volume={10},
number={1},	956	956	number={1},
pages={49-63},	957	957	pages={49-63},
abstract={This paper presents the inference and reasoning methods in a Bayesian supported knowledge-intensive case-based reasoning (CBR) system called BNCreek. The inference and reasoning process in this system is a combination of three methods. The semantic network inference methods and the CBR method are employed to handle the difficulties of inferencing and reasoning in uncertain domains. The Bayesian network inference methods are employed to make the process more accurate. An experiment from oil well drilling as a complex and uncertain application domain is conducted. The system is evaluated against expert estimations and compared with seven other corresponding systems. The normalized discounted cumulative gain (NDCG) as a rank-based metric, the weighted error (WE), and root-square error (RSE) as the statistical metrics are employed to evaluate different aspects of the system capabilities. The results show the efficiency of the developed inference and reasoning methods.},	958	958	abstract={This paper presents the inference and reasoning methods in a Bayesian supported knowledge-intensive case-based reasoning (CBR) system called BNCreek. The inference and reasoning process in this system is a combination of three methods. The semantic network inference methods and the CBR method are employed to handle the difficulties of inferencing and reasoning in uncertain domains. The Bayesian network inference methods are employed to make the process more accurate. An experiment from oil well drilling as a complex and uncertain application domain is conducted. The system is evaluated against expert estimations and compared with seven other corresponding systems. The normalized discounted cumulative gain (NDCG) as a rank-based metric, the weighted error (WE), and root-square error (RSE) as the statistical metrics are employed to evaluate different aspects of the system capabilities. The results show the efficiency of the developed inference and reasoning methods.},
issn={2192-6360},	959	959	issn={2192-6360},
doi={10.1007/s13748-020-00223-1},	960	960	doi={10.1007/s13748-020-00223-1},
url={https://doi.org/10.1007/s13748-020-00223-1}	961	961	url={https://doi.org/10.1007/s13748-020-00223-1}
}	962	962	}
	963	963
@article{PRESCOTT2024112577,	964	964	@article{PRESCOTT2024112577,
title = {Efficient multifidelity likelihood-free Bayesian inference with adaptive computational resource allocation},	965	965	title = {Efficient multifidelity likelihood-free Bayesian inference with adaptive computational resource allocation},
journal = {Journal of Computational Physics},	966	966	journal = {Journal of Computational Physics},
volume = {496},	967	967	volume = {496},
pages = {112577},	968	968	pages = {112577},
year = {2024},	969	969	year = {2024},
issn = {0021-9991},	970	970	issn = {0021-9991},
doi = {https://doi.org/10.1016/j.jcp.2023.112577},	971	971	doi = {https://doi.org/10.1016/j.jcp.2023.112577},
url = {https://www.sciencedirect.com/science/article/pii/S0021999123006721},	972	972	url = {https://www.sciencedirect.com/science/article/pii/S0021999123006721},
author = {Thomas P. Prescott and David J. Warne and Ruth E. Baker},	973	973	author = {Thomas P. Prescott and David J. Warne and Ruth E. Baker},
keywords = {Likelihood-free Bayesian inference, Multifidelity approaches},	974	974	keywords = {Likelihood-free Bayesian inference, Multifidelity approaches},
abstract = {Likelihood-free Bayesian inference algorithms are popular methods for inferring the parameters of complex stochastic models with intractable likelihoods. These algorithms characteristically rely heavily on repeated model simulations. However, whenever the computational cost of simulation is even moderately expensive, the significant burden incurred by likelihood-free algorithms leaves them infeasible for many practical applications. The multifidelity approach has been introduced in the context of approximate Bayesian computation to reduce the simulation burden of likelihood-free inference without loss of accuracy, by using the information provided by simulating computationally cheap, approximate models in place of the model of interest. In this work we demonstrate that multifidelity techniques can be applied in the general likelihood-free Bayesian inference setting. Analytical results on the optimal allocation of computational resources to simulations at different levels of fidelity are derived, and subsequently implemented practically. We provide an adaptive multifidelity likelihood-free inference algorithm that learns the relationships between models at different fidelities and adapts resource allocation accordingly, and demonstrate that this algorithm produces posterior estimates with near-optimal efficiency.}	975	975	abstract = {Likelihood-free Bayesian inference algorithms are popular methods for inferring the parameters of complex stochastic models with intractable likelihoods. These algorithms characteristically rely heavily on repeated model simulations. However, whenever the computational cost of simulation is even moderately expensive, the significant burden incurred by likelihood-free algorithms leaves them infeasible for many practical applications. The multifidelity approach has been introduced in the context of approximate Bayesian computation to reduce the simulation burden of likelihood-free inference without loss of accuracy, by using the information provided by simulating computationally cheap, approximate models in place of the model of interest. In this work we demonstrate that multifidelity techniques can be applied in the general likelihood-free Bayesian inference setting. Analytical results on the optimal allocation of computational resources to simulations at different levels of fidelity are derived, and subsequently implemented practically. We provide an adaptive multifidelity likelihood-free inference algorithm that learns the relationships between models at different fidelities and adapts resource allocation accordingly, and demonstrate that this algorithm produces posterior estimates with near-optimal efficiency.}
}	976	976	}
	977	977
@article{RISTIC202030,	978	978	@article{RISTIC202030,
title = {A tutorial on uncertainty modeling for machine reasoning},	979	979	title = {A tutorial on uncertainty modeling for machine reasoning},
journal = {Information Fusion},	980	980	journal = {Information Fusion},
volume = {55},	981	981	volume = {55},
pages = {30-44},	982	982	pages = {30-44},
year = {2020},	983	983	year = {2020},
issn = {1566-2535},	984	984	issn = {1566-2535},
doi = {https://doi.org/10.1016/j.inffus.2019.08.001},	985	985	doi = {https://doi.org/10.1016/j.inffus.2019.08.001},
url = {https://www.sciencedirect.com/science/article/pii/S1566253519301976},	986	986	url = {https://www.sciencedirect.com/science/article/pii/S1566253519301976},
author = {Branko Ristic and Christopher Gilliam and Marion Byrne and Alessio Benavoli},	987	987	author = {Branko Ristic and Christopher Gilliam and Marion Byrne and Alessio Benavoli},
keywords = {Information fusion, Uncertainty, Imprecision, Model based classification, Bayesian, Random sets, Belief function theory, Possibility functions, Imprecise probability},	988	988	keywords = {Information fusion, Uncertainty, Imprecision, Model based classification, Bayesian, Random sets, Belief function theory, Possibility functions, Imprecise probability},
abstract = {Increasingly we rely on machine intelligence for reasoning and decision making under uncertainty. This tutorial reviews the prevalent methods for model-based autonomous decision making based on observations and prior knowledge, primarily in the context of classification. Both observations and the knowledge-base available for reasoning are treated as being uncertain. Accordingly, the central themes of this tutorial are quantitative modeling of uncertainty, the rules required to combine such uncertain information, and the task of decision making under uncertainty. The paper covers the main approaches to uncertain knowledge representation and reasoning, in particular, Bayesian probability theory, possibility theory, reasoning based on belief functions and finally imprecise probability theory. The main feature of the tutorial is that it illustrates various approaches with several testing scenarios, and provides MATLAB solutions for them as a supplementary material for an interested reader.}	989	989	abstract = {Increasingly we rely on machine intelligence for reasoning and decision making under uncertainty. This tutorial reviews the prevalent methods for model-based autonomous decision making based on observations and prior knowledge, primarily in the context of classification. Both observations and the knowledge-base available for reasoning are treated as being uncertain. Accordingly, the central themes of this tutorial are quantitative modeling of uncertainty, the rules required to combine such uncertain information, and the task of decision making under uncertainty. The paper covers the main approaches to uncertain knowledge representation and reasoning, in particular, Bayesian probability theory, possibility theory, reasoning based on belief functions and finally imprecise probability theory. The main feature of the tutorial is that it illustrates various approaches with several testing scenarios, and provides MATLAB solutions for them as a supplementary material for an interested reader.}
}	990	990	}
	991	991
@article{CICIRELLO2022108619,	992	992	@article{CICIRELLO2022108619,
title = {Machine learning based optimization for interval uncertainty propagation},	993	993	title = {Machine learning based optimization for interval uncertainty propagation},
journal = {Mechanical Systems and Signal Processing},	994	994	journal = {Mechanical Systems and Signal Processing},
volume = {170},	995	995	volume = {170},
pages = {108619},	996	996	pages = {108619},
year = {2022},	997	997	year = {2022},
issn = {0888-3270},	998	998	issn = {0888-3270},
doi = {https://doi.org/10.1016/j.ymssp.2021.108619},	999	999	doi = {https://doi.org/10.1016/j.ymssp.2021.108619},
url = {https://www.sciencedirect.com/science/article/pii/S0888327021009493},	1000	1000	url = {https://www.sciencedirect.com/science/article/pii/S0888327021009493},
author = {Alice Cicirello and Filippo Giunta},	1001	1001	author = {Alice Cicirello and Filippo Giunta},
keywords = {Bounded uncertainty, Bayesian optimization, Expensive-to-evaluate deterministic computer models, Gaussian process, Communicating uncertainty},	1002	1002	keywords = {Bounded uncertainty, Bayesian optimization, Expensive-to-evaluate deterministic computer models, Gaussian process, Communicating uncertainty},
abstract = {Two non-intrusive uncertainty propagation approaches are proposed for the performance analysis of engineering systems described by expensive-to-evaluate deterministic computer models with parameters defined as interval variables. These approaches employ a machine learning based optimization strategy, the so-called Bayesian optimization, for evaluating the upper and lower bounds of a generic response variable over the set of possible responses obtained when each interval variable varies independently over its range. The lack of knowledge caused by not evaluating the response function for all the possible combinations of the interval variables is accounted for by developing a probabilistic description of the response variable itself by using a Gaussian Process regression model. An iterative procedure is developed for selecting a small number of simulations to be evaluated for updating this statistical model by using well-established acquisition functions and to assess the response bounds. In both approaches, an initial training dataset is defined. While one approach builds iteratively two distinct training datasets for evaluating separately the upper and lower bounds of the response variable, the other one builds iteratively a single training dataset. Consequently, the two approaches will produce different bound estimates at each iteration. The upper and lower response bounds are expressed as point estimates obtained from the mean function of the posterior distribution. Moreover, a confidence interval on each estimate is provided for effectively communicating to engineers when these estimates are obtained at a combination of the interval variables for which no deterministic simulation has been run. Finally, two metrics are proposed to define conditions for assessing if the predicted bound estimates can be considered satisfactory. The applicability of these two approaches is illustrated with two numerical applications, one focusing on vibration and the other on vibro-acoustics.}	1003	1003	abstract = {Two non-intrusive uncertainty propagation approaches are proposed for the performance analysis of engineering systems described by expensive-to-evaluate deterministic computer models with parameters defined as interval variables. These approaches employ a machine learning based optimization strategy, the so-called Bayesian optimization, for evaluating the upper and lower bounds of a generic response variable over the set of possible responses obtained when each interval variable varies independently over its range. The lack of knowledge caused by not evaluating the response function for all the possible combinations of the interval variables is accounted for by developing a probabilistic description of the response variable itself by using a Gaussian Process regression model. An iterative procedure is developed for selecting a small number of simulations to be evaluated for updating this statistical model by using well-established acquisition functions and to assess the response bounds. In both approaches, an initial training dataset is defined. While one approach builds iteratively two distinct training datasets for evaluating separately the upper and lower bounds of the response variable, the other one builds iteratively a single training dataset. Consequently, the two approaches will produce different bound estimates at each iteration. The upper and lower response bounds are expressed as point estimates obtained from the mean function of the posterior distribution. Moreover, a confidence interval on each estimate is provided for effectively communicating to engineers when these estimates are obtained at a combination of the interval variables for which no deterministic simulation has been run. Finally, two metrics are proposed to define conditions for assessing if the predicted bound estimates can be considered satisfactory. The applicability of these two approaches is illustrated with two numerical applications, one focusing on vibration and the other on vibro-acoustics.}
}	1004	1004	}
	1005	1005
@INPROCEEDINGS{9278071,	1006	1006	@INPROCEEDINGS{9278071,
author={Petit, Maxime and Dellandrea, Emmanuel and Chen, Liming},	1007	1007	author={Petit, Maxime and Dellandrea, Emmanuel and Chen, Liming},
booktitle={2020 Joint IEEE 10th International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob)},	1008	1008	booktitle={2020 Joint IEEE 10th International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob)},
title={Bayesian Optimization for Developmental Robotics with Meta-Learning by Parameters Bounds Reduction},	1009	1009	title={Bayesian Optimization for Developmental Robotics with Meta-Learning by Parameters Bounds Reduction},
year={2020},	1010	1010	year={2020},
volume={},	1011	1011	volume={},
number={},	1012	1012	number={},
pages={1-8},	1013	1013	pages={1-8},
keywords={Optimization;Robots;Task analysis;Bayes methods;Visualization;Service robots;Cognition;developmental robotics;long-term memory;meta learning;hyperparmeters automatic optimization;case-based reasoning},	1014	1014	keywords={Optimization;Robots;Task analysis;Bayes methods;Visualization;Service robots;Cognition;developmental robotics;long-term memory;meta learning;hyperparmeters automatic optimization;case-based reasoning},
doi={10.1109/ICDL-EpiRob48136.2020.9278071}	1015	1015	doi={10.1109/ICDL-EpiRob48136.2020.9278071}
}	1016	1016	}
	1017	1017
@article{LI2023477,	1018	1018	@article{LI2023477,
title = {Hierarchical and partitioned planning strategy for closed-loop devices in low-voltage distribution network based on improved KMeans partition method},	1019	1019	title = {Hierarchical and partitioned planning strategy for closed-loop devices in low-voltage distribution network based on improved KMeans partition method},
journal = {Energy Reports},	1020	1020	journal = {Energy Reports},
volume = {9},	1021	1021	volume = {9},
pages = {477-485},	1022	1022	pages = {477-485},
year = {2023},	1023	1023	year = {2023},
note = {2022 The 3rd International Conference on Power and Electrical Engineering},	1024	1024	note = {2022 The 3rd International Conference on Power and Electrical Engineering},
issn = {2352-4847},	1025	1025	issn = {2352-4847},
doi = {https://doi.org/10.1016/j.egyr.2023.05.161},	1026	1026	doi = {https://doi.org/10.1016/j.egyr.2023.05.161},
url = {https://www.sciencedirect.com/science/article/pii/S2352484723009137},	1027	1027	url = {https://www.sciencedirect.com/science/article/pii/S2352484723009137},
author = {Jingqi Li and Junlin Li and Dan Wang and Chengxiong Mao and Zhitao Guan and Zhichao Liu and Miaomiao Du and Yuanzhuo Qi and Lexiang Wang and Wenge Liu and Pengfei Tang},	1028	1028	author = {Jingqi Li and Junlin Li and Dan Wang and Chengxiong Mao and Zhitao Guan and Zhichao Liu and Miaomiao Du and Yuanzhuo Qi and Lexiang Wang and Wenge Liu and Pengfei Tang},
keywords = {Closed-loop device, Distribution network partition, Device planning, Hierarchical planning, Improved KMeans partition method},	1029	1029	keywords = {Closed-loop device, Distribution network partition, Device planning, Hierarchical planning, Improved KMeans partition method},
abstract = {To improve the reliability of power supply, this paper proposes a hierarchical and partitioned planning strategy for closed-loop devices in low-voltage distribution network. Based on the geographic location and load situation of the distribution network area, an improved KMeans partition method is used to partition the area in the upper layer. In the lower layer, an intelligent algorithm is adopted to decide the numbers and placement locations of mobile low-voltage contact boxes and mobile seamless closed-loop load transfer devices in each partition with the goal of the highest closed-loop safety, the greatest improvement in annual power outage amount and the lowest cost. Finally, the feasibility and effectiveness of the proposed strategy are proved by an example.}	1030	1030	abstract = {To improve the reliability of power supply, this paper proposes a hierarchical and partitioned planning strategy for closed-loop devices in low-voltage distribution network. Based on the geographic location and load situation of the distribution network area, an improved KMeans partition method is used to partition the area in the upper layer. In the lower layer, an intelligent algorithm is adopted to decide the numbers and placement locations of mobile low-voltage contact boxes and mobile seamless closed-loop load transfer devices in each partition with the goal of the highest closed-loop safety, the greatest improvement in annual power outage amount and the lowest cost. Finally, the feasibility and effectiveness of the proposed strategy are proved by an example.}
}	1031	1031	}
	1032	1032
@article{SAXENA2024100838,	1033	1033	@article{SAXENA2024100838,
title = {Hybrid KNN-SVM machine learning approach for solar power forecasting},	1034	1034	title = {Hybrid KNN-SVM machine learning approach for solar power forecasting},
journal = {Environmental Challenges},	1035	1035	journal = {Environmental Challenges},
volume = {14},	1036	1036	volume = {14},
pages = {100838},	1037	1037	pages = {100838},
year = {2024},	1038	1038	year = {2024},
issn = {2667-0100},	1039	1039	issn = {2667-0100},
doi = {https://doi.org/10.1016/j.envc.2024.100838},	1040	1040	doi = {https://doi.org/10.1016/j.envc.2024.100838},
url = {https://www.sciencedirect.com/science/article/pii/S2667010024000040},	1041	1041	url = {https://www.sciencedirect.com/science/article/pii/S2667010024000040},
author = {Nishant Saxena and Rahul Kumar and Yarrapragada K S S Rao and Dilbag Singh Mondloe and Nishikant Kishor Dhapekar and Abhishek Sharma and Anil Singh Yadav},	1042	1042	author = {Nishant Saxena and Rahul Kumar and Yarrapragada K S S Rao and Dilbag Singh Mondloe and Nishikant Kishor Dhapekar and Abhishek Sharma and Anil Singh Yadav},
keywords = {Solar power forecasting, Hybrid model, KNN, Optimization, Solar energy, SVM},	1043	1043	keywords = {Solar power forecasting, Hybrid model, KNN, Optimization, Solar energy, SVM},
abstract = {Predictions about solar power will have a significant impact on large-scale renewable energy plants. Photovoltaic (PV) power generation forecasting is particularly sensitive to measuring the uncertainty in weather conditions. Although several conventional techniques like long short-term memory (LSTM), support vector machine (SVM), etc. are available, but due to some restrictions, their application is limited. To enhance the precision of forecasting solar power from solar farms, a hybrid machine learning model that includes blends of the K-Nearest Neighbor (KNN) machine learning technique with the SVM to increase reliability for power system operators is proposed in this investigation. The conventional LSTM technique is also implemented to compare the performance of the proposed hybrid technique. The suggested hybrid model is improved by the use of structural diversity and data diversity in KNN and SVM, respectively. For the solar power predictions, the suggested method was tested on the Jodhpur real-time series dataset obtained from the data centers of weather stations using Meteonorm. The data set includes metrics such as Hourly Average Temperature (HAT), Hourly Total Sunlight Duration (HTSD), Hourly Total Global Solar Radiation (HTGSR), and Hourly Total Photovoltaic Energy Generation (HTPEG). The collated data has been segmented into training data, validation data, and testing data. Furthermore, the proposed technique performed better when evaluated on the three performance indices, viz., accuracy, sensitivity, and specificity. Compared with the conventional LSTM technique, the hybrid technique improved the prediction with 98\% accuracy.}	1044	1044	abstract = {Predictions about solar power will have a significant impact on large-scale renewable energy plants. Photovoltaic (PV) power generation forecasting is particularly sensitive to measuring the uncertainty in weather conditions. Although several conventional techniques like long short-term memory (LSTM), support vector machine (SVM), etc. are available, but due to some restrictions, their application is limited. To enhance the precision of forecasting solar power from solar farms, a hybrid machine learning model that includes blends of the K-Nearest Neighbor (KNN) machine learning technique with the SVM to increase reliability for power system operators is proposed in this investigation. The conventional LSTM technique is also implemented to compare the performance of the proposed hybrid technique. The suggested hybrid model is improved by the use of structural diversity and data diversity in KNN and SVM, respectively. For the solar power predictions, the suggested method was tested on the Jodhpur real-time series dataset obtained from the data centers of weather stations using Meteonorm. The data set includes metrics such as Hourly Average Temperature (HAT), Hourly Total Sunlight Duration (HTSD), Hourly Total Global Solar Radiation (HTGSR), and Hourly Total Photovoltaic Energy Generation (HTPEG). The collated data has been segmented into training data, validation data, and testing data. Furthermore, the proposed technique performed better when evaluated on the three performance indices, viz., accuracy, sensitivity, and specificity. Compared with the conventional LSTM technique, the hybrid technique improved the prediction with 98\% accuracy.}
}	1045	1045	}
	1046	1046
@article{RAKESH2023100898,	1047	1047	@article{RAKESH2023100898,
title = {Moving object detection using modified GMM based background subtraction},	1048	1048	title = {Moving object detection using modified GMM based background subtraction},
journal = {Measurement: Sensors},	1049	1049	journal = {Measurement: Sensors},
volume = {30},	1050	1050	volume = {30},
pages = {100898},	1051	1051	pages = {100898},
year = {2023},	1052	1052	year = {2023},
issn = {2665-9174},	1053	1053	issn = {2665-9174},
doi = {https://doi.org/10.1016/j.measen.2023.100898},	1054	1054	doi = {https://doi.org/10.1016/j.measen.2023.100898},
url = {https://www.sciencedirect.com/science/article/pii/S2665917423002349},	1055	1055	url = {https://www.sciencedirect.com/science/article/pii/S2665917423002349},
author = {S. Rakesh and Nagaratna P. Hegde and M. {Venu Gopalachari} and D. Jayaram and Bhukya Madhu and Mohd Abdul Hameed and Ramdas Vankdothu and L.K. {Suresh Kumar}},	1056	1056	author = {S. Rakesh and Nagaratna P. Hegde and M. {Venu Gopalachari} and D. Jayaram and Bhukya Madhu and Mohd Abdul Hameed and Ramdas Vankdothu and L.K. {Suresh Kumar}},
keywords = {Background subtraction, Gaussian mixture models, Intelligent video surveillance, Object detection},	1057	1057	keywords = {Background subtraction, Gaussian mixture models, Intelligent video surveillance, Object detection},
abstract = {Academics have become increasingly interested in creating cutting-edge technologies to enhance Intelligent Video Surveillance (IVS) performance in terms of accuracy, speed, complexity, and deployment. It has been noted that precise object detection is the only way for IVS to function well in higher level applications including event interpretation, tracking, classification, and activity recognition. Through the use of cutting-edge techniques, the current study seeks to improve the performance accuracy of object detection techniques based on Gaussian Mixture Models (GMM). It is achieved by developing crucial phases in the object detecting process. In this study, it is discussed how to model each pixel as a mixture of Gaussians and how to update the model using an online k-means approximation. The adaptive mixture model's Gaussian distributions are then analyzed to identify which ones are more likely to be the product of a background process. Each pixel is categorized according to whether the background model is thought to include the Gaussian distribution that best depicts it.}	1058	1058	abstract = {Academics have become increasingly interested in creating cutting-edge technologies to enhance Intelligent Video Surveillance (IVS) performance in terms of accuracy, speed, complexity, and deployment. It has been noted that precise object detection is the only way for IVS to function well in higher level applications including event interpretation, tracking, classification, and activity recognition. Through the use of cutting-edge techniques, the current study seeks to improve the performance accuracy of object detection techniques based on Gaussian Mixture Models (GMM). It is achieved by developing crucial phases in the object detecting process. In this study, it is discussed how to model each pixel as a mixture of Gaussians and how to update the model using an online k-means approximation. The adaptive mixture model's Gaussian distributions are then analyzed to identify which ones are more likely to be the product of a background process. Each pixel is categorized according to whether the background model is thought to include the Gaussian distribution that best depicts it.}
}	1059	1059	}
	1060	1060
@article{JIAO2022540,	1061	1061	@article{JIAO2022540,
title = {Interpretable fuzzy clustering using unsupervised fuzzy decision trees},	1062	1062	title = {Interpretable fuzzy clustering using unsupervised fuzzy decision trees},
journal = {Information Sciences},	1063	1063	journal = {Information Sciences},
volume = {611},	1064	1064	volume = {611},
pages = {540-563},	1065	1065	pages = {540-563},
year = {2022},	1066	1066	year = {2022},
issn = {0020-0255},	1067	1067	issn = {0020-0255},
doi = {https://doi.org/10.1016/j.ins.2022.08.077},	1068	1068	doi = {https://doi.org/10.1016/j.ins.2022.08.077},
url = {https://www.sciencedirect.com/science/article/pii/S0020025522009872},	1069	1069	url = {https://www.sciencedirect.com/science/article/pii/S0020025522009872},
author = {Lianmeng Jiao and Haoyu Yang and Zhun-ga Liu and Quan Pan},	1070	1070	author = {Lianmeng Jiao and Haoyu Yang and Zhun-ga Liu and Quan Pan},
keywords = {Fuzzy clustering, Interpretable clustering, Unsupervised decision tree, Cluster merging},	1071	1071	keywords = {Fuzzy clustering, Interpretable clustering, Unsupervised decision tree, Cluster merging},
abstract = {In clustering process, fuzzy partition performs better than hard partition when the boundaries between clusters are vague. Whereas, traditional fuzzy clustering algorithms produce less interpretable results, limiting their application in security, privacy, and ethics fields. To that end, this paper proposes an interpretable fuzzy clustering algorithm—fuzzy decision tree-based clustering which combines the flexibility of fuzzy partition with the interpretability of the decision tree. We constructed an unsupervised multi-way fuzzy decision tree to achieve the interpretability of clustering, in which each cluster is determined by one or several paths from the root to leaf nodes. The proposed algorithm comprises three main modules: feature and cutting point-selection, node fuzzy splitting, and cluster merging. The first two modules are repeated to generate an initial unsupervised decision tree, and the final module is designed to combine similar leaf nodes to form the final compact clustering model. Our algorithm optimizes an internal clustering validation metric to automatically determine the number of clusters without their initial positions. The synthetic and benchmark datasets were used to test the performance of the proposed algorithm. Furthermore, we provided two examples demonstrating its interest in solving practical problems.}	1072	1072	abstract = {In clustering process, fuzzy partition performs better than hard partition when the boundaries between clusters are vague. Whereas, traditional fuzzy clustering algorithms produce less interpretable results, limiting their application in security, privacy, and ethics fields. To that end, this paper proposes an interpretable fuzzy clustering algorithm—fuzzy decision tree-based clustering which combines the flexibility of fuzzy partition with the interpretability of the decision tree. We constructed an unsupervised multi-way fuzzy decision tree to achieve the interpretability of clustering, in which each cluster is determined by one or several paths from the root to leaf nodes. The proposed algorithm comprises three main modules: feature and cutting point-selection, node fuzzy splitting, and cluster merging. The first two modules are repeated to generate an initial unsupervised decision tree, and the final module is designed to combine similar leaf nodes to form the final compact clustering model. Our algorithm optimizes an internal clustering validation metric to automatically determine the number of clusters without their initial positions. The synthetic and benchmark datasets were used to test the performance of the proposed algorithm. Furthermore, we provided two examples demonstrating its interest in solving practical problems.}
}	1073	1073	}
	1074	1074
@article{ARNAUGONZALEZ2023101516,	1075	1075	@article{ARNAUGONZALEZ2023101516,
title = {A methodological approach to enable natural language interaction in an Intelligent Tutoring System},	1076	1076	title = {A methodological approach to enable natural language interaction in an Intelligent Tutoring System},
journal = {Computer Speech and Language},	1077	1077	journal = {Computer Speech and Language},
volume = {81},	1078	1078	volume = {81},
pages = {101516},	1079	1079	pages = {101516},
year = {2023},	1080	1080	year = {2023},
issn = {0885-2308},	1081	1081	issn = {0885-2308},
doi = {https://doi.org/10.1016/j.csl.2023.101516},	1082	1082	doi = {https://doi.org/10.1016/j.csl.2023.101516},
url = {https://www.sciencedirect.com/science/article/pii/S0885230823000359},	1083	1083	url = {https://www.sciencedirect.com/science/article/pii/S0885230823000359},
author = {Pablo Arnau-González and Miguel Arevalillo-Herráez and Romina Albornoz-De Luise and David Arnau},	1084	1084	author = {Pablo Arnau-González and Miguel Arevalillo-Herráez and Romina Albornoz-De Luise and David Arnau},
keywords = {Intelligent tutoring systems (ITS), Interactive learning environments (ILE), Conversational agents, Rasa, Natural language understanding (NLU), Natural language processing (NLP)},	1085	1085	keywords = {Intelligent tutoring systems (ITS), Interactive learning environments (ILE), Conversational agents, Rasa, Natural language understanding (NLU), Natural language processing (NLP)},
abstract = {In this paper, we present and evaluate the recent incorporation of a conversational agent into an Intelligent Tutoring System (ITS), using the open-source machine learning framework Rasa. Once it has been appropriately trained, this tool is capable of identifying the intention of a given text input and extracting the relevant entities related to the message content. We describe both the generation of a realistic training set in Spanish language that enables the creation of the required Natural Language Understanding (NLU) models and the evaluation of the resulting system. For the generation of the training set, we have followed a methodology that can be easily exported to other ITS. The model evaluation shows that the conversational agent can correctly identify the majority of the user intents, reporting an f1-score above 95%, and cooperate with the ITS to produce a consistent dialogue flow that makes interaction more natural.}	1086	1086	abstract = {In this paper, we present and evaluate the recent incorporation of a conversational agent into an Intelligent Tutoring System (ITS), using the open-source machine learning framework Rasa. Once it has been appropriately trained, this tool is capable of identifying the intention of a given text input and extracting the relevant entities related to the message content. We describe both the generation of a realistic training set in Spanish language that enables the creation of the required Natural Language Understanding (NLU) models and the evaluation of the resulting system. For the generation of the training set, we have followed a methodology that can be easily exported to other ITS. The model evaluation shows that the conversational agent can correctly identify the majority of the user intents, reporting an f1-score above 95%, and cooperate with the ITS to produce a consistent dialogue flow that makes interaction more natural.}
}	1087	1087	}
	1088	1088
@article{MAO20224065,	1089	1089	@article{MAO20224065,
title = {An Exploratory Approach to Intelligent Quiz Question Recommendation},	1090	1090	title = {An Exploratory Approach to Intelligent Quiz Question Recommendation},
journal = {Procedia Computer Science},	1091	1091	journal = {Procedia Computer Science},
volume = {207},	1092	1092	volume = {207},
pages = {4065-4074},	1093	1093	pages = {4065-4074},
year = {2022},	1094	1094	year = {2022},
note = {Knowledge-Based and Intelligent Information and Engineering Systems: Proceedings of the 26th International Conference KES2022},	1095	1095	note = {Knowledge-Based and Intelligent Information and Engineering Systems: Proceedings of the 26th International Conference KES2022},
issn = {1877-0509},	1096	1096	issn = {1877-0509},
doi = {https://doi.org/10.1016/j.procs.2022.09.469},	1097	1097	doi = {https://doi.org/10.1016/j.procs.2022.09.469},
url = {https://www.sciencedirect.com/science/article/pii/S1877050922013631},	1098	1098	url = {https://www.sciencedirect.com/science/article/pii/S1877050922013631},
author = {Kejie Mao and Qiwen Dong and Ye Wang and Daocheng Honga},	1099	1099	author = {Kejie Mao and Qiwen Dong and Ye Wang and Daocheng Honga},
keywords = {question recommendation, two-sided recommender systems, reinforcement learning, intelligent tutoring},	1100	1100	keywords = {question recommendation, two-sided recommender systems, reinforcement learning, intelligent tutoring},
abstract = {With the rapid advancement of ICT, the digital transformation on education is greatly accelerating in various applications. As a particularly prominent application of digital education, quiz question recommendation is playing a vital role in precision teaching, smart tutoring, and personalized learning. However, the looming challenge of quiz question recommender for students is to satisfy the question diversity demands for students ZPD (the zone of proximal development) stage dynamically online. Therefore, we propose to formalize quiz question recommendation with a novel approach of reinforcement learning based two-sided recommender system. We develop a recommendation framework RTR (Reinforcement-Learning based Two-sided Recommender Systems) for taking into account the interests of both sides of the system, learning and adapting to those interests in real time, and resulting in more satisfactory recommended content. This established recommendation framework captures question characters and student dynamic preferences by considering the emergence of both sides of the system, and it yields a better learning experience in the context of practical quiz question generation.}	1101	1101	abstract = {With the rapid advancement of ICT, the digital transformation on education is greatly accelerating in various applications. As a particularly prominent application of digital education, quiz question recommendation is playing a vital role in precision teaching, smart tutoring, and personalized learning. However, the looming challenge of quiz question recommender for students is to satisfy the question diversity demands for students ZPD (the zone of proximal development) stage dynamically online. Therefore, we propose to formalize quiz question recommendation with a novel approach of reinforcement learning based two-sided recommender system. We develop a recommendation framework RTR (Reinforcement-Learning based Two-sided Recommender Systems) for taking into account the interests of both sides of the system, learning and adapting to those interests in real time, and resulting in more satisfactory recommended content. This established recommendation framework captures question characters and student dynamic preferences by considering the emergence of both sides of the system, and it yields a better learning experience in the context of practical quiz question generation.}
}	1102	1102	}
	1103	1103
@article{CLEMENTE2022118171,	1104	1104	@article{CLEMENTE2022118171,
title = {A proposal for an adaptive Recommender System based on competences and ontologies},	1105	1105	title = {A proposal for an adaptive Recommender System based on competences and ontologies},
journal = {Expert Systems with Applications},	1106	1106	journal = {Expert Systems with Applications},
volume = {208},	1107	1107	volume = {208},
pages = {118171},	1108	1108	pages = {118171},
year = {2022},	1109	1109	year = {2022},
issn = {0957-4174},	1110	1110	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2022.118171},	1111	1111	doi = {https://doi.org/10.1016/j.eswa.2022.118171},
url = {https://www.sciencedirect.com/science/article/pii/S0957417422013392},	1112	1112	url = {https://www.sciencedirect.com/science/article/pii/S0957417422013392},
author = {Julia Clemente and Héctor Yago and Javier {de Pedro-Carracedo} and Javier Bueno},	1113	1113	author = {Julia Clemente and Héctor Yago and Javier {de Pedro-Carracedo} and Javier Bueno},
keywords = {Recommender system, , Ontology network, Methodological development, Student modeling},	1114	1114	keywords = {Recommender system, , Ontology network, Methodological development, Student modeling},
abstract = {Context:	1115	1115	abstract = {Context:
Competences represent an interesting pedagogical support in many processes like diagnosis or recommendation. From these, it is possible to infer information about the progress of the student to provide help targeted both, trainers who must make adaptive tutoring decisions for each learner, and students to detect and correct their learning weaknesses. For the correct development of any of these tasks, it is important to have a suitable student model that allows the representation of the most significant information possible about the student. Additionally, it would be very advantageous for this modeling to incorporate mechanisms from which it would be possible to infer more information about the student’s state of knowledge.	1116	1116	Competences represent an interesting pedagogical support in many processes like diagnosis or recommendation. From these, it is possible to infer information about the progress of the student to provide help targeted both, trainers who must make adaptive tutoring decisions for each learner, and students to detect and correct their learning weaknesses. For the correct development of any of these tasks, it is important to have a suitable student model that allows the representation of the most significant information possible about the student. Additionally, it would be very advantageous for this modeling to incorporate mechanisms from which it would be possible to infer more information about the student’s state of knowledge.
Objective:	1117	1117	Objective:
To facilitate this goal, in this paper a new approach to develop an adaptive competence-based recommender system is proposed.	1118	1118	To facilitate this goal, in this paper a new approach to develop an adaptive competence-based recommender system is proposed.
Method:	1119	1119	Method:
We present a methodological development guide as well as a set of ontological and non-ontological resources to develop and adapt the prototype of the proposed recommender system.	1120	1120	We present a methodological development guide as well as a set of ontological and non-ontological resources to develop and adapt the prototype of the proposed recommender system.
Results:	1121	1121	Results:
A modular flexible ontology network previously built for this purpose has been extended, which is responsible for recording the instructional design and student information. Furthermore, we describe a case study based on a first aid learning experience to assess the prototype with the proposed methodology.	1122	1122	A modular flexible ontology network previously built for this purpose has been extended, which is responsible for recording the instructional design and student information. Furthermore, we describe a case study based on a first aid learning experience to assess the prototype with the proposed methodology.
Conclusions:	1123	1123	Conclusions:
We highlight the relevance of flexibility and adaptability in learning modeling and recommendation processes. In order to promote improvement in the personalized learning of students, we present a Recommender System prototype taking advantages of ontologies, with a methodological guide, a broad taxonomy of recommendation criteria and the nature of competences. Future lines of research lines, including a more comprehensive evaluation of the system, will allow us to demonstrate in depth its adaptability according to the characteristics of the student, flexibility and extensibility for its integration in various environments and domains.}	1124	1124	We highlight the relevance of flexibility and adaptability in learning modeling and recommendation processes. In order to promote improvement in the personalized learning of students, we present a Recommender System prototype taking advantages of ontologies, with a methodological guide, a broad taxonomy of recommendation criteria and the nature of competences. Future lines of research lines, including a more comprehensive evaluation of the system, will allow us to demonstrate in depth its adaptability according to the characteristics of the student, flexibility and extensibility for its integration in various environments and domains.}
}	1125	1125	}
	1126	1126
@article{https://doi.org/10.1155/2023/2578286,	1127	1127	@article{https://doi.org/10.1155/2023/2578286,
author = {Li, Linqing and Wang, Zhifeng},	1128	1128	author = {Li, Linqing and Wang, Zhifeng},
title = {Knowledge Graph-Enhanced Intelligent Tutoring System Based on Exercise Representativeness and Informativeness},	1129	1129	title = {Knowledge Graph-Enhanced Intelligent Tutoring System Based on Exercise Representativeness and Informativeness},
journal = {International Journal of Intelligent Systems},	1130	1130	journal = {International Journal of Intelligent Systems},
volume = {2023},	1131	1131	volume = {2023},
number = {1},	1132	1132	number = {1},
pages = {2578286},	1133	1133	pages = {2578286},
doi = {https://doi.org/10.1155/2023/2578286},	1134	1134	doi = {https://doi.org/10.1155/2023/2578286},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1155/2023/2578286},	1135	1135	url = {https://onlinelibrary.wiley.com/doi/abs/10.1155/2023/2578286},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1155/2023/2578286},	1136	1136	eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1155/2023/2578286},
abstract = {In the realm of online tutoring intelligent systems, e-learners are exposed to a substantial volume of learning content. The extraction and organization of exercises and skills hold significant importance in establishing clear learning objectives and providing appropriate exercise recommendations. Presently, knowledge graph-based recommendation algorithms have garnered considerable attention among researchers. However, these algorithms solely consider knowledge graphs with single relationships and do not effectively model exercise-rich features, such as exercise representativeness and informativeness. Consequently, this paper proposes a framework, namely, the Knowledge Graph Importance-Exercise Representativeness and Informativeness Framework, to address these two issues. The framework consists of four intricate components and a novel cognitive diagnosis model called the Neural Attentive Cognitive Diagnosis model to recommend the proper exercises. These components encompass the informativeness component, exercise representation component, knowledge importance component, and exercise representativeness component. The informativeness component evaluates the informational value of each exercise and identifies the candidate exercise set (EC) that exhibits the highest exercise informativeness. Moreover, the exercise representation component utilizes a graph neural network to process student records. The output of the graph neural network serves as the input for exercise-level attention and skill-level attention, ultimately generating exercise embeddings and skill embeddings. Furthermore, the skill embeddings are employed as input for the knowledge importance component. This component transforms a one-dimensional knowledge graph into a multidimensional one through four class relations and calculates skill importance weights based on novelty and popularity. Subsequently, the exercise representativeness component incorporates exercise weight knowledge coverage to select exercises from the candidate exercise set for the tested exercise set. Lastly, the cognitive diagnosis model leverages exercise representation and skill importance weights to predict student performance on the test set and estimate their knowledge state. To evaluate the effectiveness of our selection strategy, extensive experiments were conducted on two types of publicly available educational datasets. The experimental results demonstrate that our framework can recommend appropriate exercises to students, leading to improved student performance.},	1137	1137	abstract = {In the realm of online tutoring intelligent systems, e-learners are exposed to a substantial volume of learning content. The extraction and organization of exercises and skills hold significant importance in establishing clear learning objectives and providing appropriate exercise recommendations. Presently, knowledge graph-based recommendation algorithms have garnered considerable attention among researchers. However, these algorithms solely consider knowledge graphs with single relationships and do not effectively model exercise-rich features, such as exercise representativeness and informativeness. Consequently, this paper proposes a framework, namely, the Knowledge Graph Importance-Exercise Representativeness and Informativeness Framework, to address these two issues. The framework consists of four intricate components and a novel cognitive diagnosis model called the Neural Attentive Cognitive Diagnosis model to recommend the proper exercises. These components encompass the informativeness component, exercise representation component, knowledge importance component, and exercise representativeness component. The informativeness component evaluates the informational value of each exercise and identifies the candidate exercise set (EC) that exhibits the highest exercise informativeness. Moreover, the exercise representation component utilizes a graph neural network to process student records. The output of the graph neural network serves as the input for exercise-level attention and skill-level attention, ultimately generating exercise embeddings and skill embeddings. Furthermore, the skill embeddings are employed as input for the knowledge importance component. This component transforms a one-dimensional knowledge graph into a multidimensional one through four class relations and calculates skill importance weights based on novelty and popularity. Subsequently, the exercise representativeness component incorporates exercise weight knowledge coverage to select exercises from the candidate exercise set for the tested exercise set. Lastly, the cognitive diagnosis model leverages exercise representation and skill importance weights to predict student performance on the test set and estimate their knowledge state. To evaluate the effectiveness of our selection strategy, extensive experiments were conducted on two types of publicly available educational datasets. The experimental results demonstrate that our framework can recommend appropriate exercises to students, leading to improved student performance.},
year = {2023}	1138	1138	year = {2023}
}	1139	1139	}
	1140	1140
@inproceedings{badier:hal-04092828,	1141	1141	@inproceedings{badier:hal-04092828,
TITLE = {{Comprendre les usages et effets d'un syst{\`e}me de recommandations p{\'e}dagogiques en contexte d'apprentissage non-formel}},	1142	1142	TITLE = {{Comprendre les usages et effets d'un syst{\`e}me de recommandations p{\'e}dagogiques en contexte d'apprentissage non-formel}},
AUTHOR = {Badier, Ana{\"e}lle and Lefort, Mathieu and Lefevre, Marie},	1143	1143	AUTHOR = {Badier, Ana{\"e}lle and Lefort, Mathieu and Lefevre, Marie},
URL = {https://hal.science/hal-04092828},	1144	1144	URL = {https://hal.science/hal-04092828},
BOOKTITLE = {{EIAH'23}},	1145	1145	BOOKTITLE = {{EIAH'23}},
ADDRESS = {Brest, France},	1146	1146	ADDRESS = {Brest, France},
YEAR = {2023},	1147	1147	YEAR = {2023},
MONTH = Jun,	1148	1148	MONTH = Jun,
HAL_ID = {hal-04092828},	1149	1149	HAL_ID = {hal-04092828},
HAL_VERSION = {v1},	1150	1150	HAL_VERSION = {v1},
}	1151	1151	}
	1152	1152
@article{BADRA2023108920,	1153	1153	@article{BADRA2023108920,
title = {Case-based prediction – A survey},	1154	1154	title = {Case-based prediction – A survey},
journal = {International Journal of Approximate Reasoning},	1155	1155	journal = {International Journal of Approximate Reasoning},
volume = {158},	1156	1156	volume = {158},
pages = {108920},	1157	1157	pages = {108920},
year = {2023},	1158	1158	year = {2023},
issn = {0888-613X},	1159	1159	issn = {0888-613X},
doi = {https://doi.org/10.1016/j.ijar.2023.108920},	1160	1160	doi = {https://doi.org/10.1016/j.ijar.2023.108920},
url = {https://www.sciencedirect.com/science/article/pii/S0888613X23000440},	1161	1161	url = {https://www.sciencedirect.com/science/article/pii/S0888613X23000440},
author = {Fadi Badra and Marie-Jeanne Lesot},	1162	1162	author = {Fadi Badra and Marie-Jeanne Lesot},
keywords = {Case-based prediction, Analogical transfer, Similarity},	1163	1163	keywords = {Case-based prediction, Analogical transfer, Similarity},
abstract = {This paper clarifies the relation between case-based prediction and analogical transfer. Case-based prediction consists in predicting the outcome associated with a new case directly from its comparison with a set of cases retrieved from a case base, by relying solely on a structured memory and some similarity measures. Analogical transfer is a cognitive process that allows to derive some new information about a target situation by applying a plausible inference principle, according to which if two situations are similar with respect to some criteria, then it is plausible that they are also similar with respect to other criteria. Case-based prediction algorithms are known to apply analogical transfer to make predictions, but the existing approaches are diverse, and developing a unified theory of case-based prediction remains a challenge. In this paper, we show that a common principle underlying case-based prediction methods is that they interpret the plausible inference as a transfer of similarity knowledge from a situation space to an outcome space. Among all potential outcomes, the predicted outcome is the one that optimizes this transfer, i.e., that makes the similarities in the outcome space most compatible with the observed similarities in the situation space. Based on this observation, a systematic analysis of the different theories of case-based prediction is presented, where the approaches are distinguished according to the type of knowledge used to measure the compatibility between the two sets of similarity relations.}	1164	1164	abstract = {This paper clarifies the relation between case-based prediction and analogical transfer. Case-based prediction consists in predicting the outcome associated with a new case directly from its comparison with a set of cases retrieved from a case base, by relying solely on a structured memory and some similarity measures. Analogical transfer is a cognitive process that allows to derive some new information about a target situation by applying a plausible inference principle, according to which if two situations are similar with respect to some criteria, then it is plausible that they are also similar with respect to other criteria. Case-based prediction algorithms are known to apply analogical transfer to make predictions, but the existing approaches are diverse, and developing a unified theory of case-based prediction remains a challenge. In this paper, we show that a common principle underlying case-based prediction methods is that they interpret the plausible inference as a transfer of similarity knowledge from a situation space to an outcome space. Among all potential outcomes, the predicted outcome is the one that optimizes this transfer, i.e., that makes the similarities in the outcome space most compatible with the observed similarities in the situation space. Based on this observation, a systematic analysis of the different theories of case-based prediction is presented, where the approaches are distinguished according to the type of knowledge used to measure the compatibility between the two sets of similarity relations.}
}	1165	1165	}
	1166	1166
	1167	1167
@Article{jmse11050890 ,	1168	1168	@Article{jmse11050890 ,
AUTHOR = {Louvros, Panagiotis and Stefanidis, Fotios and Boulougouris, Evangelos and Komianos, Alexandros and Vassalos, Dracos},	1169	1169	AUTHOR = {Louvros, Panagiotis and Stefanidis, Fotios and Boulougouris, Evangelos and Komianos, Alexandros and Vassalos, Dracos},
TITLE = {Machine Learning and Case-Based Reasoning for Real-Time Onboard Prediction of the Survivability of Ships},	1170	1170	TITLE = {Machine Learning and Case-Based Reasoning for Real-Time Onboard Prediction of the Survivability of Ships},
JOURNAL = {Journal of Marine Science and Engineering},	1171	1171	JOURNAL = {Journal of Marine Science and Engineering},
VOLUME = {11},	1172	1172	VOLUME = {11},
YEAR = {2023},	1173	1173	YEAR = {2023},
NUMBER = {5},	1174	1174	NUMBER = {5},
ARTICLE-NUMBER = {890},	1175	1175	ARTICLE-NUMBER = {890},
URL = {https://www.mdpi.com/2077-1312/11/5/890},	1176	1176	URL = {https://www.mdpi.com/2077-1312/11/5/890},
ISSN = {2077-1312},	1177	1177	ISSN = {2077-1312},
ABSTRACT = {The subject of damaged stability has greatly profited from the development of new tools and techniques in recent history. Specifically, the increased computational power and the probabilistic approach have transformed the subject, increasing accuracy and fidelity, hence allowing for a universal application and the inclusion of the most probable scenarios. Currently, all ships are evaluated for their stability and are expected to survive the dangers they will most likely face. However, further advancements in simulations have made it possible to further increase the fidelity and accuracy of simulated casualties. Multiple time domain and, to a lesser extent, Computational Fluid dynamics (CFD) solutions have been suggested as the next “evolutionary” step for damage stability. However, while those techniques are demonstrably more accurate, the computational power to utilize them for the task of probabilistic evaluation is not there yet. In this paper, the authors present a novel approach that aims to serve as a stopgap measure for introducing the time domain simulations in the existing framework. Specifically, the methodology presented serves the purpose of a fast decision support tool which is able to provide information regarding the ongoing casualty utilizing prior knowledge gained from simulations. This work was needed and developed for the purposes of the EU-funded project SafePASS.},	1178	1178	ABSTRACT = {The subject of damaged stability has greatly profited from the development of new tools and techniques in recent history. Specifically, the increased computational power and the probabilistic approach have transformed the subject, increasing accuracy and fidelity, hence allowing for a universal application and the inclusion of the most probable scenarios. Currently, all ships are evaluated for their stability and are expected to survive the dangers they will most likely face. However, further advancements in simulations have made it possible to further increase the fidelity and accuracy of simulated casualties. Multiple time domain and, to a lesser extent, Computational Fluid dynamics (CFD) solutions have been suggested as the next “evolutionary” step for damage stability. However, while those techniques are demonstrably more accurate, the computational power to utilize them for the task of probabilistic evaluation is not there yet. In this paper, the authors present a novel approach that aims to serve as a stopgap measure for introducing the time domain simulations in the existing framework. Specifically, the methodology presented serves the purpose of a fast decision support tool which is able to provide information regarding the ongoing casualty utilizing prior knowledge gained from simulations. This work was needed and developed for the purposes of the EU-funded project SafePASS.},
DOI = {10.3390/jmse11050890}	1179	1179	DOI = {10.3390/jmse11050890}
}	1180	1180	}
	1181	1181
	1182	1182
@Article{su14031366,	1183	1183	@Article{su14031366,
AUTHOR = {Chun, Se-Hak and Jang, Jae-Won},	1184	1184	AUTHOR = {Chun, Se-Hak and Jang, Jae-Won},
TITLE = {A New Trend Pattern-Matching Method of Interactive Case-Based Reasoning for Stock Price Predictions},	1185	1185	TITLE = {A New Trend Pattern-Matching Method of Interactive Case-Based Reasoning for Stock Price Predictions},
JOURNAL = {Sustainability},	1186	1186	JOURNAL = {Sustainability},
VOLUME = {14},	1187	1187	VOLUME = {14},
YEAR = {2022},	1188	1188	YEAR = {2022},
NUMBER = {3},	1189	1189	NUMBER = {3},
ARTICLE-NUMBER = {1366},	1190	1190	ARTICLE-NUMBER = {1366},
URL = {https://www.mdpi.com/2071-1050/14/3/1366},	1191	1191	URL = {https://www.mdpi.com/2071-1050/14/3/1366},
ISSN = {2071-1050},	1192	1192	ISSN = {2071-1050},
ABSTRACT = {In this paper, we suggest a new case-based reasoning method for stock price predictions using the knowledge of traders to select similar past patterns among nearest neighbors obtained from a traditional case-based reasoning machine. Thus, this method overcomes the limitation of conventional case-based reasoning, which does not consider how to retrieve similar neighbors from previous patterns in terms of a graphical pattern. In this paper, we show how the proposed method can be used when traders find similar time series patterns among nearest cases. For this, we suggest an interactive prediction system where traders can select similar patterns with individual knowledge among automatically recommended neighbors by case-based reasoning. In this paper, we demonstrate how traders can use their knowledge to select similar patterns using a graphical interface, serving as an exemplar for the target. These concepts are investigated against the backdrop of a practical application involving the prediction of three individual stock prices, i.e., Zoom, Airbnb, and Twitter, as well as the prediction of the Dow Jones Industrial Average (DJIA). The verification of the prediction results is compared with a random walk model based on the RMSE and Hit ratio. The results show that the proposed technique is more effective than the random walk model but it does not statistically surpass the random walk model.},	1193	1193	ABSTRACT = {In this paper, we suggest a new case-based reasoning method for stock price predictions using the knowledge of traders to select similar past patterns among nearest neighbors obtained from a traditional case-based reasoning machine. Thus, this method overcomes the limitation of conventional case-based reasoning, which does not consider how to retrieve similar neighbors from previous patterns in terms of a graphical pattern. In this paper, we show how the proposed method can be used when traders find similar time series patterns among nearest cases. For this, we suggest an interactive prediction system where traders can select similar patterns with individual knowledge among automatically recommended neighbors by case-based reasoning. In this paper, we demonstrate how traders can use their knowledge to select similar patterns using a graphical interface, serving as an exemplar for the target. These concepts are investigated against the backdrop of a practical application involving the prediction of three individual stock prices, i.e., Zoom, Airbnb, and Twitter, as well as the prediction of the Dow Jones Industrial Average (DJIA). The verification of the prediction results is compared with a random walk model based on the RMSE and Hit ratio. The results show that the proposed technique is more effective than the random walk model but it does not statistically surpass the random walk model.},
DOI = {10.3390/su14031366}	1194	1194	DOI = {10.3390/su14031366}
}	1195	1195	}
	1196	1196
@Article{fire7040107,	1197	1197	@Article{fire7040107,
AUTHOR = {Pei, Qiuyan and Jia, Zhichao and Liu, Jia and Wang, Yi and Wang, Junhui and Zhang, Yanqi},	1198	1198	AUTHOR = {Pei, Qiuyan and Jia, Zhichao and Liu, Jia and Wang, Yi and Wang, Junhui and Zhang, Yanqi},
TITLE = {Prediction of Coal Spontaneous Combustion Hazard Grades Based on Fuzzy Clustered Case-Based Reasoning},	1199	1199	TITLE = {Prediction of Coal Spontaneous Combustion Hazard Grades Based on Fuzzy Clustered Case-Based Reasoning},
JOURNAL = {Fire},	1200	1200	JOURNAL = {Fire},
VOLUME = {7},	1201	1201	VOLUME = {7},
YEAR = {2024},	1202	1202	YEAR = {2024},
NUMBER = {4},	1203	1203	NUMBER = {4},
ARTICLE-NUMBER = {107},	1204	1204	ARTICLE-NUMBER = {107},
URL = {https://www.mdpi.com/2571-6255/7/4/107},	1205	1205	URL = {https://www.mdpi.com/2571-6255/7/4/107},
ISSN = {2571-6255},	1206	1206	ISSN = {2571-6255},
ABSTRACT = {Accurate prediction of the coal spontaneous combustion hazard grades is of great significance to ensure the safe production of coal mines. However, traditional coal temperature prediction models have low accuracy and do not predict the coal spontaneous combustion hazard grades. In order to accurately predict coal spontaneous combustion hazard grades, a prediction model of coal spontaneous combustion based on principal component analysis (PCA), case-based reasoning (CBR), fuzzy clustering (FM), and the snake optimization (SO) algorithm was proposed in this manuscript. Firstly, based on the change rule of the concentration of signature gases in the process of coal warming, a new method of classifying the risk of spontaneous combustion of coal was established. Secondly, MeanRadius-SMOTE was adopted to balance the data structure. The weights of the prediction indicators were calculated through PCA to enhance the prediction precision of the CBR model. Then, by employing FM in the case base, the computational cost of CBR was reduced and its computational efficiency was improved. The SO algorithm was used to determine the hyperparameters in the PCA-FM-CBR model. In addition, multiple comparative experiments were conducted to verify the superiority of the model proposed in this manuscript. The results indicated that SO-PCA-FM-CBR possesses good prediction performance and also improves computational efficiency. Finally, the authors of this manuscript adopted the Random Balance Designs—Fourier Amplitude Sensitivity Test (RBD-FAST) to explain the output of the model and analyzed the global importance of input variables. The results demonstrated that CO is the most important variable affecting the coal spontaneous combustion hazard grades.},	1207	1207	ABSTRACT = {Accurate prediction of the coal spontaneous combustion hazard grades is of great significance to ensure the safe production of coal mines. However, traditional coal temperature prediction models have low accuracy and do not predict the coal spontaneous combustion hazard grades. In order to accurately predict coal spontaneous combustion hazard grades, a prediction model of coal spontaneous combustion based on principal component analysis (PCA), case-based reasoning (CBR), fuzzy clustering (FM), and the snake optimization (SO) algorithm was proposed in this manuscript. Firstly, based on the change rule of the concentration of signature gases in the process of coal warming, a new method of classifying the risk of spontaneous combustion of coal was established. Secondly, MeanRadius-SMOTE was adopted to balance the data structure. The weights of the prediction indicators were calculated through PCA to enhance the prediction precision of the CBR model. Then, by employing FM in the case base, the computational cost of CBR was reduced and its computational efficiency was improved. The SO algorithm was used to determine the hyperparameters in the PCA-FM-CBR model. In addition, multiple comparative experiments were conducted to verify the superiority of the model proposed in this manuscript. The results indicated that SO-PCA-FM-CBR possesses good prediction performance and also improves computational efficiency. Finally, the authors of this manuscript adopted the Random Balance Designs—Fourier Amplitude Sensitivity Test (RBD-FAST) to explain the output of the model and analyzed the global importance of input variables. The results demonstrated that CO is the most important variable affecting the coal spontaneous combustion hazard grades.},
DOI = {10.3390/fire7040107}	1208	1208	DOI = {10.3390/fire7040107}
}	1209	1209	}
	1210	1210
@Article{Desmarais2012,	1211	1211	@Article{Desmarais2012,
author={Desmarais, Michel C.	1212	1212	author={Desmarais, Michel C.
and Baker, Ryan S. J. d.},	1213	1213	and Baker, Ryan S. J. d.},
title={A review of recent advances in learner and skill modeling in intelligent learning environments},	1214	1214	title={A review of recent advances in learner and skill modeling in intelligent learning environments},
journal={User Modeling and User-Adapted Interaction},	1215	1215	journal={User Modeling and User-Adapted Interaction},
year={2012},	1216	1216	year={2012},
month={Apr},	1217	1217	month={Apr},
day={01},	1218	1218	day={01},
volume={22},	1219	1219	volume={22},
number={1},	1220	1220	number={1},
pages={9-38},	1221	1221	pages={9-38},
abstract={In recent years, learner models have emerged from the research laboratory and research classrooms into the wider world. Learner models are now embedded in real world applications which can claim to have thousands, or even hundreds of thousands, of users. Probabilistic models for skill assessment are playing a key role in these advanced learning environments. In this paper, we review the learner models that have played the largest roles in the success of these learning environments, and also the latest advances in the modeling and assessment of learner skills. We conclude by discussing related advancements in modeling other key constructs such as learner motivation, emotional and attentional state, meta-cognition and self-regulated learning, group learning, and the recent movement towards open and shared learner models.},	1222	1222	abstract={In recent years, learner models have emerged from the research laboratory and research classrooms into the wider world. Learner models are now embedded in real world applications which can claim to have thousands, or even hundreds of thousands, of users. Probabilistic models for skill assessment are playing a key role in these advanced learning environments. In this paper, we review the learner models that have played the largest roles in the success of these learning environments, and also the latest advances in the modeling and assessment of learner skills. We conclude by discussing related advancements in modeling other key constructs such as learner motivation, emotional and attentional state, meta-cognition and self-regulated learning, group learning, and the recent movement towards open and shared learner models.},
issn={1573-1391},	1223	1223	issn={1573-1391},
doi={10.1007/s11257-011-9106-8},	1224	1224	doi={10.1007/s11257-011-9106-8},
url={https://doi.org/10.1007/s11257-011-9106-8}	1225	1225	url={https://doi.org/10.1007/s11257-011-9106-8}
}	1226	1226	}
	1227	1227
@article{Eide,	1228	1228	@article{Eide,
title={Dynamic slate recommendation with gated recurrent units and Thompson sampling},	1229	1229	title={Dynamic slate recommendation with gated recurrent units and Thompson sampling},
author={Eide, Simen and Leslie, David S. and Frigessi, Arnoldo},	1230	1230	author={Eide, Simen and Leslie, David S. and Frigessi, Arnoldo},
language={English},	1231	1231	language={English},
type={article},	1232	1232	type={article},
volume = {36},	1233	1233	volume = {36},
year = {2022},	1234	1234	year = {2022},
issn = {1573-756X},	1235	1235	issn = {1573-756X},
doi = {https://doi.org/10.1007/s10618-022-00849-w},	1236	1236	doi = {https://doi.org/10.1007/s10618-022-00849-w},
url = {https://doi.org/10.1007/s10618-022-00849-w},	1237	1237	url = {https://doi.org/10.1007/s10618-022-00849-w},
abstract={We consider the problem of recommending relevant content to users of an internet platform in the form of lists of items, called slates. We introduce a variational Bayesian Recurrent Neural Net recommender system that acts on time series of interactions between the internet platform and the user, and which scales to real world industrial situations. The recommender system is tested both online on real users, and on an offline dataset collected from a Norwegian web-based marketplace, FINN.no, that is made public for research. This is one of the first publicly available datasets which includes all the slates that are presented to users as well as which items (if any) in the slates were clicked on. Such a data set allows us to move beyond the common assumption that implicitly assumes that users are considering all possible items at each interaction. Instead we build our likelihood using the items that are actually in the slate, and evaluate the strengths and weaknesses of both approaches theoretically and in experiments. We also introduce a hierarchical prior for the item parameters based on group memberships. Both item parameters and user preferences are learned probabilistically. Furthermore, we combine our model with bandit strategies to ensure learning, and introduce ‘in-slate Thompson sampling’ which makes use of the slates to maximise explorative opportunities. We show experimentally that explorative recommender strategies perform on par or above their greedy counterparts. Even without making use of exploration to learn more effectively, click rates increase simply because of improved diversity in the recommended slates.}	1238	1238	abstract={We consider the problem of recommending relevant content to users of an internet platform in the form of lists of items, called slates. We introduce a variational Bayesian Recurrent Neural Net recommender system that acts on time series of interactions between the internet platform and the user, and which scales to real world industrial situations. The recommender system is tested both online on real users, and on an offline dataset collected from a Norwegian web-based marketplace, FINN.no, that is made public for research. This is one of the first publicly available datasets which includes all the slates that are presented to users as well as which items (if any) in the slates were clicked on. Such a data set allows us to move beyond the common assumption that implicitly assumes that users are considering all possible items at each interaction. Instead we build our likelihood using the items that are actually in the slate, and evaluate the strengths and weaknesses of both approaches theoretically and in experiments. We also introduce a hierarchical prior for the item parameters based on group memberships. Both item parameters and user preferences are learned probabilistically. Furthermore, we combine our model with bandit strategies to ensure learning, and introduce ‘in-slate Thompson sampling’ which makes use of the slates to maximise explorative opportunities. We show experimentally that explorative recommender strategies perform on par or above their greedy counterparts. Even without making use of exploration to learn more effectively, click rates increase simply because of improved diversity in the recommended slates.}
}	1239	1239	}
	1240	1240
@InProceedings{10.1007/978-3-031-09680-8_14,	1241	1241	@InProceedings{10.1007/978-3-031-09680-8_14,
author={Sablayrolles, Louis	1242	1242	author={Sablayrolles, Louis
and Lefevre, Marie	1243	1243	and Lefevre, Marie
and Guin, Nathalie	1244	1244	and Guin, Nathalie
and Broisin, Julien},	1245	1245	and Broisin, Julien},
editor={Crossley, Scott	1246	1246	editor={Crossley, Scott
and Popescu, Elvira},	1247	1247	and Popescu, Elvira},
title={Design and Evaluation of a Competency-Based Recommendation Process},	1248	1248	title={Design and Evaluation of a Competency-Based Recommendation Process},
booktitle={Intelligent Tutoring Systems},	1249	1249	booktitle={Intelligent Tutoring Systems},
year={2022},	1250	1250	year={2022},
publisher={Springer International Publishing},	1251	1251	publisher={Springer International Publishing},
address={Cham},	1252	1252	address={Cham},
pages={148--160},	1253	1253	pages={148--160},
abstract={The purpose of recommending activities to learners is to provide them with resources adapted to their needs, to facilitate the learning process. However, when teachers face a large number of students, it is difficult for them to recommend a personalized list of resources to each learner. In this paper, we are interested in the design of a system that automatically recommends resources to learners using their cognitive profile expressed in terms of competencies, but also according to a specific strategy defined by teachers. Our contributions relate to (1) a competency-based pedagogical strategy allowing to express the teacher's expertise, and (2) a recommendation process based on this strategy. This process has been experimented and assessed with students learning Shell programming in a first-year computer science degree. The first results show that (i) the items selected by our system from the set of possible items were relevant according to the experts; (ii) our system provided recommendations in a reasonable time; (iii) the recommendations were consulted by the learners but lacked usability.},	1254	1254	abstract={The purpose of recommending activities to learners is to provide them with resources adapted to their needs, to facilitate the learning process. However, when teachers face a large number of students, it is difficult for them to recommend a personalized list of resources to each learner. In this paper, we are interested in the design of a system that automatically recommends resources to learners using their cognitive profile expressed in terms of competencies, but also according to a specific strategy defined by teachers. Our contributions relate to (1) a competency-based pedagogical strategy allowing to express the teacher's expertise, and (2) a recommendation process based on this strategy. This process has been experimented and assessed with students learning Shell programming in a first-year computer science degree. The first results show that (i) the items selected by our system from the set of possible items were relevant according to the experts; (ii) our system provided recommendations in a reasonable time; (iii) the recommendations were consulted by the learners but lacked usability.},
isbn={978-3-031-09680-8}	1255	1255	isbn={978-3-031-09680-8}
}	1256	1256	}
	1257	1257
@inproceedings{10.1145/3578337.3605122,	1258	1258	@inproceedings{10.1145/3578337.3605122,
author = {Xu, Shuyuan and Ge, Yingqiang and Li, Yunqi and Fu, Zuohui and Chen, Xu and Zhang, Yongfeng},	1259	1259	author = {Xu, Shuyuan and Ge, Yingqiang and Li, Yunqi and Fu, Zuohui and Chen, Xu and Zhang, Yongfeng},
title = {Causal Collaborative Filtering},	1260	1260	title = {Causal Collaborative Filtering},
year = {2023},	1261	1261	year = {2023},
isbn = {9798400700736},	1262	1262	isbn = {9798400700736},
publisher = {Association for Computing Machinery},	1263	1263	publisher = {Association for Computing Machinery},
address = {New York, NY, USA},	1264	1264	address = {New York, NY, USA},
url = {https://doi.org/10.1145/3578337.3605122},	1265	1265	url = {https://doi.org/10.1145/3578337.3605122},
doi = {10.1145/3578337.3605122},	1266	1266	doi = {10.1145/3578337.3605122},
abstract = {Many of the traditional recommendation algorithms are designed based on the fundamental idea of mining or learning correlative patterns from data to estimate the user-item correlative preference. However, pure correlative learning may lead to Simpson's paradox in predictions, and thus results in sacrificed recommendation performance. Simpson's paradox is a well-known statistical phenomenon, which causes confusions in statistical conclusions and ignoring the paradox may result in inaccurate decisions. Fortunately, causal and counterfactual modeling can help us to think outside of the observational data for user modeling and personalization so as to tackle such issues. In this paper, we propose Causal Collaborative Filtering (CCF) --- a general framework for modeling causality in collaborative filtering and recommendation. We provide a unified causal view of CF and mathematically show that many of the traditional CF algorithms are actually special cases of CCF under simplified causal graphs. We then propose a conditional intervention approach for do-operations so that we can estimate the user-item causal preference based on the observational data. Finally, we further propose a general counterfactual constrained learning framework for estimating the user-item preferences. Experiments are conducted on two types of real-world datasets---traditional and randomized trial data---and results show that our framework can improve the recommendation performance and reduce the Simpson's paradox problem of many CF algorithms.},	1267	1267	abstract = {Many of the traditional recommendation algorithms are designed based on the fundamental idea of mining or learning correlative patterns from data to estimate the user-item correlative preference. However, pure correlative learning may lead to Simpson's paradox in predictions, and thus results in sacrificed recommendation performance. Simpson's paradox is a well-known statistical phenomenon, which causes confusions in statistical conclusions and ignoring the paradox may result in inaccurate decisions. Fortunately, causal and counterfactual modeling can help us to think outside of the observational data for user modeling and personalization so as to tackle such issues. In this paper, we propose Causal Collaborative Filtering (CCF) --- a general framework for modeling causality in collaborative filtering and recommendation. We provide a unified causal view of CF and mathematically show that many of the traditional CF algorithms are actually special cases of CCF under simplified causal graphs. We then propose a conditional intervention approach for do-operations so that we can estimate the user-item causal preference based on the observational data. Finally, we further propose a general counterfactual constrained learning framework for estimating the user-item preferences. Experiments are conducted on two types of real-world datasets---traditional and randomized trial data---and results show that our framework can improve the recommendation performance and reduce the Simpson's paradox problem of many CF algorithms.},
booktitle = {Proceedings of the 2023 ACM SIGIR International Conference on Theory of Information Retrieval},	1268	1268	booktitle = {Proceedings of the 2023 ACM SIGIR International Conference on Theory of Information Retrieval},
pages = {235–245},	1269	1269	pages = {235–245},
numpages = {11},	1270	1270	numpages = {11},
keywords = {recommender systems, counterfactual reasoning, collaborative filtering, causal analysis, Simpson's paradox},	1271	1271	keywords = {recommender systems, counterfactual reasoning, collaborative filtering, causal analysis, Simpson's paradox},
location = {Taipei, Taiwan},	1272	1272	location = {Taipei, Taiwan},
series = {ICTIR '23}	1273	1273	series = {ICTIR '23}
}	1274	1274	}
	1275	1275
@inproceedings{10.1145/3583780.3615048,	1276	1276	@inproceedings{10.1145/3583780.3615048,
author = {Zhu, Zheqing and Van Roy, Benjamin},	1277	1277	author = {Zhu, Zheqing and Van Roy, Benjamin},
title = {Scalable Neural Contextual Bandit for Recommender Systems},	1278	1278	title = {Scalable Neural Contextual Bandit for Recommender Systems},
year = {2023},	1279	1279	year = {2023},
isbn = {9798400701245},	1280	1280	isbn = {9798400701245},
publisher = {Association for Computing Machinery},	1281	1281	publisher = {Association for Computing Machinery},
address = {New York, NY, USA},	1282	1282	address = {New York, NY, USA},
url = {https://doi.org/10.1145/3583780.3615048},	1283	1283	url = {https://doi.org/10.1145/3583780.3615048},
doi = {10.1145/3583780.3615048},	1284	1284	doi = {10.1145/3583780.3615048},
abstract = {High-quality recommender systems ought to deliver both innovative and relevant content through effective and exploratory interactions with users. Yet, supervised learning-based neural networks, which form the backbone of many existing recommender systems, only leverage recognized user interests, falling short when it comes to efficiently uncovering unknown user preferences. While there has been some progress with neural contextual bandit algorithms towards enabling online exploration through neural networks, their onerous computational demands hinder widespread adoption in real-world recommender systems. In this work, we propose a scalable sample-efficient neural contextual bandit algorithm for recommender systems. To do this, we design an epistemic neural network architecture, Epistemic Neural Recommendation (ENR), that enables Thompson sampling at a large scale. In two distinct large-scale experiments with real-world tasks, ENR significantly boosts click-through rates and user ratings by at least 9\% and 6\% respectively compared to state-of-the-art neural contextual bandit algorithms. Furthermore, it achieves equivalent performance with at least 29\% fewer user interactions compared to the best-performing baseline algorithm. Remarkably, while accomplishing these improvements, ENR demands orders of magnitude fewer computational resources than neural contextual bandit baseline algorithms.},	1285	1285	abstract = {High-quality recommender systems ought to deliver both innovative and relevant content through effective and exploratory interactions with users. Yet, supervised learning-based neural networks, which form the backbone of many existing recommender systems, only leverage recognized user interests, falling short when it comes to efficiently uncovering unknown user preferences. While there has been some progress with neural contextual bandit algorithms towards enabling online exploration through neural networks, their onerous computational demands hinder widespread adoption in real-world recommender systems. In this work, we propose a scalable sample-efficient neural contextual bandit algorithm for recommender systems. To do this, we design an epistemic neural network architecture, Epistemic Neural Recommendation (ENR), that enables Thompson sampling at a large scale. In two distinct large-scale experiments with real-world tasks, ENR significantly boosts click-through rates and user ratings by at least 9\% and 6\% respectively compared to state-of-the-art neural contextual bandit algorithms. Furthermore, it achieves equivalent performance with at least 29\% fewer user interactions compared to the best-performing baseline algorithm. Remarkably, while accomplishing these improvements, ENR demands orders of magnitude fewer computational resources than neural contextual bandit baseline algorithms.},
booktitle = {Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},	1286	1286	booktitle = {Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
pages = {3636–3646},	1287	1287	pages = {3636–3646},
numpages = {11},	1288	1288	numpages = {11},
keywords = {contextual bandits, decision making under uncertainty, exploration vs exploitation, recommender systems, reinforcement learning},	1289	1289	keywords = {contextual bandits, decision making under uncertainty, exploration vs exploitation, recommender systems, reinforcement learning},
location = {Birmingham, United Kingdom},	1290	1290	location = {Birmingham, United Kingdom},
series = {CIKM '23}	1291	1291	series = {CIKM '23}
}	1292	1292	}
	1293	1293
@ARTICLE{10494875,	1294	1294	@ARTICLE{10494875,
author={Ghoorchian, Saeed and Kortukov, Evgenii and Maghsudi, Setareh},	1295	1295	author={Ghoorchian, Saeed and Kortukov, Evgenii and Maghsudi, Setareh},
journal={IEEE Open Journal of Signal Processing},	1296	1296	journal={IEEE Open Journal of Signal Processing},
title={Non-Stationary Linear Bandits With Dimensionality Reduction for Large-Scale Recommender Systems},	1297	1297	title={Non-Stationary Linear Bandits With Dimensionality Reduction for Large-Scale Recommender Systems},
year={2024},	1298	1298	year={2024},
volume={5},	1299	1299	volume={5},
number={},	1300	1300	number={},
pages={548-558},	1301	1301	pages={548-558},
keywords={Vectors;Recommender systems;Decision making;Runtime;Signal processing algorithms;Covariance matrices;Robustness;Decision-making;multi-armed bandit;non-stationary environment;online learning;recommender systems},	1302	1302	keywords={Vectors;Recommender systems;Decision making;Runtime;Signal processing algorithms;Covariance matrices;Robustness;Decision-making;multi-armed bandit;non-stationary environment;online learning;recommender systems},
doi={10.1109/OJSP.2024.3386490}	1303	1303	doi={10.1109/OJSP.2024.3386490}
}	1304	1304	}
	1305	1305
@article{GIANNIKIS2024111752,	1306	1306	@article{GIANNIKIS2024111752,
title = {Reinforcement learning for addressing the cold-user problem in recommender systems},	1307	1307	title = {Reinforcement learning for addressing the cold-user problem in recommender systems},
journal = {Knowledge-Based Systems},	1308	1308	journal = {Knowledge-Based Systems},
volume = {294},	1309	1309	volume = {294},
pages = {111752},	1310	1310	pages = {111752},
year = {2024},	1311	1311	year = {2024},
issn = {0950-7051},	1312	1312	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2024.111752},	1313	1313	doi = {https://doi.org/10.1016/j.knosys.2024.111752},
url = {https://www.sciencedirect.com/science/article/pii/S0950705124003873},	1314	1314	url = {https://www.sciencedirect.com/science/article/pii/S0950705124003873},
author = {Stelios Giannikis and Flavius Frasincar and David Boekestijn},	1315	1315	author = {Stelios Giannikis and Flavius Frasincar and David Boekestijn},
keywords = {Recommender systems, Reinforcement learning, Active learning, Cold-user problem},	1316	1316	keywords = {Recommender systems, Reinforcement learning, Active learning, Cold-user problem},
abstract = {Recommender systems are widely used in webshops because of their ability to provide users with personalized recommendations. However, the cold-user problem (i.e., recommending items to new users) is an important issue many webshops face. With the recent General Data Protection Regulation in Europe, the use of additional user information such as demographics is not possible without the user’s explicit consent. Several techniques have been proposed to solve the cold-user problem. Many of these techniques utilize Active Learning (AL) methods, which let cold users rate items to provide better recommendations for them. In this research, we propose two novel approaches that combine reinforcement learning with AL to elicit the users’ preferences and provide them with personalized recommendations. We compare reinforcement learning approaches that are either AL-based or item-based, where the latter predicts users’ ratings of an item by using their ratings of similar items. Differently than many of the existing approaches, this comparison is made based on implicit user information. Using a large real-world dataset, we show that the item-based strategy is more accurate than the AL-based strategy as well as several existing AL strategies.}	1317	1317	abstract = {Recommender systems are widely used in webshops because of their ability to provide users with personalized recommendations. However, the cold-user problem (i.e., recommending items to new users) is an important issue many webshops face. With the recent General Data Protection Regulation in Europe, the use of additional user information such as demographics is not possible without the user’s explicit consent. Several techniques have been proposed to solve the cold-user problem. Many of these techniques utilize Active Learning (AL) methods, which let cold users rate items to provide better recommendations for them. In this research, we propose two novel approaches that combine reinforcement learning with AL to elicit the users’ preferences and provide them with personalized recommendations. We compare reinforcement learning approaches that are either AL-based or item-based, where the latter predicts users’ ratings of an item by using their ratings of similar items. Differently than many of the existing approaches, this comparison is made based on implicit user information. Using a large real-world dataset, we show that the item-based strategy is more accurate than the AL-based strategy as well as several existing AL strategies.}
}	1318	1318	}
	1319	1319
@article{IFTIKHAR2024121541,	1320	1320	@article{IFTIKHAR2024121541,
title = {A reinforcement learning recommender system using bi-clustering and Markov Decision Process},	1321	1321	title = {A reinforcement learning recommender system using bi-clustering and Markov Decision Process},
journal = {Expert Systems with Applications},	1322	1322	journal = {Expert Systems with Applications},
volume = {237},	1323	1323	volume = {237},
pages = {121541},	1324	1324	pages = {121541},
year = {2024},	1325	1325	year = {2024},
issn = {0957-4174},	1326	1326	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2023.121541},	1327	1327	doi = {https://doi.org/10.1016/j.eswa.2023.121541},
url = {https://www.sciencedirect.com/science/article/pii/S0957417423020432},	1328	1328	url = {https://www.sciencedirect.com/science/article/pii/S0957417423020432},
author = {Arta Iftikhar and Mustansar Ali Ghazanfar and Mubbashir Ayub and Saad {Ali Alahmari} and Nadeem Qazi and Julie Wall},	1329	1329	author = {Arta Iftikhar and Mustansar Ali Ghazanfar and Mubbashir Ayub and Saad {Ali Alahmari} and Nadeem Qazi and Julie Wall},
keywords = {Reinforcement learning, Markov Decision Process, Bi-clustering, Q-learning, Policy},	1330	1330	keywords = {Reinforcement learning, Markov Decision Process, Bi-clustering, Q-learning, Policy},
abstract = {Collaborative filtering (CF) recommender systems are static in nature and does not adapt well with changing user preferences. User preferences may change after interaction with a system or after buying a product. Conventional CF clustering algorithms only identifies the distribution of patterns and hidden correlations globally. However, the impossibility of discovering local patterns by these algorithms, headed to the popularization of bi-clustering algorithms. Bi-clustering algorithms can analyze all dataset dimensions simultaneously and consequently, discover local patterns that deliver a better understanding of the underlying hidden correlations. In this paper, we modelled the recommendation problem as a sequential decision-making problem using Markov Decision Processes (MDP). To perform state representation for MDP, we first converted user-item votings matrix to a binary matrix. Then we performed bi-clustering on this binary matrix to determine a subset of similar rows and columns. A bi-cluster merging algorithm is designed to merge similar and overlapping bi-clusters. These bi-clusters are then mapped to a squared grid (SG). RL is applied on this SG to determine best policy to give recommendation to users. Start state is determined using Improved Triangle Similarity (ITR similarity measure. Reward function is computed as grid state overlapping in terms of users and items in current and prospective next state. A thorough comparative analysis was conducted, encompassing a diverse array of methodologies, including RL-based, pure Collaborative Filtering (CF), and clustering methods. The results demonstrate that our proposed method outperforms its competitors in terms of precision, recall, and optimal policy learning.}	1331	1331	abstract = {Collaborative filtering (CF) recommender systems are static in nature and does not adapt well with changing user preferences. User preferences may change after interaction with a system or after buying a product. Conventional CF clustering algorithms only identifies the distribution of patterns and hidden correlations globally. However, the impossibility of discovering local patterns by these algorithms, headed to the popularization of bi-clustering algorithms. Bi-clustering algorithms can analyze all dataset dimensions simultaneously and consequently, discover local patterns that deliver a better understanding of the underlying hidden correlations. In this paper, we modelled the recommendation problem as a sequential decision-making problem using Markov Decision Processes (MDP). To perform state representation for MDP, we first converted user-item votings matrix to a binary matrix. Then we performed bi-clustering on this binary matrix to determine a subset of similar rows and columns. A bi-cluster merging algorithm is designed to merge similar and overlapping bi-clusters. These bi-clusters are then mapped to a squared grid (SG). RL is applied on this SG to determine best policy to give recommendation to users. Start state is determined using Improved Triangle Similarity (ITR similarity measure. Reward function is computed as grid state overlapping in terms of users and items in current and prospective next state. A thorough comparative analysis was conducted, encompassing a diverse array of methodologies, including RL-based, pure Collaborative Filtering (CF), and clustering methods. The results demonstrate that our proposed method outperforms its competitors in terms of precision, recall, and optimal policy learning.}
}	1332	1332	}
	1333	1333
@article{Soto2,	1334	1334	@article{Soto2,
author={Soto-Forero, Daniel and Ackermann, Simha and Betbeder, Marie-Laure and Henriet, Julien},	1335	1335	author={Soto-Forero, Daniel and Ackermann, Simha and Betbeder, Marie-Laure and Henriet, Julien},
title={Automatic Real-Time Adaptation of Training Session Difficulty Using Rules and Reinforcement Learning in the AI-VT ITS},	1336	1336	title={Automatic Real-Time Adaptation of Training Session Difficulty Using Rules and Reinforcement Learning in the AI-VT ITS},
journal = {International Journal of Modern Education and Computer Science(IJMECS)},	1337	1337	journal = {International Journal of Modern Education and Computer Science(IJMECS)},
volume = {16},	1338	1338	volume = {16},
pages = {56-71},	1339	1339	pages = {56-71},
year = {2024},	1340	1340	year = {2024},
issn = {2075-0161},	1341	1341	issn = {2075-0161},
doi = { https://doi.org/10.5815/ijmecs.2024.03.05},	1342	1342	doi = { https://doi.org/10.5815/ijmecs.2024.03.05},
url = {https://www.mecs-press.org/ijmecs/ijmecs-v16-n3/v16n3-5.html},	1343	1343	url = {https://www.mecs-press.org/ijmecs/ijmecs-v16-n3/v16n3-5.html},
keywords={Real Time Adaptation, Intelligent Training System, Thompson Sampling, Case-Based Reasoning, Automatic Adaptation},	1344	1344	keywords={Real Time Adaptation, Intelligent Training System, Thompson Sampling, Case-Based Reasoning, Automatic Adaptation},
abstract={Some of the most common and typical issues in the field of intelligent tutoring systems (ITS) are (i) the correct identification of learners’ difficulties in the learning process, (ii) the adaptation of content or presentation of the system according to the difficulties encountered, and (iii) the ability to adapt without initial data (cold-start). In some cases, the system tolerates modifications after the realization and assessment of competences. Other systems require complicated real-time adaptation since only a limited number of data can be captured. In that case, it must be analyzed properly and with a certain precision in order to obtain the appropriate adaptations. Generally, for the adaptation step, the ITS gathers common learners together and adapts their training similarly. Another type of adaptation is more personalized, but requires acquired or estimated information about each learner (previous grades, probability of success, etc.). Some of these parameters may be difficult to obtain, and others are imprecise and can lead to misleading adaptations. The adaptation using machine learning requires prior training with a lot of data. This article presents a model for the real time automatic adaptation of a predetermined session inside an ITS called AI-VT. This adaptation process is part of a case-based reasoning global model. The characteristics of the model proposed in this paper (i) require a limited number of data in order to generate a personalized adaptation, (ii) do not require training, (iii) are based on the correlation to complexity levels, and (iv) are able to adapt even at the cold-start stage. The proposed model is presented with two different configurations, deterministic and stochastic. The model has been tested with a database of 1000 learners, corresponding to different knowledge levels in three different scenarios. The results show the dynamic adaptation of the proposed model in both versions, with the adaptations obtained helping the system to evolve more rapidly and identify learner weaknesses in the different levels of complexity as well as the generation of pertinent recommendations in specific cases for each learner capacity.}	1345	1345	abstract={Some of the most common and typical issues in the field of intelligent tutoring systems (ITS) are (i) the correct identification of learners’ difficulties in the learning process, (ii) the adaptation of content or presentation of the system according to the difficulties encountered, and (iii) the ability to adapt without initial data (cold-start). In some cases, the system tolerates modifications after the realization and assessment of competences. Other systems require complicated real-time adaptation since only a limited number of data can be captured. In that case, it must be analyzed properly and with a certain precision in order to obtain the appropriate adaptations. Generally, for the adaptation step, the ITS gathers common learners together and adapts their training similarly. Another type of adaptation is more personalized, but requires acquired or estimated information about each learner (previous grades, probability of success, etc.). Some of these parameters may be difficult to obtain, and others are imprecise and can lead to misleading adaptations. The adaptation using machine learning requires prior training with a lot of data. This article presents a model for the real time automatic adaptation of a predetermined session inside an ITS called AI-VT. This adaptation process is part of a case-based reasoning global model. The characteristics of the model proposed in this paper (i) require a limited number of data in order to generate a personalized adaptation, (ii) do not require training, (iii) are based on the correlation to complexity levels, and (iv) are able to adapt even at the cold-start stage. The proposed model is presented with two different configurations, deterministic and stochastic. The model has been tested with a database of 1000 learners, corresponding to different knowledge levels in three different scenarios. The results show the dynamic adaptation of the proposed model in both versions, with the adaptations obtained helping the system to evolve more rapidly and identify learner weaknesses in the different levels of complexity as well as the generation of pertinent recommendations in specific cases for each learner capacity.}
}	1346	1346	}
	1347	1347
@InProceedings{10.1007/978-3-031-63646-2_11 ,	1348	1348	@InProceedings{10.1007/978-3-031-63646-2_11 ,
author={Soto-Forero, Daniel and Betbeder, Marie-Laure and Henriet, Julien},	1349	1349	author={Soto-Forero, Daniel and Betbeder, Marie-Laure and Henriet, Julien},
editor={Recio-Garcia, Juan A. and Orozco-del-Castillo, Mauricio G. and Bridge, Derek},	1350	1350	editor={Recio-Garcia, Juan A. and Orozco-del-Castillo, Mauricio G. and Bridge, Derek},
title={Ensemble Stacking Case-Based Reasoning for Regression},	1351	1351	title={Ensemble Stacking Case-Based Reasoning for Regression},
booktitle={Case-Based Reasoning Research and Development},	1352	1352	booktitle={Case-Based Reasoning Research and Development},
year={2024},	1353	1353	year={2024},
publisher={Springer Nature Switzerland},	1354	1354	publisher={Springer Nature Switzerland},
address={Cham},	1355	1355	address={Cham},
pages={159--174},	1356	1356	pages={159--174},
abstract={This paper presents a case-based reasoning algorithm with a two-stage iterative double stacking to find approximate solutions to one and multidimensional regression problems. This approach does not require training, so it can work with dynamic data at run time. The solutions are generated using stochastic algorithms in order to allow exploration of the solution space. The evaluation is performed by transforming the regression problem into an optimization problem with an associated objective function. The algorithm has been tested in comparison with nine classical regression algorithms on ten different regression databases extracted from the UCI site. The results show that the proposed algorithm generates solutions in most cases quite close to the real solutions. According to the RMSE, the proposed algorithm globally among the four best algorithms, according to MAE, to the fourth best algorithms of the ten evaluated, suggesting that the results are reasonably good.},	1357	1357	abstract={This paper presents a case-based reasoning algorithm with a two-stage iterative double stacking to find approximate solutions to one and multidimensional regression problems. This approach does not require training, so it can work with dynamic data at run time. The solutions are generated using stochastic algorithms in order to allow exploration of the solution space. The evaluation is performed by transforming the regression problem into an optimization problem with an associated objective function. The algorithm has been tested in comparison with nine classical regression algorithms on ten different regression databases extracted from the UCI site. The results show that the proposed algorithm generates solutions in most cases quite close to the real solutions. According to the RMSE, the proposed algorithm globally among the four best algorithms, according to MAE, to the fourth best algorithms of the ten evaluated, suggesting that the results are reasonably good.},
isbn={978-3-031-63646-2}	1358	1358	isbn={978-3-031-63646-2}
}	1359	1359	}
	1360	1360
@article{ZHANG2018189,	1361	1361	@article{ZHANG2018189,
title = {A three learning states Bayesian knowledge tracing model},	1362	1362	title = {A three learning states Bayesian knowledge tracing model},
journal = {Knowledge-Based Systems},	1363	1363	journal = {Knowledge-Based Systems},
volume = {148},	1364	1364	volume = {148},
pages = {189-201},	1365	1365	pages = {189-201},
year = {2018},	1366	1366	year = {2018},
issn = {0950-7051},	1367	1367	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2018.03.001},	1368	1368	doi = {https://doi.org/10.1016/j.knosys.2018.03.001},
url = {https://www.sciencedirect.com/science/article/pii/S0950705118301199},	1369	1369	url = {https://www.sciencedirect.com/science/article/pii/S0950705118301199},
author = {Kai Zhang and Yiyu Yao},	1370	1370	author = {Kai Zhang and Yiyu Yao},
keywords = {Bayesian knowledge tracing, Three-way decisions},	1371	1371	keywords = {Bayesian knowledge tracing, Three-way decisions},
abstract = {This paper proposes a Bayesian knowledge tracing model with three learning states by extending the original two learning states. We divide a learning process into three sections by using an evaluation function for three-way decisions. Advantages of such a trisection over traditional bisection are demonstrated by comparative experiments. We develop a three learning states model based on the trisection of the learning process. We apply the model to a series of comparative experiments with the original model. Qualitative and quantitative analyses of the experimental results indicate the superior performance of the proposed model over the original model in terms of prediction accuracies and related statistical measures.}	1372	1372	abstract = {This paper proposes a Bayesian knowledge tracing model with three learning states by extending the original two learning states. We divide a learning process into three sections by using an evaluation function for three-way decisions. Advantages of such a trisection over traditional bisection are demonstrated by comparative experiments. We develop a three learning states model based on the trisection of the learning process. We apply the model to a series of comparative experiments with the original model. Qualitative and quantitative analyses of the experimental results indicate the superior performance of the proposed model over the original model in terms of prediction accuracies and related statistical measures.}
}	1373	1373	}
	1374	1374
@article{Li_2024,	1375	1375	@article{Li_2024,
doi = {10.3847/1538-4357/ad3215},	1376	1376	doi = {10.3847/1538-4357/ad3215},
url = {https://dx.doi.org/10.3847/1538-4357/ad3215},	1377	1377	url = {https://dx.doi.org/10.3847/1538-4357/ad3215},
year = {2024},	1378	1378	year = {2024},
month = {apr},	1379	1379	month = {apr},
publisher = {The American Astronomical Society},	1380	1380	publisher = {The American Astronomical Society},
volume = {965},	1381	1381	volume = {965},
number = {2},	1382	1382	number = {2},
pages = {125},	1383	1383	pages = {125},
author = {Zhigang Li and Zhejie Ding and Yu Yu and Pengjie Zhang},	1384	1384	author = {Zhigang Li and Zhejie Ding and Yu Yu and Pengjie Zhang},
title = {The Kullback–Leibler Divergence and the Convergence Rate of Fast Covariance Matrix Estimators in Galaxy Clustering Analysis},	1385	1385	title = {The Kullback–Leibler Divergence and the Convergence Rate of Fast Covariance Matrix Estimators in Galaxy Clustering Analysis},
journal = {The Astrophysical Journal},	1386	1386	journal = {The Astrophysical Journal},
abstract = {We present a method to quantify the convergence rate of the fast estimators of the covariance matrices in the large-scale structure analysis. Our method is based on the Kullback–Leibler (KL) divergence, which describes the relative entropy of two probability distributions. As a case study, we analyze the delete-d jackknife estimator for the covariance matrix of the galaxy correlation function. We introduce the information factor or the normalized KL divergence with the help of a set of baseline covariance matrices to diagnose the information contained in the jackknife covariance matrix. Using a set of quick particle mesh mock catalogs designed for the Baryon Oscillation Spectroscopic Survey DR11 CMASS galaxy survey, we find that the jackknife resampling method succeeds in recovering the covariance matrix with 10 times fewer simulation mocks than that of the baseline method at small scales (s ≤ 40 h −1 Mpc). However, the ability to reduce the number of mock catalogs is degraded at larger scales due to the increasing bias on the jackknife covariance matrix. Note that the analysis in this paper can be applied to any fast estimator of the covariance matrix for galaxy clustering measurements.}	1387	1387	abstract = {We present a method to quantify the convergence rate of the fast estimators of the covariance matrices in the large-scale structure analysis. Our method is based on the Kullback–Leibler (KL) divergence, which describes the relative entropy of two probability distributions. As a case study, we analyze the delete-d jackknife estimator for the covariance matrix of the galaxy correlation function. We introduce the information factor or the normalized KL divergence with the help of a set of baseline covariance matrices to diagnose the information contained in the jackknife covariance matrix. Using a set of quick particle mesh mock catalogs designed for the Baryon Oscillation Spectroscopic Survey DR11 CMASS galaxy survey, we find that the jackknife resampling method succeeds in recovering the covariance matrix with 10 times fewer simulation mocks than that of the baseline method at small scales (s ≤ 40 h −1 Mpc). However, the ability to reduce the number of mock catalogs is degraded at larger scales due to the increasing bias on the jackknife covariance matrix. Note that the analysis in this paper can be applied to any fast estimator of the covariance matrix for galaxy clustering measurements.}
}	1388	1388	}
	1389	1389
@Article{Kim2024,	1390	1390	@Article{Kim2024,
author={Kim, Wonjik},	1391	1391	author={Kim, Wonjik},
title={A Random Focusing Method with Jensen--Shannon Divergence for Improving Deep Neural Network Performance Ensuring Architecture Consistency},	1392	1392	title={A Random Focusing Method with Jensen--Shannon Divergence for Improving Deep Neural Network Performance Ensuring Architecture Consistency},
journal={Neural Processing Letters},	1393	1393	journal={Neural Processing Letters},
year={2024},	1394	1394	year={2024},
month={Jun},	1395	1395	month={Jun},
day={17},	1396	1396	day={17},
volume={56},	1397	1397	volume={56},
number={4},	1398	1398	number={4},
pages={199},	1399	1399	pages={199},
abstract={Multiple hidden layers in deep neural networks perform non-linear transformations, enabling the extraction of meaningful features and the identification of relationships between input and output data. However, the gap between the training and real-world data can result in network overfitting, prompting the exploration of various preventive methods. The regularization technique called 'dropout' is widely used for deep learning models to improve the training of robust and generalized features. During the training phase with dropout, neurons in a particular layer are randomly selected to be ignored for each input. This random exclusion of neurons encourages the network to depend on different subsets of neurons at different times, fostering robustness and reducing sensitivity to specific neurons. This study introduces a novel approach called random focusing, departing from complete neuron exclusion in dropout. The proposed random focusing selectively highlights random neurons during training, aiming for a smoother transition between training and inference phases while keeping network architecture consistent. This study also incorporates Jensen--Shannon Divergence to enhance the stability and efficacy of the random focusing method. Experimental validation across tasks like image classification and semantic segmentation demonstrates the adaptability of the proposed methods across different network architectures, including convolutional neural networks and transformers.},	1400	1400	abstract={Multiple hidden layers in deep neural networks perform non-linear transformations, enabling the extraction of meaningful features and the identification of relationships between input and output data. However, the gap between the training and real-world data can result in network overfitting, prompting the exploration of various preventive methods. The regularization technique called 'dropout' is widely used for deep learning models to improve the training of robust and generalized features. During the training phase with dropout, neurons in a particular layer are randomly selected to be ignored for each input. This random exclusion of neurons encourages the network to depend on different subsets of neurons at different times, fostering robustness and reducing sensitivity to specific neurons. This study introduces a novel approach called random focusing, departing from complete neuron exclusion in dropout. The proposed random focusing selectively highlights random neurons during training, aiming for a smoother transition between training and inference phases while keeping network architecture consistent. This study also incorporates Jensen--Shannon Divergence to enhance the stability and efficacy of the random focusing method. Experimental validation across tasks like image classification and semantic segmentation demonstrates the adaptability of the proposed methods across different network architectures, including convolutional neural networks and transformers.},
issn={1573-773X},	1401	1401	issn={1573-773X},
doi={10.1007/s11063-024-11668-z},	1402	1402	doi={10.1007/s11063-024-11668-z},
url={https://doi.org/10.1007/s11063-024-11668-z}	1403	1403	url={https://doi.org/10.1007/s11063-024-11668-z}
}	1404	1404	}
	1405	1405
@InProceedings{pmlr-v238-ou24a,	1406	1406	@InProceedings{pmlr-v238-ou24a,
title = {Thompson Sampling Itself is Differentially Private},	1407	1407	title = {Thompson Sampling Itself is Differentially Private},
author = {Ou, Tingting and Cummings, Rachel and Avella Medina, Marco},	1408	1408	author = {Ou, Tingting and Cummings, Rachel and Avella Medina, Marco},
booktitle = {Proceedings of The 27th International Conference on Artificial Intelligence and Statistics},	1409	1409	booktitle = {Proceedings of The 27th International Conference on Artificial Intelligence and Statistics},
pages = {1576--1584},	1410	1410	pages = {1576--1584},
year = {2024},	1411	1411	year = {2024},
editor = {Dasgupta, Sanjoy and Mandt, Stephan and Li, Yingzhen},	1412	1412	editor = {Dasgupta, Sanjoy and Mandt, Stephan and Li, Yingzhen},
volume = {238},	1413	1413	volume = {238},
series = {Proceedings of Machine Learning Research},	1414	1414	series = {Proceedings of Machine Learning Research},
month = {02--04 May},	1415	1415	month = {02--04 May},
publisher = {PMLR},	1416	1416	publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v238/ou24a/ou24a.pdf},	1417	1417	pdf = {https://proceedings.mlr.press/v238/ou24a/ou24a.pdf},
url = {https://proceedings.mlr.press/v238/ou24a.html},	1418	1418	url = {https://proceedings.mlr.press/v238/ou24a.html},
abstract = {In this work we first show that the classical Thompson sampling algorithm for multi-arm bandits is differentially private as-is, without any modification. We provide per-round privacy guarantees as a function of problem parameters and show composition over $T$ rounds; since the algorithm is unchanged, existing $O(\sqrt{NT\log N})$ regret bounds still hold and there is no loss in performance due to privacy. We then show that simple modifications – such as pre-pulling all arms a fixed number of times, increasing the sampling variance – can provide tighter privacy guarantees. We again provide privacy guarantees that now depend on the new parameters introduced in the modification, which allows the analyst to tune the privacy guarantee as desired. We also provide a novel regret analysis for this new algorithm, and show how the new parameters also impact expected regret. Finally, we empirically validate and illustrate our theoretical findings in two parameter regimes and demonstrate that tuning the new parameters substantially improve the privacy-regret tradeoff.}	1419	1419	abstract = {In this work we first show that the classical Thompson sampling algorithm for multi-arm bandits is differentially private as-is, without any modification. We provide per-round privacy guarantees as a function of problem parameters and show composition over $T$ rounds; since the algorithm is unchanged, existing $O(\sqrt{NT\log N})$ regret bounds still hold and there is no loss in performance due to privacy. We then show that simple modifications – such as pre-pulling all arms a fixed number of times, increasing the sampling variance – can provide tighter privacy guarantees. We again provide privacy guarantees that now depend on the new parameters introduced in the modification, which allows the analyst to tune the privacy guarantee as desired. We also provide a novel regret analysis for this new algorithm, and show how the new parameters also impact expected regret. Finally, we empirically validate and illustrate our theoretical findings in two parameter regimes and demonstrate that tuning the new parameters substantially improve the privacy-regret tradeoff.}
}	1420	1420	}
	1421	1421
@Article{math12111758,	1422	1422	@Article{math12111758,
AUTHOR = {Uguina, Antonio R. and Gomez, Juan F. and Panadero, Javier and Martínez-Gavara, Anna and Juan, Angel A.},	1423	1423	AUTHOR = {Uguina, Antonio R. and Gomez, Juan F. and Panadero, Javier and Martínez-Gavara, Anna and Juan, Angel A.},
TITLE = {A Learnheuristic Algorithm Based on Thompson Sampling for the Heterogeneous and Dynamic Team Orienteering Problem},	1424	1424	TITLE = {A Learnheuristic Algorithm Based on Thompson Sampling for the Heterogeneous and Dynamic Team Orienteering Problem},
JOURNAL = {Mathematics},	1425	1425	JOURNAL = {Mathematics},
VOLUME = {12},	1426	1426	VOLUME = {12},
YEAR = {2024},	1427	1427	YEAR = {2024},
NUMBER = {11},	1428	1428	NUMBER = {11},
ARTICLE-NUMBER = {1758},	1429	1429	ARTICLE-NUMBER = {1758},
URL = {https://www.mdpi.com/2227-7390/12/11/1758},	1430	1430	URL = {https://www.mdpi.com/2227-7390/12/11/1758},
ISSN = {2227-7390},	1431	1431	ISSN = {2227-7390},
ABSTRACT = {The team orienteering problem (TOP) is a well-studied optimization challenge in the field of Operations Research, where multiple vehicles aim to maximize the total collected rewards within a given time limit by visiting a subset of nodes in a network. With the goal of including dynamic and uncertain conditions inherent in real-world transportation scenarios, we introduce a novel dynamic variant of the TOP that considers real-time changes in environmental conditions affecting reward acquisition at each node. Specifically, we model the dynamic nature of environmental factors—such as traffic congestion, weather conditions, and battery level of each vehicle—to reflect their impact on the probability of obtaining the reward when visiting each type of node in a heterogeneous network. To address this problem, a learnheuristic optimization framework is proposed. It combines a metaheuristic algorithm with Thompson sampling to make informed decisions in dynamic environments. Furthermore, we conduct empirical experiments to assess the impact of varying reward probabilities on resource allocation and route planning within the context of this dynamic TOP, where nodes might offer a different reward behavior depending upon the environmental conditions. Our numerical results indicate that the proposed learnheuristic algorithm outperforms static approaches, achieving up to 25% better performance in highly dynamic scenarios. Our findings highlight the effectiveness of our approach in adapting to dynamic conditions and optimizing decision-making processes in transportation systems.},	1432	1432	ABSTRACT = {The team orienteering problem (TOP) is a well-studied optimization challenge in the field of Operations Research, where multiple vehicles aim to maximize the total collected rewards within a given time limit by visiting a subset of nodes in a network. With the goal of including dynamic and uncertain conditions inherent in real-world transportation scenarios, we introduce a novel dynamic variant of the TOP that considers real-time changes in environmental conditions affecting reward acquisition at each node. Specifically, we model the dynamic nature of environmental factors—such as traffic congestion, weather conditions, and battery level of each vehicle—to reflect their impact on the probability of obtaining the reward when visiting each type of node in a heterogeneous network. To address this problem, a learnheuristic optimization framework is proposed. It combines a metaheuristic algorithm with Thompson sampling to make informed decisions in dynamic environments. Furthermore, we conduct empirical experiments to assess the impact of varying reward probabilities on resource allocation and route planning within the context of this dynamic TOP, where nodes might offer a different reward behavior depending upon the environmental conditions. Our numerical results indicate that the proposed learnheuristic algorithm outperforms static approaches, achieving up to 25% better performance in highly dynamic scenarios. Our findings highlight the effectiveness of our approach in adapting to dynamic conditions and optimizing decision-making processes in transportation systems.},
DOI = {10.3390/math12111758}	1433	1433	DOI = {10.3390/math12111758}
}	1434	1434	}
	1435	1435
@inproceedings{NEURIPS2023_9d8cf124,	1436	1436	@inproceedings{NEURIPS2023_9d8cf124,
author = {Abel, David and Barreto, Andre and Van Roy, Benjamin and Precup, Doina and van Hasselt, Hado P and Singh, Satinder},	1437	1437	author = {Abel, David and Barreto, Andre and Van Roy, Benjamin and Precup, Doina and van Hasselt, Hado P and Singh, Satinder},
booktitle = {Advances in Neural Information Processing Systems},	1438	1438	booktitle = {Advances in Neural Information Processing Systems},
editor = {A. Oh and T. Naumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},	1439	1439	editor = {A. Oh and T. Naumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},
pages = {50377--50407},	1440	1440	pages = {50377--50407},
publisher = {Curran Associates, Inc.},	1441	1441	publisher = {Curran Associates, Inc.},
title = {A Definition of Continual Reinforcement Learning},	1442	1442	title = {A Definition of Continual Reinforcement Learning},
url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/9d8cf1247786d6dfeefeeb53b8b5f6d7-Paper-Conference.pdf},	1443	1443	url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/9d8cf1247786d6dfeefeeb53b8b5f6d7-Paper-Conference.pdf},
volume = {36},	1444	1444	volume = {36},
year = {2023}	1445	1445	year = {2023}
}	1446	1446	}
	1447	1447
@article{NGUYEN2024111566,	1448	1448	@article{NGUYEN2024111566,
title = {Dynamic metaheuristic selection via Thompson Sampling for online optimization},	1449	1449	title = {Dynamic metaheuristic selection via Thompson Sampling for online optimization},
journal = {Applied Soft Computing},	1450	1450	journal = {Applied Soft Computing},
volume = {158},	1451	1451	volume = {158},
pages = {111566},	1452	1452	pages = {111566},
year = {2024},	1453	1453	year = {2024},
issn = {1568-4946},	1454	1454	issn = {1568-4946},
doi = {https://doi.org/10.1016/j.asoc.2024.111566},	1455	1455	doi = {https://doi.org/10.1016/j.asoc.2024.111566},
url = {https://www.sciencedirect.com/science/article/pii/S1568494624003405},	1456	1456	url = {https://www.sciencedirect.com/science/article/pii/S1568494624003405},
author = {Alain Nguyen},	1457	1457	author = {Alain Nguyen},
keywords = {Selection hyper-heuristic, Multi-armed-bandit, Thompson Sampling, Online optimization},	1458	1458	keywords = {Selection hyper-heuristic, Multi-armed-bandit, Thompson Sampling, Online optimization},
abstract = {It is acknowledged that no single heuristic can outperform all the others in every optimization problem. This has given rise to hyper-heuristic methods for providing solutions to a wider range of problems. In this work, a set of five non-competing low-level heuristics is proposed in a hyper-heuristic framework. The multi-armed bandit problem analogy is efficiently leveraged and Thompson Sampling is used to actively select the best heuristic for online optimization. The proposed method is compared against ten population-based metaheuristic algorithms on the well-known CEC’05 optimizing benchmark consisting of 23 functions of various landscapes. The results show that the proposed algorithm is the only one able to find the global minimum of all functions with remarkable consistency.}	1459	1459	abstract = {It is acknowledged that no single heuristic can outperform all the others in every optimization problem. This has given rise to hyper-heuristic methods for providing solutions to a wider range of problems. In this work, a set of five non-competing low-level heuristics is proposed in a hyper-heuristic framework. The multi-armed bandit problem analogy is efficiently leveraged and Thompson Sampling is used to actively select the best heuristic for online optimization. The proposed method is compared against ten population-based metaheuristic algorithms on the well-known CEC’05 optimizing benchmark consisting of 23 functions of various landscapes. The results show that the proposed algorithm is the only one able to find the global minimum of all functions with remarkable consistency.}
}	1460	1460	}
	1461	1461
@Article{Malladi2024,	1462	1462	@Article{Malladi2024,
author={Malladi, Rama K.},	1463	1463	author={Malladi, Rama K.},
title={Application of Supervised Machine Learning Techniques to Forecast the COVID-19 U.S. Recession and Stock Market Crash},	1464	1464	title={Application of Supervised Machine Learning Techniques to Forecast the COVID-19 U.S. Recession and Stock Market Crash},
journal={Computational Economics},	1465	1465	journal={Computational Economics},
year={2024},	1466	1466	year={2024},
month={Mar},	1467	1467	month={Mar},
day={01},	1468	1468	day={01},
volume={63},	1469	1469	volume={63},
number={3},	1470	1470	number={3},
pages={1021-1045},	1471	1471	pages={1021-1045},
abstract={Machine learning (ML), a transformational technology, has been successfully applied to forecasting events down the road. This paper demonstrates that supervised ML techniques can be used in recession and stock market crash (more than 20{\%} drawdown) forecasting. After learning from strictly past monthly data, ML algorithms detected the Covid-19 recession by December 2019, six months before the official NBER announcement. Moreover, ML algorithms foresaw the March 2020 S{\&}P500 crash two months before it happened. The current labor market and housing are harbingers of a future U.S. recession (in 3 months). Financial factors have a bigger role to play in stock market crashes than economic factors. The labor market appears as a top-two feature in predicting both recessions and crashes. ML algorithms detect that the U.S. exited recession before December 2020, even though the official NBER announcement has not yet been made. They also do not anticipate a U.S. stock market crash before March 2021. ML methods have three times higher false discovery rates of recessions compared to crashes.},	1472	1472	abstract={Machine learning (ML), a transformational technology, has been successfully applied to forecasting events down the road. This paper demonstrates that supervised ML techniques can be used in recession and stock market crash (more than 20{\%} drawdown) forecasting. After learning from strictly past monthly data, ML algorithms detected the Covid-19 recession by December 2019, six months before the official NBER announcement. Moreover, ML algorithms foresaw the March 2020 S{\&}P500 crash two months before it happened. The current labor market and housing are harbingers of a future U.S. recession (in 3 months). Financial factors have a bigger role to play in stock market crashes than economic factors. The labor market appears as a top-two feature in predicting both recessions and crashes. ML algorithms detect that the U.S. exited recession before December 2020, even though the official NBER announcement has not yet been made. They also do not anticipate a U.S. stock market crash before March 2021. ML methods have three times higher false discovery rates of recessions compared to crashes.},
issn={1572-9974},	1473	1473	issn={1572-9974},
doi={10.1007/s10614-022-10333-8},	1474	1474	doi={10.1007/s10614-022-10333-8},
url={https://doi.org/10.1007/s10614-022-10333-8}	1475	1475	url={https://doi.org/10.1007/s10614-022-10333-8}
}	1476	1476	}
	1477	1477
@INPROCEEDINGS{10493943,	1478	1478	@INPROCEEDINGS{10493943,
author={Raaa Subha and Naaa Gayathri and Saaa Sasireka and Raaa Sathiyabanu and Baaa Santhiyaa and Baaa Varshini},	1479	1479	author={Raaa Subha and Naaa Gayathri and Saaa Sasireka and Raaa Sathiyabanu and Baaa Santhiyaa and Baaa Varshini},
booktitle={2024 5th International Conference on Mobile Computing and Sustainable Informatics (ICMCSI)},	1480	1480	booktitle={2024 5th International Conference on Mobile Computing and Sustainable Informatics (ICMCSI)},
title={Intelligent Tutoring Systems using Long Short-Term Memory Networks and Bayesian Knowledge Tracing},	1481	1481	title={Intelligent Tutoring Systems using Long Short-Term Memory Networks and Bayesian Knowledge Tracing},
year={2024},	1482	1482	year={2024},
volume={0},	1483	1483	volume={0},
number={0},	1484	1484	number={0},
pages={24-29},	1485	1485	pages={24-29},
keywords={Knowledge engineering;Filtering;Estimation;Transforms;Real-time systems;Bayes methods;Problem-solving;Intelligent Tutoring System (ITS);Long Short-Term Memory (LSTM);Bayesian Knowledge Tracing (BKT);Reinforcement Learning},	1486	1486	keywords={Knowledge engineering;Filtering;Estimation;Transforms;Real-time systems;Bayes methods;Problem-solving;Intelligent Tutoring System (ITS);Long Short-Term Memory (LSTM);Bayesian Knowledge Tracing (BKT);Reinforcement Learning},
doi={10.1109/ICMCSI61536.2024.00010}	1487	1487	doi={10.1109/ICMCSI61536.2024.00010}
}	1488	1488	}
	1489	1489
@article{https://doi.org/10.1155/2024/4067721,	1490	1490	@article{https://doi.org/10.1155/2024/4067721,
author = {Ahmed, Esmael},	1491	1491	author = {Ahmed, Esmael},
title = {Student Performance Prediction Using Machine Learning Algorithms},	1492	1492	title = {Student Performance Prediction Using Machine Learning Algorithms},
journal = {Applied Computational Intelligence and Soft Computing},	1493	1493	journal = {Applied Computational Intelligence and Soft Computing},
volume = {2024},	1494	1494	volume = {2024},
number = {1},	1495	1495	number = {1},
pages = {4067721},	1496	1496	pages = {4067721},
doi = {https://doi.org/10.1155/2024/4067721},	1497	1497	doi = {https://doi.org/10.1155/2024/4067721},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1155/2024/4067721},	1498	1498	url = {https://onlinelibrary.wiley.com/doi/abs/10.1155/2024/4067721},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1155/2024/4067721},	1499	1499	eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1155/2024/4067721},
abstract = {Education is crucial for a productive life and providing necessary resources. With the advent of technology like artificial intelligence, higher education institutions are incorporating technology into traditional teaching methods. Predicting academic success has gained interest in education as a strong academic record improves a university’s ranking and increases student employment opportunities. Modern learning institutions face challenges in analyzing performance, providing high-quality education, formulating strategies for evaluating students’ performance, and identifying future needs. E-learning is a rapidly growing and advanced form of education, where students enroll in online courses. Platforms like Intelligent Tutoring Systems (ITS), learning management systems (LMS), and massive open online courses (MOOC) use educational data mining (EDM) to develop automatic grading systems, recommenders, and adaptative systems. However, e-learning is still considered a challenging learning environment due to the lack of direct interaction between students and course instructors. Machine learning (ML) is used in developing adaptive intelligent systems that can perform complex tasks beyond human abilities. Some areas of applications of ML algorithms include cluster analysis, pattern recognition, image processing, natural language processing, and medical diagnostics. In this research work, K-means, a clustering data mining technique using Davies’ Bouldin method, obtains clusters to find important features affecting students’ performance. The study found that the SVM algorithm had the best prediction results after parameter adjustment, with a 96\% accuracy rate. In this paper, the researchers have examined the functions of the Support Vector Machine, Decision Tree, naive Bayes, and KNN classifiers. The outcomes of parameter adjustment greatly increased the accuracy of the four prediction models. Naïve Bayes model’s prediction accuracy is the lowest when compared to other prediction methods, as it assumes a strong independent relationship between features.},	1500	1500	abstract = {Education is crucial for a productive life and providing necessary resources. With the advent of technology like artificial intelligence, higher education institutions are incorporating technology into traditional teaching methods. Predicting academic success has gained interest in education as a strong academic record improves a university’s ranking and increases student employment opportunities. Modern learning institutions face challenges in analyzing performance, providing high-quality education, formulating strategies for evaluating students’ performance, and identifying future needs. E-learning is a rapidly growing and advanced form of education, where students enroll in online courses. Platforms like Intelligent Tutoring Systems (ITS), learning management systems (LMS), and massive open online courses (MOOC) use educational data mining (EDM) to develop automatic grading systems, recommenders, and adaptative systems. However, e-learning is still considered a challenging learning environment due to the lack of direct interaction between students and course instructors. Machine learning (ML) is used in developing adaptive intelligent systems that can perform complex tasks beyond human abilities. Some areas of applications of ML algorithms include cluster analysis, pattern recognition, image processing, natural language processing, and medical diagnostics. In this research work, K-means, a clustering data mining technique using Davies’ Bouldin method, obtains clusters to find important features affecting students’ performance. The study found that the SVM algorithm had the best prediction results after parameter adjustment, with a 96\% accuracy rate. In this paper, the researchers have examined the functions of the Support Vector Machine, Decision Tree, naive Bayes, and KNN classifiers. The outcomes of parameter adjustment greatly increased the accuracy of the four prediction models. Naïve Bayes model’s prediction accuracy is the lowest when compared to other prediction methods, as it assumes a strong independent relationship between features.},
year = {2024}	1501	1501	year = {2024}
}	1502	1502	}
	1503	1503
@article{HAZEM,	1504	1504	@article{HAZEM,
author = {Hazem A. Alrakhawi and Nurullizam Jamiat and Samy S. Abu-Naser},	1505	1505	author = {Hazem A. Alrakhawi and Nurullizam Jamiat and Samy S. Abu-Naser},
title = {Intelligent Tutoring Systems in education: A systematic review of usage, tools, effects and evaluation},	1506	1506	title = {Intelligent Tutoring Systems in education: A systematic review of usage, tools, effects and evaluation},
journal = {Journal of Theoretical and Applied Information Technology},	1507	1507	journal = {Journal of Theoretical and Applied Information Technology},
volume = {2023},	1508	1508	volume = {2023},
number = {4},	1509	1509	number = {4},
pages = {4067721},	1510	1510	pages = {4067721},
doi = {},	1511	1511	doi = {},
url = {},	1512	1512	url = {},
abstract = {},	1513	1513	abstract = {},
year = {2023}	1514	1514	year = {2023}
}	1515	1515	}
	1516	1516
@Article{Liu2023,	1517	1517	@Article{Liu2023,
author={Liu, Mengchi	1518	1518	author={Liu, Mengchi
and Yu, Dongmei},	1519	1519	and Yu, Dongmei},
title={Towards intelligent E-learning systems},	1520	1520	title={Towards intelligent E-learning systems},
journal={Education and Information Technologies},	1521	1521	journal={Education and Information Technologies},
year={2023},	1522	1522	year={2023},
month={Jul},	1523	1523	month={Jul},
day={01},	1524	1524	day={01},
volume={28},	1525	1525	volume={28},
number={7},	1526	1526	number={7},
pages={7845-7876},	1527	1527	pages={7845-7876},
abstract={The prevalence of e-learning systems has made educational resources more accessible, interactive and effective to learners without the geographic and temporal boundaries. However, as the number of users increases and the volume of data grows, current e-learning systems face some technical and pedagogical challenges. This paper provides a comprehensive review on the efforts of applying new information and communication technologies to improve e-learning services. We first systematically investigate current e-learning systems in terms of their classification, architecture, functions, challenges, and current trends. We then present a general architecture for big data based e-learning systems to meet the ever-growing demand for e-learning. We also describe how to use data generated in big data based e-learning systems to support more flexible and customized course delivery and personalized learning.},	1528	1528	abstract={The prevalence of e-learning systems has made educational resources more accessible, interactive and effective to learners without the geographic and temporal boundaries. However, as the number of users increases and the volume of data grows, current e-learning systems face some technical and pedagogical challenges. This paper provides a comprehensive review on the efforts of applying new information and communication technologies to improve e-learning services. We first systematically investigate current e-learning systems in terms of their classification, architecture, functions, challenges, and current trends. We then present a general architecture for big data based e-learning systems to meet the ever-growing demand for e-learning. We also describe how to use data generated in big data based e-learning systems to support more flexible and customized course delivery and personalized learning.},
issn={1573-7608},	1529	1529	issn={1573-7608},
doi={10.1007/s10639-022-11479-6},	1530	1530	doi={10.1007/s10639-022-11479-6},
url={https://doi.org/10.1007/s10639-022-11479-6}	1531	1531	url={https://doi.org/10.1007/s10639-022-11479-6}
}	1532	1532	}
	1533	1533
@InProceedings{10.1007/978-3-031-63646-2_13,	1534	1534	@InProceedings{10.1007/978-3-031-63646-2_13,
author="Soto-Forero, Daniel	1535	1535	author="Soto-Forero, Daniel
and Ackermann, Simha	1536	1536	and Ackermann, Simha
and Betbeder, Marie-Laure	1537	1537	and Betbeder, Marie-Laure
and Henriet, Julien",	1538	1538	and Henriet, Julien",
editor="Recio-Garcia, Juan A.	1539	1539	editor="Recio-Garcia, Juan A.
and Orozco-del-Castillo, Mauricio G.	1540	1540	and Orozco-del-Castillo, Mauricio G.
and Bridge, Derek",	1541	1541	and Bridge, Derek",
title="The Intelligent Tutoring System AI-VT with Case-Based Reasoning and Real Time Recommender Models",	1542	1542	title="The Intelligent Tutoring System AI-VT with Case-Based Reasoning and Real Time Recommender Models",
booktitle="Case-Based Reasoning Research and Development",	1543	1543	booktitle="Case-Based Reasoning Research and Development",
year="2024",	1544	1544	year="2024",
publisher="Springer Nature Switzerland",	1545	1545	publisher="Springer Nature Switzerland",
address="Cham",	1546	1546	address="Cham",
pages="191--205",	1547	1547	pages="191--205",
abstract="This paper presents a recommendation model coupled on an existing CBR system model through a new modular architecture designed to integrate multiple services in a learning system called AI-VT (Artificial Intelligence Training System). The recommendation model provides a semi-automatic review of the CBR, two variants of the recommendation model have been implemented: deterministic and stochastic. The model has been tested with 1000 simulated learners, and compared with an original CBR system and BKT (Bayesian Knowledge Tracing) recommender system. The results show that the proposed model identifies learners' weaknesses correctly and revises the content of the ITS (Intelligent Tutoring System) better than the original ITS with CBR. Compared to BKT, the results at each level of complexity are variable, but overall the proposed stochastic model obtains better results.",	1548	1548	abstract="This paper presents a recommendation model coupled on an existing CBR system model through a new modular architecture designed to integrate multiple services in a learning system called AI-VT (Artificial Intelligence Training System). The recommendation model provides a semi-automatic review of the CBR, two variants of the recommendation model have been implemented: deterministic and stochastic. The model has been tested with 1000 simulated learners, and compared with an original CBR system and BKT (Bayesian Knowledge Tracing) recommender system. The results show that the proposed model identifies learners' weaknesses correctly and revises the content of the ITS (Intelligent Tutoring System) better than the original ITS with CBR. Compared to BKT, the results at each level of complexity are variable, but overall the proposed stochastic model obtains better results.",
isbn="978-3-031-63646-2"	1549	1549	isbn="978-3-031-63646-2"
}	1550	1550	}
	1551	1551
@article{doi:10.1137/23M1592420,	1552	1552	@article{doi:10.1137/23M1592420,
author = {Minsker, Stanislav and Strawn, Nate},	1553	1553	author = {Minsker, Stanislav and Strawn, Nate},
title = {The Geometric Median and Applications to Robust Mean Estimation},	1554	1554	title = {The Geometric Median and Applications to Robust Mean Estimation},
journal = {SIAM Journal on Mathematics of Data Science},	1555	1555	journal = {SIAM Journal on Mathematics of Data Science},
volume = {6},	1556	1556	volume = {6},
number = {2},	1557	1557	number = {2},
pages = {504-533},	1558	1558	pages = {504-533},
year = {2024},	1559	1559	year = {2024},
doi = {10.1137/23M1592420},	1560	1560	doi = {10.1137/23M1592420},
URL = { https://doi.org/10.1137/23M1592420},	1561	1561	URL = { https://doi.org/10.1137/23M1592420},
eprint = {https://doi.org/10.1137/23M1592420},	1562	1562	eprint = {https://doi.org/10.1137/23M1592420},
abstract = { Abstract.This paper is devoted to the statistical and numerical properties of the geometric median and its applications to the problem of robust mean estimation via the median of means principle. Our main theoretical results include (a) an upper bound for the distance between the mean and the median for general absolutely continuous distributions in $\mathbb R^d$, and examples of specific classes of distributions for which these bounds do not depend on the ambient dimension $d$; (b) exponential deviation inequalities for the distance between the sample and the population versions of the geometric median, which again depend only on the trace-type quantities and not on the ambient dimension. As a corollary, we deduce improved bounds for the (geometric) median of means estimator that hold for large classes of heavy-tailed distributions. Finally, we address the error of numerical approximation, which is an important practical aspect of any statistical estimation procedure. We demonstrate that the objective function minimized by the geometric median satisfies a “local quadratic growth” condition that allows one to translate suboptimality bounds for the objective function to the corresponding bounds for the numerical approximation to the median itself and propose a simple stopping rule applicable to any optimization method which yields explicit error guarantees. We conclude with the numerical experiments, including the application to estimation of mean values of log-returns for S\&P 500 data. }	1563	1563	abstract = { Abstract.This paper is devoted to the statistical and numerical properties of the geometric median and its applications to the problem of robust mean estimation via the median of means principle. Our main theoretical results include (a) an upper bound for the distance between the mean and the median for general absolutely continuous distributions in $\mathbb R^d$, and examples of specific classes of distributions for which these bounds do not depend on the ambient dimension $d$; (b) exponential deviation inequalities for the distance between the sample and the population versions of the geometric median, which again depend only on the trace-type quantities and not on the ambient dimension. As a corollary, we deduce improved bounds for the (geometric) median of means estimator that hold for large classes of heavy-tailed distributions. Finally, we address the error of numerical approximation, which is an important practical aspect of any statistical estimation procedure. We demonstrate that the objective function minimized by the geometric median satisfies a “local quadratic growth” condition that allows one to translate suboptimality bounds for the objective function to the corresponding bounds for the numerical approximation to the median itself and propose a simple stopping rule applicable to any optimization method which yields explicit error guarantees. We conclude with the numerical experiments, including the application to estimation of mean values of log-returns for S\&P 500 data. }
}	1564	1564	}
	1565	1565
@article{lei2024analysis,	1566	1566	@article{lei2024analysis,
title={Analysis of Simpson’s Paradox and Its Applications},	1567	1567	title={Analysis of Simpson’s Paradox and Its Applications},
author={Lei, Zhihao},	1568	1568	author={Lei, Zhihao},
journal={Highlights in Science, Engineering and Technology},	1569	1569	journal={Highlights in Science, Engineering and Technology},
volume={88},	1570	1570	volume={88},
pages={357--362},	1571	1571	pages={357--362},
year={2024}	1572	1572	year={2024}
}	1573	1573	}
	1574	1574
@InProceedings{pmlr-v108-seznec20a,	1575	1575	@InProceedings{pmlr-v108-seznec20a,
title = {A single algorithm for both restless and rested rotting bandits},	1576	1576	title = {A single algorithm for both restless and rested rotting bandits},
author = {Seznec, Julien and Menard, Pierre and Lazaric, Alessandro and Valko, Michal},	1577	1577	author = {Seznec, Julien and Menard, Pierre and Lazaric, Alessandro and Valko, Michal},
booktitle = {Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics},	1578	1578	booktitle = {Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics},
pages = {3784--3794},	1579	1579	pages = {3784--3794},
year = {2020},	1580	1580	year = {2020},
editor = {Chiappa, Silvia and Calandra, Roberto},	1581	1581	editor = {Chiappa, Silvia and Calandra, Roberto},
volume = {108},	1582	1582	volume = {108},
series = {Proceedings of Machine Learning Research},	1583	1583	series = {Proceedings of Machine Learning Research},
month = {26--28 Aug},	1584	1584	month = {26--28 Aug},
publisher = {PMLR},	1585	1585	publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v108/seznec20a/seznec20a.pdf},	1586	1586	pdf = {http://proceedings.mlr.press/v108/seznec20a/seznec20a.pdf},
url = {https://proceedings.mlr.press/v108/seznec20a.html},	1587	1587	url = {https://proceedings.mlr.press/v108/seznec20a.html},
abstract = {In many application domains (e.g., recommender systems, intelligent tutoring systems), the rewards associated to the available actions tend to decrease over time. This decay is either caused by the actions executed in the past (e.g., a user may get bored when songs of the same genre are recommended over and over) or by an external factor (e.g., content becomes outdated). These two situations can be modeled as specific instances of the rested and restless bandit settings, where arms are rotting (i.e., their value decrease over time). These problems were thought to be significantly different, since Levine et al. (2017) showed that state-of-the-art algorithms for restless bandit perform poorly in the rested rotting setting. In this paper, we introduce a novel algorithm, Rotting Adaptive Window UCB (RAW-UCB), that achieves near-optimal regret in both rotting rested and restless bandit, without any prior knowledge of the setting (rested or restless) and the type of non-stationarity (e.g., piece-wise constant, bounded variation). This is in striking contrast with previous negative results showing that no algorithm can achieve similar results as soon as rewards are allowed to increase. We confirm our theoretical findings on a number of synthetic and dataset-based experiments.}	1588	1588	abstract = {In many application domains (e.g., recommender systems, intelligent tutoring systems), the rewards associated to the available actions tend to decrease over time. This decay is either caused by the actions executed in the past (e.g., a user may get bored when songs of the same genre are recommended over and over) or by an external factor (e.g., content becomes outdated). These two situations can be modeled as specific instances of the rested and restless bandit settings, where arms are rotting (i.e., their value decrease over time). These problems were thought to be significantly different, since Levine et al. (2017) showed that state-of-the-art algorithms for restless bandit perform poorly in the rested rotting setting. In this paper, we introduce a novel algorithm, Rotting Adaptive Window UCB (RAW-UCB), that achieves near-optimal regret in both rotting rested and restless bandit, without any prior knowledge of the setting (rested or restless) and the type of non-stationarity (e.g., piece-wise constant, bounded variation). This is in striking contrast with previous negative results showing that no algorithm can achieve similar results as soon as rewards are allowed to increase. We confirm our theoretical findings on a number of synthetic and dataset-based experiments.}
}	1589	1589	}
	1590	1590
@article{doi:10.3233/AIC-1994-7104,	1591	1591	@article{doi:10.3233/AIC-1994-7104,
author = {Agnar Aamodt and Enric Plaza},	1592	1592	author = {Agnar Aamodt and Enric Plaza},
title ={Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches},	1593	1593	title ={Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches},
journal = {AI Communications},	1594	1594	journal = {AI Communications},
volume = {7},	1595	1595	volume = {7},
number = {1},	1596	1596	number = {1},
pages = {39-59},	1597	1597	pages = {39-59},
year = {1994},	1598	1598	year = {1994},
doi = {10.3233/AIC-1994-7104},	1599	1599	doi = {10.3233/AIC-1994-7104},
URL = {	1600	1600	URL = {
https://journals.sagepub.com/doi/abs/10.3233/AIC-1994-7104	1601	1601	https://journals.sagepub.com/doi/abs/10.3233/AIC-1994-7104
},	1602	1602	},
eprint = {	1603	1603	eprint = {
https://journals.sagepub.com/doi/pdf/10.3233/AIC-1994-7104	1604	1604	https://journals.sagepub.com/doi/pdf/10.3233/AIC-1994-7104
},	1605	1605	},
abstract = { Case-based reasoning is a recent approach to problem solving and learning that has got a lot of attention over the last few years. Originating in the US, the basic idea and underlying theories have spread to other continents, and we are now within a period of highly active research in case-based reasoning in Europe as well. This paper gives an overview of the foundational issues related to case-based reasoning, describes some of the leading methodological approaches within the field, and exemplifies the current state through pointers to some systems. Initially, a general framework is defined, to which the subsequent descriptions and discussions will refer. The framework is influenced by recent methodologies for knowledge level descriptions of intelligent systems. The methods for case retrieval, reuse, solution testing, and learning are summarized, and their actual realization is discussed in the light of a few example systems that represent different CBR approaches. We also discuss the role of case-based methods as one type of reasoning and learning method within an integrated system architecture. }	1606	1606	abstract = { Case-based reasoning is a recent approach to problem solving and learning that has got a lot of attention over the last few years. Originating in the US, the basic idea and underlying theories have spread to other continents, and we are now within a period of highly active research in case-based reasoning in Europe as well. This paper gives an overview of the foundational issues related to case-based reasoning, describes some of the leading methodological approaches within the field, and exemplifies the current state through pointers to some systems. Initially, a general framework is defined, to which the subsequent descriptions and discussions will refer. The framework is influenced by recent methodologies for knowledge level descriptions of intelligent systems. The methods for case retrieval, reuse, solution testing, and learning are summarized, and their actual realization is discussed in the light of a few example systems that represent different CBR approaches. We also discuss the role of case-based methods as one type of reasoning and learning method within an integrated system architecture. }
}	1607	1607	}
	1608	1608
@Book{schank+abelson77,	1609	1609	@Book{schank+abelson77,
author = {Roger C. Schank and Robert P. Abelson},	1610	1610	author = {Roger C. Schank and Robert P. Abelson},
title = {Scripts, Plans, Goals and Understanding: an Inquiry into Human Knowledge Structures},	1611	1611	title = {Scripts, Plans, Goals and Understanding: an Inquiry into Human Knowledge Structures},
publisher = {L. Erlbaum},	1612	1612	publisher = {L. Erlbaum},
year = {1977},	1613	1613	year = {1977},
address = {Hillsdale, NJ},	1614	1614	address = {Hillsdale, NJ},
keywords = {PAM, SAM, TALE-SPIN, causality, conceptual dependency, goals, plans, scripts, semantic primitive, text understanding}	1615	1615	keywords = {PAM, SAM, TALE-SPIN, causality, conceptual dependency, goals, plans, scripts, semantic primitive, text understanding}
}	1616	1616	}
	1617	1617
@article{KOLODNER1983281,	1618	1618	@article{KOLODNER1983281,
title = {Reconstructive memory: A computer model},	1619	1619	title = {Reconstructive memory: A computer model},
journal = {Cognitive Science},	1620	1620	journal = {Cognitive Science},
volume = {7},	1621	1621	volume = {7},
number = {4},	1622	1622	number = {4},
pages = {281-328},	1623	1623	pages = {281-328},
year = {1983},	1624	1624	year = {1983},
issn = {0364-0213},	1625	1625	issn = {0364-0213},
doi = {https://doi.org/10.1016/S0364-0213(83)80002-0},	1626	1626	doi = {https://doi.org/10.1016/S0364-0213(83)80002-0},
url = {https://www.sciencedirect.com/science/article/pii/S0364021383800020},	1627	1627	url = {https://www.sciencedirect.com/science/article/pii/S0364021383800020},
author = {Janet L. Kolodner},	1628	1628	author = {Janet L. Kolodner},
abstract = {This study presents a process model of very long-term episodic memory. The process presented is a reconstructive process. The process involves application of three kinds of reconstructive strategies—component-to-context instantiation strategies, component-instantiation strategies, and context-to-context instantiation strategies. The first is used to direct search to appropriate conceptual categories in memory. The other two are used to direct search within the chosen conceptual category. A fourth type of strategy, called executive search strategies, guide search for concepts related to the one targeted for retrieval. A conceptual memory organization implied by human reconstructive memory is presented along with examples which motivate it. A basic retrieval algorithm is presented for traversing that stucture. Retrieval strategies arise from failures in that algorithm. The memory organization and retrieval processes are implemented in a computer program called CYRUS which stores events in the lives of former Secretaries of State Cyrus Vance and Edmund Muskie and answers questions posed in English concerning that information. Examples which motivate the process model are drawn from protocols of human memory search. Examples of CYRUS'S behavior demonstrate the implemented process model. Conclusions are drawn concerning retrieval failures and the relationship of episodic and semantic memory.}	1629	1629	abstract = {This study presents a process model of very long-term episodic memory. The process presented is a reconstructive process. The process involves application of three kinds of reconstructive strategies—component-to-context instantiation strategies, component-instantiation strategies, and context-to-context instantiation strategies. The first is used to direct search to appropriate conceptual categories in memory. The other two are used to direct search within the chosen conceptual category. A fourth type of strategy, called executive search strategies, guide search for concepts related to the one targeted for retrieval. A conceptual memory organization implied by human reconstructive memory is presented along with examples which motivate it. A basic retrieval algorithm is presented for traversing that stucture. Retrieval strategies arise from failures in that algorithm. The memory organization and retrieval processes are implemented in a computer program called CYRUS which stores events in the lives of former Secretaries of State Cyrus Vance and Edmund Muskie and answers questions posed in English concerning that information. Examples which motivate the process model are drawn from protocols of human memory search. Examples of CYRUS'S behavior demonstrate the implemented process model. Conclusions are drawn concerning retrieval failures and the relationship of episodic and semantic memory.}
}	1630	1630	}
	1631	1631
@Book{Riesbeck1989,	1632	1632	@Book{Riesbeck1989,
author = {Riesbeck C.K. and Schank R.C.},	1633	1633	author = {Riesbeck C.K. and Schank R.C.},
year = {1989},	1634	1634	year = {1989},
title = {Inside Case-Based Reasoning},	1635	1635	title = {Inside Case-Based Reasoning},
publisher = {Psychology Press},	1636	1636	publisher = {Psychology Press},
url = {https://doi.org/10.4324/9780203781821}	1637	1637	url = {https://doi.org/10.4324/9780203781821}
}	1638	1638	}
	1639	1639
@article{ALABDULRAHMAN2021114061,	1640	1640	@article{ALABDULRAHMAN2021114061,
title = {Catering for unique tastes: Targeting grey-sheep users recommender systems through one-class machine learning},	1641	1641	title = {Catering for unique tastes: Targeting grey-sheep users recommender systems through one-class machine learning},
journal = {Expert Systems with Applications},	1642	1642	journal = {Expert Systems with Applications},
volume = {166},	1643	1643	volume = {166},
pages = {114061},	1644	1644	pages = {114061},
year = {2021},	1645	1645	year = {2021},
issn = {0957-4174},	1646	1646	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2020.114061},	1647	1647	doi = {https://doi.org/10.1016/j.eswa.2020.114061},
url = {https://www.sciencedirect.com/science/article/pii/S0957417420308241},	1648	1648	url = {https://www.sciencedirect.com/science/article/pii/S0957417420308241},
author = {Rabaa Alabdulrahman and Herna Viktor},	1649	1649	author = {Rabaa Alabdulrahman and Herna Viktor},
keywords = {Recommender systems, Model-based systems, Machine learning, Grey-sheep, One-class classification},	1650	1650	keywords = {Recommender systems, Model-based systems, Machine learning, Grey-sheep, One-class classification},
abstract = {In recommendation systems, the grey-sheep problem refers to users with unique preferences and tastes that make it difficult to develop accurate profiles. That is, the similarity search approach typically followed during the recommendation process fails to yield good results. Most research does not focus on such users and thus fails to cater to more exotic tastes and emerging trends, leading to a subsequent loss in revenue and marketing opportunities. One suggested solution is to use one-class classification to generate a prediction list for these users, where decision boundaries are learned that distinguish between normal and grey-sheep users. In this paper, we present the grey-sheep one-class recommendation (GSOR) framework designed to create accurate prediction models while taking both regular and grey-sheep users into account. In addition, we introduce a novel grey-sheep movie recommendation benchmark to be used by current and future researchers. When evaluating our GSOR framework against this benchmark, our results indicate the value of combining cluster analysis, outlier detection, and one-class learning to generate relevant and timely recommendation lists from data sets that contain grey-sheep users. Specifically, by employing one-class decision tree algorithms, our GSOR framework was able to outperform traditional collaborative filtering-based recommendation systems in both accuracy and model construction time. Furthermore, we report that having grey-sheep users in the system often had a positive impact on the learning and recommendation processes.}	1651	1651	abstract = {In recommendation systems, the grey-sheep problem refers to users with unique preferences and tastes that make it difficult to develop accurate profiles. That is, the similarity search approach typically followed during the recommendation process fails to yield good results. Most research does not focus on such users and thus fails to cater to more exotic tastes and emerging trends, leading to a subsequent loss in revenue and marketing opportunities. One suggested solution is to use one-class classification to generate a prediction list for these users, where decision boundaries are learned that distinguish between normal and grey-sheep users. In this paper, we present the grey-sheep one-class recommendation (GSOR) framework designed to create accurate prediction models while taking both regular and grey-sheep users into account. In addition, we introduce a novel grey-sheep movie recommendation benchmark to be used by current and future researchers. When evaluating our GSOR framework against this benchmark, our results indicate the value of combining cluster analysis, outlier detection, and one-class learning to generate relevant and timely recommendation lists from data sets that contain grey-sheep users. Specifically, by employing one-class decision tree algorithms, our GSOR framework was able to outperform traditional collaborative filtering-based recommendation systems in both accuracy and model construction time. Furthermore, we report that having grey-sheep users in the system often had a positive impact on the learning and recommendation processes.}
}	1652	1652	}
	1653	1653
@article{HU2025127130,	1654	1654	@article{HU2025127130,
title = {A social importance and category enhanced cold-start user recommendation system},	1655	1655	title = {A social importance and category enhanced cold-start user recommendation system},
journal = {Expert Systems with Applications},	1656	1656	journal = {Expert Systems with Applications},
volume = {277},	1657	1657	volume = {277},
pages = {127130},	1658	1658	pages = {127130},
year = {2025},	1659	1659	year = {2025},
issn = {0957-4174},	1660	1660	issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.127130},	1661	1661	doi = {https://doi.org/10.1016/j.eswa.2025.127130},
url = {https://www.sciencedirect.com/science/article/pii/S0957417425007523},	1662	1662	url = {https://www.sciencedirect.com/science/article/pii/S0957417425007523},
author = {Bin Hu and Yinghong Ma and Zhiyuan Liu and Hong Wang},	1663	1663	author = {Bin Hu and Yinghong Ma and Zhiyuan Liu and Hong Wang},
keywords = {Social recommendation, Graph neural network, Cold-start users, Social importance, Category information},	1664	1664	keywords = {Social recommendation, Graph neural network, Cold-start users, Social importance, Category information},
abstract = {Social recommendation, which utilizes social relations to enhance recommender systems, has gained increasing attention with the rapid development of online social platforms. Although numerous studies have underscored the efficacy of integrating personal social information to bolster the performance of such systems, social recommendations still face several problems. Firstly, the cold-start problem for items persists in recommendation tasks leveraging social information. Secondly, the importance of users within social networks is often disregarded, leading to biases in recommendation tasks utilizing social information. Thirdly, the lack of utilization of item category information makes learning representations of items and users insufficient. Hence, this paper proposes a novel social recommendation model, Social Importance and Category Enhanced Cold-Start User Recommendation System (SICERec). At first, potential preference information for cold-start users is incorporated into similar user modules, extracting user preference information from historical interaction data between users and items. After that, the significance of users within social networks is considered by integrating their centrality attributes, thereby enriching the semantic representation of users. Finally, category information of user historical interaction items is incorporated into the modeling process to enrich the semantics of items. Extensive experimental results demonstrate the significant advantages of our SICERec method. Our model exhibits a minimum improvement of 15.1% in RMSE and at least 26.2% in MAE compared to state-of-the-art models when evaluated on two real datasets. Additionally, ablation experiments are conducted to validate each module’s effectiveness and provide further insights into how users’ social attributes and preferences influence their choices. We release our code at https://github.com/BinHu129/SICERec.}	1665	1665	abstract = {Social recommendation, which utilizes social relations to enhance recommender systems, has gained increasing attention with the rapid development of online social platforms. Although numerous studies have underscored the efficacy of integrating personal social information to bolster the performance of such systems, social recommendations still face several problems. Firstly, the cold-start problem for items persists in recommendation tasks leveraging social information. Secondly, the importance of users within social networks is often disregarded, leading to biases in recommendation tasks utilizing social information. Thirdly, the lack of utilization of item category information makes learning representations of items and users insufficient. Hence, this paper proposes a novel social recommendation model, Social Importance and Category Enhanced Cold-Start User Recommendation System (SICERec). At first, potential preference information for cold-start users is incorporated into similar user modules, extracting user preference information from historical interaction data between users and items. After that, the significance of users within social networks is considered by integrating their centrality attributes, thereby enriching the semantic representation of users. Finally, category information of user historical interaction items is incorporated into the modeling process to enrich the semantics of items. Extensive experimental results demonstrate the significant advantages of our SICERec method. Our model exhibits a minimum improvement of 15.1% in RMSE and at least 26.2% in MAE compared to state-of-the-art models when evaluated on two real datasets. Additionally, ablation experiments are conducted to validate each module’s effectiveness and provide further insights into how users’ social attributes and preferences influence their choices. We release our code at https://github.com/BinHu129/SICERec.}
}	1666	1666	}
	1667	1667
@inproceedings{wolf2024keep,	1668	1668	@inproceedings{wolf2024keep,
title={Keep the faith: Faithful explanations in convolutional neural networks for case-based reasoning},	1669	1669	title={Keep the faith: Faithful explanations in convolutional neural networks for case-based reasoning},
author={Wolf, Tom Nuno and Bongratz, Fabian and Rickmann, Anne-Marie and P{\"o}lsterl, Sebastian and Wachinger, Christian},	1670	1670	author={Wolf, Tom Nuno and Bongratz, Fabian and Rickmann, Anne-Marie and P{\"o}lsterl, Sebastian and Wachinger, Christian},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},	1671	1671	booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},	1672	1672	volume={38},
pages={5921--5929},	1673	1673	pages={5921--5929},
year={2024}	1674	1674	year={2024}
}	1675	1675	}
	1676	1676
@article{PAREJASLLANOVARCED2024111469,	1677	1677	@article{PAREJASLLANOVARCED2024111469,
title = {Case-based selection of explanation methods for neural network image classifiers},	1678	1678	title = {Case-based selection of explanation methods for neural network image classifiers},
journal = {Knowledge-Based Systems},	1679	1679	journal = {Knowledge-Based Systems},
volume = {288},	1680	1680	volume = {288},
pages = {111469},	1681	1681	pages = {111469},
year = {2024},	1682	1682	year = {2024},
issn = {0950-7051},	1683	1683	issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2024.111469},	1684	1684	doi = {https://doi.org/10.1016/j.knosys.2024.111469},
url = {https://www.sciencedirect.com/science/article/pii/S0950705124001047},	1685	1685	url = {https://www.sciencedirect.com/science/article/pii/S0950705124001047},
author = {Humberto Parejas-Llanovarced and Marta Caro-Martínez and Mauricio G. Orozco-del-Castillo and Juan A. Recio-García},	1686	1686	author = {Humberto Parejas-Llanovarced and Marta Caro-Martínez and Mauricio G. Orozco-del-Castillo and Juan A. Recio-García},

This is BibTeX, Version 0.99d (TeX Live 2023)	1	1	This is BibTeX, Version 0.99d (TeX Live 2023)
Capacity: max_strings=200000, hash_size=200000, hash_prime=170003	2	2	Capacity: max_strings=200000, hash_size=200000, hash_prime=170003
The top-level auxiliary file: main.aux	3	3	The top-level auxiliary file: main.aux
A level-1 auxiliary file: ./chapters/contexte2.aux	4	4	A level-1 auxiliary file: ./chapters/contexte2.aux
A level-1 auxiliary file: ./chapters/EIAH.aux	5	5	A level-1 auxiliary file: ./chapters/EIAH.aux
A level-1 auxiliary file: ./chapters/CBR.aux	6	6	A level-1 auxiliary file: ./chapters/CBR.aux
A level-1 auxiliary file: ./chapters/Architecture.aux	7	7	A level-1 auxiliary file: ./chapters/Architecture.aux
A level-1 auxiliary file: ./chapters/ESCBR.aux	8	8	A level-1 auxiliary file: ./chapters/ESCBR.aux
A level-1 auxiliary file: ./chapters/TS.aux	9	9	A level-1 auxiliary file: ./chapters/TS.aux
A level-1 auxiliary file: ./chapters/Conclusions.aux	10	10	A level-1 auxiliary file: ./chapters/Conclusions.aux
A level-1 auxiliary file: ./chapters/Publications.aux	11	11	A level-1 auxiliary file: ./chapters/Publications.aux
The style file: apalike.bst	12	12	The style file: apalike.bst
Database file #1: main.bib	13	13	Database file #1: main.bib
Warning--entry type for "Daubias2011" isn't style-file defined	14	14	Warning--entry type for "Daubias2011" isn't style-file defined
--line 693 of file main.bib	15	15	--line 693 of file main.bib
Warning--to sort, need author or key in UCI	16	16	Warning--to sort, need author or key in UCI
You've used 84 entries,	17	17	Warning--to sort, need author or key in Data
		18	You've used 85 entries,
1935 wiz_defined-function locations,	18	19	1935 wiz_defined-function locations,
1000 strings with 21030 characters,	19	20	1005 strings with 21157 characters,
and the built_in function-call counts, 37959 in all, are:	20	21	and the built_in function-call counts, 38168 in all, are:
= -- 3636	21	22	= -- 3659
> -- 1794	22	23	> -- 1794
< -- 56	23	24	< -- 56
+ -- 656	24	25	+ -- 656
- -- 602	25	26	- -- 602
* -- 3257	26	27	* -- 3272
:= -- 6493	27	28	:= -- 6527
add.period$ -- 270	28	29	add.period$ -- 274
call.type$ -- 84	29	30	call.type$ -- 85
change.case$ -- 696	30	31	change.case$ -- 701
chr.to.int$ -- 82	31	32	chr.to.int$ -- 83
cite$ -- 86	32	33	cite$ -- 89
duplicate$ -- 1432	33	34	duplicate$ -- 1442
empty$ -- 2563	34	35	empty$ -- 2581
format.name$ -- 731	35	36	format.name$ -- 731
if$ -- 7559	36	37	if$ -- 7600
int.to.chr$ -- 3	37	38	int.to.chr$ -- 3
int.to.str$ -- 0	38	39	int.to.str$ -- 0
missing$ -- 88	39	40	missing$ -- 88
newline$ -- 424	40	41	newline$ -- 430
num.names$ -- 279	41	42	num.names$ -- 279
pop$ -- 629	42	43	pop$ -- 632
preamble$ -- 1	43	44	preamble$ -- 1
purify$ -- 701	44	45	purify$ -- 706
quote$ -- 0	45	46	quote$ -- 0
skip$ -- 1094	46	47	skip$ -- 1101
stack$ -- 0	47	48	stack$ -- 0
substring$ -- 2550	48	49	substring$ -- 2563
swap$ -- 270	49	50	swap$ -- 270
text.length$ -- 24	50	51	text.length$ -- 24
text.prefix$ -- 0	51	52	text.prefix$ -- 0
top$ -- 0	52	53	top$ -- 0
type$ -- 492	53	54	type$ -- 498
warning$ -- 1	54	55	warning$ -- 2
while$ -- 284	55	56	while$ -- 284
width$ -- 0	56	57	width$ -- 0
write$ -- 1122	57	58	write$ -- 1135
(There were 2 warnings)	58	59	(There were 3 warnings)
	59	60

No preview for this file type

\relax	1	1	\relax
\providecommand\babel@aux[2]{}	2	2	\providecommand\babel@aux[2]{}
\@nameuse{bbl@beforestart}	3	3	\@nameuse{bbl@beforestart}
\catcode `:\active	4	4	\catcode `:\active
\catcode `;\active	5	5	\catcode `;\active
\catcode `!\active	6	6	\catcode `!\active
\catcode `?\active	7	7	\catcode `?\active
\providecommand\hyper@newdestlabel[2]{}	8	8	\providecommand\hyper@newdestlabel[2]{}
\providecommand\HyperFirstAtBeginDocument{\AtBeginDocument}	9	9	\providecommand\HyperFirstAtBeginDocument{\AtBeginDocument}
\HyperFirstAtBeginDocument{\ifx\hyper@anchor\@undefined	10	10	\HyperFirstAtBeginDocument{\ifx\hyper@anchor\@undefined
\global\let\oldnewlabel\newlabel	11	11	\global\let\oldnewlabel\newlabel
\gdef\newlabel#1#2{\newlabelxx{#1}#2}	12	12	\gdef\newlabel#1#2{\newlabelxx{#1}#2}
\gdef\newlabelxx#1#2#3#4#5#6{\oldnewlabel{#1}{{#2}{#3}}}	13	13	\gdef\newlabelxx#1#2#3#4#5#6{\oldnewlabel{#1}{{#2}{#3}}}
\AtEndDocument{\ifx\hyper@anchor\@undefined	14	14	\AtEndDocument{\ifx\hyper@anchor\@undefined
\let\newlabel\oldnewlabel	15	15	\let\newlabel\oldnewlabel
\fi}	16	16	\fi}
\fi}	17	17	\fi}
\global\let\hyper@last\relax	18	18	\global\let\hyper@last\relax
\gdef\HyperFirstAtBeginDocument#1{#1}	19	19	\gdef\HyperFirstAtBeginDocument#1{#1}
\providecommand\HyField@AuxAddToFields[1]{}	20	20	\providecommand\HyField@AuxAddToFields[1]{}
\providecommand\HyField@AuxAddToCoFields[2]{}	21	21	\providecommand\HyField@AuxAddToCoFields[2]{}
\providecommand \oddpage@label [2]{}	22	22	\providecommand \oddpage@label [2]{}
\babel@aux{french}{}	23	23	\babel@aux{french}{}
\@writefile{toc}{\contentsline {part}{I\hspace {1em}Contexte et Problématiques}{1}{part.1}\protected@file@percent }	24	24	\@writefile{toc}{\contentsline {part}{I\hspace {1em}Contexte et Problématiques}{1}{part.1}\protected@file@percent }
\citation{Nkambou}	25	25	\citation{Nkambou}
\citation{doi:10.1177/1754337116651013}	26	26	\citation{doi:10.1177/1754337116651013}
\@writefile{toc}{\contentsline {chapter}{\numberline {1}Introduction}{3}{chapter.1}\protected@file@percent }	27	27	\@writefile{toc}{\contentsline {chapter}{\numberline {1}Introduction}{3}{chapter.1}\protected@file@percent }
\@writefile{lof}{\addvspace {10\p@ }}	28	28	\@writefile{lof}{\addvspace {10\p@ }}
\@writefile{lot}{\addvspace {10\p@ }}	29	29	\@writefile{lot}{\addvspace {10\p@ }}
\@writefile{toc}{\contentsline {section}{\numberline {1.1}Contributions Principales}{4}{section.1.1}\protected@file@percent }	30	30	\@writefile{toc}{\contentsline {section}{\numberline {1.1}Contributions Principales}{4}{section.1.1}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {1.2}Plan de la thèse}{5}{section.1.2}\protected@file@percent }	31	31	\@writefile{toc}{\contentsline {section}{\numberline {1.2}Plan de la thèse}{5}{section.1.2}\protected@file@percent }
\@input{./chapters/contexte2.aux}	32	32	\@input{./chapters/contexte2.aux}
\@writefile{toc}{\contentsline {part}{II\hspace {1em}État de l'art}{21}{part.2}\protected@file@percent }	33	33	\@writefile{toc}{\contentsline {part}{II\hspace {1em}État de l'art}{21}{part.2}\protected@file@percent }
\@input{./chapters/EIAH.aux}	34	34	\@input{./chapters/EIAH.aux}
\@input{./chapters/CBR.aux}	35	35	\@input{./chapters/CBR.aux}
\@writefile{toc}{\contentsline {part}{III\hspace {1em}Contributions}{39}{part.3}\protected@file@percent }	36	36	\@writefile{toc}{\contentsline {part}{III\hspace {1em}Contributions}{39}{part.3}\protected@file@percent }
\@input{./chapters/Architecture.aux}	37	37	\@input{./chapters/Architecture.aux}
\@input{./chapters/ESCBR.aux}	38	38	\@input{./chapters/ESCBR.aux}
\@input{./chapters/TS.aux}	39	39	\@input{./chapters/TS.aux}
\@input{./chapters/Conclusions.aux}	40	40	\@input{./chapters/Conclusions.aux}
\@input{./chapters/Publications.aux}	41	41	\@input{./chapters/Publications.aux}
\bibstyle{apalike}	42	42	\bibstyle{apalike}
\bibdata{main.bib}	43	43	\bibdata{main.bib}
		44	\bibcite{Data}{Dat, 2023}
\bibcite{UCI}{UCI, 2024}	44	45	\bibcite{UCI}{UCI, 2024}
\bibcite{doi:10.3233/AIC-1994-7104}{Aamodt and Plaza, 1994}	45	46	\bibcite{doi:10.3233/AIC-1994-7104}{Aamodt and Plaza, 1994}
\bibcite{NEURIPS2023_9d8cf124}{Abel et~al., 2023}	46	47	\bibcite{NEURIPS2023_9d8cf124}{Abel et~al., 2023}
\bibcite{ALABDULRAHMAN2021114061}{Alabdulrahman and Viktor, 2021}	47	48	\bibcite{ALABDULRAHMAN2021114061}{Alabdulrahman and Viktor, 2021}
\bibcite{Arthurs}{Arthurs et~al., 2019}	48	49	\bibcite{Arthurs}{Arthurs et~al., 2019}
\bibcite{Auer}{Auer et~al., 2021}	49	50	\bibcite{Auer}{Auer et~al., 2021}
\bibcite{badier:hal-04092828}{Badier et~al., 2023}	50	51	\bibcite{badier:hal-04092828}{Badier et~al., 2023}
\bibcite{BAKUROV2021100913}{Bakurov et~al., 2021}	51	52	\bibcite{BAKUROV2021100913}{Bakurov et~al., 2021}
\bibcite{10.1007/978-3-642-15973-2_50}{Butdee and Tichkiewitch, 2011}	52	53	\bibcite{10.1007/978-3-642-15973-2_50}{Butdee and Tichkiewitch, 2011}
\bibcite{CHIU2023100118}{Chiu et~al., 2023}	53	54	\bibcite{CHIU2023100118}{Chiu et~al., 2023}
\bibcite{cmc.2023.033417}{Choi et~al., 2023}	54	55	\bibcite{cmc.2023.033417}{Choi et~al., 2023}
\bibcite{Riesbeck1989}{C.K. and R.C., 1989}	55	56	\bibcite{Riesbeck1989}{C.K. and R.C., 1989}
\bibcite{10.1145/3459665}{Cunningham and Delany, 2021}	56	57	\bibcite{10.1145/3459665}{Cunningham and Delany, 2021}
\bibcite{DIDDEN2023338}{Didden et~al., 2023}	57	58	\bibcite{DIDDEN2023338}{Didden et~al., 2023}
\bibcite{EZALDEEN2022100700}{Ezaldeen et~al., 2022}	58	59	\bibcite{EZALDEEN2022100700}{Ezaldeen et~al., 2022}
\bibcite{10.1007/978-3-030-58342-2_5}{Feely et~al., 2020}	59	60	\bibcite{10.1007/978-3-030-58342-2_5}{Feely et~al., 2020}
\bibcite{10.1007/978-3-319-47096-2_11}{Grace et~al., 2016}	60	61	\bibcite{10.1007/978-3-319-47096-2_11}{Grace et~al., 2016}
\bibcite{9434422}{Gupta et~al., 2021}	61	62	\bibcite{9434422}{Gupta et~al., 2021}
\bibcite{hajduk2019cognitive}{Hajduk et~al., 2019}	62	63	\bibcite{hajduk2019cognitive}{Hajduk et~al., 2019}
\bibcite{doi:10.1177/1754337116651013}{Henriet et~al., 2017}	63	64	\bibcite{doi:10.1177/1754337116651013}{Henriet et~al., 2017}
\bibcite{10.1007/978-3-030-01081-2_9}{Henriet and Greffier, 2018}	64	65	\bibcite{10.1007/978-3-030-01081-2_9}{Henriet and Greffier, 2018}
\bibcite{HIPOLITO2023103510}{Hipólito and Kirchhoff, 2023}	65	66	\bibcite{HIPOLITO2023103510}{Hipólito and Kirchhoff, 2023}
\bibcite{Hoang}{Hoang, 2018}	66	67	\bibcite{Hoang}{Hoang, 2018}
\bibcite{HU2025127130}{Hu et~al., 2025}	67	68	\bibcite{HU2025127130}{Hu et~al., 2025}
\bibcite{HUANG2023104684}{Huang et~al., 2023}	68	69	\bibcite{HUANG2023104684}{Huang et~al., 2023}
\bibcite{INGKAVARA2022100086}{Ingkavara et~al., 2022}	69	70	\bibcite{INGKAVARA2022100086}{Ingkavara et~al., 2022}
\bibcite{Daubias2011}{Jean-Daubias, 2011}	70	71	\bibcite{Daubias2011}{Jean-Daubias, 2011}
\bibcite{JUNG20095695}{Jung et~al., 2009}	71	72	\bibcite{JUNG20095695}{Jung et~al., 2009}
\bibcite{KAMALI2023110242}{Kamali et~al., 2023}	72	73	\bibcite{KAMALI2023110242}{Kamali et~al., 2023}
\bibcite{Kim2024}{Kim, 2024}	73	74	\bibcite{Kim2024}{Kim, 2024}
\bibcite{KOLODNER1983281}{Kolodner, 1983}	74	75	\bibcite{KOLODNER1983281}{Kolodner, 1983}
\bibcite{Kuzilek2017}{Kuzilek et~al., 2017}	75	76	\bibcite{Kuzilek2017}{Kuzilek et~al., 2017}
\bibcite{LALITHA2020583}{Lalitha and Sreeja, 2020}	76	77	\bibcite{LALITHA2020583}{Lalitha and Sreeja, 2020}
\bibcite{lei2024analysis}{Lei, 2024}	77	78	\bibcite{lei2024analysis}{Lei, 2024}
\bibcite{min8100434}{Leikola et~al., 2018}	78	79	\bibcite{min8100434}{Leikola et~al., 2018}
\bibcite{10.1007/978-3-030-58342-2_20}{Lepage et~al., 2020}	79	80	\bibcite{10.1007/978-3-030-58342-2_20}{Lepage et~al., 2020}
\bibcite{Li_2024}{Li et~al., 2024}	80	81	\bibcite{Li_2024}{Li et~al., 2024}
\bibcite{10.3389/fgene.2021.600040}{Liang et~al., 2021}	81	82	\bibcite{10.3389/fgene.2021.600040}{Liang et~al., 2021}
\bibcite{9870279}{Lin, 2022}	82	83	\bibcite{9870279}{Lin, 2022}
\bibcite{Liu2023}{Liu and Yu, 2023}	83	84	\bibcite{Liu2023}{Liu and Yu, 2023}
\bibcite{jmse11050890}{Louvros et~al., 2023}	84	85	\bibcite{jmse11050890}{Louvros et~al., 2023}
\bibcite{10.1007/978-3-319-61030-6_1}{Maher and Grace, 2017}	85	86	\bibcite{10.1007/978-3-319-61030-6_1}{Maher and Grace, 2017}
\bibcite{10.1007/978-3-031-63646-2_4}{Malburg et~al., 2024}	86	87	\bibcite{10.1007/978-3-031-63646-2_4}{Malburg et~al., 2024}
\bibcite{Liang}{Mang et~al., 2021}	87	88	\bibcite{Liang}{Mang et~al., 2021}
\bibcite{doi:10.1137/23M1592420}{Minsker and Strawn, 2024}	88	89	\bibcite{doi:10.1137/23M1592420}{Minsker and Strawn, 2024}
\bibcite{MUANGPRATHUB2020e05227}{Muangprathub et~al., 2020}	89	90	\bibcite{MUANGPRATHUB2020e05227}{Muangprathub et~al., 2020}
\bibcite{Muller}{Müller and Bergmann, 2015}	90	91	\bibcite{Muller}{Müller and Bergmann, 2015}
\bibcite{NGUYEN2024111566}{Nguyen, 2024}	91	92	\bibcite{NGUYEN2024111566}{Nguyen, 2024}
\bibcite{Nkambou}{Nkambou et~al., 2010}	92	93	\bibcite{Nkambou}{Nkambou et~al., 2010}
\bibcite{Obeid}{Obeid et~al., 2022}	93	94	\bibcite{Obeid}{Obeid et~al., 2022}
\bibcite{10.1007/978-3-319-24586-7_20}{Onta{\~{n}}{\'o}n et~al., 2015}	94	95	\bibcite{10.1007/978-3-319-24586-7_20}{Onta{\~{n}}{\'o}n et~al., 2015}
\bibcite{pmlr-v238-ou24a}{Ou et~al., 2024}	95	96	\bibcite{pmlr-v238-ou24a}{Ou et~al., 2024}
\bibcite{PAREJASLLANOVARCED2024111469}{Parejas-Llanovarced et~al., 2024}	96	97	\bibcite{PAREJASLLANOVARCED2024111469}{Parejas-Llanovarced et~al., 2024}
\bibcite{PETROVIC201617}{Petrovic et~al., 2016}	97	98	\bibcite{PETROVIC201617}{Petrovic et~al., 2016}
\bibcite{Richter2013}{Richter and Weber, 2013}	98	99	\bibcite{Richter2013}{Richter and Weber, 2013}
\bibcite{RICHTER20093}{Richter, 2009}	99	100	\bibcite{RICHTER20093}{Richter, 2009}
\bibcite{Robertson2014ARO}{Robertson and Watson, 2014}	100	101	\bibcite{Robertson2014ARO}{Robertson and Watson, 2014}
\bibcite{ROLDANREYES20151}{{Roldan Reyes} et~al., 2015}	101	102	\bibcite{ROLDANREYES20151}{{Roldan Reyes} et~al., 2015}
\bibcite{Sadeghi}{Sadeghi~Moghadam et~al., 2024}	102	103	\bibcite{Sadeghi}{Sadeghi~Moghadam et~al., 2024}
\bibcite{schank+abelson77}{Schank and Abelson, 1977}	103	104	\bibcite{schank+abelson77}{Schank and Abelson, 1977}
\bibcite{pmlr-v108-seznec20a}{Seznec et~al., 2020}	104	105	\bibcite{pmlr-v108-seznec20a}{Seznec et~al., 2020}
\bibcite{9072123}{Sinaga and Yang, 2020}	105	106	\bibcite{9072123}{Sinaga and Yang, 2020}
\bibcite{skittou2024recommender}{Skittou et~al., 2024}	106	107	\bibcite{skittou2024recommender}{Skittou et~al., 2024}
\bibcite{10.1007/978-3-030-01081-2_25}{Smyth and Cunningham, 2018}	107	108	\bibcite{10.1007/978-3-030-01081-2_25}{Smyth and Cunningham, 2018}
\bibcite{10.1007/978-3-030-58342-2_8}{Smyth and Willemsen, 2020}	108	109	\bibcite{10.1007/978-3-030-58342-2_8}{Smyth and Willemsen, 2020}
\bibcite{Soto2}{Soto-Forero et~al., 2024a}	109	110	\bibcite{Soto2}{Soto-Forero et~al., 2024a}
\bibcite{10.1007/978-3-031-63646-2_13}{Soto-Forero et~al., 2024b}	110	111	\bibcite{10.1007/978-3-031-63646-2_13}{Soto-Forero et~al., 2024b}
\bibcite{10.1007/978-3-031-63646-2_11}{Soto-Forero et~al., 2024c}	111	112	\bibcite{10.1007/978-3-031-63646-2_11}{Soto-Forero et~al., 2024c}
\bibcite{SU2022109547}{Su et~al., 2022}	112	113	\bibcite{SU2022109547}{Su et~al., 2022}
\bibcite{8495930}{Supic, 2018}	113	114	\bibcite{8495930}{Supic, 2018}
\bibcite{math12111758}{Uguina et~al., 2024}	114	115	\bibcite{math12111758}{Uguina et~al., 2024}
\bibcite{buildings13030651}{Uysal and Sonmez, 2023}	115	116	\bibcite{buildings13030651}{Uysal and Sonmez, 2023}
\bibcite{WANG2021331}{Wang et~al., 2021}	116	117	\bibcite{WANG2021331}{Wang et~al., 2021}
\bibcite{wolf2024keep}{Wolf et~al., 2024}	117	118	\bibcite{wolf2024keep}{Wolf et~al., 2024}
\bibcite{9627973}{Xu et~al., 2021}	118	119	\bibcite{9627973}{Xu et~al., 2021}
\bibcite{10.1145/3578337.3605122}{Xu et~al., 2023}	119	120	\bibcite{10.1145/3578337.3605122}{Xu et~al., 2023}
\bibcite{YU2023110163}{Yu and Li, 2023}	120	121	\bibcite{YU2023110163}{Yu and Li, 2023}
\bibcite{YU2024123745}{Yu et~al., 2024}	121	122	\bibcite{YU2024123745}{Yu et~al., 2024}
\bibcite{ZHANG2021100025}{Zhang. and Aslan, 2021}	122	123	\bibcite{ZHANG2021100025}{Zhang. and Aslan, 2021}
\bibcite{ZHANG2018189}{Zhang and Yao, 2018}	123	124	\bibcite{ZHANG2018189}{Zhang and Yao, 2018}
\bibcite{ZHANG2023110564}{Zhang et~al., 2023}	124	125	\bibcite{ZHANG2023110564}{Zhang et~al., 2023}
\bibcite{ZHAO2023118535}{Zhao et~al., 2023}	125	126	\bibcite{ZHAO2023118535}{Zhao et~al., 2023}
\bibcite{Zhou2021}{Zhou and Wang, 2021}	126	127	\bibcite{Zhou2021}{Zhou and Wang, 2021}
\bibcite{jmse10040464}{Zuluaga et~al., 2022}	127	128	\bibcite{jmse10040464}{Zuluaga et~al., 2022}
\gdef \@abspage@last{122}	128	129	\gdef \@abspage@last{122}

This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.5.31) 11 JUL 2025 22:56	1	1	This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.5.31) 11 JUL 2025 23:11
entering extended mode	2	2	entering extended mode
restricted \write18 enabled.	3	3	restricted \write18 enabled.
%&-line parsing enabled.	4	4	%&-line parsing enabled.
**main.tex	5	5	**main.tex
(./main.tex	6	6	(./main.tex
LaTeX2e <2022-11-01> patch level 1	7	7	LaTeX2e <2022-11-01> patch level 1
L3 programming layer <2023-05-22> (./spimufcphdthesis.cls	8	8	L3 programming layer <2023-05-22> (./spimufcphdthesis.cls
Document Class: spimufcphdthesis 2022/02/10	9	9	Document Class: spimufcphdthesis 2022/02/10
	10	10
(/usr/local/texlive/2023/texmf-dist/tex/latex/upmethodology/upmethodology-docum	11	11	(/usr/local/texlive/2023/texmf-dist/tex/latex/upmethodology/upmethodology-docum
ent.cls	12	12	ent.cls
Document Class: upmethodology-document 2022/10/04	13	13	Document Class: upmethodology-document 2022/10/04
(./upmethodology-p-common.sty	14	14	(./upmethodology-p-common.sty
Package: upmethodology-p-common 2015/04/24	15	15	Package: upmethodology-p-common 2015/04/24
	16	16
(/usr/local/texlive/2023/texmf-dist/tex/latex/base/ifthen.sty	17	17	(/usr/local/texlive/2023/texmf-dist/tex/latex/base/ifthen.sty
Package: ifthen 2022/04/13 v1.1d Standard LaTeX ifthen package (DPC)	18	18	Package: ifthen 2022/04/13 v1.1d Standard LaTeX ifthen package (DPC)
)	19	19	)
(/usr/local/texlive/2023/texmf-dist/tex/latex/tools/xspace.sty	20	20	(/usr/local/texlive/2023/texmf-dist/tex/latex/tools/xspace.sty
Package: xspace 2014/10/28 v1.13 Space after command names (DPC,MH)	21	21	Package: xspace 2014/10/28 v1.13 Space after command names (DPC,MH)
)	22	22	)
(/usr/local/texlive/2023/texmf-dist/tex/latex/xcolor/xcolor.sty	23	23	(/usr/local/texlive/2023/texmf-dist/tex/latex/xcolor/xcolor.sty
Package: xcolor 2022/06/12 v2.14 LaTeX color extensions (UK)	24	24	Package: xcolor 2022/06/12 v2.14 LaTeX color extensions (UK)
	25	25
(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics-cfg/color.cfg	26	26	(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics-cfg/color.cfg
File: color.cfg 2016/01/02 v1.6 sample color configuration	27	27	File: color.cfg 2016/01/02 v1.6 sample color configuration
)	28	28	)
Package xcolor Info: Driver file: pdftex.def on input line 227.	29	29	Package xcolor Info: Driver file: pdftex.def on input line 227.
	30	30
(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics-def/pdftex.def	31	31	(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics-def/pdftex.def
File: pdftex.def 2022/09/22 v1.2b Graphics/color driver for pdftex	32	32	File: pdftex.def 2022/09/22 v1.2b Graphics/color driver for pdftex
)	33	33	)
(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics/mathcolor.ltx)	34	34	(/usr/local/texlive/2023/texmf-dist/tex/latex/graphics/mathcolor.ltx)
Package xcolor Info: Model `cmy' substituted by `cmy0' on input line 1353.	35	35	Package xcolor Info: Model `cmy' substituted by `cmy0' on input line 1353.
Package xcolor Info: Model `hsb' substituted by `rgb' on input line 1357.	36	36	Package xcolor Info: Model `hsb' substituted by `rgb' on input line 1357.
Package xcolor Info: Model `RGB' extended on input line 1369.	37	37	Package xcolor Info: Model `RGB' extended on input line 1369.
Package xcolor Info: Model `HTML' substituted by `rgb' on input line 1371.	38	38	Package xcolor Info: Model `HTML' substituted by `rgb' on input line 1371.
Package xcolor Info: Model `Hsb' substituted by `hsb' on input line 1372.	39	39	Package xcolor Info: Model `Hsb' substituted by `hsb' on input line 1372.
Package xcolor Info: Model `tHsb' substituted by `hsb' on input line 1373.	40	40	Package xcolor Info: Model `tHsb' substituted by `hsb' on input line 1373.