Dans de nombreux domaines, les données peuvent être de grande dimension. Ça pose le problème de la réduction de dimension. Les techniques de réduction de dimension peuvent être classées en fonction de leur but : techniques pour la représentation optimale et techniques pour la classification, ainsi qu'en fonction de leur stratégie : la sélection et l'extraction des caractéristiques. L'ensemble des caractéristiques résultant des méthodes d'extraction est non interprétable. Ainsi, la première problématique scientifique de la thèse est comment extraire des caractéristiques latentes interprétables? La réduction de dimension pour la classification vise à améliorer la puissance de classification du sous-ensemble sélectionné. Nous voyons le développement de la tâche de classification comme la tâche d'identification des facteurs déclencheurs, c'est-à-dire des facteurs qui peuvent influencer le transfert d'éléments de données d'une classe à l'autre. La deuxième problématique scientifique de cette thèse est comment identifier automatiquement ces facteurs déclencheurs? Nous visons à résoudre les deux problématiques scientifiques dans le domaine d'application des systèmes de recommandation. Nous proposons d'interpréter les caractéristiques latentes de systèmes de recommandation basés sur la factorisation de matrices comme des utilisateurs réels. Nous concevons un algorithme d'identification automatique des facteurs déclencheurs basé sur les concepts d'analyse par contraste. Au travers d'expérimentations, nous montrons que les motifs définis peuvent être considérés comme des facteurs déclencheurs / In many application areas, data elements can be high-dimensional. This raises the problem of dimensionality reduction. The dimensionality reduction techniques can be classified based on their aim: dimensionality reduction for optimal data representation and dimensionality reduction for classification, as well as based on the adopted strategy: feature selection and feature extraction. The set of features resulting from feature extraction methods is usually uninterpretable. Thereby, the first scientific problematic of the thesis is how to extract interpretable latent features? The dimensionality reduction for classification aims to enhance the classification power of the selected subset of features. We see the development of the task of classification as the task of trigger factors identification that is identification of those factors that can influence the transfer of data elements from one class to another. The second scientific problematic of this thesis is how to automatically identify these trigger factors? We aim at solving both scientific problematics within the recommender systems application domain. We propose to interpret latent features for the matrix factorization-based recommender systems as real users. We design an algorithm for automatic identification of trigger factors based on the concepts of contrast analysis. Through experimental results, we show that the defined patterns indeed can be considered as trigger factors

Les systèmes de recommandation se sont imposés comme étant des outils indispensables face à une quantité de données qui ne cesse chaque jour de croître depuis l'avènement d'Internet. Leur objectif est de proposer aux utilisateurs des items susceptibles de les intéresser sans que ces derniers n'aient besoin d'agir pour les obtenir. Après s'être majoritairement focalisés sur la précision de la prédiction d'intérêt, ces systèmes ont évolué pour prendre en compte d'autres critères dans leur processus de recommandation, tels que les facteurs humains inhérents à la prise de décision, afin d'améliorer la qualité et l'utilité des recommandations. Cependant, la prise en compte de certains facteurs humains tels que la diversité et le contexte demeure critiquable. Alors que le contexte des utilisateurs est inféré sur la base d'informations collectées à l'insu de leur vie privée, la prise en compte de la diversité est quant à elle réduite à une dimension qu'un système se doit de maximiser. Or, certains travaux récents démontrent que la diversité correspond à un besoin évoluant dynamiquement au cours du temps, et dont la proportion à insuffler dans les recommandations est dépendante de la tâche effectuée (i.e du contexte). Partant du postulat inverse selon lequel l'analyse de l'évolution de la diversité au cours du temps permet de définir le contexte de l'utilisateur, nous proposons dans ce manuscrit une nouvelle approche de modélisation contextuelle basée sur la diversité. En effet, nous soutenons qu'une variation de diversité remarquable peut être la conséquence d'un changement de contexte et qu'il faut alors adapter la stratégie de recommandation en conséquence. Nous présentons la première approche de la littérature permettant de modéliser en temps réel l'évolution de la diversité, ainsi qu'une nouvelle famille de contextes dits implicites n'exploitant aucune donnée sensible. La possibilité de remplacer les contextes traditionnels (explicites) par les contextes implicites est confirmée de plusieurs manières. Premièrement, nous démontrons sur deux corpus issus d'applications réelles qu'il existe un fort recouvrement entre les changements de contextes explicites et les changements de contextes implicites. Deuxièmement, une étude utilisateur impliquant de nombreux participants nous permet de démontrer l'existence de liens entre les contextes explicites et les caractéristiques des items consultés dans ces derniers. Fort de ces constats et du potentiel offert par nos modèles, nous présentons également plusieurs approches de recommandation et de prise en compte des besoins des utilisateurs / Recommender Systems (RS) have become essential tools to deal with an endless increasing amount of data available on the Internet. Their goal is to provide items that may interest users before they have to find them by themselves. After being exclusively focused on the precision of users' interests prediction task, RS had to evolve by taking into account other criteria like human factors involved in the decision-making process while computing recommendations, so as to improve their quality and usefulness of recommendations. Nevertheless, the way some human factors, such as context and diversity needs, are managed remains open to criticism. While context-aware recommendations relies on exploiting data that are collected without any consideration for users' privacy, diversity has been coming down to a dimension which has to be maximized. However recent studies demonstrate that diversity corresponds to a need which evolves dynamically over time. In addition, the optimal amount of diversity to provide in the recommendations depends on the on-going task of users (i.e their contexts). Thereby, we argue that analyzing the evolution of diversity over time would be a promising way to define a user's context, under the condition that context is now defined by item attributes. Indeed, we support the idea that a sudden variation of diversity can reflect a change of user's context which requires to adapt the recommendation strategy. We present in this manuscript the first approach to model the evolution of diversity over time and a new kind of context, called ``implicit contexts'', that are respectful of privacy (in opposition to explicit contexts). We confirm the benefits of implicit contexts compared to explicit contexts from several points of view. As a first step, using two large music streaming datasets we demonstrate that explicit and implicit context changes are highly correlated. As a second step, a user study involving many participants allowed us to demonstrate the links between the explicit contexts and the characteristics of the items consulted in the meantime. Based on these observations and the advantages offered by our models, we also present several approaches to provide privacy-preserving context-aware recommendations and to take into account user's needs

<div>Over the past decade, universities had a significant improvement in using online learning tools. A standard learning management system provides fundamental functionalities to satisfy the basic needs of its users. The new generation of learning management systems have introduced a novel system that provides social networking features. An unprecedented number of users use the social aspects of such platforms to create their profile, collaborate with other users, and find their desired career path. Nowadays there are many learning systems which provide learning materials, certificates, and course management systems. This allows us to utilize such information to help the students and the instructors in their academic life. </div><div><br></div><div>The presented research work's primary goal is to focus on creating an intelligent personal assistant within the social learning systems. The proposed personal assistant has a human-like persona, learns about the users, and recommends useful and meaningful materials for them. The designed system offers a set of features for both institutions and members to achieve their goal within the learning system. It recommends jobs and friends for the users based on their profile. The proposed agent also prioritizes the messages and shows the most important message to the user. </div><div><br></div><div>The developed software supports model-controller-view architecture and provides a set of RESTful APIs which allows the institutions to integrate the proposed intelligent agent with their learning system. <br></div>

"The goal of this thesis is to extend the theory and practice of matrix completion algorithms, and how they can be utilized, improved, and scaled up to handle large data sets. Matrix completion involves predicting missing entries in real-world data matrices using the modeling assumption that the fully observed matrix is low-rank. Low-rank matrices appear across a broad selection of domains, and such a modeling assumption is similar in spirit to Principal Component Analysis. Our focus is on large scale problems, where the matrices have millions of rows and columns. In this thesis we provide new analysis for the convergence rates of matrix completion techniques using convex nuclear norm relaxation. In addition, we validate these results on both synthetic data and data from two real-world domains (recommender systems and Internet tomography). The results we obtain show that with an empirical, data-inspired understanding of various parameters in the algorithm, this matrix completion problem can be solved more efficiently than some previous theory suggests, and therefore can be extended to much larger problems with greater ease. "

Recommender Systems (RS) support users to find relevant content, such as movies, books, songs, and other products based on their preferences. Such preferences are gathered by analyzing past users interactions, however, data collected for this purpose are typically prone to sparsity and high dimensionality. Clustering-based techniques have been proposed to handle these problems effectively and efficiently by segmenting the data into a number of similar groups based on predefined characteristics. Although these techniques have gained increasing attention in the recommender systems community, they are usually bound to a particular recommender system and/or require critical parameters, such as the number of clusters. In this work, we present three variants of a general-purpose method to optimally extract users groups from a hierarchical clustering algorithm specifically targeting RS problems. The proposed extraction methods do not require critical parameters and can be applied prior to any recommendation system. Our experiments have shown promising recommendation results in the context of nine well-known public datasets from different domains. / Sistemas de Recomendação auxiliam usuários a encontrar conteúdo relevante, como filmes, livros, músicas entre outros produtos baseando-se em suas preferências. Tais preferências são obtidas ao analisar interações passadas dos usuários, no entanto, dados coletados com esse propósito tendem a tipicamente possuir alta dimensionalidade e esparsidade. Técnicas baseadas em agrupamento de dados têm sido propostas para lidar com esses problemas de foma eficiente e eficaz ao dividir os dados em grupos similares baseando-se em características pré-definidas. Ainda que essas técnicas tenham recebido atenção crescente na comunidade de sistemas de recomendação, tais técnicas são usualmente atreladas a um algoritmo de recomendação específico e/ou requerem parâmetros críticos, como número de grupos. Neste trabalho, apresentamos três variantes de um método de propósitvo geral de extração ótima de grupos em uma hierarquia, atacando especificamente problemas em Sistemas de Recomendação. Os métodos de extração propostos não requerem parâmetros críticos e podem ser aplicados antes de qualquer sistema de recomendação. Os experimentos mostraram resultados promissores no contexto de nove bases de dados públicas conhecidas em diferentes domínios.

São muitas as pesquisas nos dias de hoje que buscam por métodos e ferramentas para aumentar a autonomia do aluno na condução dos processos de aprendizagem, uma vez que os métodos tradicionais de ensino nem sempre se mostram eficazes na formação de estudantes com capacidade crítica, coerente com as necessidades do mundo atual. O presente trabalho tem como objetivo investigar de que modo uma atividade de Aprendizagem Baseada em Projetos (ABPr), apoiada por um ambiente tecnológico desenvolvido para este fim, pode contribuir no desenvolvimento de projetos por meio de recursos de recomendação de conteúdo e ferramentas de colaboração entre pares. Para isto é utilizado uma abordagem ativa de aprendizagem, a ABPr, definida como um método de aprendizagem centrado no aluno e que enfatiza atividades para o desenvolvimento de projetos. Durante este processo os alunos podem tomar suas próprias decisões e agir sozinhos ou em grupos. Para a aplicação do método proposto foi estruturada uma atividade educacional, que consiste no desenvolvimento de um projeto a partir das investigações dos alunos em relação a um tema proposto pelo professor. O desenvolvimento deste projeto se inicia e termina em sala de aula, entretanto as fases intermediárias podem ocorrer em outros locais. Para a execução da atividade foi desenvolvida uma ferramenta que incentiva a colaboração entre os alunos. Isto permite uma maior interação entre os participantes e também a possibilidade dos alunos colaborarem nos projetos uns dos outros. Durante o desenvolvimento de seus projetos, a ferramenta sugere materiais complementares relacionados ao assunto tratado, como forma auxiliar os alunos em seus processos investigativos. Para a avaliação do trabalho proposto foi estruturada uma pesquisa quali-quantitativa, na modalidade estudo de caso, com coleta de dados por meio da análise de projetos, registro de atividades, questionários e entrevistas. Os resultados obtidos através dos experimentos realizados demonstraram que a atividade educacional proposta por este trabalho contribuiu de forma significativa para o desenvolvimento de projetos e para uma maior interação entre os alunos. / Many research works focus on the development of methods and tools to increase student autonomy in the conduct of learning processes, as traditional teaching methods are not always effective in training students with critical skills, in accordance with the needs of today's world. This study aims to investigate how a Project-based Learning (PBL) activity, supported by a technological environment developed for this purpose, can contribute to the development of projects by means of content recommendation resources and collaboration tools among peers. For this reason, an active learning approach is used, PBL, defined as a student-centered learning method that emphasizes activities for project development. During this process students can make their own decisions and act alone or in groups. For the application of the proposed method an educational activity was structured consisting in the development of a project based on students' investigations related to a topic proposed by the teacher. The development of this project starts and ends in the classroom, but the intermediate stages can occur in other places. For the execution of the activity, a tool was developed for fostering collaboration between students. This allows a higher interaction between participants and the possibility of students to collaborate on each other's projects. During the development of their projects, the tool suggests additional materials related to the subject at hand, as a way to assist students in their research processes. For the evaluation of the proposed work a quali-quantitative study was structured, with data collection performed from project analysis, activity logging, questionnaires and interviews. Results from the experiments performed showed that the educational activity proposed by this work contributed significantly to the development of projects and for a higher interaction among students.

Sistemas de recomendação foram propostos no início da década de 1990 com o objetivo de auxiliar seus usuários a lidar com a sobrecarga cognitiva criada com o advento da internet e o aumento constante de documentos. De lá para cá tais sistemas passaram a assumir vários outros papéis, tais como “auxiliar usuários a explorar”, “melhorar a tomada de decisão”, ou até mesmo “entreter”. Para atingir tais novos objetivos, o sistema necessita olhar para características do usuário que auxiliem no entendimento da tarefa desempenhada pelo usuário e como a recomendação pode auxiliar tal tarefa. Nesse sentido, propõe-se nessa tese uma integração entre estratégias de recomendação e de adaptação para criar um novo processo de recomendação adaptativa. É mostrado que tal integração pode melhorar a acurácia da recomendação, e dar bons resultados na retenção de usuários, e na interação destes com os sistemas. Para validar a abordagem, é implementado um protótipo para recomendação de filmes a serem utilizados em sala de aula. São também coletadas estatísticas de 78 usuários que participaram do experimento de avaliação da abordagem. / Recommender systems were proposed in early 90’s with the goal to help users deal with cognitive overload brought by the internet and the constant increase of documents. From there to now such systems have assumed many other roles like “help users to explore”, “improve decision making”, or even “entertain”. To accomplish such new goals, the system needs to look to user characteristics that help in understand what the user task is and how to adapt the recommendation to support such task. In this direction, it is proposed in this thesis an integration between recommender and adaptive strategies into a new process of adaptive recommendation. It is shown that such integration can improve recommendation accuracy and give good results to user retention, and interaction with the systems. To validate the approach, it is implemented a prototype to recommend movies to be used in a classroom. It is also collected some statistics about the 78 users who have participated of the experiment for evaluation of the new approach.

As one of the social relations, "distrust" also performs an important role in online Web sites. We also observe that distrust information can also be incorporated to improve recommendation quality. Hence, the last part of this thesis studies the problem on how to improve recommender system by considering explicit distrust information among users. We make the assumption that users' distrust relations can be interpreted as the "dissimilar" relations since user ui distrusts user ud indicates that user ui disagrees with most of the opinions issued by user ud. Based on this intuition, the distrust relations between users can be easily modeled by adding the regularization term into the objective functions of the user-item matrix factorization. The experiments on the Epinions dataset indicate that distrust information is at least as important as trust information. / However, the data sparsity problem of the involved user-item matrix seriously affects the recommendation quality. Many existing approaches to recommender systems cannot easily deal with users who have made very few ratings. The objective of this thesis is to study how to build effective and efficient approaches to improve the recommendation performance. / In this thesis, we first propose two collaborative filtering methods which only utilize the user-item matrix for recommendations. The first method is a neighborhood-based collaborative filtering method which designs an effective missing data prediction algorithm to improve recommendation quality, while the second one is a model-based collaborative filtering method which employs matrix factorization technique to make the recommendation more accurate. / In view of the exponential growth of information generated by online users, social contextual information analysis is becoming important for many Web applications. Hence, based on the assumption that users can be easily influenced by the friends they trust and prefer their friends' recommendations, we propose two recommendation algorithms by incorporating users' social trust information. These two methods are based on probabilistic matrix factorization. The complexity analysis indicates that our approaches can be applied to very large datasets since they scale linearly with the number of observations, while the experimental results show that our methods perform better than the state-of-the-art approaches. / Recommender Systems are becoming increasingly indispensable nowadays since they focus on solving the information overload problem, by providing users with more proactive and personalized information services. Typically, recommender systems are based on Collaborative Filtering, which is a technique that automatically predicts the interest of an active user by collecting rating information from other similar users or items. Due to their potential commercial values and the associated great research challenges, Recommender systems have been extensively studied by both academia and industry recently. / Ma, Hao. / "December 2009." / Advisers: Irwin King; Michael R. Lyu. / Source: Dissertation Abstracts International, Volume: 72-01, Section: B, page: . / Thesis (Ph.D.)--Chinese University of Hong Kong, 2010. / Includes bibliographical references (leaves 136-154). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Electronic reproduction. Ann Arbor, MI : ProQuest Information and Learning Company, [200-] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstract also in Chinese.

Mobile devices, such as smartphones and wearable devices, have become a popular platform to deliver recommendations and interact with users. To learn the decision rule of assigning recommendations, i.e. policy, neither one homogeneous policy for all users nor completely heterogeneous policy for each user is appropriate. Many attempts have been made to learn a policy for making recommendations using observational mobile health (mHealth) data. The majority of them focuses on a homogeneous policy, that is a one-fit-to-all policy for all users. It is a fair starting point for mHealth study, but it ignores the underlying user heterogeneity. Users with similar behavior pattern may have unobservable underlying heterogeneity. To solve this problem, we develop a personalized learning framework that models both population and personalized effect simultaneously. In the first part of this dissertation, we address the personalized policy learning problem using longitudinal mHealth application usage data. Personalized policy represents a paradigm shift from developing a single policy that may prescribe personalized decisions by tailoring. Specifically, we aim to develop the best policy, one per user, based on estimating random effects under generalized linear mixed model. With many random effects, we consider new estimation method and penalized objective to circumvent high-dimensional integrals for marginal likelihood approximation. We establish consistency and optimality of our method with endogenous application usage. We apply our method to develop personalized prompt schedules in 294 application users, with a goal to maximize the prompt response rate given past application usage and other contextual factors. We found the best push schedule given the same covariates varied among the users, thus calling for personalized policies. Using the estimated personalized policies would have achieved a mean prompt response rate of 23% in these users at 16 weeks or later: this is a remarkable improvement on the observed rate (11%), while the literature suggests 3%-15% user engagement at 3 months after download. The proposed method compares favorably to existing estimation methods including using the R function glmer in a simulation study. In the second part of this dissertation, we aim to solve a practical problem in the mHealth area. Low response rate has been a major issue that blocks researchers from collecting high quality mHealth data. Therefore, developing a prompting system is important to keep user engagement and increase response rate. We aim to learn personalized prompting time for users in order to gain a high response rate. An extension of the personalized learning algorithm is applied on the Intellicare data that incorporates penalties of the population effect parameters and personalized effect parameters into learning the personalized decision rule of sending prompts. The number of personalized policy parameters increases with sample size. Since there is a large number of users in the Intellicare data, it is challenging to estimate such high dimensional parameters. To solve the computational issue, we employ a bagging method that first bootstraps subsamples and then ensembles parameters learned from each subsample. The analysis of Intellicare data shows that sending prompts at a personalized hour helps achieve a higher response rate compared to a one-fit-to-all prompting hour.

With the dramatic growth of the Internet, it is much easier for us to acquire information than before. It is, however, relatively difficult to extract desired information through the huge information pool. One method is to rely on the search engines by analyzing the queried keywords to locate the relevant information. The other one is to recommend users what they may be interested in via recommender systems that analyze the users¡¦ past preferences or other users with similar interests to lessen our information processing loadings. Typical recommendation techniques are classified into content-based filtering technique and collaborative filtering (CF) technique. Several research works in literature have indicated that the performance of collaborative filtering is superior to that of content-based filtering in that it is subject to neither the content format nor users¡¦ past experiences. The collaborative filtering technique, however, has its own limitation of the sparsity problem. To relieve such a problem, researchers proposed several CF-typed variants, including item-based CF and trust-based CF. Few works in literature, however, focus on their performance comparison. The objective of this research is thus to evaluate both approaches under different settings such as the sparsity degrees, data scales, and number of neighbors to make recommendations. We conducted two experiments to examine their performance. The results show that trust-based CF is generally better than item-based CF in sparsity problem. Their difference, however, becomes insignificant with the sparsity decreasing. In addition, the computational time for trust-based CF increases more quickly than that for item-based CF, even though both exhibit exponential growths. Finally, the optimal number of nearest neighbors in both approaches does not heavily depend on the data scale but displays steady robustness.