L'objectif principal de la thèse est de proposer une méthode de recommandation prenant en compte la vie privée des utilisateurs ainsi que l'évolutivité du système. Pour atteindre cet objectif, une technique hybride basée sur le filtrage par contenu et le filtrage collaboratif est utilisée pour atteindre un modèle précis de recommandation, sous la pression des mécanismes visant à maintenir la vie privée des utilisateurs. Les contributions de la thèse sont trois : Tout d'abord, un modèle de filtrage collaboratif est défini en utilisant agent côté client qui interagit avec l'information sur les éléments, cette information est stockée du côté du système de recommandation. Ce modèle est augmenté d’un modèle hybride qui comprend une stratégie basée sur le filtrage par contenu. En utilisant un modèle de la connaissance basée sur des mots clés qui décrivent le domaine de l'article filtré, l'approche hybride augmente la performance de prédiction des modèles sans élever l’effort de calcul, dans un scenario du réglage de démarrage à froid. Finalement, certaines stratégies pour améliorer la protection de la vie privée du système de recommandation sont introduites : la génération de bruit aléatoire est utilisée pour limiter les conséquences éventuelles d'une attaque lorsque l'on observe en permanence l'interaction entre l'agent côté client et le serveur, et une stratégie basée sur la liste noire est utilisée pour s’abstenir de révéler au serveur des interactions avec des articles que l'utilisateur considère comme pouvant transgresser sa vie privée. L'utilisation du modèle hybride atténue l'impact négatif que ces stratégies provoquent sur la performance prédictive des recommandations. / The main objective of this thesis is to propose a recommendation method that keeps in mind the privacy of users as well as the scalability of the system. To achieve this goal, an hybrid technique using content-based and collaborative filtering paradigms is used in order to attain an accurate model for recommendation, under the strain of mechanisms designed to keep user privacy, particularly designed to reduce the user exposure risk. The thesis contributions are threefold : First, a Collaborative Filtering model is defined by using client-side agent that interacts with public information about items kept on the recommender system side. Later, this model is extended into an hybrid approach for recommendation that includes a content-based strategy for content recommendation. Using a knowledge model based on keywords that describe the item domain, the hybrid approach increases the predictive performance of the models without much computational effort on the cold-start setting. Finally, some strategies to improve the recommender system's provided privacy are introduced: Random noise generation is used to limit the possible inferences an attacker can make when continually observing the interaction between the client-side agent and the server, and a blacklisted strategy is used to refrain the server from learning interactions that the user considers violate her privacy. The use of the hybrid model mitigates the negative impact these strategies cause on the predictive performance of the recommendations.

Cette thèse présente les résultats d'un projet de recherche multidisciplinaire (Agorantic) sur les systèmes de recommandation. Le but de ce travail était de proposer de nouvelles fonctionnalités qui peuvent rendre les systèmes de recommandations (RS) plus attrayants que ceux existants. Nous proposons également une nouvelle approche et une réflexion sur l'évaluation. Dans la conception du système, nous avons voulu répondre aux préoccupations suivantes: 1. Les gens s'habituent à recevoir des recommandations. Néanmoins, après quelques mauvaises recommandations, les utilisateurs ne seront plus convaincus par les RS. 2. En outre, si ces suggestions viennent sans explication, pourquoi les gens devraient les suivre ? 3. Le fait que la perception, les goûts et les humeurs des utilisateurs goûts varient au fil du temps est bien connue. Pourtant, la plupart des systèmes de recommandation ne parviennent pas à offrir le bon niveau de «réactivité» que les utilisateurs attendent, c'est à dire la capacité de détecter et d'intégrer des changements dans les besoins, les préférences, la popularité, etc. Recommander un film une semaine après sa sortie pourrait être trop tard. 4. L'utilisateur pourrait être intéressé par des articles moins populaires (dans la «longue traine»), c'est à dire des recommandations moins systématiques. Pour répondre à ces questions clés, nous avons conçu un nouveau système de recommandation sémantique et adaptatif (SRAS), comportant trois fonctionnalités innovantes, à savoir l'argumentation, l'adaptation dynamique et un algorithme d'appariement. • Adaptation dynamique: le système est mis à jour de façon continue, à chaque nouvelle note / évènement. (Chapitre 4) • Argumentation: chaque recommandation présente les raisons qui ont conduit à cette recommandation. Cela peut être considéré comme une première étape vers une argumentation plus sophistiqué. Notre volonté est de rendre les utilisateurs plus responsables de leur choix, en leur donnant le maximum d'informations. (Chapitre 5) • Algorithme d'appariement: permet aux articles les moins populaires d'être recommandés aux utilisateurs. (Chapitre 6) Nous avons conçu un nouveau système de recommandation capable de générer des recommandations textuellement bien argumentées dans lequel l'utilisateur final aura plusieurs éléments pour faire un choix éclairé. En outre, les paramètres du système sont dynamiquement et continuellement mis à jour, afin de fournir des recommandations et des arguments en la phase avec le passé très récent. Nous avons inclus un niveau sémantique, c'est à dire les mots, termes et expressions comme ils sont naturellement exprimés dans les commentaires utilisateurs. Nous n'utilisons pas d'étiquettes ou lexique pré-déterminé. Les performances de notre système sont comparables à l'état de l'art. En outre, le fait qu'il génère un argumentaire le rend encore plus attrayant et pourrait renforcer la fidélité des utilisateurs / This thesis presents the results of a multidisciplinary research project (Agorantic) on Recommender Systems. The goal of this work was to propose new features that may render recommender systems (RS) more attractive than the existing ones. We also propose a new approach to and a reflection about evaluation. In designing the system, we wanted to address the following concerns: 1. People are getting used to receive recommendations. Nevertheless, after a few bad recommendations, users will not be convinced anymore by the RS. 2. Moreover, if these suggestions come without explanations, why people should trust it? 3. The fact that item perception and user tastes and moods vary over time is well known. Still, most recommender systems fail to offer the right level of “reactivity” that users are expecting, i.e. the ability to detect and to integrate changes in needs, preferences, popularity, etc. Suggesting a movie a week after its release might be too late. In the same vein, it could take only a few ratings to make an item go from not advisable to advisable, or the other way around. 4. Users might be interested in less popular items (in the ” long tail”) and want less systematic recommendations. To answer these key issues, we have designed a new semantic and adaptive recommender system (SARS) including three innovative features, namely Argumentation, Dynamic Adaptation and a Matching Algorithm. • Dynamic Adaptation: the system is updated in a continuous way, as each new review/rating is posted. (Chapter 4) • Argumentation: each recommendation relies on and comes along with some keywords, providing the reasons that led to that recommendation. This can be seen as a first step towards a more sophisticated argumentation. We believe that, by making users more responsible for their choices, it will prevent them from losing confidence in the system. (Chapter 5) • Matching Algorithm: allows less popular items to be recommended by applying a match- ing game to users and items preferences. (Chapter 6) The system should be sensed as less intrusive thanks to relevant arguments (well-chosen words) and less responsible to unsatisfaction of the customers. We have designed a new recommender system intending to provide textually well-argued recommendations in which the end user will have more elements to make a well-informed choice. Moreover, the system parameters are dynamically and continuously updated, in order to pro- vide recommendations and arguments in phase with the very recent past. We have included a semantic level, i.e words, terms and phrases as they are naturally expressed in reviews about items. We do not use tags or pre-determined lexicon. The performances of our system are comparable to the state of the art. In addition, the fact that it provides argumentations makes it even more attractive and could enhance customers loyalty

Les systèmes de recommandation se sont révélés être des outils efficaces pour aider les utilisateurs à faire face à la surcharge informationnelle. D’importants progrès ont été réalisés dans le domaine durant les deux dernières décennies, menant en particulier à l’exploitation de l’information contextuelle pour modéliser l’aspect dynamique des utilisateurs et des articles. La définition traditionnelle du contexte, adoptée dans la plupart des systèmes de recommandation contextuels, ne répond pas à plusieurs contraintes rencontrées dans les applications du monde réel. Dans cette thèse, nous abordons les problèmes de recommandation en présence d’informations contextuelles partiellement observables et d’informations contextuelles non observables dans deux applications particulières, la recommandation d’hôtels et la recommandation en ligne, remettant en question plusieurs aspects de la définition traditionnelle du contexte, notamment l'accessibilité, la pertinence, l'acquisition et la modélisation.La première partie de la thèse étudie le problème de recommandation d’hôtels qui souffre du démarrage à froid continu, limitant la performance des approches classiques de recommandation. Le voyage n’est pas une activité fréquente et les utilisateurs ont tendance à adopter des comportements diversifiés en fonction de leurs situations spécifiques. Après une analyse du comportement des utilisateurs dans ce domaine, nous proposons de nouvelles approches de recommandation intégrant des informations contextuelles partiellement observables affectant les utilisateurs. Nous montrons comment cela contribue à améliorer la qualité des recommandations.La deuxième partie de la thèse aborde le problème de recommandation en ligne en présence de flux de données où les observations apparaissent continûment à haute fréquence. Nous considérons que les utilisateurs et les articles reposent sur des informations contextuelles non observables par le système et évoluent de façons différentes à des rythmes différents. Nous proposons alors d’effectuer de la détection active de changements et d’assurer la mise à jour des modèles en temps réel. Nous concevons de nouvelles méthodes qui s’adaptent aux changements qui apparaissent au niveau des préférences des utilisateurs et des perceptions et descriptions des articles, et montrons l’importance de la recommandation adaptative en ligne pour garantir de bonnes performances au cours du temps. / Recommender systems have proven to be valuable tools to help users overcome the information overload, and significant advances have been made in the field over the last two decades. In particular, contextual information has been leveraged to model the dynamics occurring within users and items. Context is a complex notion and its traditional definition, which is adopted in most recommender systems, fails to cope with several issues occurring in real-world applications. In this thesis, we address the problems of partially observable and unobservable contexts in two particular applications, hotel recommendation and online recommendation, challenging several aspects of the traditional definition of context, including accessibility, relevance, acquisition, and modeling.The first part of the thesis investigates the problem of hotel recommendation which suffers from the continuous cold-start issue, limiting the performance of classical approaches for recommendation. Traveling is not a frequent activity and users tend to have multifaceted behaviors depending on their specific situation. Following an analysis of the user behavior in this domain, we propose novel recommendation approaches integrating partially observable context affecting users and we show how it contributes in improving the recommendation quality.The second part of the thesis addresses the problem of online adaptive recommendation in streaming environments where data is continuously generated. Users and items may depend on some unobservable context and can evolve in different ways and at different rates. We propose to perform online recommendation by actively detecting drifts and updating models accordingly in real-time. We design novel methods adapting to changes occurring in user preferences, item perceptions, and item descriptions, and show the importance of online adaptive recommendation to ensure a good performance over time.

Research on social networks and understanding the interactions of the users can be modeled as a task of graph mining, such as predicting nodes and edges in networks.Dealing with such unstructured data in large social networks has been a challenge for researchers in several years. Neural Networks have recently proven very successful in performing predictions on number of speech, image, and text data and have become the de facto method when dealing with such data in a large volume. Graph NeuralNetworks, however, have only recently become mature enough to be used in real large-scale graph prediction tasks, and require proper structure and data modeling to be viable and successful. In this research, we provide a new modeling of the social network which captures the attributes of the nodes from various dimensions. We also introduce the Neural Network architecture that is required for optimally utilizing the new data structure. Finally, in order to provide a hot-start for our model, we initialize the weights of the neural network using a pre-trained graph embedding method. We have also developed a new graph embedding algorithm. We will first explain how previous graph embedding methods are not optimal for all types of graphs, and then provide a solution on how to combat those limitations and come up with a new graph embedding method.

Cette thèse s'articule autour des problèmes d'optimisation à grande échelle, et plus particulièrement autour des méthodes de factorisation matricielle sur des problèmes de grandes tailles. L'objectif des méthodes de factorisation de grandes matrices est d'extraire des variables latentes qui permettent d'expliquer les données dans un espace de dimension réduite. Nous nous sommes intéressés au domaine d'application de la recommandation et plus particulièrement au problème de prédiction de préférences d'utilisateurs.Dans une contribution, nous nous sommes intéressés à l'application de méthodes de factorisation dans un environnement de recommandation contextuelle et notamment dans un contexte social.Dans une seconde contribution, nous nous sommes intéressés au problème de sélection de modèle pour la factorisation où l'on cherche à déterminer de façon automatique le rang de la factorisation par estimation de risque. / This thesis focuses on large scale optimization problems and especially on matrix factorization methods for large scale problems. The purpose of such methods is to extract some latent variables which will explain the data in smaller dimension space. We use our methods to address the problem of preference prediction in the framework of the recommender systems. Our first contribution focuses on matrix factorization methods applied in context-aware recommender systems problems, and particularly in socially-aware recommandation.We also address the problem of model selection for matrix factorization which ails to automatically determine the rank of the factorization.

Social shopping is one of the latest trends on the Internet. Websites dedicated to social networking with a focus on shopping have been emerging on the web for a few years. The basic idea is that consumers are looking for product information on the Internet and social shopping sites provide a place for consumers to find this information from other consumers. These sites provide a place for their users to engage in socialization and shopping simultaneously, sometimes following recommendations of premier users, who are labeled from other users. However, purchases aren't made through these sites. So, there may still be something missing from the experience. For these sites, social pricing mechanisms may be implemented to provide revenue. Major ecommerce websites have begun focusing on increasing social features throughout the transaction process. For example, more websites are including ratings, reviews and recommendations of products and services by other consumers. However, pure ecommerce websites do not provide functionality that allows consumers to communicate in real time. Hence, there are some features missing from the social experience. Also, the social functionality included in pure e-commerce websites, tends to be utilized for the benefit of the Web site, as opposed to the consumers. Both social shopping sites and ecommerce sites have seen independently successful though few sites have been able to truly integrate these together at this point. It may be more beneficial to the end user if these sites could work in unison. This thesis is an exploratory study of the emerging social shopping phenomenon. The contributions of this work include analysis of the social shopping phenomenon and identifying metrics and Web sites that incorporate social shopping, a survey of academic literature related to social shopping and social pricing and a review of current recommender system algorithms with a discussion on how to incorporate social networking data into the algorithms to improve recommendations. Improvement suggestions include incorporating customer purchase history with social networking information. Potential future research ideas are included.

Recommender systems are an integral part of almost all modern e-commerce companies. They contribute significantly to the overall customer satisfaction by helping the user discover new and relevant items, which consequently leads to higher sales and stronger customer retention. It is, therefore, not surprising that large e-commerce shops like Amazon or streaming platforms like Netflix and Spotify even use multiple recommender systems to further increase user engagement. Finding the most relevant items for each user is a difficult task that is critically dependent on the available user feedback information. However, most users typically interact with products only through noisy implicit feedback, such as clicks or purchases, rather than providing explicit information about their preferences, such as product ratings. This usually makes large amounts of behavioural user data necessary to infer accurate user preferences. One popular approach to make the most use of both forms of feedback is called collaborative filtering. Here, the main idea is to compare individual user behaviour with the behaviour of all known users. Although there are many different collaborative filtering techniques, matrix factorization models are among the most successful ones. In contrast, while neural networks are nowadays the state-of-the-art method for tasks such as image recognition or natural language processing, they are still not very popular for collaborative filtering tasks. Therefore, the main focus of this thesis is the derivation of multiple wide neural network architectures to mimic and extend matrix factorization models for various collaborative filtering problems and to gain insights into the connection between these models. The basics of the proposed architecture are wide and shallow feedforward neural networks, which will be established for rating prediction tasks on explicit feedback datasets. These networks consist of large input and output layers, which allow them to capture user and item representation similar to matrix factorization models. By deriving all weight updates and comparing the structure of both models, it is proven that a simplified version of the proposed network can mimic common matrix factorization models: a result that has not been shown, as far as we know, in this form before. Additionally, various extensions are thoroughly evaluated. The new findings of this evaluation can also easily be transferred to other matrix factorization models. This neural network architecture can be extended to be used for personalized ranking tasks on implicit feedback datasets. For these problems, it is necessary to rank products according to individual preferences using only the provided implicit feedback. One of the most successful and influential approaches for personalized ranking tasks is Bayesian Personalized Ranking, which attempts to learn pairwise item rankings and can also be used in combination with matrix factorization models. It is shown, how the introduction of an additional ranking layer forces the network to learn pairwise item rankings. In addition, similarities between this novel neural network architecture and a matrix factorization model trained with Bayesian Personalized Ranking are proven. To the best of our knowledge, this is the first time that these connections have been shown. The state-of-the-art performance of this network is demonstrated in a detailed evaluation. The most comprehensive feedback datasets consist of a mixture of explicit as well as implicit feedback information. Here, the goal is to predict if a user will like an item, similar to rating prediction tasks, even if this user has never given any explicit feedback at all: a problem, that has not been covered by the collaborative filtering literature yet. The network to solve this task is composed out of two networks: one for the explicit and one for the implicit feedback. Additional item features are learned using the implicit feedback, which capture all information necessary to rank items. Afterwards, these features are used to improve the explicit feedback prediction. Both parts of this combined network have different optimization goals, are trained simultaneously and, therefore, influence each other. A detailed evaluation shows that this approach is helpful to improve the network's overall predictive performance especially for ranking metrics.

The research aims to focus on Recurrent Neural Networks (RNN) and its application to the session-aware recommendations empowered by implicit user feedback and content-based metadata. To investigate the promising architecture of RNN, we implement seven different models utilizing various types of implicit feedback and content information. Our results showed that using RNN with complex implicit feedback increases the next-item prediction comparing the baseline models like Cosine Similarity, Doc2Vec, and Item2Vec.

In modern times, live video streaming events in companies has become an increasingly relevantmethod for communications. As a platform provider for these events, being able to deliverrelevant recommendations for event scheduling times to users is an important feature. A systemproviding relevant recommendations to users can be described as a recommender system.Recommender systems usually face issues such as having to be trained purely offline, astraining the system online can be costly or time-consuming, requiring manual user feedback.While many solutions and advancements have been made in recommender systems over theyears, such as contributions in the Netflix Prize, it still continues to be an active research topic.This work aims at designing a recommender system which observes users' past sequentialscheduling behavior to provide relevant recommendations for scheduling upcoming live videoevents. The developed recommender system uses reinforcement learning as a model, withcomponents such as a generative model to help it learn from offline data.

The abundance of information available on the internet necessitates means of quickly finding what is relevant for the individual user. To this end, there has been much research concerning recommender systems and lately specifically methods using deep learning for such systems. This work proposes a Multi-Armed Bandit as a recommender for suggested sites on a browser start page. The system is compared to a pre-existing baseline and does not manage to outperform it in the setting used in controlled experiments. A Neural Collaborative Filtering system is then constructed using a stacked autoencoder and is used to produce user preference vectors that are inserted in the bandit in the hope of improving its performance. Analysis indicates that the bandit solution works better as the number of items grows. The user-informed initialisation used in this work shows a trend of improving over a randomly-initialised bandit, but results are inconclusive. This work also contributes an analysis of the problem domain including which factors impact the performance on the model training for preference vectors, and the performance of the bandit algorithms.