In this thesis a number of models for recommender systems are explored, all using collaborative filtering to produce their recommendations. Extra focus is put on two models: Matrix Factorization, which is a linear model and Multi-Layer Perceptron, which is a non-linear model. With an additional purpose of training the models without collecting any sensitive data from the users, both models were implemented with a learning technique that does not require the server's knowledge of the users' data, called federated learning. The federated version of Matrix Factorization is already well-researched, and has proven not to protect the users' data at all; the data is derivable from the information that the users communicate to the server that is necessary for the learning of the model. However, on the federated Multi-Layer Perceptron model, no research could be found. In this thesis, such a model is therefore designed and presented. Arguments are put forth in support of the privacy preservability of the model, along with a proof of the user data not being analytically derivable for the central server.    In addition, new ways to further put the protection of the users' data on the test are discussed. All models are evaluated on two different data sets. The first data set contains data on ratings of movies and is called MovieLens 1M. The second is a data set that consists of anonymized fund transactions, provided by the Swedish bank SEB for this thesis. Test results suggest that the federated versions of the models can achieve similar recommendation performance as their non-federated counterparts.

Collaborative filtering is a popular technique to use behavior data in the form of user’s interactions with, or ratings of, items in a system to provide personalized recommendations of items to the user. This study compares three different state-of-the-art Recommendation System models that implement this technique, Matrix Factorization, Multi-layer Perceptron and Neural Matrix Factorization, using behavior data from a digital book platform for children. The field of Recommendation Systems is growing, and many platforms can benefit of personalizing the user experience and simplifying the use of the platforms. To perform a more complex comparison and introduce a new take on the models, this study proposes a new way to represent the behavior data as input to the models, i.e., to use the Term Frequency-Inverse Document Frequency (TFIDF) of occurrences of interactions between users and books, as opposed to the traditional binary representation (positive if there has been any interaction and negative otherwise). The performance is measured by extracting the last book read for each user, and evaluating how the models would rank that book for recommendations to the user. To assess the value of the models for the children’s reading platform, the models are also compared to the existing Recommendation System on the digital book platform. The results indicate that the Matrix Factorization model performs best out of the three models when using children’s reading behavior data. However, due to the long training process and larger set of hyperparameters to tune for the other two models, these may not have reached an optimal hyperparameter tuning, thereby affecting the comparison among the three state-of-the-art models. This limitation is further discussed in the study. All three models perform significantly better than the current system on the digital book platform. The models with the proposed representation using TF-IDF values show notable promise, performing better than the binary representation in almost all numerical metrics for all models. These results can suggest future research work on more ways of representing behavior data as input to these types of models. / Kollaborativ filtrering är en populär teknik för att använda beteendedata från användare i form av t.ex. interaktioner med, eller betygsättning av, objekt i ett system för att ge användaren personliga rekommendationer om objekt. I den här studien jämförs tre olika modeller av moderna rekommendationssystem som tillämpar denna teknik, matrisfaktorisering, flerlagersperceptron och neural matrisfaktorisering, med hjälp av beteendedata från en digital läsplattform för barn. Rekommendationssystem är ett växande område, och många plattformar kan dra nytta av att anpassa användarupplevelsen utifrån individen och förenkla användningen av plattformen. För att utföra en mer komplex jämförelse och introducera en ny variant av modellerna, föreslår denna studie ett nytt sätt att representera beteendedata som indata till modellerna, d.v.s. att använda termfrekvens med omvänd dokumentfrekvens (TF- IDF) av förekomster av interaktioner mellan användare och böcker, i motsats till den traditionella binära representationen (positiv om en tidigare interaktion existerar och negativ i annat fall). Prestandan mäts genom att extrahera den senaste boken som lästs för varje användare, och utvärdera hur högt modellerna skulle rangordna den boken i rekommendationer till användaren. För att värdesätta modellerna för plattformen med digitala böcker, så jämförs modellerna också med det befintliga rekommendationssystemet på plattformen. Resultaten tyder på att matrisfaktorisering-modellen presterar bäst utav de tre modellerna när man använder data från barns läsbeteende. På grund av den långa träningstiden och fler hyperparametrar att optimera för de andra två modellerna, kan det dock vara så att de inte har nått en optimal hyperparameterinställning, vilket påverkar jämförelsen mellan de tre moderna modellerna. Denna begränsning diskuteras ytterligare i studien. Alla tre modellerna presterar betydligt bättre än det nuvarande systemet på läsplattformen. Modellerna med den föreslagna representationen av TFIDF-värden visar sig mycket lovande och presterar bättre än den binära representationen i nästan alla numeriska mått för alla modeller. Dessa resultat kan ge skäl för framtida forskning av fler sätt att representera beteendedata som indata till denna typ av modeller.

Pendant la dernière décennie nous avons vu une transformation incroyable du monde de la musique qui est passé des cassettes et disques compacts à la musique numérique en ligne. Avec l'explosion de la musique numérique, nous avons besoin de systèmes de recommandation de musique pour choisir les chansons susceptibles d’être appréciés à partir de ces énormes bases de données en ligne ou personnelles. Actuellement, la plupart des systèmes de recommandation de musique utilisent l’algorithme de filtrage collaboratif ou celui du filtrage à base de contenu. Dans ce mémoire, nous proposons un algorithme hybride et original qui combine le filtrage collaboratif avec le filtrage basé sur étiquetage, amélioré par la technique de filtrage basée sur le contexte d’utilisation afin de produire de meilleures recommandations. Notre approche suppose que les préférences de l'utilisateur changent selon le contexte d'utilisation. Par exemple, un utilisateur écoute un genre de musique en conduisant vers son travail, un autre type en voyageant avec la famille en vacances, un autre pendant une soirée romantique ou aux fêtes. De plus, si la sélection a été générée pour plus d'un utilisateur (voyage en famille, fête) le système proposera des chansons en fonction des préférences de tous ces utilisateurs. L'objectif principal de notre système est de recommander à l'utilisateur de la musique à partir de sa collection personnelle ou à partir de la collection du système, les nouveautés et les prochains concerts. Un autre objectif de notre système sera de collecter des données provenant de sources extérieures, en s'appuyant sur des techniques de crawling et sur les flux RSS pour offrir des informations reliées à la musique tels que: les nouveautés, les prochains concerts, les paroles et les artistes similaires. Nous essayerons d’unifier des ensembles de données disponibles gratuitement sur le Web tels que les habitudes d’écoute de Last.fm, la base de données de la musique de MusicBrainz et les étiquettes des MusicStrands afin d'obtenir des identificateurs uniques pour les chansons, les albums et les artistes. / In the last decade we have seen an incredible transformation of the world of music, from cassette tapes and compact discs to online digital music. With the explosion of the digital music we need music recommender systems to select those probably preferred songs from these huge online or personal databases. Currently, most music recommender systems use either collaborative (social) filtering or content-based algorithms. In this work we propose an original hybrid collaborative and tag-based algorithm improved by the context-of-use filtering technique in order to generate better playlists. Our approach assumes that user preferences change depending on the context of use. For example, a user listen one kind of music while driving to work, another type while traveling with the family on vacation, another one in a romantic evening or at parties. Additionally, if the playlist was generated for more than one user (family trip, party) the system will propose songs depending on the preferences of all these users. The main goal of our system is to recommend music to the user from the personal music collection or new music from system music collection, new releases and incoming concerts. Another goal of our system it will be to gather data from external sources based on crawling techniques and RSS Feeds to provide music related content like: new releases, incoming concerts, lyrics, similar artists. We’ll try to interlink some free available datasets on the web like listening habits from Last.fm, music database from MusicBrainz and tags from MusicStrands in order to obtain unique identifiers for songs, albums and artists.

Cette thèse porte sur l'étude des Systèmes de Recommandation dans un cadre séquentiel, où les retours des utilisateurs sur des articles arrivent dans le système l'un après l'autre. Après chaque retour utilisateur, le système doit le prendre en compte afin d'améliorer les recommandations futures. De nombreuses techniques de recommandation ou méthodologies d'évaluation ont été proposées par le passé pour les problèmes de recommandation. Malgré cela, l'évaluation séquentielle, qui est pourtant plus réaliste et se rapproche davantage du cadre d'évaluation d'un vrai système de recommandation, a été laissée de côté. Le contexte séquentiel nécessite de prendre en considération différents aspects non visibles dans un contexte fixe. Le premier de ces aspects est le dilemme dit d'exploration vs. exploitation: le modèle effectuant les recommandations doit trouver le bon compromis entre recueillir de l'information sur les goûts des utilisateurs à travers des étapes d'exploration, et exploiter la connaissance qu'il a à l'heure actuelle pour maximiser le feedback reçu. L'importance de ce premier point est mise en avant à travers une première évaluation, et nous proposons une approche à la fois simple et efficace, basée sur la Factorisation de Matrice et un algorithme de Bandit Manchot, pour produire des recommandations appropriées. Le second aspect pouvant apparaître dans le cadre séquentiel surgit dans le cas où une liste ordonnée d'articles est recommandée au lieu d'un seul article. Dans cette situation, le feedback donné par l'utilisateur est multiple: la partie explicite concerne la note donnée par l'utilisateur concernant l'article choisi, tandis que la partie implicite concerne les articles cliqués (ou non cliqués) parmi les articles de la liste. En intégrant les deux parties du feedback dans un modèle d'apprentissage, nous proposons une approche basée sur la Factorisation de Matrice, qui peut recommander de meilleures listes ordonnées d'articles, et nous évaluons cette approche dans un contexte séquentiel particulier pour montrer son efficacité. / This thesis is dedicated to the study of Recommendation Systems under a sequential setting, where the feedback given by users on items arrive one after another in the system. After each feedback, the system has to integrate it and try to improve future recommendations. Many techniques or evaluation methods have already been proposed to study the recommendation problem. Despite that, such sequential setting, which is more realistic and represent a closer framework to a real Recommendation System evaluation, has surprisingly been left aside. Under a sequential context, recommendation techniques need to take into consideration several aspects which are not visible for a fixed setting. The first one is the exploration-exploitation dilemma: the model making recommendations needs to find a good balance between gathering information about users' tastes or items through exploratory recommendation steps, and exploiting its current knowledge of the users and items to try to maximize the feedback received. We highlight the importance of this point through the first evaluation study and propose a simple yet efficient approach to make effective recommendation, based on Matrix Factorization and Multi-Armed Bandit algorithms. The second aspect emphasized by the sequential context appears when a list of items is recommended to the user instead of a single item. In such a case, the feedback given by the user includes two parts: the explicit feedback as the rating, but also the implicit feedback given by clicking (or not clicking) on other items of the list. By integrating both feedback into a Matrix Factorization model, we propose an approach which can suggest better ranked list of items, and we evaluate it in a particular setting.

Pendant la dernière décennie nous avons vu une transformation incroyable du monde de la musique qui est passé des cassettes et disques compacts à la musique numérique en ligne. Avec l'explosion de la musique numérique, nous avons besoin de systèmes de recommandation de musique pour choisir les chansons susceptibles d’être appréciés à partir de ces énormes bases de données en ligne ou personnelles. Actuellement, la plupart des systèmes de recommandation de musique utilisent l’algorithme de filtrage collaboratif ou celui du filtrage à base de contenu. Dans ce mémoire, nous proposons un algorithme hybride et original qui combine le filtrage collaboratif avec le filtrage basé sur étiquetage, amélioré par la technique de filtrage basée sur le contexte d’utilisation afin de produire de meilleures recommandations. Notre approche suppose que les préférences de l'utilisateur changent selon le contexte d'utilisation. Par exemple, un utilisateur écoute un genre de musique en conduisant vers son travail, un autre type en voyageant avec la famille en vacances, un autre pendant une soirée romantique ou aux fêtes. De plus, si la sélection a été générée pour plus d'un utilisateur (voyage en famille, fête) le système proposera des chansons en fonction des préférences de tous ces utilisateurs. L'objectif principal de notre système est de recommander à l'utilisateur de la musique à partir de sa collection personnelle ou à partir de la collection du système, les nouveautés et les prochains concerts. Un autre objectif de notre système sera de collecter des données provenant de sources extérieures, en s'appuyant sur des techniques de crawling et sur les flux RSS pour offrir des informations reliées à la musique tels que: les nouveautés, les prochains concerts, les paroles et les artistes similaires. Nous essayerons d’unifier des ensembles de données disponibles gratuitement sur le Web tels que les habitudes d’écoute de Last.fm, la base de données de la musique de MusicBrainz et les étiquettes des MusicStrands afin d'obtenir des identificateurs uniques pour les chansons, les albums et les artistes. / In the last decade we have seen an incredible transformation of the world of music, from cassette tapes and compact discs to online digital music. With the explosion of the digital music we need music recommender systems to select those probably preferred songs from these huge online or personal databases. Currently, most music recommender systems use either collaborative (social) filtering or content-based algorithms. In this work we propose an original hybrid collaborative and tag-based algorithm improved by the context-of-use filtering technique in order to generate better playlists. Our approach assumes that user preferences change depending on the context of use. For example, a user listen one kind of music while driving to work, another type while traveling with the family on vacation, another one in a romantic evening or at parties. Additionally, if the playlist was generated for more than one user (family trip, party) the system will propose songs depending on the preferences of all these users. The main goal of our system is to recommend music to the user from the personal music collection or new music from system music collection, new releases and incoming concerts. Another goal of our system it will be to gather data from external sources based on crawling techniques and RSS Feeds to provide music related content like: new releases, incoming concerts, lyrics, similar artists. We’ll try to interlink some free available datasets on the web like listening habits from Last.fm, music database from MusicBrainz and tags from MusicStrands in order to obtain unique identifiers for songs, albums and artists.

推薦系統為網站提升使用者滿意度、減少使用者所花費的時間並且替網站提供方提升銷售，是現在網站中不可或缺的要素，而推薦系統的研究集中在娛樂項目，學術研究論文推薦系統的研究有限。若能給予有價值的相關文獻，提供協助，無疑是加速進步的速度。 在過去的研究中，為了達到個人化目的所使用的方法，都有不可避免或未解決的缺點，2002年美國研究圖書館協會提出布達佩斯開放獲取計劃(Budapest Open Access Initiative)，不要求使用者註冊帳號與支付款項就能取得研究論文全文，這樣的做法使期刊走向開放的風氣開始盛行，時至今日，開放獲取對學術期刊網站帶來重大的影響。在這樣的時空背景之下，本研究提出一個適用於學術論文之推薦機制，以FP-Growth演算法與協同過濾做為推薦方法的基礎，消弭過去研究之缺點，並具個人化推薦的優點，經實驗驗證後，證實本研究所提出的推薦架構具有良好的成效。 / Recommendation system is used in many field like movie, music, electric commerce and library. It’s not only save customers’ time but also raise organizations’ efficient. Recommended system is an essential element in a website. Some methods have been developed for recommended system, but they are primarily focused on content or collaboration-based mechanisms. For academic research, it is very important that relevant literature can be provided to researchers when they conduct literature review. Previous research indicates that there are inevitable or unsolved shortcomings in existing methods such as cold starts. Association of Research Libraries purpose “Budapest Open Access Initiative” that is advocate open access concept. Open access means that users can get full paper without register and pay fee. It’s a major impact to academic journal website. In this space-time background, we propose a hybrid recommendation mechanism that takes into consideration the nature of recommendation academic papers to mitigate the shortcomings of existing methods.

Cette thèse traite des systèmes de recommandation automatiques. Les moteurs de recommandation automatique sont des systèmes qui permettent, par des techniques de data mining, de recommander automatiquement à des clients, en fonction de leurs consommations passées, des produits susceptibles de les intéresser. Ces systèmes permettent par exemple d'augmenter les ventes sur des sites web marchands : le site Amazon a une stratégie marketing en grande partie basée sur la recommandation automatique. Amazon a popularisé l'usage de la recommandation automatique par la célèbre fonction de recommandation que nous qualifions d'item-to-items, le fameux : " les personnes qui ont vu/acheté cet articles ont aussi vu/acheté ces articles. La contribution centrale de cette thèse est d'analyser les systèmes de recommandation automatiques dans le contexte industriel, et notamment des besoins marketing, et de croiser cette analyse avec les travaux académiques. / This thesis deals with automatic recommendation systems. Automatic recommendation systems are systems that allow, through data mining techniques, to recommend automatically to users, based on their past consumption, items that may interest them. These systems allow for example to increase sales on e-commerce websites: the Amazon site has a marketing strategy based mainly on the recommendation. Amazon has popularized the use of automatic recommendation based on the recommendation function that we call item-to-items, the famous "people who have seen / bought this product have also seen / bought these articles". The central contribution of this thesis is to analyze the automatic recommendation systems in the industrial context, including marketing needs, and to cross this analysis with academic works.

Mängden data som transporteras över Internet idag är stor. Vilket innebär att det finns ett överflöd av information och ett behov för att urskilja relevant innehåll mot irrelevant. För att uppnå detta används rekommendationssystem som i sin tur använder olika filtreringstekniker. Det finns tre huvudtyper av tekniker, innehållsbaserad kollaborativ filtrering och hybrid tekniker. Syftet med studien är att jämföra olika filtreringstekniker och undersöka hur teknikerna förhåller sig till träffsäkerhet mot testset, beräkningsbelasning och användarnöjdhet. För att genomföra detta på ett nyanserat sätt har en strukturerad litteraturstudie genomförts där sju olika steg för inhämtning och analys av dokument gjorts. En kvalitativ metanalys genomfördes på de 28 utvalda tidskriftsartiklarna. IFT rekommendationssystem var den rekommendationsteknik som anses minst effektiv utifrån den definition som studien utgått från. KFT och hybrid rekommendationssystem är de som är mest effektiva enligt denna studie. Hybrid system kan ta vara på fördelar från alla andra tekniker och kan synergiskt också motverka vissa av teknikernas förknippade brister, mot kostnaden av att implementations-komplexiteten ökar.

In this thesis, various machine learning domains have been combined in order to build a video recommender system that is based on object detection. The work combines two extensively studied research fields, recommender systems and computer vision, that also are rapidly growing and popular techniques on commercial markets. To investigate the performance of the approach, three different content-based recommender systems have been implemented at Spotify, which are based on the following video features: object detections, titles and descriptions, and user preferences. These systems have then been evaluated and compared against each other together with their hybridized result. Two algorithms have been implemented, the prediction and the top-N algorithm, where the former is the more reliable source for evaluating the system's performance. The evaluation of the system shows that the overall performance scores for predicting values of the users' liked and disliked videos are in the range from about 40 % to 70 % for the prediction algorithm and from about 15 % to 70 % for the top-N algorithm. The approach based on object detection performs worse in comparison to the other approaches. Hence, there seems to be is a low correlation between the user preferences and the video contents in terms of object detection data. Therefore, this data is not very suitable for describing the content of videos and using it in the recommender system. However, the results of this study cannot be generalized to apply for other systems before the approach has been evaluated in other environments and for various data sets. Moreover, there are plenty of room for refinements and improvements to the system, as well as there are many interesting research areas for future work.

Conselho Nacional de Desenvolvimento Científico e Tecnológico / Hoje em dia, a quantidade de usuários que utilizam sites de comércio eletrônico para realizar compras está aumentando muito, principalmente devido à facilidade e rapidez. Muitos sites de comércio eletrônico, diferentemente das lojas físicas, disponibilizam aos seus usuários uma grande variedade de produtos e serviços, e os usuários podem ter muita dificuldade em encontrar produtos de sua preferência. Normalmente, a preferência por um produto pode ser influenciada pela aparência visual da imagem do produto. Neste contexto, os Sistemas de Recomendação de produtos que estão associados a Imagens (SRI) tornaram-se indispensáveis para ajudar os usuários a encontrar produtos que podem ser, eventualmente, agradáveis ou úteis para eles. Geralmente, os SRI usam o comportamento passado dos usuários (cliques, compras, críticas, avaliações, etc.) e/ou atributos de produtos para definirem as preferências dos usuários. Um dos principais desafios enfrentados em SRI é a necessidade de o usuário fornecer algumas informações sobre suas preferências sobre os produtos, a fim de obter novas recomendações do sistema. Infelizmente, os usuários nem sempre estão dispostos a fornecer tais informações de forma explícita. Assim, a fim de lidar com esse desafio, os métodos para obtenção de informações de forma implícita do usuário são desejáveis. Neste trabalho, propõe-se investigar em que medida informações sobre atenção visual do usuário podem ajudar a melhorar a predição de avaliação e consequentemente produzir SRI mais precisos. É também objetivo deste trabalho o desenvolvimento de dois novos métodos, um método baseado em Filtragem Colaborativa (FC) que combina avaliações e dados de atenção visual para representar o comportamento passado dos usuários, e outro método baseado no conteúdo dos itens, que combina atributos textuais, características visuais e dados de atenção visual para compor o perfil dos itens. Os métodos propostos foram avaliados em uma base de imagens de pinturas e uma base de imagens de roupas. Os resultados experimentais mostram que os métodos propostos neste trabalho possuem ganhos significativos em predição de avaliação e precisão na recomendação quando comparados ao estado-da-arte. Vale ressaltar que as técnicas propostas são flexíveis, podendo ser utilizadas em outros cenários que exploram a atenção visual dos itens recomendados. / Nowadays, the amount of users using e-commerce sites for shopping is greatly increasing, mainly due to the easiness and rapidity of this way of consumption. Many e-commerce sites, differently from physical stores, can offer their users a wide range of products and services, and the users can find it very difficult to find products of your preference. Typically, your preference for a product can be influenced by the visual appearance of the product image. In this context, Image Recommendation Systems (IRS) have become indispensable to help users to find products that may possibly pleasant or be useful to them. Generally, IRS use past behavior of users (clicks, purchases, reviews, ratings, etc.) and/or attributes of the products to define the preferences of users. One of the main challenges faced by IRS is the need of the user to provide some information about his / her preferences on products in order to get further recommendations from the system. Unfortunately, users are not always willing to provide such information explicitly. So, in order to cope with this challenge, methods for obtaining user’s implicit feedback are desirable. In this work, the author propose an investigation to discover to which extent information concerning user visual attention can help improve the rating prediction hence produce more accurate IRS. This work proposes to develop two new methods, a method based on Collaborative Filtering (CF) which combines ratings and data visual attention to represent the past behavior of users, and another method based on the content of the items, which combines textual attributes, visual features and visual attention data to compose the profile of the items. The proposed methods were evaluated in a painting dataset and a clothing dataset. The experimental results show significant improvements in rating prediction and precision in recommendation when compared to the state-of-the-art. It is worth mentioning that the proposed techniques are flexible and can be applied in other scenarios that exploits the visual attention of the recommended items. / Tese (Doutorado)