Algorithms for repeated (or “online”) decision-making are predominantly studied under the assumption that feedback is either statistical (determined by fixed probability distributions) or adversarial (changing over time in a potentially worst-case manner). Both of these assumptions ignore a phenomenon commonly present in repeated interactions with other agents, in which the space of our possible future outcomes is shaped in a structured and potentially predictable manner by our history of prior decisions. In this thesis, we consider online decision problems where the feedback model is adaptive rather than purely statistical or adversarial. One such example is a repeated game played against an opponent who uses a learning algorithm of their own; here, we give a characterization of possible outcome spaces which unifies disparate equilibrium notions, and serves as a basis for designing new algorithms. We then consider the task of providing recommendations to an agent whose preferences adapt based on the recommendation history, where we explore algorithmic tradeoffs in terms of the structure of this adaptivity pattern. We conclude by offering a general framework and algorithmic toolkit for approaching adaptive problems of this form.

Für viele Krankheitsbilder und Indikationen ist ein breites Spektrum an Arzneimitteln und Arzneimittelkombinationen verfügbar. Darüber hinaus stellen Therapieziele oft Kompromisse zwischen medizinischen Zielstellungen und Präferenzen und Erwartungen von Patienten dar, um Zufriedenheit und Adhärenz zu gewährleisten. Die Auswahl der optimalen Therapieoption kann daher eine große Herausforderung für den behandelnden Arzt darstellen. Klinische Entscheidungsunterstützungssysteme, die Wirksamkeit oder Risiken unerwünschter Arzneimittelwirkung für Behandlungsoptionen vorhersagen, können diesen Entscheidungsprozess unterstützen und \linebreak Leitlinien-basierte Empfehlungen ergänzen, wenn Leitlinien oder wissenschaftliche Literatur fehlen oder ungeeignet sind. Bis heute sind keine derartigen Systeme verfügbar. Im Rahmen dieser Arbeit wird die Anwendung von Methoden aus der Domäne der Recommender Systems (RS) und des Maschinellen Lernens (ML) in solchen Unterstützungssystemen untersucht. Aufgrund ihres erfolgreichen Einsatzes in anderen Empfehlungssystemen und der einfachen Interpretierbarkeit werden zum einen Nachbarschafts-basierte Collaborative Filter (CF) an die besonderen Anforderungen und Herausforderungen der Therapieempfehlung angepasst. Zum anderen werden ein Modell-basierter CF-Ansatz (SLIM) und ein ML Algorithmus (GBM) erprobt. Alle genannten Ansätze werden anhand eines exemplarischen Therapieempfehlungssystems evaluiert, das auf die Behandlung der Autoimmunkrankheit Psoriasis abzielt. Um das Risiko der Empfehlung kontraindizierter oder gar gesundheitsgefährdender Medikamente zu reduzieren, werden Regeln aus evidenzbasierten Leitlinien und Expertenempfehlungen implementiert, um solche Therapieoptionen aus den Empfehlungslisten herauszufiltern. Insbesondere die Nachbarschafts-basierten CF-Algorithmen zeigen insgesamt kleine durchschnittliche Abweichungen zwischen geschätztem und tatsächlichem Therapie-Outcome. Auch die aus den Outcome-Schätzungen abgeleiteten Empfehlungen zeigen eine hohe Übereinstimmung mit der tatsächlich angewandten Behandlung. Die Modell-basierten Ansätze sind den Nachbarschafts-basierten Ansätzen insgesamt unterlegen, was auf den begrenzten Umfang der verfügbaren Trainingsdaten zurückzuführen ist und die Generalisierungsfähigkeit der Modelle erschwert. Im Vergleich mit menschlichen Experten sind alle untersuchten Algorithmen jedoch hinsichtlich Übereinstimmung mit der tatsächlich angewandten Therapie unterlegen. Eine objektive und effiziente Bewertung des Behandlungserfolgs kann als Voraussetzung für ein erfolgreiches ``Krankheitsmanagement'' angesehen werden. Daher wird in weiteren Untersuchungen für ausgwählten klinische Anwendungen der Einsatz von ML Methoden zur automatischen Quantifizierung von Gesunheitszustand und Therapie-Outcome erprobt. Zusätzlich, als weitere Quelle für Informationen über Therapiewirksamkeiten, wird der Einsatz von Sentiment Analysis Methoden zur Extraktion solcher Informationen aus Medikamenten-Bewertungen untersucht. / Under most medical conditions and indications, a great variety of pharmaceutical drugs and drug combinations are available. Beyond that, trade-offs need to be found between the medical requirements and the patients' preferences and expectations in order to support patients’ satisfaction and adherence to treatments. As a consequence, the selection of an optimal therapy option for an individual patient poses a challenging task to prescribers. Clinical Decision Support Systems (CDSSs), which predict outcome as effectiveness and risk of adverse effects for available treatment options, can support this decision-making process and complement guideline-based decision-making where evidence from scientific literature is missing or inappropriate. To date, no such systems are available. Within this work, the application of methods from the Recommender Systems (RS) domain and Machine Learning (ML) in such decision support systems is studied. Due to their successful application in other recommender systems and good interpretability, neighborhood-based CF algorithms are transferred to the medical domain and are adapted to meet the requirements and challenges of the therapy recommendation task. Moreover, a model-based CF method (SLIM) and a state of the art ML algorithm (GBM) are employed. All algorithms are evaluated in an exemplary therapy recommender system, targeting the treatment of the autoimmune skin disease Psoriasis. In order to reduce the risk of recommending contraindicated or even health-endangering drugs, rules derived from evidence-based guidelines and expert recommendations are implemented to filter such options from the recommendation lists. Especially the neighborhood-based CF algorithms show small average errors between estimated and observed outcome. Also, the recommendations derived from outcome estimates show high agreement with the ground truth. The performance of both model-based approaches is inferior to the neighborhood-based recommender. This is primarily assumed to be due to the limited training data sizes, which renders generalizability of the learned models difficult. Compared with recommendations provided by various experts, all proposed approaches are, however, inferior in terms of agreement with the ground truth. An objective and efficient assessment of treatment response can be regarded a prerequisite for successful ``disease management''. Therefore, the use of ML methods for the automatic quantification of health status and therapy outcome for selected clinical applications is investigated in further experiments. Moreover, as additional source of information about drug effectiveness, the use of Sentiment Analysis, in order to extract such information from drug reviews, is investigated.

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal.

This work deals with the issue of poviding recommendations for individual users of e-shop based on the obtained user preferences. The work includes an overview of existing recommender systems, their methods of getting user preferences, the methods of using objects' content and recommender algorithms. An integral part of this work is design and implementated for independent software component for Content-based recommendation. Component is able to receive various user preferences and various forms of object's input data. The component also contains various processing methods for implicit feedback and various methods for making recommendations. Component is written in the Java programming language and uses a PostgreSQL database. The thesis also includes experiments that was carried out with usage of component designed on datasets slantour.cz and antikvariat-ichtys.cz e-shops.

Recommender systems are programs that aim to present items like songs or books that are likely to be interesting for a user. These systems have become increasingly popular and are intensively studied by research groups all over the world. In web systems, like e-shops or community servers there are usually multiple data sources we can use for recommending, as user and item attributes, user-item rating or implicit feedback from user behaviour. In the thesis, we present a concept of a Universal Recommender System (Unresyst) that can use these data sources and is domain-independent at the same time. We propose how Unresyst can be used. From the contemporary methods of recommending, we choose a knowledge based algorithm combined with collaborative filtering as the most appropriate algorithm for Unresyst. We analyze data sources in various systems and generalize them to be domain-independent. We design the architecture of Unresyst, describe its interfaces and methods for processing the data sources. We adapt Unresyst to three real-world data sets, evaluate the recommendation accuracy results and compare them to a contemporary collaborative filtering recommender. The comparison shows that combining multiple data sources can improve the accuracy of collaborative filtering algorithms and can be used in systems where...

Dans cette thèse nous considérons une plate-forme de filtrage collaboratif distribué dans laquelle chaque pair héberge ses informations personnelles, comme les URL qu'il a aimé, les pages d'information qui ont attiré son intérêt ou les vidéos qu'il a regardé, sur sa propre machine. Ensuite, sans reposer sur un tiers fiable, le pair s'engage dans un protocole distribué combinant ses données personnelles avec celles d'autres pairs afin d'exécuter un filtrage collaboratif. L'objectif principal est d'être capable de recevoir des recommandations personnalisées ou d'autres services comme un moteur de recherche distribué et personnalisé. Les protocoles de filtrage collaboratif basés sur les utilisateurs, qui demandent de calculer des similarités entre les utilisateurs, ont été appliqués aux systèmes distribués. Comme le calcul de similarité entre les utilisateurs nécessite d'utiliser leurs profils privés personnels, cela lève de sérieuses inquiétudes quant à la vie privée. Dans cette thèse nous adressons le problème du calcul de similarités entre les pairs sur une plate forme collaborative. Notre travail fournit une primitive préservant la vie privée pour le calcul de similarité qui peut rendre les protocoles collaboratifs respectueux de la vie privée. Nous adressons le challenge unique qui est d'appliquer des techniques de préservation de la vie privée pour le calcul de similarité aux systèmes dynamiques à large échelle. En particulier, nous introduisons un protocole cryptographique bipartite qui assure la differential privacy, une notion forte de préservation de la vie privée. De plus, nous résolvons le problème du budget de vie privée qui empêcherait les pairs de calculer leur similarité plus d'un nombre fixe de fois en introduisant la notion de canaux anonymes bidirectionnels. Nous développons aussi une variante hétérogène de la differential privacy qui peut apporter différent niveaux de confidentialité pour différent utilisateurs, et même différent niveaux de confidentialité pour différents items du profile d'un même utilisateur, prenant ainsi en compte différentes attentes en terme de préservation de la vie privée. De plus, nous proposons un protocole non interactif qui est très efficace pour publier une représentation compacte et préservant la vie privée des profiles des utilisateurs qui peut être utilisée pour estimer la similarité. Enfin, nous étudions le problème de choisir un paramètre de vie privée à la fois théoriquement et empiriquement en créant plusieurs attaques par inférence qui démontrent pour quelles valeurs du paramètre de vie privée le niveau de vie privée fournis est acceptable. / In this thesis, we consider a distributed collaborative platform in which each peer hosts his private information, such as the URLs he liked or the news articles that grabbed his interest or videos he watched, on his own machine. Then, without relying on a trusted third party, the peer engages in a distributed protocol, combining his private data with other peers' private data to perform collaborative filtering. The main objective is to be able to receive personalized recommendations or other services such as a personalized distributed search engine. User-based collaborative filtering protocols, which depend on computing user-to-user similarity, have been applied to distributed systems. As computing the similarity between users requires the use of their private profiles, this raises serious privacy concerns. In this thesis, we address the problem of privately computing similarities between peers in collaborative platforms. Our work provides a private primitive for similarity computation that can make collaborative protocols privacy-friendly. We address the unique challenges associated with applying privacy-preserving techniques for similarity computation to dynamic large scale systems. In particular, we introduce a two-party cryptographic protocol that ensures differential privacy, a strong notion of privacy. Moreover, we solve the privacy budget issue that would prevent peers from computing their similarities more than a fixed number of times by introducing the notion of bidirectional anonymous channel. We also develop a heterogeneous variant of differential privacy that can provide different level of privacy to different users, and even different level of privacy to different items within a single user's profile, thus taking into account different privacy expectations. Moreover, we propose a non-interactive protocol that is very efficient for releasing a small and private representation of peers' profiles that can be used to estimate similarity. Finally, we study the problem of choosing an appropriate privacy parameter both theoretically and empirically by creating several inference attacks that demonstrate for which values of the privacy parameter the privacy level provided is acceptable.

Com o propósito de auxiliar os usuários no processo de tomada de decisão, diversos tipos de sistemas Web passaram a incorporar sistemas de recomendação. As abordagens mais utilizadas são a filtragem baseada em conteúdo, que recomenda itens com base nos seus atributos, a filtragem colaborativa, que recomenda itens de acordo com o comportamento de usuários similares, e os sistemas híbridos, que combinam duas ou mais técnicas. A abordagem baseada em conteúdo apresenta o problema de análise limitada de conteúdo, o qual pode ser reduzido com a utilização de informações semânticas. A filtragem colaborativa, por sua vez, apresenta o problema da partida fria, esparsidade e alta dimensionalidade dos dados. Dentre as técnicas de filtragem colaborativa, as baseadas em fatoração de matrizes são geralmente mais eficazes porque permitem descobrir as características subjacentes às interações entre usuários e itens. Embora sistemas de recomendação usufruam de diversas técnicas de recomendação, a maioria das técnicas apresenta falta de informações semânticas para representarem os itens do acervo. Estudos na área de sistemas de recomendação têm analisado a utilização de dados abertos conectados provenientes da Web dos Dados como fonte de informações semânticas. Dessa maneira, este trabalho tem como objetivo investigar como relações semânticas computadas a partir das bases de conhecimentos disponíveis na Web dos Dados podem beneficiar sistemas de recomendação. Este trabalho explora duas questões neste contexto: como a similaridade de itens pode ser calculada com base em informações semânticas e; como semelhanças entre os itens podem ser combinadas em uma técnica de fatoração de matrizes, de modo que o problema da partida fria de itens possa ser efetivamente amenizado. Como resultado, originou-se uma métrica de similaridade semântica que aproveita a hierarquia das bases de conhecimento e obteve um desempenho superior às outras métricas na maioria das bases de dados. E também o algoritmo Item-MSMF que utiliza informações semânticas para amenizar o problema de partida fria e obteve desempenho superior em todas as bases de dados avaliadas no cenário de partida fria. / In order to assist users in the decision-making process, several types of web systems started to incorporate recommender systems. The most commonly used approaches are content-based filtering, which recommends items based on their attributes; collaborative filtering, which recommends items according to the behavior of similar users; and hybrid systems that combine both techniques. The content-based approach presents the problem of limited content analysis, which can be reduced by using semantic information. The collaborative filtering, presents the problem of cold start, sparsity and high dimensionality of the data. Among the techniques of collaborative filtering, those based on matrix factorization are generally more effective because they allow us to discover the underlying characteristics of interactions between users and items. Although recommender systems have several techniques, most of them lack semantic information to represent the items in the collection. Studies in this area have analyzed linked open data from the Web of data as source of semantic information. In this way, this work aims to investigate how semantic relationships computed from the knowledge bases available in the Data Web can benefit recommendation systems. This work explores two questions in this context: how the similarity of items can be calculated based on semantic information and; as similarities between items can be combined in a matrix factorization technique, so that the cold start problem of items can be effectively softened. As a result, a semantic similarity metric was developed that leverages the knowledge base hierarchy and outperformed other metrics in most databases. Also the Item-MSMF algorithm that uses semantic information to soften the cold start problem and obtained superior performance in all databases evaluated in the cold start scenario.

A crescente oferta de programas de código aberto na rede mundial de computadores expõe potenciais usuários a muitas possibilidades de escolha. Em face da pluralidade de interesses desses indivíduos, mecanismos eficientes que os aproximem daquilo que buscam trazem benefícios para eles próprios, assim como para os desenvolvedores dos programas. Este trabalho apresenta o AppRecommender, um recomendador de aplicativos GNU/Linux que realiza uma filtragem no conjunto de programas disponíveis e oferece sugestões individualizadas para os usuários. Tal feito é alcançado por meio da análise de perfis e descoberta de padrões de comportamento na população estudada, de sorte que apenas os aplicativos considerados mais suscetíveis a aceitação sejam oferecidos aos usuários. / The increasing availability of open source software on the World Wide Web exposes potential users to a wide range of choices. Given the individuals plurality of interests, mechanisms that get them close to what they are looking for would benefit users and software developers. This work presents AppRecommender, a recommender system for GNU/Linux applications which performs a filtering on the set of available software and individually offers suggestions to users. This is achieved by analyzing profiles and discovering patterns of behavior of the studied population, in a way that only those applications considered most prone to acceptance are presented to users.

Sistemas de recomendação surgiram da necessidade de selecionar e apresentar conteúdo relevante a usuários de acordo com suas preferências. Dentre os diversos métodos existentes, aqueles baseados em conteúdo faz em uso exclusivo da informação inerente aos itens. Estas informações podem ser criadas a partir de técnicas de indexação automática e manual. Enquanto que as abordagens automáticas necessitam de maiores recursos computacionais e são limitadas á tarefa específica que desempenham, os métodos manuais são caros e propensos a erros. Por outro lado, com a expansão da Web e a possibilidade de usuários comuns criarem novos conteúdos e anotações sobre diferentes itens e produtos, uma alternativa é obter esses metadados criados colaborativamente pelos próprios usuários. Entretanto, essas informações, em especial revisões e comentários, podem conter ruídos, além de estarem em uma forma desestruturada. Deste modo, este trabalho1 tem como objetivo desenvolver métodos de construção de representações de itens baseados em descrições colaborativas para um sistema de recomendação. Objetiva-se analisar o impacto que diferentes técnicas de extração de características, aliadas à análise de sentimento, causam na precisão da geração de sugestões, avaliando-se os resultados em dois cenários de recomendação: predição de notas e geração de ranques. Dentre as técnicas analisadas, observa-se que a melhor apresenta um ganho no poder descritivo dos itens, ocasionando uma melhora no sistema de recomendação. / Recommender systems arose from the need to select and present relevant content to users according to their preferences. Among several existent methods, those based on content make exclusive use of information inherent to the items. This information can be created through automatic and manual indexing techniques. While automa-tic approaches require greater computing resources and are limited to the specific task they perform, manual methods are expensive and prone to errors. On the other hand, with the expansion of theWeb and the possibility of common users to create new content and descriptions about different items and products, an alternative is to get these metadata created collaboratively by the users. However, this information, especially reviews and comments, may contain noise, be- sides being in a unstructured fashion. Thus, this study aims to develop methods for the construction of items representations based on collaborative descriptions for a recommender system. This study aims to analyze the impact that different feature extraction techniques, combined with sentiment analysis, caused in the accuracy of the generated suggestions, evaluating the results in both recommendations cenarios: rating prediction and ranking generation. Among the analyzed techniques, it is observed that the best is able to describe items in a more effcient manner, resulting in an improvement in the recommendation system.

Un des premiers modèles d'indexation de documents qui utilise des termes comme descripteurs était une structure de treillis, cela une vingtaine d'années avant l'arrivée de l'analyse formelle de concepts (FCA pour "Formal Concept Analysis"), qui s'affirme maintenant comme un formalisme théorique important et solide pour l'analyse de données et la découverte de connaissances. Actuellement, la communauté en recherche d'information (RI) s'intéresse particulièrement à des techniques avancées pour la recherche des documents qui relèvent des probabilités et des statistiques. En parallèle, l'intérêt de la communauté FCA au développement de techniques qui font avancer l'état de l'art en RI tout en offrant des fonctionnalités sémantiques lui est toujours bien vivant. Dans cette thèse, nous présentons un ensemble de contributions sur ce que nous avons appelé les systèmes FCA de recherche d'information ("FCA-based IR systems''). Nous avons divisé nos contributions en deux parties, à savoir l'extraction et l'indexation. Pour la récupération, nous proposons une nouvelle technique qui exploite les relations sémantiques entre les descripteurs dans un corpus de documents. Pour l'indexation, nous proposons un nouveau modèle qui permet de mettre en oeuvre un modèle vectoriel d'indexation des documents s'appuyant sur un treillis de concepts (ou treillis de Galois). En outre, nous proposons un modèle perfectionné pour l'indexation hétérogène dans lequel nous combinons le modèle vectoriel et le modèle de recherche booléen. Finalement, nous présentons une technique de fouille de données inspiré de l'indexation des documents, à savoir un modèle d'énumération exhaustive des biclusters en utilisant la FCA. Le biclustering est une nouvelle technique d'analyse de données dans laquelle les objets sont liés via la similitude dans certains attributs de l'espace de description, et non pas par tous les attributs comme dans le "clustering'' standard. En traduisant ce problème en termes d'analyse formelle de concepts, nous pouvons exploiter l'algorithmique associée à la FCA pour développer une technique d'extraction de biclusters de valeurs similaires. Nous montrons le très bon comportement de notre technique, qui fonctionne mieux que les techniques actuelles de biclustering avec énumération exhaustive / One of the first models ever to be considered as an index for documents using terms as descriptors, was a lattice structure, a couple of decades before the arrival of Formal Concept Analysis (FCA) as a solid theory for data mining and knowledge discovery.While the Information Retrieval (IR) community has shifted to more advanced techniques for document retrieval, like probabilistic and statistic paradigms, the interest of the FCA community on developing techniques that would improve the state-of-the-art in IR while providing relevance feedback and semantic based features, never decayed. In this thesis we present a set of contributions on what we call FCA-based IR systems. We have divided our contributions in two sets, namely retrieval and indexing. For retrieval, we propose a novel technique that exploits semantic relations among descriptors in a document corpus and a new concept lattice navigation strategy (called cousin concepts), enabling us to support classification-based reasoning to provide better results compared with state-of-the-art retrieval techniques. The basic notion in our strategy is supporting query modification using "term replacements'' using the lattice structure and semantic similarity. For indexing, we propose a new model that allows supporting the vector space model of retrieval using concept lattices. One of the main limitations of current FCA-based IR systems is related to the binary nature of the input data required for FCA to generate a concept lattice. We propose the use of pattern structures, an extension of FCA to deal with complex object descriptions, in order to support more advanced retrieval paradigms like the vector space model. In addition, we propose an advanced model for heterogeneous indexing through which we can combine the vector space model and the Boolean retrieval model. The main advantage of this approach is the ability of supporting indexing of convex regions in an arbitrary vectorial space built from a document collection. Finally, we move forward to a mining model associated with document indexing, namely exhaustive bicluster enumeration using FCA. Biclustering is an emerging data analysis technique in which objects are related by similarity under certain attributes of the description space, instead of the whole description space like in standard clustering. By translating this problem to the framework of FCA, we are able to exploit the robust machinery associated with the computation of concept lattices to provide an algorithm for mining biclusters based on similar values. We show how our technique performs better than current exhaustive enumeration biclustering techniques.