O crescimento acelerado da internet proporcionou uma quantidade grande de informações acessíveis aos usuários. Ainda que tal quantidade possua algumas vantagens, os usuários que possuem pouca ou nenhuma experiência para escolher uma alternativa dentre as várias apresentadas terão dificuldades em encontrar informações (ou itens, considerando o escopo deste trabalho) úteis e que atendam às suas necessidades. Devido a esse contexto, os sistemas de recomendação foram desenvolvidos para auxiliar os usuários a encontrar itens relevantes e personalizados. Tais sistemas são divididos em diversas arquiteturas. Como exemplo estão as arquiteturas baseadas em: conteúdo, filtro colaborativo e conhecimento. Para este trabalho, a primeira arquitetura foi explorada. A arquitetura baseada em conteúdo recomenda itens ao usuário com base na similaridade desses aos itens que o usuário mostrou interesse no passado. Por consequência, essa arquitetura possui a limitação de, geralmente, realizar recomendações com baixa serendipidade, uma vez que os itens recomendados tendem a ser semelhantes àqueles observados pelo o usuário e, portanto, não apresentam novidade ou surpresa. Diante desta limitação, o aspecto de serendipidade tem destaque nas discussões apresentadas neste trabalho. Assim, o objetivo deste trabalho é minimizar o problema da baixa serendipidade das recomendações por meio da utilização da análise de similaridades parciais implementada usando ensemble de agrupamentos. Para alcançar este objetivo, estratégias de recomendação baseadas em conteúdo implementadas usando agrupamento e ensemble de agrupamento foram propostas e avaliadas neste trabalho. A avaliação contou com análises qualitativas sobre as recomendações produzidas e com um estudo com usuários. Nesse estudo, quatro estratégias de recomendação de notícias foram avaliadas, incluindo as duas propostas neste trabalhos, uma estratégia baseada em recomendação aleatória, e uma estratégia baseada em coagrupamento. As avaliações consideraram aspectos de relevância, surpresa e serendipidade de recomendações. Esse último aspecto é descrito como itens que apresentam tanto surpresa quanto relevância ao usuário. Os resultados de ambas análises mostraram a viabilidade da utilização de agrupamento como base de recomendação, uma vez que o ensemble de agrupamentos obteve resultados satisfatórios em todos os aspectos, principalmente em surpresa, enquanto a estratégia baseada em agrupamento simples obteve os melhores resultados em relevância e serendipidade / The accelerated growth of the internet has provided a large amount of information accessible to users. Although this amount of information has some advantages, users who have little or no experience in choosing one of several alternatives will find it difficulty to find useful information (or items, considering the scope of this work) that meets their needs. Due to this context, recommender systems have been developed to help users find relevant and personalized items. Such systems are divided into several architectures as content-based, collaborative filtering and knowledge-based. The first architecture was explored in this work. The content-based architecture recommends items to the user based on their similarity to items that the user has shown interest in the past. Consequently, this architecture has the limitation of generally making recommendations with low serendipity, since the recommended items tend to be similar to those observed by the user and, therefore, do not present novelty or surprise. Given this limitation, the aspect of serendipity is highlighted in the discussions presented in this work. Thus, the objective of this work is to minimize the problem of the low serendipity of the recommendations through the use of the partial similarity analysis implemented using cluster ensemble. To achieve this goal, content-based recommendation strategies implemented using clustering and cluster ensemble were proposed and evaluated. The evaluation involved qualitative analysis of the recommendations and a study with users. In such a study, four news recommendation strategies were evaluated including the two strategies proposed in this work, a strategy based on random recommendation, and a strategy based on co-clustering. The evaluations considered aspects of relevance, surprise and serendipity of recommendations. This last aspect is described as items that present both surprise and relevance to the user. The results of both analyzes showed the feasibility of using clustering as the basis of recommendation, since cluster ensemble had satisfactory results in all aspects, mainly in surprise, whereas the simple clustering-based strategy obtained the best results in relevance and serendipity

Cette thèse est consacrée à l’étude des systèmes de recommandation basés sur des réseaux de neurones artificiels appris pour faire de l'ordonnancement de produits avec des retours implicites (sous forme de clics). Dans ce sens, nous proposons un nouveau modèle neuronal qui apprend conjointement la représentation des utilisateurs et des produits dans un espace latent, ainsi que la relation de préférence des utilisateurs sur les produits. Nous montrons que le modèle proposé est apprenable au sens du principe de la minimisation du risque empirique et performant par rapport aux autres modèles de l'état de l'art sur plusieurs collections. En outre, nous contribuons à la création de deux nouvelles collections, produites grâce aux enregistrements des comportements de clients de Kelkoo (https://www.kelkoo.com/); le leader européen de la publicité programmatique et de Purch (http://www.purch.com/). Les deux jeux de données recueillent des retours implicites des utilisateurs sur des produits, ainsi qu’un grand nombre d'informations contextuelles concernant à la fois les clients et les produits. La collections de données de Purch contient en plus une information sur la popularité des produits ainsi que des commentaires textuelles associés. Nous proposons, une stratégie simple et efficace sur la manière de prendre en compte le biais de la popularité ainsi qu'un modèle probabiliste latent temporel pour extraire automatiquement les thèmes des textes des commentaires.Mots clés. Systèmes de recommandation, apprentissage d'ordonnancement, réseaux de neurones, recommandations avec des retours implicites, Modèles probabilistes latents temporels / This thesis is dedicated to the study of Recommendation Systems for implicit feedback (clicks) mostly using Learning-to-rank and neural network based approaches. In this line, we derive a novel Neural-Network model that jointly learns a new representation of users and items in an embedded space as well as the preference relation of users over the pairs of items and give theoretical analysis. In addition we contribute to the creation of two novel, publicly available, collections for recommendations that record the behavior of customers of European Leaders in eCommerce advertising, Kelkoofootnote{url{https://www.kelkoo.com/}} and Purchfootnote{label{purch}url{http://www.purch.com/}}. Both datasets gather implicit feedback, in form of clicks, of users, along with a rich set of contextual features regarding both customers and offers. Purch's dataset, is affected by popularity bias. Therefore, we propose a simple yet effective strategy on how to overcome the popularity bias introduced while designing an efficient and scalable recommendation algorithm by introducing diversity based on an appropriate representation of items. Further, this collection contains contextual information about offers in form of text. We make use of this textual information in novel time-aware topic models and show the use of topics as contextual information in Factorization Machines that improves performance. In this vein and in conjunction with a detailed description of the datasets, we show the performance of six state-of-the-art recommender models.Keywords. Recommendation Systems, Data Sets, Learning-to-Rank, Neural Network, Popularity Bias, Diverse Recommendations, Contextual information, Topic Model.

Analyzing ranking data is an essential component in a wide range of important applications including web-search and recommendation systems. Rankings are difficult to visualize or model due to the computational difficulties associated with the large number of items. On the other hand, partial or incomplete rankings induce more difficulties since approaches that adapt well to typical types of rankings cannot apply generally to all types. While analyzing ranking data has a long history in statistics, construction of an efficient framework to analyze incomplete ranking data (with or without ties) is currently an open problem. This thesis addresses the problem of scalability for visualizing and modeling partial incomplete rankings. In particular, we propose a distance measure for top-k rankings with the following three properties: (1) metric, (2) emphasis on top ranks, and (3) computational efficiency. Given the distance measure, the data can be projected into a low dimensional continuous vector space via multi-dimensional scaling (MDS) for easy visualization. We further propose a non-parametric model for estimating distributions of partial incomplete rankings. For the non-parametric estimator, we use a triangular kernel that is a direct analogue of the Euclidean triangular kernel. The computational difficulties for large n are simplified using combinatorial properties and generating functions associated with symmetric groups. We show that our estimator is computational efficient for rankings of arbitrary incompleteness and tie structure. Moreover, we propose an efficient learning algorithm to construct a preference elicitation system from partial incomplete rankings, which can be used to solve the cold-start problems in ranking recommendations. The proposed approaches are examined in experiments with real search engine and movie recommendation data.

Consumers face several challenges using the Internet to fill health-related needs. (1) In many cases, they face a language gap as they look for information that is written in unfamiliar technical language. (2) Medical information in social media is of variable quality and may be appealing even when it is dangerous. (3) Discussion groups provide valuable social support for necessary lifestyle changes, but are variable in their levels of activity. (4) Finding less popular groups is tedious. We present solutions to these challenges. We use a novel adaptation of topic models to address the language gap. Conventional topic models discover a set of unrelated topics that together explain the combinations of words in a collection of documents. We add additional structure that provides relationships between topics corresponding to relationships between consumer and technical medical topics. This allows us to support search for technical information using informal consumer medical questions. We also analyze social media related to eating disorders. A third of these videos promote eating disorders and consumers are twice as engaged by these dangerous videos. We study the interactions of two communities in a photo-sharing site. There, a community that encourages recovery from eating disorders interacts with the pro-eating disorder community in an attempt to persuade them, but we found that this attempt entrenches the pro-eating disorder community more firmly in its position. We study the process by which consumers participate in discussion groups in an online diabetes community. We develop novel event history analysis techniques to identify the characteristics of groups in a diabetes community that are correlated with consumer activity. This analysis reveals that uniformly advertise the popular groups to all consumers impairs the diversity of the groups and limits their value to the community. To help consumers find interesting discussion groups, we develop a system for personalized recommendation for social connections. We extend matrix factorization techniques that are effective for product recommendation so that they become suitable for implicit power-law-distributed social ratings. We identify the best approaches for recommendation of a variety of social connections involving consumers, discussion groups and discussions.

This thesis investigates both theoretical and practical aspects of the design and analysis of iterative algorithms for trust and reputation management and recommender systems. It also studies the application of iterative trust and reputation management mechanisms in ad-hoc networks and P2P systems. First, an algebraic and iterative trust and reputation management scheme (ITRM) is proposed. The proposed ITRM can be applied to centralized schemes, in which a central authority collects the reports and forms the reputations of the service providers (sellers) as well as report/rating trustworthiness of the (service) consumers (buyers). It is shown that ITRM is robust in filtering out the peers who provide unreliable ratings. Next, the first application of Belief Propagation algorithm, a fully iterative probabilistic algorithm, on trust and reputation management (BP-ITRM) is proposed. In BP-ITRM, the reputation management problem is formulated as an inference problem, and it is described as computing marginal likelihood distributions from complicated global functions of many variables. However, it is observed that computing the marginal probability functions is computationally prohibitive for large scale reputation systems. Therefore, the belief propagation algorithm is utilized to efficiently (in linear complexity) compute these marginal probability distributions. In BP-ITRM, the reputation system is modeled by using a factor graph and reputation values of the service providers (sellers) are computed by iterative probabilistic message passing between the factor and variable nodes on the graph. It is shown that BP-ITRM is reliable in filtering out malicious/unreliable reports. It is proven that BP-ITRM iteratively reduces the error in the reputation values of service providers due to the malicious raters with a high probability. Further, comparison of BP-ITRM with some well-known and commonly used reputation management techniques (e.g., Averaging Scheme, Bayesian Approach and Cluster Filtering) indicates the superiority of the proposed scheme both in terms of robustness against attacks and efficiency. The introduction of the belief propagation and iterative message passing methods onto trust and reputation management has opened up several research directions. Thus, next, the first application of the belief propagation algorithm in the design of recommender systems (BPRS) is proposed. In BPRS, recommendations (predicted ratings) for each active user are iteratively computed by probabilistic message passing between variable and factor nodes in a factor graph. It is shown that as opposed to the previous recommender algorithms, BPRS does not require solving the recommendation problem for all users if it wishes to update the recommendations for only a single active user using the most recent data (ratings). Further, BPRS computes the recommendations for each user with linear complexity, without requiring a training period while it remains comparable to the state of art methods such as Correlation-based neighborhood model (CorNgbr) and Singular Value Decomposition (SVD) in terms of rating and precision accuracy. This work also explores fundamental research problems related to application of iterative and probabilistic reputation management systems in various fields (such as ad-hoc networks and P2P systems). A distributed malicious node detection mechanism is proposed for delay tolerant networks (DTNs) using ITRM which enables every node to evaluate other nodes based on their past behavior, without requiring a central authority. Further, for the first time. the belief propagation algorithm is utilized in the design and evaluation of distributed trust and reputation management systems for P2P networks. Several schemes are extensively simulated and are compared to demonstrate the effectiveness of the iterative algorithms and belief propagation on these applications.

Retail managers have been interested in learning about cross-category purchase behavior of their customers for a fairly long time. More recently, the task of inferring cross-category relationship patterns among retail assortments is gaining attraction due to its promotional potential within recommender systems used in online environments. Collaborative filtering algorithms are frequently used in such settings for the prediction of choices, preferences and/or ratings of online users. This paper investigates the suitability of such methods for situations when only binary pick-any customer information (i.e., choice/nonchoice of items, such as shopping basket data) is available. We present an extension of collaborative filtering algorithms for such data situations and apply it to a real-world retail transaction dataset. The new method is benchmarked against more conventional algorithms and can be shown to deliver superior results in terms of predictive accuracy. (author's abstract) / Series: Report Series SFB "Adaptive Information Systems and Modelling in Economics and Management Science"

Over the last decade, excess amount of information is being provided on the web and information filtering systems such as recommender systems have become one of the most important technologies to overcome the &bdquo / Information Overload&amp / #8223 / problem by providing personalized services to users. Several researches have been made to improve quality of recommendations and provide maximum user satisfaction within a single domain based on the domain specific knowledge. However, the current infrastructures of the recommender systems cannot provide the complete mechanisms to meet user needs in several domains and recommender systems show poor performance in cross-domain item recommendations. Within this thesis work, a dynamic framework is proposed which differs from the previous works as it focuses on the easy development of knowledge-based recommenders and it proposes an intensive cross domain capability with the help of domain knowledge. The framework has a generic and flexible structure that data models and user interfaces are generated based on ontologies. New recommendation domains can be integrated to the framework easily in order to improve recommendation diversity. The cross-domain recommendation is accomplished via an abstraction in domain features if the direct matching of the domain features is not possible when the domains are not very close to each other.

As the Internet became widespread all over the world, people started to share great amount of data on the web and almost every people joined different data networks in order to have a quick access to data shared among people and survive against the information overload on the web. Recommender systems are created to provide users more personalized information services and to make data available for people without an extra effort. Most of these systems aim to get or learn user preferences, explicitly or implicitly depending to the system, and guess &ldquo / preferable data&rdquo / that has not already been consumed by the user. Traditional approaches use user/item similarity or item content information to filter items for the active user / however most of the recent approaches also consider the trustworthiness of users. By using trustworthiness, only reliable users according to the target user opinion will be considered during information retrieval. Within this thesis work, a content boosted method of using trust data in recommender systems is proposed. It is aimed to be shown that people who trust the active user and the people, whom the active user trusts, also have correlated opinions with the active user. This results the fact that the rated items by these people can also be used while offering new items to users. For this research, www.epinions.com site is crawled, in order to access user trust relationships, product content information and review ratings which are ratings given by users to product reviews that are written by other users.

Data analysis has become a very important area for both companies and researchers as a consequence of the technological developments in recent years. Companies are trying to increase their profit by analyzing the existing data about their customers and making decisions for the future according to the results of these analyses. Parallel to the need of companies, researchers are investigating different methodologies to analyze data more accurately with high performance. Recommender systems are one of the most popular and widespread data analysis tools. A recommender system applies knowledge discovery techniques to the existing data and makes personalized product recommendations during live customer interaction. However, the huge growth of customers and products especially on the internet, poses some challenges for recommender systems, producing high quality recommendations and performing millions of recommendations per second. In order to improve the performance of recommender systems, researchers have proposed many different methods. Singular Value Decomposition (SVD) technique based on dimension reduction is one of these methods which produces high quality recommendations, but has to undergo very expensive matrix calculations. In this thesis, we propose and experimentally validate some contributions to SVD technique which are based on the user and the item categorization. Besides, we adopt tags to classical 2D (User-Item) SVD technique and report the results of experiments. Results are promising to make more accurate and scalable recommender systems.

In the last decade, with the extensive use of mobile electronic and wireless communication devices, there is a growing need for context aware applications and many pervasive computing applications have become integral parts of our daily lives. Context aware recommender systems are one of the popular ones in this area. Such systems surround the users and integrate with the environment / hence, they are aware of the users&#039 / context and use that information to deliver personalized recommendations about everyday tasks. In this manner, predicting user&rsquo / s next activity preferences with high accuracy improves the personalized service quality of context aware recommender systems and naturally provides user satisfaction. Predicting activities of people is useful and the studies on this issue in ubiquitous environment are considerably insufficient. Thus, this thesis proposes an activity prediction model to forecast a user&rsquo / s next activity preference using past preferences of the user in certain contexts and current contexts of user in ubiquitous environment. The proposed model presents a new approach for activity prediction by taking advantage of ontology. A prototype application is implemented to demonstrate the applicability of this proposed model and the obtained outputs of a sample case on this application revealed that the proposed model can reasonably predict the next activities of the users.