This report describes a distributed approach to social filtering based on the agent metaphor. Firstly, previous approaches are described, such as cognitive filtering and social filtering. Then a couple of previously implemented systems are presented and then a new system design is proposed. The main goal is to give the requirements and design of an agent-based system that recommends web-documents. The presented approach combines cognitive and social filtering to get the advantages from both techniques. Finally, a prototype implementation called WebCondor is described and results of testing the system are reported and discussed.

[pt] A recomendação de itens a partir do feedback implícito dos usuários consiste em identificar padrões no interesse dos usuários por estes itens a partir de ações dos usuários, tais como cliques, interações ou o consumo de conteúdos específicos. Isso, de forma a prover sugestões personalizadas que se adéquem ao gosto destes usuários. Nesta dissertação, avaliamos a performance de alguns modelos de fatoração matricial otimizados para a tarefa de recomendação a partir de dados implícitos no consumo das ofertas de vídeos da Globo.com. Propusemos tratar estes dados de consumo como indicativos de intenção de um usuário em assistir um vídeo. Além disso, avaliamos como os vieses únicos dos usuários e vídeos, e sua variação temporal impactam o resultado das recomendações. Também sugerimos a utilização de um modelo de fatoração incremental otimizado para este problema, que escala linearmente com o tamanho da entrada, isto é, com os dados de visualizações e quantidade de variáveis latentes. Na tarefa de prever a intenção dos usuários em consumir um conteúdo novo, nosso melhor modelo de fatoração apresenta um RMSE de 0,0524 usando o viés de usuários e vídeos, assim como sua variação temporal. / [en] Item recommendation from implicit feedback datasets consists of passively tracking different sorts of user behavior, such as purchase history, watching habits and browsing activities in order to improve customer experience through providing personalized recommendations that fits into users taste. In this work we evaluate the performance of different matrix factorization models tailored for the recommendation task for the implicit feedback dataset extracted from Globo.com s video site s access logs. We propose treating the data as indication of a positive preference from a user regarding the video watched. Besides that we evaluated the impact of effects associated with either users or items, known as biases or intercepts, independent of any interactions and its time changing behavior throughout the life span of the data in the result of recommendations. We also suggest a scalable and incremental procedure, which scales linearly with the input data size. In trying to predict the intention of the users for consuming new videos our best factorization models achieves a RMSE of 0,0524 using user s and video s bias as well as its temporal dynamics.

Recommender systems emerged in the mid '90s with the objective of helping users select items or products most suited for them. Whether it is Facebook recommending people you might know, Spotify recommending songs you might like or Youtube recommending videos you might want to watch, recommender systems can now be found in every corner of the internet. In order to handle the immense increase of data online, the development of sophisticated recommender systems is crucial for filtering out information, enhancing web services by tailoring them according to the preferences of the user. This thesis aims to improve the accuracy of recommendations produced by a classical collaborative filtering recommender system by utilizing temporal properties, more precisely the date on which an item was rated by a user. Three different time-weighted implementations are presented and evaluated: time-weighted prediction approach, time-weighted similarity approach and our proposed approach, weighting the mean rating of a user on time. The different approaches are evaluated using the well known MovieLens 100k dataset. Results show that it is possible to slightly increase the accuracy of recommendations by utilizing temporal properties.

In this thesis, we present an innovative and novel approach to personalise maps/geo-spatial services for mobile users. With the proposed map personalisation approach, only relevant data will be extracted from detailed maps/geo-spatial services on the fly, based on a user’s current location, preferences and requirements. This would result in dramatic improvements in the legibility of maps on mobile device screens, as well as significant reductions in the amount of data being transmitted; which, in turn, would reduce the download time and cost of transferring the required geo-spatial data across mobile networks. Furthermore, the proposed map personalisation approach has been implemented into a working system, based on a four-tier client server architecture, wherein fully detailed maps/services are stored on the server, and upon a user’s request personalised maps/services, extracted from the fully detailed maps/services based on the user’s current location, preferences, are sent to the user’s mobile device through mobile networks. By using open and standard system development tools, our system is open to everyday mobile devices rather than smart phones and Personal Digital Assistants (PDA) only, as is prevalent in most current map personalisation systems. The proposed map personalisation approach combines content-based information filtering and collaborative information filtering techniques into an algorithmic solution, wherein content-based information filtering is used for regular users having a user profile stored on the system, and collaborative information filtering is used for new/occasional users having no user profile stored on the system. Maps/geo-spatial services are personalised for regular users by analysing the user’s spatial feature preferences automatically collected and stored in their user profile from previous usages, whereas, map personalisation for new/occasional users is achieved through analysing the spatial feature preferences of like-minded users in the system in order to make an inference for the target user. Furthermore, with the use of association rule mining, an advanced inference technique, the spatial features retrieved for new/occasional users through collaborative filtering can be attained. The selection of spatial features through association rule mining is achieved by finding interesting and similar patterns in the spatial features most commonly retrieved by different user groups, based on their past transactions or usage sessions with the system.

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

With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

The Internet represents a unique opportunity for people to interact with each other across time and space, and online communities have existed long before the Internet's solidification in everyday living. There are two inherent challenges that online communities continue to contend with: motivating participation and organizing information. An online community's success or failure rests on the content generated by its users. Specifically, users need to continually participate by contributing new content and organizing existing content for others to be attracted and retained. I propose both participation and organization can be enhanced if users have an explicit awareness of the implicit social network which results from their online interactions. My approach makes this normally ``hidden" social network visible and shows users that these intangible relations have an impact on satisfying their information needs and vice versa. That is, users can more readily situate their information needs within social processes, understanding that the value of information they receive and give is influenced and has influence on the mostly incidental relations they have formed with others. First, I describe how to model a social network within an online discussion forum and visualize the subsequent relationships in a way that motivates participation. Second, I show that social networks can also be modeled to generate recommendations of information items and that, through an interactive visualization, users can make direct adjustments to the model in order to improve their personal recommendations. I conclude that these modeling and visualization techniques are beneficial to online communities as their social capital is enhanced by "weaving" users more tightly together.

With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

The Internet represents a unique opportunity for people to interact with each other across time and space, and online communities have existed long before the Internet's solidification in everyday living. There are two inherent challenges that online communities continue to contend with: motivating participation and organizing information. An online community's success or failure rests on the content generated by its users. Specifically, users need to continually participate by contributing new content and organizing existing content for others to be attracted and retained. I propose both participation and organization can be enhanced if users have an explicit awareness of the implicit social network which results from their online interactions. My approach makes this normally ``hidden" social network visible and shows users that these intangible relations have an impact on satisfying their information needs and vice versa. That is, users can more readily situate their information needs within social processes, understanding that the value of information they receive and give is influenced and has influence on the mostly incidental relations they have formed with others. First, I describe how to model a social network within an online discussion forum and visualize the subsequent relationships in a way that motivates participation. Second, I show that social networks can also be modeled to generate recommendations of information items and that, through an interactive visualization, users can make direct adjustments to the model in order to improve their personal recommendations. I conclude that these modeling and visualization techniques are beneficial to online communities as their social capital is enhanced by "weaving" users more tightly together.

Despite recommender systems being useful, for some applications it is hard to accumulate all the required information needed for the recommendation. In today‟s ubiquitous environment, mobile devices with different characteristics are widely available. Our work focuses on the recommendation service built on mobile environment to support tourists‟ traveling need. When tourists visit a new attraction, their recommender systems can exchange data with the attraction system to help obtain rating information of people with similar tastes. Such asynchronous rating exchange mechanisms allow a tourist to receive ratings from other people even though they may not collocate at the same time. We proposed four data exchange methods between a user and an attraction system. Our recommendation mechanism incorporates other users‟ opinions to provide recommendations once the user has collected enough ratings. Every method is compared under four conditions which attraction systems carry different amount of existing data. Then we compare these methods under different amount of existing rating data and shed the light on their advantages and disadvantages. Finally, we compare our proposed asynchronous methods with other synchronous data exchange methods proposed previously.