Thesis (B.A.)--Haverford College, Dept. of Computer Science, 2009. / Includes bibliographical references.

Thesis (Ph. D.)--University of Washington, 2006. / Vita. Includes bibliographical references (leaves 155-163).

The emergence of the content overloaded internet creates a lot of new challengesfor users and service providers a like. To minimize the displayed amount of contentlike movies, music, or other products service providers like Netflix or Amazonare using recommender systems which aim to guide the user trough the availableinformation. These systems collect knowledge about the user and try to deliver personalized experiences. Most of the state-of-the-art recommender systems are using acontent focused approach but often fail to grasp the nature of users’ desires. Therefore,a mood-as-input model is developed which combines the existing research onhuman mood identification and the emotion classification of content in the domainof movies. In order to match these two components different machine learning modelsare evaluated and a Random Forest is selected as the main matching algorithm.The results of this study indicate that the mood of a user can be used to create personalizedcontent recommendations and that it can perform better than an Arbitrarysystem.

This thesis presents a domain knowledge based similarity measure for recommender systems, using Systembolaget's open API with product information as input data. The project includes the development of the similarity measure, implementing it in a content based recommender engine as well as evaluating the model and comparing it to an existing model which uses a bag-of-words based approach. The developed similarity measure uses domain knowledge to calculate the similarity of three feature, grapes, wine regions and production year, to attempt to improve the quality of recommendations. The result shows that the bag-of-words based model performs slightly better than the domain knowledge based model, in terms of coverage, diversity and correctness. However, the results are not conclusive enough to discourage from more investigation into using domain knowledge in recommender systems.

Made available in DSpace on 2014-06-12T15:59:57Z (GMT). No. of bitstreams: 2 arquivo5808_1.pdf: 2179927 bytes, checksum: f047e667d6364f038f0a1fdd91b757b2 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2011 / A atividade de teste de software pode ser bastante árdua e custosa. No contexto de testes manuais, todo o esforço com o objetivo de reduzir o tempo de execução dos testes e aumentar a contenção de defeitos é bem-vindo. Uma possível estratégia é alocar os casos de teste de acordo com o perfil do testador de forma a maximizar a produtividade. Entretanto, otimizar a alocação de casos de teste não é uma tarefa trivial: em grandes companhias, gerentes de teste são responsáveis por alocar centenas de casos de teste aos testadores disponíveis ao início de uma nova execução. Neste trabalho nós propomos dois algoritmos para a alocação automática de casos de teste e três perfis para os testadores baseados em sistemas de recomendação (o mesmo tipo de sistema que recomenda, por exemplo, um livro na Amazon.com ou um filme no Netflix.com). Cada um dos algoritmos de alocação pode ser combinado com os três perfis de testador, resultando em seis sistemas de alocação possíveis: Exp-Manager, Exp-Blind, MO-Manager, MO-Blind, Eff-Manager, e Eff-Blind. Nossos sistemas de alocação consideram a efetividade (defeitos válidos encontrados no passado) e experiência do testador (habilidade em executar testes com determinadas características). Com o objetivo de comparar os nossos sistemas de alocação com a alocação do gerente e com alocações aleatórias, um experimento controlado, utilizando 100 alocações com pelo menos 50 casos de teste cada uma, foi realizado em um cenário industrial real. Os sistemas de alocação foram avaliados através das métricas de precisão, recall e taxa de não-alocação (percentual de casos de teste não alocados). Em nosso experimento, a aplicação da ANOVA (uma técnica estatística utilizada para verificar se as amostras de dois ou mais grupos são oriundas de populações com médias iguais) e do teste de Tukey (um procedimento de comparações múltiplas para identificar quais médias são significativamente diferentes entre si) mostraram que o Exp-Manager supera os demais sistemas de alocação com respeito às métricas de precisão e recall. Todos os sistemas de alocação mostraram-se superiores ao algoritmo randômico. A precisão média (entre os sistemas de alocação) variou de 39.32% a 64.83% enquanto o recall médio variou de 39.19% a 64.83%; para a métrica de não-alocação, três sistemas de alocação (Exp-Manager, Exp-Blind e MOBlind) apresentaram um melhor desempenho alcançando taxa zero de não-alocação para todas as alocações de testes. A taxa média de não-alocação variou de 0% a 2.34% (para a métrica não-alocação, quanto menor, melhor). No cenário industrial real onde o nosso trabalho foi realizado, gerentes de teste gastam de 16 a 30 dias de trabalho por ano com a atividade de alocação de casos de teste. Nossos sistemas de alocação podem ajudá-los a realizar esta atividade de forma mais rápida e mais eficaz

This thesis describes the process of conceptualizing and developing a recommendersystem for a peer-to-peer commerce application. The application in question is calledPlick and is a vintage clothes marketplace where private persons and smaller vintageretailers buy and sell secondhand clothes from each other. Recommender systems is arelatively young field of research but has become more popular in recent years withthe advent of big data applications such as Netflix and Amazon. Examples ofrecommender systems being used in e-marketplace applications are however stillsparse and the main contribution of this thesis is insight into this sub-problem inrecommender system research. The three main families of recommender algorithmsare analyzed and two of them are deemed unfitting for the e-marketplace scenario.Out of the third family, collaborative filtering, three algorithms are described,implemented and tested on a large subset of data collected in Plick that consistsmainly of clicks made by users on items in the system. By using both traditional andnovel evaluation techniques it is further shown that a user-based collaborative filteringalgorithm yields the most accurate recommendations when compared to actual userbehavior. This represents a divergence from recommender systems commonly usedin e-commerce applications. The paper concludes with a discussion on the cause andsignificance of this difference and the impact of certain data-preprocessing techniqueson the results.

Collaboration and sharing of information are the basis of modern social web system. Users in the social web systems are establishing and joining online communities, in order to collectively share their content with a group of people having common topic of interest. Group or community activities have increased exponentially in modern social Web systems. With the explosive growth of social communities, users of social Web systems have experienced considerable difficulty with discovering communities relevant to their interests. In this study, we address the problem of recommending communities to individual users. Recommender techniques that are based solely on community affiliation, may fail to find a wide range of proper communities for users when their available data are insufficient. We regard this problem as tag-based personalized searches. Based on social tags used by members of communities, we first represent communities in a low-dimensional space, the so-called latent semantic space, by using Latent Semantic Analysis. Then, for recommending communities to a given user, we capture how each community is relevant to both user’s personal tag usage and other community members’ tagging patterns in the latent space. We specially focus on the challenging problem of recommending communities to users who have joined very few communities or having no prior community membership. Our evaluation on two heterogeneous datasets shows that our approach can significantly improve the recommendation quality.

One of the most important challenges facing us today is to personalize services based on user preferences. In order to achieve this objective, the design of Recommender Systems (RSs), which are systems designed to aid the users through different decision-making processes by providing recommendations to them, have been an active area of research. RSs may produce personalized and non-personalized recommendations. Non-personalized RSs provide general suggestions to a user, based on the number of times an item has been selected in the past. Personalized RSs, on the other hand, aim to predict the most suitable items for a specific user, based on the user’s preferences and constraints. The latter are the focus of this thesis. While Recommender Systems have been successful in many domains, a number of challenges remain. For example, most implementations consider only single criteria ratings, and consequently are unable to identify why a user prefers an item over others. Many systems classify the user into one single group or cluster which is an unrealistic approach, since in real world users share commonalities in different degrees with diverse types of users. Others require a large amount of previously gathered data about users’ interactions and preferences, in order to be successfully applied. In this study, we introduce a methodology for the creation of Personalized Multi Criteria Context Aware Recommender Systems that aims to overcome these shortcomings. Our methodology incorporates the user’s current context information, and techniques from the Multiple Criteria Decision Analysis (MCDA) field of study to analyze and model the user preferences. To this end, we create a multi criteria user preference model to assess the utility of each item for a specific user, to then recommend the items with the highest utility. The criteria considered when creating the user preference model are the user’s location, mobility level and user profile. The latter is obtained by considering the user specific needs, and generalizing the user data from a large scale demographic database. We present a case study where we applied our methodology into PeRS, a personal Recommender System to recommend events that will take place within the Ottawa/Gatineau Region. Furthermore, we conduct an offline experiment performed to evaluate our methodology, as implemented in our case study. From the experimental results we conclude that our RS is capable to accurately narrow down, and identify, the groups from a demographic database where a user may belong, and subsequently generate highly accurate recommendation lists of items that match with his/her preferences. This means that the system has the ability to understand and typify the user. Moreover, the results show that the obtained system accuracy doesn’t depend on the user profile. Therefore, the system is potentially capable to produce equally accurate recommendations for a wide range of the population.

Personalisation of content is a frequently used technique intended to improve user engagement and provide more value to users. Systems designed to provide recommendations to users are called recommender systems and are used in many different industries. This study evaluates the potential of personalisation in a media group primarily publishing local news, and studies how information stored by the group may be used for recommending content. Specifically, the study focuses primarily on content-based filtering by article tags and user grouping by demographics. This study first analyses the data stored by a media group to evaluate what information, data structures, and trends have potential use in recommender systems. These insights are then applied in the implementation of recommender systems, leveraging that data to perform personalised recommendations. When evaluating the performance of these recommender systems, it was found that tag-based content selection and demographic grouping each contribute to accurately recommending content, but that neither method is sufficient for providing fully accurate recommendations.

Living in the booming age of information, we have to rely on powerful information retrieval tools to seek the unique piece of desired knowledge from such a big data world, like using personalized search engine and recommendation systems. As one of the core components, ranking model can appear in almost everywhere as long as we need a relative order of desired/relevant entities. Based on the most general and intuitive assumption that entities without user actions (e.g., clicks, purchase, comments) are of less interest than those with user actions, the objective function of pairwise ranking models is formulated by measuring the contrast between positive (with actions) and negative (without actions) entities. This contrastive relationship is the core of pairwise ranking models. The construction of these positive-negative pairs has great influence on the model inference accuracy. Especially, it is challenging to explore the entity relationships in heterogeneous information network. In this thesis, we aim at advancing the development of the methodologies and principles of mining heterogeneous information network through learning entity relations from a pairwise learning to rank optimization perspective. More specifically we first show the connections of different relation learning objectives modified from different ranking metrics including both pairwise and list-wise objectives. We prove that most of popular ranking metrics can be optimized in the same lower bound. Secondly, we propose the class-imbalance problem imposed by entity relation comparison in ranking objectives, and prove that class-imbalance problem can lead to frequency 5 clustering and gradient vanishment problems. As a response, we indicate out that developing a fast adaptive sampling method is very essential to boost the pairwise ranking model. To model the entity dynamic dependency, we propose to unify the individual-level interaction and union-level interactions, and result in a multi-order attentive ranking model to improve the preference inference from multiple views.