Research On Recommender Systems Based On Collaborative Filtering

The fast development of Web2.0technology sparked a new revolution of the in-ternet. Users now play a new role in the world of internet, they take the initiative to generate information instead of simply getting information from the web. As the rapid growth of the users’population, the user-centric information generation mode leads to the exponential growth of the available information in internet, which cause the infor-mation overload problem. The information overload problem refers that people can not quickly and accurately locate the information they need. Currently, the technology to solve information overload problem can be classified into to two categories. The first technology is information retrieval represented by the search engine and the second is information filtering represented by recommender systems. The most important differ-ence between these two technologies is that search engines need queries formatted by the user and recommender systems need no queries. Thus the quality of the results of search engines depend on how users describe their information needs. Recommender systems however, filter out the information that the user is interested in by exploiting users’profile data and historical activities(watching,listening,buying etc.). So, recom-mender systems can play an very important role in the situation that uses’can not tell their information need precisely.Many recommendation algorithms have been proposed by both academia and in-dustry, collaborative filtering is one of the most effective recommendation algorithms. Collaborative filtering algorithm has been successfully applied to many commercial recommender system, but there are still issues such as the data sparsity problem and the cold start problem to be solved. With the rapid rise of social media, user-centric social networking web sites generate vast amounts of data which may reflects users’interests, how to leverage these data to improve the performance of the recommendation algorith-m has become a very hot research area. In view of the above key issues, this dissertation launched a study of the following aspects.First, research on the similarity model of collaborative filtering. User/item simi-larity calculation is the most critical issue in the memory-based collaborative filtering algorithms, sparsity of the rating matrix and unbalance of negative and positive ratings causes inaccurate similarity computation, thus limit the recommendation quality. In this dissertation, we introduce a weighting scheme and a penalty function to address the above issue. Experiment results show that improved similarity model can significantly improve the recommendation accuracy.Second, Integrating social information into collaborative filtering. The rich social information brings great opportunities for recommendation system. How to effectively leverage the abundant social network information to improve the accuracy of recom-mendation systems is the core issue of the research on social recommendation systems. In this dissertation, we build an user similarity model based on Tencent micro-blogging users’ real social network information, and effectively combine the social information based similarity model and the rating information based similarity model. Experiment results show that the proposed approach can effectively ease the data sparsity problem and improve the recommendation quality.Third, combining user-based and item-based collaborative algorithms using stacked regression. Collaborative filtering algorithms can be classified into user-based meth-ods and item-based methods according to different assumptions. In this dissertation, we studied the advantages and disadvantages of both methods and propose a two level machine learning framework to effectively combine both based on stacked regression. Experiment results show that the proposed framework can effectively ease the data s-parsity problem and improve the recommendation quality.Fourth, research on global and local model combing. In this dissertation, we claim that different users and items have different preference over user-based and item-based methods. According to above point of view, we use machine learning algorithm to auto-matically discover users and items preference information over these two methods, and use the preference information to locally combine the predictions. Experiment results show that the performance of the local combing model is significantly better than the global combing model in literature.