Research Of Personalized Recommendation Technology For Collaborative Filtering

The recommendation system is a kind of powerful tool to tackle the information-overload problem. A recommendation system is able to learn users’ preferences from their historical behaviors and rating information, thereby help them find their new interests. Compared with other information retrieval tools such as search engine, recommendation system can provide more personalized service for users. In recent years, the widely application of recommendation system in various fields, especially in e-commerce field, leads to the academia and the industry paying great attention to it.Collaborative filtering(CF) is the most prevailing and successful recommendation technique. Compared with the content-based technique, CF can find users’ new preferences and does not rely on the content of items, thus, it can be deployed in any kind of recommendation system. However, CF still faces many challenges, such as data sparsity and algorithms’ scalability. The accuracy of CF still needs to be further improved.To alleviate the problems of CF, some efforts have been done in this thesis:1) Comprehensively study the current CF techniques and introduce the principles of several popular algorithms in detail: then compare them by a series of experiments on different data sparse level: finally the differences of their performances are analyzed.2) An improved Slope One collaborative filtering algorithm using fuzzy clustering is proposed. The improved algorithm has better performances than original one in terms of prediction accuracy and scalability.3) To get more accurate predictions, a Two-Stage model ensemble framework is proposed.Using this ensemble framework, several kinds of CF models can be fused and better prediction can be gotten.