With the rapid development of the Internet, the online data volume is increasing extremely fast, which marks that we have entered the "big data era". Nevertheless, users' capacity of retrieving and processing information is not yet improving with the growth of high volume data. Hence the problem of "information overload" arises. Recommender systems, as an indispensable service for information filtering, are playing an critical role in coping with the problem. Among different recommendation algorithms, collaborative filtering has been steadily receiving more attention, which is also the focus of our research in this thesis. Aiming at three most important issues in collaborative filtering, namely cold-start recommendation, recommendation with implicit feedbacks and multi-domain recommendation, we have the following contribution to this area: (1) We propose a community discovering guided cold-start recommendation method. Recommendation for new users is a key challenge due to the lack of prior information from them, which is the cold-start problem. Preference elicitation framework has been proposed as an efficient solution. In this thesis, we exploit the community as an effective information which is not fully used in former approaches. By integrating community discovery into preference elicitation, our model can successfully solve the cold start problem. To perform community discovering process, the model utilizes rating similarity graph and social network as a graph regularization. Experimental results on Flixster and Douban datasets demonstrate that our method outperforms traditional preference elicitation methods. (2) With the implicit feedbacks from users, we define a new concept named "item group", and then propose a item group based personalized ranking approach. The prior preference of a user will affect his behaviors in future and a user's nearest neighbor set usually has most similar behaviors as the the user. So we expect to exploit this prior information from one's nearest neighbors. With the above consideration, we define a new concept, "item group", with the cumulative implicit feedbacks from a user's nearest neighbor set. In order to capture one's prior preference, we propose a novel method to model the pairwise ranking relation of different item groups. Moreover, we incorporate the item group based pairwise preference of a user into item based pairwise preference to obtain a novel framework. Experimental results demonstrate the proposed ...
修改评论