| 英文摘要 | Faced with the explosive growth of the mixed online information, how to select the potential enjoyed contents and filter out uninterested ones for online users have become an urgent problem in the era of the Internet. To this end, recommender system was proposed and became another significant tool of solving information overload following search engine. Different with search engine, the essential goal of recommender system is to build a prediction model without any description of user needs. Instead, by analyzing user features, recommender system can connect users with their possible interested information contents(items), and automatically make recommendations through the mined connection. Therefore, with its powerful user experience, recommender systems have been widely used in electronic commerce, social networks, resource sharing, and many other Internet applications. Nevertheless, real-world data of user behaviors often suffer from missing and sparsity issues which make user interest modeling and personalized recommendation fairly challenging. Placing emphasis on data sparsity and cold start problems, we propose multi-domain based user interests profiling, modeling and optimizing approaches in this paper. Taking advantage of domain prior information and knowledge transfer among domains, we provide several solutions for shortcomings in traditional algorithms, and then improve the recommendation performance effectively. Our works and contributions could be summarized as: Firstly, we propose a user interest profiling method based on semi-supervised probabilistic topic model, which can cluster users into multiple domains, and construct domain-specific personalized recommendation. On the one hand, existing collaborative filtering algorithms ignore the behavior variety across different domains. That is, users have similar tastes in one domain could not infer that they have similar tastes in other domains; On the other hand, when it comes to extremely sparse user-item interaction data, ``Collaborative Effect'' is inclined to suggest well-known items rather than those long-tailed ones, and the user interests over different topics are largely under-explored. To overcome above two problems, we provide a domain-specific recommendation framework which is called as TopRec. In particular, the framework firstly analyzes and models user interests by experts-guided community topic mining process, and then implements domain-specific collaborative filtering algori... |
修改评论