User-based Collaborative Filtering On Tagging Data

User-based collaborative filtering is one of the most widely-used recommender methods. It recommends items to a user according to her similar users’opinions. The key point of user-based collaborative filtering is to compute users’similarities. In tradi-tional user-based collaborative filtering, the similarity between two users is determined by their ratings to co-rated items. In some cases, two users rate few common items, such that the similarity between them may be inaccurate and it results in misleading recommendations.With the rapid development of social tagging systems, social tagging data pos-es new challenges opportunities for recommender systems. Many researchers have proposed different methods to exploit tagging data to improve the performance of rec-ommender systems. In this paper, we propose a new approach to compute users’sim-ilarities using the inferred tag ratings. A user’s preference for a tag t can be inferred based on her ratings of items tagged with t. A user rates too few such items, then her inferred rating to t may be inaccurate. Hence the relationships among tags are used to infer her preference for t based on all her item ratings, such that the preference of the user could be accurate.As the number of tags may be very large, the computation of using the inferred tag ratings to compute users’similarities may have a high complexity. In this paper, we also propose an approach to compute users’similarities using the inferred topic ratings. The topic can be generated from the tagging data using clustering or proba-bility models. In this paper, we generate topics using hierarchical clustering and LDA respectively. The result of experiment shows that LDA is more suitable, so we select LDA at last.Experiments were done on the MovieLens data set to evaluate the performance of our approaches. The results show that our approaches outperform traditional user-based collaborative filtering.