Improved user-based collaborative filtering algorithm with topic model and time tag

Abstract

Collaborative filtering algorithms make use of interaction rates between users and items for generating recommendations. Similarity among users is calculated based on rating mostly, without considering explicit properties of users involved. Considering the number of tags of a user can direct response the user preference to some extent, we propose a collaborative filtering algorithm using topic model called user-item-tag latent Dirichlet allocation (UITLDA) in this paper. UITLDA model consists of two parts. The first part is active user with its item. The second part is active user with its tag. We form topic model from these two parts respectively. The two topics constrain each other and integrate into a new topic distribution. This model not only increases the user's similarity, but also reduces the density of the matrix. In prediction computation, we also introduce time delay function to increase the precision. The experiments showed that the proposed algorithm achieved better performance compared with baseline on MovieLens datasets.