A Bayesian Nonparametric Topic Model for User Interest Modeling

Web users display their preferences implicitly by a sequence of pages they navigated. Web recommendation systems use methods to extract useful knowledge about user interests from such data. We propose a Bayesian nonparametric approach to the problem of modeling user interests in recommender systems using implicit feedback like user navigations and clicks on items. Our approach is based on the discovery of a set of latent interests that are shared among users in the system and make a key assumption that each user activity is motivated only by several interests amongst user interest profile which is quite different from most of the existing recommendation algorithms. By using a beta process and a Dirichlet prior, the number of hidden interests and the relationships between interests and items are both inferred from the data. In order to model the sequential information on user's visits, we make a Markovian assumption on each user's navigated item sequence. We develop a Markov chain Monte Carlo inference method based on the Indian buffet process representation of the beta process. We validate our sampling algorithm using synthetic data and real world datasets to demonstrate promising results on recovering the hidden user interests.