Internet e-services now provide so many items that effective recommendation systems are crucial to users to search for desire items. In this paper, we present a new recommendation method which is based on theoretical graphical models. We incorporate the concept of pairwise learning into the latent Dirichlet allocation model to discover user preferences which differentiate users’ precedence on items. A voting mechanism applied to the learned user preferences is devised so that favorite items are suggested to the users. Preliminary experiments based on a real-world dataset demonstrate that the discovered user preferences are effective in item recommendations. Also, incorporating pairwise learning successfully enhances the LDA-based recommendation method in terms of the recommendation precision and coverage rate at H.