Abstract

We propose a framework for collaborative fila tering based on Restricted Boltzmann Mat chines (RBM), which extends previous RBMa based approaches in several important direce tions. First, while previous RBM research has focused on modeling the correlation beC tween item ratings, we model both user-user a and item-item correlations in a unified hybrid i non-IID framework. We further use real valo ues in the visible layer as opposed to multit nomial variables, thus taking advantage of t the natural order between user-item ratings. z Finally, we explore the potential of combinC ing the original training data with data genh erated by the RBM-based model itself in a l bootstrapping fashion. The evaluation on a two MovieLens datasets (with 100K and 1M n user-item ratings, respectively), shows that u our RBM model rivals the best previouslyT proposed approaches.