Abstract: Accurate modeling of ratings and text reviews is at the core of successful recommender systems. While neural networks have been remarkably successful in modeling images and natural language, they have been largely unexplored in recommender system research. In this paper, we provide a neural network model that combines ratings, reviews, and temporal patterns to learn highly accurate recommendations. We co-train for prediction on both numerical ratings and natural language reviews, as well as using a recurrent architecture to capture the dynamic components of users' and items' states. We demonstrate that incorporating text reviews and temporal dynamic gives state-of-the-art results over the IMDb dataset.
Conflicts: cs.utexas.edu, google.com, cs.cmu.edu