One-shot Collaborative Filtering

We propose a new one-shot collaborative filtering method. In contrast to the conventional methods, which predict unobserved ratings individually and independently, our method predicts all unobserved ratings simultaneously and with mutual dependence. With the proposed method, first for observed ratings, we compute empirical marginal distributions of the ratings over users and/or items. Then, for unrated data, these marginal distributions are represented as a function of unknown ratings, and the unknown ratings are predicted by minimizing the Kullback-Leibler (KL) divergence between both the rated and unrated rating distributions. We evaluate the prediction performance and the computational time of our method by using real movie rating data. We confirmed that the proposed method could provide prediction errors comparable to those provided by the conventional top-level methods, but could significantly reduce the computational time