A Generalized Probabilistic Framework and its Variants for Training Top-k Recommender System

Accounting for missing ratings in available training data was recently shown [3, 17] to lead to large improvements in the top-k hit rate of recommender systems, compared to state-of-the-art approaches optimizing the popular rootmean-square-error (RMSE) on the observed ratings. In this paper, we take a Bayesian approach, which lends itself naturally to incorporating background knowledge concerning the missing-data mechanism. The resulting log posterior distribution is very similar to the objective function in [17]. We conduct elaborate experiments with real-world data, testing several variants of our approach under different hypothetical scenarios concerning the missing ratings. In the second part of this paper, we provide a generalized probabilistic framework for dealing with possibly multiple observed rating values for a user-item pair. Several practical applications are subsumed by this generalization, including aggregate recommendations (e.g., recommending artists based on ratings concerning their songs) as well as collaborative filtering of sequential data (e.g., recommendations based on TV consumption over time). We present promising preliminary experimental results on IP-TV data.