Bayesian Personalized Ranking for Non-Uniformly Sampled Items

In this paper, we describe our approach to track 2 of the KDD Cup 2011. The task was to predict which 3 out of 6 candidate songs were positively rated – instead of not rated at all – by a user. The candidate items were not sampled uniformly, but according to their general popularity. We develop an adapted version of the Bayesian Personalized Ranking (BPR) optimization criterion [9] that takes the non-uniform sampling of negative test items into account. Furthermore, we present a modified version of the generic BPR learning algorithm that maximizes the new criterion. We use it to train ranking matrix factorization models as components of an ensemble. Additionally, we combine the ranking predictions with rating prediction models to also take into account rating data. With an ensemble of such combined models, we ranked 8th (out of more than 300 teams) in track 2 of the KDD Cup 2011, without using the additional taxonomic information offered by the competition organizers.