BPR: Bayesian Personalized Ranking from Implicit Feedback

The paper addresses item recommendation—predicting a personalized ranking over a set of items such as websites, movies, or products—in the common scenario of implicit feedback like clicks and purchases. It observes that although popular methods such as matrix factorization (MF) and adaptive k-nearest-neighbor (kNN) are designed for this personalized ranking task, none of them is directly optimized for ranking. The authors present BPR-Opt, a generic optimization criterion for personalized ranking derived as the maximum posterior estimator from a Bayesian analysis of the problem, together with a generic learning algorithm based on stochastic gradient descent with bootstrap sampling.

The authors show how to apply their method to two state-of-the-art recommender models, matrix factorization and adaptive kNN. Their experiments indicate that for personalized ranking, the BPR optimization method outperforms the standard learning techniques used for MF and kNN. The results underscored the importance of optimizing models with respect to the right criterion, and BPR became a widely used objective for ranking from implicit feedback.