Weighted matrix factorization with Bayesian personalized ranking

This paper proposes an improvement to item recommendation systems based on collaborative filtering (CF) with implicit feedback data. Combined with the Bayesian Personalized Ranking (BPR) optimization approach, recommended for implicit-only feedback contexts, CF has been shown to be effective in generating accurate recommendations. The method, based on the assumption that a user prefers a consumed item to an unconsumed item, aims to maximize the difference of predicted scores between these items for each user. In most of the existing CF recommendation methods, all items are assigned the same weight, which of course is not the case in reality. In this paper, a new improved matrix factorization (MF) approach is proposed where the weights of items are allowed to vary and be reflective of items' importance or their desirability to a user. The scheme integrates these item weights as appropriate and utilizes a dynamic learning model where learning is driven by BPR. The performance of the proposed method is tested against the traditional MF. Tests confirm that better accuracy can be indeed achieved by the proposed method.