Double bayesian pairwise learning for one-class collaborative filtering

摘要

Recommender systems have become an indispensable tool for real-world applications. Only one-class feedback can be obtained in many applications. Therefore, the one-class recommendation problem has attracted much attention. Pairwise ranking methods are popular for dealing with the one-class problem. Bayesian Personalized Ranking (BPR) is one of the most popular pairwise methods, assuming users prefer the observed item to the unobserved item. The parameters in BPR are learned based on stochastic gradient descent (SGD). However, the previous work has shown that existing the vanishing gradient problem in the learning process when the preference difference between the observed item and the unobserved item is very large. In this paper, we propose a novel algorithm called Double Bayesian Pairwise Learning (DBPL). In the learning process of DBPL, the preference difference between the observed item and the unobserved item can be reduced by fusing a relatively smaller preference difference between another pair of items. Moreover, we calculate potential preference scores between users and items based on user–item interactions to measure preference differences between unobserved items of each user. Experimental results on three real-world datasets show the effectiveness of the DBPL algorithm.