Pairwise learning for personalized ranking with noisy comparisons

Abstract

Learning-to-rank from pairwise comparisons has a wide spectrum of applications in diverse domains. For personalized recommendation applications, the widely used implicit feedback is kind of incomplete data where only the interaction can be observed while the preference intensity for the interaction cannot be observed. Existing solutions degrade pairwise learning from implicit feedback by setting the preference intensity of each interaction as 1 to learn representations as model parameters. Noisy comparisons consisting of untrustful interactions (e.g. mistaken clicks) may lead to inaccurate optimization of a pairwise learning model.In this paper, we propose a new pairwise learning algorithm to learn personalized ranking from incomplete data where noisy comparisons are widely existed, called BPRAC. As prior knowledge about trustful interactions are not available, we introduce new indicators for measuring interaction trustfulness, which are to be learned together with users’ and items’ representations as model parameters in our BPRAC algorithm. We first derive the distribution of estimated item scores for trustful interactions from pairwise comparisons. The proposed BPRAC algorithm adopts the expectation-and-maximization framework: We estimate indicators using Bayesian inference in the expectation step; while learning representations for personalized ranking in the maximization step. We also analyze the convergence of our learning algorithm. Experiments on real-world datasets validate the effectiveness of our estimation of trustful interactions and the superiority of our personalized ranking over peer algorithms.