Price-aware debiased learning model for recommendation.

Abstract

Popularity bias is common in recommender systems, which makes systems overly rely on the popularity of items rather than their alignment with user interests, eventually decreasing the recommendation performance. Some studies suggest that popularity bias can be divided into a harmful part and a beneficial part, and only the harmful part should be alleviated. In this paper, we find that item prices could cause a beneficial part of popularity bias that reflects user preferences for prices and should be retained during debiasing while the harmful part of popularity bias should be reduced. However, such a beneficial part of popularity bias remains unconsidered in existing debiasing studies, which could hurt the recommendation performance. To tackle this issue, we argue that it is necessary to disentangle such a beneficial part of popularity bias from the harmful popularity bias. Firstly, we propose a causal graph and utilize causal intervention technique to explore how such price-aware popularity bias occurs. Secondly, in order to disentangle the beneficial price-aware popularity bias, we propose a Price-Aware Popularity-Debiased Model (PAPDM), which considers various factors including user's purchasing capacity and consumption willingness to model the effect of price-aware popularity. Thirdly, we propose a popularity-based negative sampling method to identify true negative samples among unobserved items mixed with potential positive samples. This issue remains overlooked by existing debiasing studies, reducing recommendation performance. Extensive experiments on widely used datasets show that PAPDM outperforms several recent debiased models in overall recommendation performance. Further analysis indicates that PAPDM effectively reduces popularity bias.