Calibrating Popularity Bias Based on Quality for Recommendation Fairness

Zhengyi Guo, Yanmin Zhu, Zhaobo Wang, Mengyuan Jing

Abstract: Recommendation fairness has become a significant topic recently. One obstacle to it is popularity bias. Previous works reduce the bias by eliminating popularity difference entirely. However, items with higher quality should be assigned with higher popularity. Completely excluding popularity will cripple recommender on recommendation accuracy and unfairly treat items with high quality. For this reason, we argue that the popularity of items should not be avoided at all costs but instead calibrated based on their quality, which is a depiction of items’ natural properties. To this end, we propose the Quality Recommendation (QR) framework. Specifically, we separate item embedding into quality embedding and non-popularity ID embedding. The former efficiently encodes quality related features and the latter eliminates stereotype of specific items’ popularity. To evaluate model vulnerability to popularity-quality discrepancy, we propose a novel evaluation method. Its core idea is simulate this discrepancy at the training stage. Experiments on datasets show the effective fairness and recommendation performance of our proposed methods.