Interactive Recommender System: Causality-based Popularity Bias and Popularity Drift
Rongtao Ye, Wai Kin Victor Chan, Yun Ye, Kai Zhang, Yuqing Miao · 2024
Popularity bias and popularity drift are common issues faced in personalized recommendations, which can affect the accuracy of recommendations. To address the former, we analyzed the impact of popularity bias using a causal graph and provided new popularity values through intervention operations of causal inference. For the latter, due to the many influencing factors of popularity drift, we used causal relationships and Fourier Transformation to retain causal factors, separate non-causal factors, and predict possible drift by reconstructing invariant intrinsic causal mechanisms. We tested the learned Transformer-based reinforcement learning strategy to prove the superiority of the proposed method on datasets and an interactive recommender system.