Causal Debiasing in Recommender Systems: Principles and Prospects

The recommender system (RS) directly influences users’ experience in e-commerce, video, social media and service platforms. However, it performs restrictively due in part to the inevitable various biases. Although traditional debiasing methods such as Inverse Propensity Score (IPS) and Double Robustness (DR) can mitigate the problems to some extent, new glitches (idealization of exposure mechanism and high variation) still have an effect on the performance of RS. Nowadays, causal inference, which can facilitate the robustness and fairness of RS models, has become one crucial method to handle such bias problems. This paper initially outlines the connotation, significance and restrictions of RS and causal inference respectively. Depending on the procedure of RS, it then classifies and analyzes some popular causal debiasing methods in the past three years. Finally, potential future prospects are provided according to the characteristics of those popular methods, aiming at giving a guide to subsequent research and study.