Popularity bias mitigation in recommender systems via contrastive decoupling learning

The traditional training paradigm of recommendation systems optimizes training loss by autonomously exposing a large number of popular items, leading to popularity bias in the recommendation results. Existing research has attempted to mitigate this bias through methods like causal embedding to decouple popularity, but these methods overlook factors such as item quality and user conformity, resulting in insufficient decoupling of popularity. To address this issue, this paper proposes a popularity decoupling method based on contrastive learning. Firstly, user-side popularity tendency features are introduced to decouple user interests from their conformity. Then, on the item side, a quality scoring task is incorporated to strengthen the decoupling between item properties and popularity. Finally, two types of recommendations are generated, a biased recommendation that considers popularity, and an unbiased recommendation that ignores popularity. Experiments conducted on two real-world datasets show that this method outperforms existing solutions.