Enhancing Disentanglement of Popularity Bias for Recommendation with Triplet Contrastive Learning

Popularity bias is a common phenomenon in the user-item interaction, which means a user interacts with the items just because of the items' popularity, but the user does not actually interest in these items. Neglecting popularity bias in recommendation systems can result in favoring popular items over personal preferences. This paper proposes a new recommendation framework for enhancing the D is E ntanglement of popularity bias based on C ontrastive L earning ( DECL ). In DECL, the interest and conformity representation sets of the users and the items are generated through a disentangled representation learning process. A contrastive learning process is then performed to optimize the distributions of the disentangled sets in the representation space. A customized loss function is designed to facilitate the parameter optimization, and the final recommendation is made based on the interest and conformity. Extensive experiments and comparison studies are conducted on three real-world datasets to validate the effectiveness of the proposed DECL framework. The experiment results show that compared with the state-of-the-art methods, DECL can achieve up to 10.69% performance improvement on the Ciao dataset. This indicates the proposed system can effectively disentangle popularity bias in recommendation and has large application potential.