DFCDR: Domain-Aware Feature Decoupling and Fusion for Cross-Domain Recommendation

Jinyue Wei, Yue Kou, Derong Shen, Tiezheng Nie, Dong Li

Abstract: Cross-Domain Recommendation (CDR) has indisputably proven its efficacy in alleviating the challenge of data sparsity in Recommender Systems. However, introducing domain-specific preferences from the source domain can introduce irrelevant information to the target domain. Furthermore, directly combining domain-general and domain-specific information may hinder the performance of the target domain. In this paper, we propose a domain-aware feature decoupling and fusion framework for CDR (DFCDR), which enables CDR more trustworthy and accurate. Specifically, we first design a user-level differential privacy method to protect users’ privacy within each domain. Then we propose a contrastive learning-based feature decoupling method that achieves two pivotal goals: disentangling users’ domain-specific preferences from their domain-general preferences, as well as differentiating between the popular and non-popular features of items. Finally, we present an adaptive feature fusion strategy that leverages a gating network to effectively fuse users’ domain-general and domain-specific features in the target domain. We conduct extensive experiments on two real-world datasets. The results demonstrate the effectiveness of our proposed method.