Cross-domain recommendation via interest-aware pseudo-overlapping user alignment

Abstract

The cold-start problem remains a classic challenge in recommender systems. Cross-Domain Recommendation, which utilizes information from auxiliary source domains to boost performance, presents an effective solution. Bridge-based cross-domain methods are especially beneficial for cold-start users, who have interactions in the source domain but not the target domain. These methods typically learn a mapping function to transfer user preferences from source to target domain. However, they face two significant challenges: (1) Dependency on overlapping users, as the mapping function’s training largely relies on the limited number of overlapping users available in practical scenarios. (2) The uniform user embeddings lack the capacity to reflect multiple interests of users in the target domain, leading to weak expression of mapped users. To tackle these challenges, we introduce a new cross-domain recommendation model. Initially, the model learns a global shared interest pool across domains using an interest activation network. It then groups users by their activated interests and matches them with pseudo-overlapping users within the same interest group. In the cross-domain transfer phase, we incorporate an interest meta-network module to create personalized interest bridges for effective preference transfer. Additionally, we enhance the model with a semi-supervised learning strategy that leverages pseudo-overlapping user data to mitigate data sparsity. Consequently, comprehensive experiments confirm that our model surpasses existing state-of-the-art methods.