Interest-Guided Adaptive Bidirectional Transfer via Graph Neural Networks for Cross-Domain Recommendation

Abstract

Cross-domain recommendation (CDR) aims to leverage rich data from multiple domains to deliver personalized recommendations. However, existing methods primarily rely on overlapping users to transfer knowledge across domains. This approach overlooks the fact that individuals may exhibit different or even conflicting preferences across domains, making it difficult to effectively address the diversity of users' cross-domain interests. According to the principle of collaborative filtering, a user can share similar preferences with other users, regardless of their domain affiliation. Therefore, cross-domain knowledge transfer should also extend to similar users, necessitating the accurate capture of latent cross-domain user associations. To overcome these limitations, this paper proposes an Interest-Guided Adaptive Bidirectional Transfer via Graph Neural Networks for Cross-Domain Recommendation(IGbtCDR). The method incorporates a bidirectional mapping network module, constructed using Multilayer Perceptrons, to establish a personalized cross-domain transfer matrix between source and target domains. It enables topologically unreachable but distantly similar users to form connections, facilitating the efficient capture and propagation of long-range cross-domain user associations while dynamically adapting to users' evolving cross-domain interests. Furthermore, an interest-guided bidirectional update module, built upon Multi-head Attention mechanisms, is introduced to dynamically mine user relationships. This component overcomes the limitations imposed by original topologies or overlapping users, thereby enhancing personalized recommendation performance. Extensive experiments on four real-world datasets demonstrate that IGbtCDR significantly outperforms state-of-the-art baselines, achieving average relative improvements of 7.14%, 15.14%, 6.57%, and 9.84% in HR@10 and 4.29%, 6.72%, 15.35%, and 13.08% in NDCG@10 across the datasets.