Balancing global and local interests in cross-domain recommendation systems

Yi Zhao, Jin Zhang, Jibing Gong, Jiquan Peng, Xindong Wu, Shishan Gong, Shuying Du

Abstract: The cornerstone of cross-domain recommendation lies in harnessing knowledge from relevant dense domains to enhance the recommendation performance in sparse domains, involving the fusion of knowledge across multiple domains. Due to substantial variations in user click data distribution across different domains, the manifestation of user interests across all domains does not always align with those observed in specific domains. Furthermore, due to differences in sparsity across various domains, the interests in dense domains often overshadow those in sparse domains, leading to suboptimal recommendation performance. Therefore, striking a balance between users’ global interests and local interests is crucial for cross-domain recommendation. To address these challenges, we propose a novel approach, BGLI-CDR, that automatically balances users’ global and local interests according to different characteristics of different users. Specially, BGLI-CDR employs a user-shared heterogeneous graph to model users’ global interests in both domains and employs an attention method to capture users’ local interests in the specific domain. Then, we propose a self-driven meta-learning-based method to balance the global and local interests, ensuring that neither the local interests nor the global interests are sacrificed. The experiments on real-world datasets demonstrate that our model outperforms state-of-art baselines.

External IDs:dblp:journals/mms/ZhaoZGPWGD25