Privacy-Aware Knowledge Transfer for Cross-Domain Recommender System

Pham Minh Thu Do, Jie Lu, Qian Zhang, Guangquan Zhang

Abstract: Cross-domain recommender systems have proven effective in alleviating data sparsity by facilitating knowledge transfer across domains. However, ensuring robust user privacy remains a significant challenge. Existing methods that incorporate privacy mechanisms often provide limited protection, restricting their practical deployment. To address these challenges, we propose a novel privacy-preserving cross-domain recommender system that enables secure and effective knowledge transfer while mitigating negative transfer. Our framework leverages federated transfer learning, treating each domain as an independent client to prevent raw data sharing and minimize privacy risks. A graph encoder is introduced to learn item, local user, and global user representations, enhanced by a consistency mechanism that promotes alignment between local and global representations while maintaining their distinctiveness. To further strengthen privacy, global user representations are protected using differential privacy before cross-domain exchanges. Additionally, we design a bidirectional transfer mechanism that eliminates the need for user-identifying information, enhancing privacy safeguards without compromising recommendation performance. To address the issue of negative transfer, we introduce a negative transfer contrastive loss that preserves the semantic integrity of user representations when integrating cross-domain knowledge. Experimental results on real-world datasets show that our approach significantly improves recommendation accuracy while providing strong privacy protection, outperforming state-of-the-art baseline methods.

External IDs:doi:10.1007/978-3-032-09318-9_12