FedPCL-CDR: A federated prototype-based contrastive learning framework for privacy-preserving cross-domain...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain recommendation (CDR) leverages data-rich domains to improve recommendations in sparse domains.
• Existing CDR methods often overlook user privacy by requiring public interaction data.
• Performance declines with few overlapping users, limiting knowledge transfer.

Purpose of the Study:

• To propose a novel Federated Prototype-based Contrastive Learning (CL) framework for privacy-preserving CDR (FedPCL-CDR).
• To address privacy concerns and improve CDR performance with non-overlapping user data.
• To enhance knowledge transfer in sparse domains through federated learning.

Main Methods:

• Federated learning framework with local client learning and global server aggregation.
• Local Differential Privacy (LDP) to learn differential prototypes from user data.
• Contrastive learning using both local and global differential prototypes for knowledge transfer.

Main Results:

• FedPCL-CDR significantly outperforms state-of-the-art (SOTA) baselines across four CDR tasks.
• Achieved average improvements of 5.76% in HR@10, 7.36% in NDCG@10, and 13.53% in MRR@10.
• Demonstrated effectiveness in utilizing non-overlapping user information while preserving privacy.

Conclusions:

• FedPCL-CDR offers a privacy-preserving and effective solution for cross-domain recommendation.
• The framework successfully handles sparse overlapping-user conditions.
• The proposed method advances the field of privacy-preserving machine learning in recommendation systems.