Federated Heterogeneous Graph Contrastive Learning for Privacy-preserving Recommendation

Recommender systems aim to predict users’ interests and needs by analyzing their historical interaction data, thereby providing personalized content and product suggestions. Traditional recommendation methods, such as collaborative filtering and hybrid systems, achieve significant success in improving user experience and driving sales but fall short in handling user privacy issues. Federated Recommendation (FedRec) emerges to address these limitations, integrating the principles of federated learning (FL), a distributed learning approach that allows multiple clients to collaboratively train models without sharing their personal data. Despite FedRec making significant progress in protecting user privacy, it also faces some performance and personalization challenges. i) the model performance deficiencies caused by limitations in data availability; ii) the heterogeneity caused by uneven distribution of client data. In this paper, we propose a novel framework, named Federated Heterogeneous Graph Contrastive Learning (FedHGCL), which utilizes heterogeneous information to construct multiple augmented views for contrastive learning (CL) to enhance FedRec. Firstly, we introduce federated heterogeneous graph contrastive learning, where each user locally constructs a small CL-based multi-view framework to enhance recommendation performance. Secondly, we design a user sampling strategy for data augmentation (DA) on the client-side and model updates on the server to balance the training data for each user. Last but not least, we prove that the CL and DA used in FedHGCL meet the requirements for privacy-preserving recommendation. Extensive experiments on three real-world datasets prove the effectiveness of FedHGCL in personalized recommendation and privacy protection.