Federated Recommendation Model Integrating Social Relationships and Knowledge Graph

Abstract

Integrating social relationships and knowledge graph (KG) in centralized recommendations is an effective method to mitigate data sparsity and cold-start problems. However, current research has only attempted to include either social relationships or a KG in federated recommendation systems (FedRS), failing to utilize all available user and item information. Motivated by this, this article proposes a federated recommendation model by integrating both social relationships and KG. However, directly applying these two strategies to FedRS could lead to privacy leakage and increased communication overhead. To address this, this work encrypts user embeddings of clients using homomorphic encryption before uploading them to the server and adds pseudo-friends and pseudo-items when uploading user social and interaction lists to prevent privacy leakage. Then, the server distributes friend embeddings and pretrained entity and relationship embeddings to clients based on the social and interaction lists, reducing communication overhead. Finally, clients employ the graph attention network to integrate social relationships and KG to improve the accuracy of the recommendation. Experiments on two real datasets demonstrate that our federated recommendation model achieves improvements of 1.8%, 2.1% in area under curve and 3.3%, 1.9% in F1 score over the last federated recommendation methods.