FedKGRec: privacy-preserving federated knowledge graph aware recommender system

Knowledge Graph(KG) aware recommendation generally incorporates KG as side information to enhance the user and item representations. Although effective in addressing data sparsity and cold start issues, these methods can raise privacy concerns and legal risks due to the centralized storage of user-item interactions. In order to solve this issue, a novel privacy-preserving framework for KG-aware recommendation named FedKGRec is introduced, which trains the recommendation model collaboratively with the orchestration of a central server. First, we design a local KG-aware Recommendation model (KGRec), crux of which are the user preference propagation module and the item neighbor expansion module, aiming to enhance the user and item representations simultaneously. Then, the local differential privacy (LDP) technique is applied to perturb the local model parameters before they are sent to the central server or aggregator, making it extremely difficult for malicious parties to extract individual sensitive information from the aggregated results. Extensive comparative experiments on three public datasets demonstrate that the proposed FedKGRec outperforms the state-of-the-art federated recommendation methods in terms of AUC, ACC and F1.