FedRKG: A Privacy-preserving Federated Recommendation Framework via Knowledge Graph Enhancement

TL;DR

FedRKG tackles privacy-preserving federated recommendation by leveraging a server-maintained knowledge graph built from public item data to enable higher-order user-item interactions. It introduces on-demand KG subgraph sampling and a request-based distribution mechanism, while employing a relation-aware GNN on the client and Local Differential Privacy with pseudo-labeling to protect interaction data and gradients. The framework demonstrates competitive performance against centralized approaches and outperforms existing federated baselines across three real-world datasets, achieving an average gain of about 4%. This approach shows that knowledge graphs can enhance federated recommendation without compromising privacy, offering practical benefits for privacy-conscious deployment in real-world systems.

Abstract

Federated Learning (FL) has emerged as a promising approach for preserving data privacy in recommendation systems by training models locally. Recently, Graph Neural Networks (GNN) have gained popularity in recommendation tasks due to their ability to capture high-order interactions between users and items. However, privacy concerns prevent the global sharing of the entire user-item graph. To address this limitation, some methods create pseudo-interacted items or users in the graph to compensate for missing information for each client. Unfortunately, these methods introduce random noise and raise privacy concerns. In this paper, we propose FedRKG, a novel federated recommendation system, where a global knowledge graph (KG) is constructed and maintained on the server using publicly available item information, enabling higher-order user-item interactions. On the client side, a relation-aware GNN model leverages diverse KG relationships. To protect local interaction items and obscure gradients, we employ pseudo-labeling and Local Differential Privacy (LDP). Extensive experiments conducted on three real-world datasets demonstrate the competitive performance of our approach compared to centralized algorithms while ensuring privacy preservation. Moreover, FedRKG achieves an average accuracy improvement of 4% compared to existing federated learning baselines.

Figures (6)

• Figure 1: Comparison of centralized learning, federated learning with enhanced user connections, and federated learning with enhanced project connections.
• Figure 2: The framework of FedRKG.
• Figure 3: Relation-aware aggregation in client.
• Figure 4: Results for top-K recommendation. The dashed line represents centralized learning, while the solid line represents federated learning. Our method surpasses all federated baselines and, furthermore, achieves competitive results compared to centralized learning.
• Figure 5: Sensitivity analysis of activated clients and receptive field depth.