Exploring Personalized Intents with Knowledge Graph for Federated Self-supervised Recommendation

In the recommendation scenario of cross-user federated graph learning, each client denotes a user’s device and locally preserves privacy-sensitive preference data. As raw data are decentralized among clients, the local graph structure is extremely sparse, and the graph-signal propagation across local graphs is also blocked under the federated privacy constraint. To the issues of local graph sparsity and client-wise feature fusion, previous studies usually devise graph expansion or clustering methods to reconstruct a global graph or append peer-to-peer collaboration. Different from existing efforts, we study the knowledge-driven recommendation with federated self-supervised learning. The knowledge graph (KG) composed of entities and relations can effectively organize side information of items to enrich plain user-item bipartite graphs. Based on the federated KG embedding procedure, we propose an intent-level pretext task using self-supervised contrast learning to improve the target recommendation task. We innovatively apply the idea of instance-level discrimination to intent-wise supervised signals. Considering the long-tail distribution of relations, the positive pair of intent features is represented locally with the fusion of relevant item knowledge, and negative samples are from other clients. Our intent-level discriminative task aims to maximize mutual information of positive pairs and disperse all intent features in the intent hyperspace. Extensive experiments on typical knowledge-driven recommendation models indicate that the proposed framework can enhance the performance of federated models and even surpass the centralized-training mode.