Guiding Federated Graph Recommendation with LLM-encoded knowledge

The paper presents a novel framework that integrates LLM-encoded knowledge into federated graph recommendation systems to enhance the aggregation of user-item interactions while preserving privacy. By leveraging a frozen LLM to summarize client interaction patterns into semantic vectors, the framework enables the central server to align structural representations across non-IID clients more effectively. Experimental results demonstrate that this approach significantly improves recommendation accuracy compared to existing federated graph methods, highlighting its potential for practitioners in developing more effective and privacy-preserving recommendation systems.