Collaborative Optimization in Federated Recommendation: Integrating User Interests and Differential Privacy

Abstract

This paper addresses the conflict between personalization effectiveness and privacy protection in federated recommendation systems. It proposes a collaborative optimization method that integrates local interest guidance with a differential privacy mechanism. The goal is to enhance both recommendation performance and security under the condition of data isolation across multiple parties. Specifically, a personalized model aggregation strategy based on local interest embeddings is designed. By incorporating user preference features into the global model update process, the model can adaptively capture individual differences among clients during aggregation. At the same time, to reduce the negative impact of privacy protection on model performance, a differential privacy-driven personalized update mechanism is introduced. This mechanism ensures the non-inferability of user data while applying a gradient-guided noise regulation strategy. It helps preserve the local model's ability to represent individual interests. Multiple comparative experiments conducted on standard recommendation datasets show that the proposed method outperforms representative federated recommendation models across various metrics. It also demonstrates strong robustness and stability under highly non-independent data distributions and high noise settings. Further ablation studies confirm the independent and joint contributions of the two key modules in enhancing the model's personalization capacity and resistance to interference. These results highlight the method's ability to achieve an effective balance between privacy protection and recommendation quality.