Communication Efficient Federated Personalized Recommendation

Abstract

Recommendation systems that can correctly predict user preferences in the information age have become an important factor for business success. However, recommendation systems require users' personal information, and centralized collection and processing of user data may lead to serious privacy risks. Good progress has been made in recent years using federated learning techniques for privacy-preserving recommendations, but several key challenges remain to be addressed: most federated recommender systems ignore communication process optimization, inequities in aggregation of federated models, and lack of personalization to users. In this paper, we propose a communication efficient and fair personalized federated recommendation approach (CFFR) to address these challenges. CFFR uses adaptive client group selection to personalize models while accelerating the training process. A fair-aware model aggregation algorithm is proposed that adaptively captures the performance and data imbalance among different clients to address the unfairness problem. Extensive experimental results demonstrate the effectiveness and efficiency of our proposed method.