Addressing data imbalance for federated recommender systems: a rebalancing framework with gradient alignment regularization

Federated recommender systems (FRSs) utilize decentralized data to offer personalized and privacy-preserving recommendations. Existing studies on FRSs overlook the issue of data imbalance, with clients possessing varying data volumes. Experimental results demonstrate that data imbalance across clients reduces the model’s performance in FRSs. In this paper, we propose BalFed for FRSs to address data imbalance. In BalFed, we devise gradient alignment regularization to reduce the negative impact of clients with excessive or insufficient data by constraining the gradient deviation between local and global models during training. Furthermore, we design an improved pseudo-labeling technique to address global data imbalance. The improved pseudo-labeling technique utilizes the inherent local and global models in FRSs to set dual thresholds, and thresholds are adaptively adjusted based on the performance of local models. We instantiate BalFed on various datasets, recommendation models, and federated algorithms. Evaluation results show an average performance improvement of 8.92% without introducing additional communication overhead, demonstrating its effectiveness.