FedMVC: A Client Selection Algorithm in Federated Learning via Multi-view Clustering

In federated learning systems, clients engage in iterative model training through coordination with a central server. A critical challenge arises from the practical constraint that not all clients can participate in every training round, necessitating efficient client selection mechanisms. Current approaches to client selection suffer from two key limitations: (1) they often fail to account for the inherent diversity in client data distributions, and (2) they may systematically exclude clients with potential contributions. To address these shortcomings, we propose a novel Multi-View Clustering-based client selection strategy specifically designed for Federated Learning with Non-IID data (FedMVC). It utilizes both gradients and losses to analyze the direction of model update and model effect, and then achieves the client classification with similar data distributions into a cluster. Meanwhile, it incorporates both accuracy rate and the historical selection time as dual criteria, systematically balancing the exploitation of high-performing clients with the exploration of potentially valuable participants. The experiments on CIFAR-10 datasets verify the effectiveness and robustness of the proposed method.