Cluster-Based Two-Stage Client Selection Strategy for Personalized Federated Learning
Ziwen Huang · 2024
Data heterogeneity poses a significant challenge in federated learning, often leading to slow convergence and reduced accuracy. To address this, we propose an adaptive clustered federated learning algorithm that uses the Silhouette Coefficient to determine the optimal number of clusters, thereby grouping clients with similar data distributions and mitigating the impact of data heterogeneity. Furthermore, we introduce a two-stage client selection strategy—comprising inter-cluster and intra-cluster selection—that enhances model performance and convergence. By considering both the discrepancy between group and global models and the utility of individual clients, our approach ensures a balanced and effective federated learning process. Extensive experiments on benchmark datasets for variable data heterogeneity levels reveal that the proposed algorithm consistently outperforms baselines in terms of higher accuracy (up to 2.91%) and is robust to the number of participants.