Clustered federated learning (CFL) addresses the challenge of data heterogeneity in federated learning (FL) by customizing models for different groups of clients. However, existing CFL methods heavily rely on indirect metrics, such as model parameters, gradient information, or loss function values, for client clustering. These approaches often fail to fully capture the diversity and intrinsic characteristics of client data distributions, leading to inaccurate representations of client data features. To address this issue, we propose a novel CFL framework called vector quantization-based CFL (VQCFL). First, we introduce a vector quantization network (VQNet), which effectively captures the intrinsic structure of client data by mapping the local feature space into discrete feature dictionary vectors. In addition, to prevent drift in the feature dictionary vectors, we propose a global feature anchor strategy that aligns feature dictionary vectors across clients, ensuring consistent updates within the same feature space. Furthermore, we present a novel cross-cluster knowledge-sharing mechanism that integrates feature information from different clusters through global aggregation of feature dictionary vectors. Combined with a personalized cross-cluster classifier weight adjustment strategy, this mechanism significantly enhances the model's generalization performance in the presence of mixed data heterogeneity. Experimental results under various settings demonstrate that VQCFL achieves superior local personalization and global generalization performance.
Chen et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: