Clustered Federated Learning with Slightly Skewed Labels

Jiaming Pei, Wei Li

Abstract: Clustered federated learning methods are proposed to realize the personalization of federated learning. The core of these methods is KMeans while it cannot identify sample from which cluster under slight non-IID data distribution. This paper proposed Gaussian mixture cluster (GMCFL) to measure the probability and uncertainty of that samples belongs to which clusters. This method efficiently aggregated the model parameters between clusters and were robust to non-IID data distribution as well. The empirical results demonstrated our method had better performance than other state-of-the art clustered federated learning methods.