FedRFC: Federated Learning with Recursive Fuzzy Clustering for improved non-IID data training

In contemporary times, artificial intelligence is extensively applied across domains, concurrently raising concerns about privacy breaches. In response, federated learning has emerged as a promising solution that allows multiple parties to collaboratively train shared models without sharing local data. Nonetheless, the prevalence of non-IID data among clients poses challenges for traditional federated learning approaches, thereby limiting their efficacy in practical applications. To address this, we propose FedRFC, a novel clustered federated learning framework. This framework employs a recursive fuzzy clustering algorithm to iteratively partition clients into overlapping clusters, thereby improving the training effectiveness of the federated learning models on non-IID data. The performance of FedRFC is evaluated using four real-world datasets and six synthetic datasets, demonstrating superior performance compared to all baseline federated learning methods when applied to non-IID datasets. The enhancements amount to around 3% on the mildly non-IID datasets, approximately 5% on the moderately non-IID datasets, and exceeding 10% on the extremely non-IID datasets. The results indicated our method could effectively utilize the local data and achieve successful learning from non-IID data.