Clustered federated learning based on fuzzy time series morphology
Dong Lv, Xingwang Li, Yuwei Wang, Yitao Li · 2025
As federated learning (FL) becomes increasingly prevalent across various fields, its advantages in addressing clustering problems with time series data are becoming more apparent. Especially in finance, meteorology, and healthcare, clustering analysis of time series data is crucial for constructing accurate global models. This paper proposes a novel FL framework, Clustered Federated Learning Based on Fuzzy Time Series Morphology (FTS-CFL), aimed at achieving more precise clustering by capturing the similar morphological characteristics of time series within clusters. In the FTS-CFL framework, we first conduct morphological analysis on the time series data from each client to extract key timeseries features. Subsequently, we utilize a fuzzy clustering algorithm to cluster these features, which can handle the uncertainty and complexity of time series data, thereby aggregating clients with similar time series morphology. In this way, our method can effectively reduce the negative impact of non-independent and identically distributed (non-IID) data on model performance without compromising data privacy. To validate the effectiveness of the FTS-CFL algorithm, we conducted extensive experiments on multiple real-world time series datasets. The experimental results show that compared to existing FL methods, FTS-CFL achieves significant improvements in model accuracy and convergence speed. Additionally, we explored the performance of FTS-CFL under different parameter settings and its robustness against varying degrees of non-IID data.