FedMixer: Heterogeneity-Resilient Federated Learning for Time Series Forecasting

Abstract

Time series forecasting (TSF) models the underlying patterns of historical and current time series data to make predictions for future trends, which is widely applied in varying areas to enhance decision-making, optimize resource allocation, and improve operational efficiency. However, traditional TSF approaches are increasingly challenged by data privacy concerns and the demand for large datasets. Federated learning (FL) is a distributed training framework that allows models to be trained locally on clients. However, the heterogeneity of time series data across clients significantly impacts model performance. Unlike well-studied domains like image classification, the issue of data heterogeneity in TSF remains underexplored. In this paper, we propose FedMixer, a heterogeneity-resilient FL method designed for TSF tasks. FedMixer employs an adaptive spectral clustering mechanism to group clients and assigns tailored heterogeneous models to each cluster, allowing personalized model configurations that suit varying data characteristics among clients. To achieve aggregation across models with heterogeneous structures, we propose a mutual distillation design to facilitate effective knowledge sharing, allowing clients from different clusters to learn from each other. Extensive experiments on four datasets demonstrate that FedMixer consistently outperforms baseline methods, offering superior prediction accuracy and robustness in data heterogeneity and scalability, while maintaining computational and communication efficiency.