Data- and distance-aware clustering for scalable wireless federated learning

Introduction Federated learning (FL) enables model training on edge devices using local data while aggregating model updates at a central server without exchanging raw data, thereby preserving privacy. However, achieving satisfactory convergence accuracy with low communication energy remains challenging. This work investigates a three-tier clustered FL (CFL) architecture to improve global training performance and communication efficiency through joint device clustering and resource scheduling. Methods We analyze how clustering strategies influence learning convergence and communication energy consumption. Based on this analysis, we propose a clustering method that jointly accounts for gradient cosine similarity and communication distance. A simplified procedure is further developed for device association and cluster-head selection, with the goals of improving intra-cluster data balance and reducing the overall communication distance to the server. Results Simulations demonstrate that the proposed method consistently improves model accuracy while reducing communication energy consumption compared with random clustering and similarity-based clustering baselines. Discussion These results indicate that jointly considering update similarity and communication distance in CFL can effectively balance learning quality and communication cost, offering a practical approach for energy-efficient federated training in edge networks.