Hierarchical Federated Learning with Adaptive Clustering on Non-IID Data

Federated learning (FL) in a mobile edge network faces challenges from both communication and learning perspectives. The typically non-i.i.d. data can lead to slow convergence and low accuracy. To ease these challenges, frequent communications between user equipments (UEs) and the central macro base station (MBS) are necessary, aggravating the communication burden. In this paper, a novel hierarchical FL framework is proposed to alleviate the biased convergence of the global model, achieving better communication and computation efficiency. Specifically, the UEs are adaptively clustered and allocated to specific small base stations (SBSs) according to channel conditions, geographic locations, and data distributions. The SBSs are further aggregated to the MBS, forming a hierarchical FL framework. The joint user clustering and wireless resource allocation optimization problem is formulated. To solve this problem, a cross entropy (CE) based method with low computational complexity is proposed. Simulation results validate that the proposed hierarchical FL system can save more than 87 percent training time under the EMNIST Letters dataset, achieving fast convergence and significantly improving the system efficiency.