Adaptive Aggregation in Clustered Federated Learning for Wireless Networks

Federated Learning (FL) utilizes local data on edge devices for model training, ensuring data privacy without the need for data transmission between devices to the central server. However, during the training iteration process, the high energy consumption cost of direct communication between devices and the central server poses significant challenges. This paper explores a two-tier wireless FL scheme named clustered federated learning (CFL), where devices within each cluster, i.e., cluster members, perform intra-cluster aggregation at its cluster head and different cluster heads perform inter-cluster aggregation at the central server. We first conduct convergence analysis to indicate that increasing the data richness of each cluster can improve global convergence accuracy. Then we employ the minimum dominating set (MDS) algorithm to select cluster heads and determine the device association strategy. An optimization problem of dynamic aggregation is also proposed, which aims to minimize the weighted sum of training loss and energy consumption. Finally, experimental results validate the efficiency of the proposed clustering algorithm.