A Novel Group Management Scheme of Clustered Federated Learning for Mobile Traffic Prediction in Mobile Edge Computing Systems

Abstract

This study developed a novel group managementscheme based on clustered federated learning (FL) for mobiletraffic prediction (referred to as FedGM) in mobile edge com-puting (MEC) systems. In FedGM, to improve the convergencetime during the FL procedure, we considered multiple MECservers to first be clustered based on their geographic locationsand augmented data patterns as references for clustering. In eachcluster, by alleviating the straggler impact owing to the hetero-geneity of MEC servers, we then designed a group managementscheme that optimizes i) the number of groups to be createdand ii) the group association of the MEC servers by minimizingtheir average idle time and group creation cost. For this purpose,we rigorously formulated analytical models for the computationtime for local training and estimated the average idle time byapplying different frequencies of local training over the MECservers. The optimization problem was designed using a non-convex problem, and thus a genetic-based heuristic approachwas devised for determining a suboptimal solution. By reducingthe average idle time, thereby increasing the workload of theMEC servers, the experimental results for two real-world mobiletraffic datasets show that FedGM surpasses previous state-of-the-art methods in terms of convergence speed with an acceptableaccuracy loss.