From Non-IID to IID: Mobility-Aware Hierarchical Federated Learning With Client-Edge Association Control

Deploying federated learning (FL) in wireless network with hierarchical client-edge-cloud architecture enables large-scale distribution collaboration without long-distance communication latency. However, the ongoing edge dynamics with uncertain client mobility and imbalanced data distributions, poses great challenge for collaboration efficiency of FL. In this work, we first model the client mobility with a Markov chain, and formulate the minimization of performance degradation as a client-edge association control problem. With the analysis of client mobility patterns, we propose ALPHA, a new client-edge association control framework for mobility-aware FL, to reshape the edge-level data distributions close to i.i.d in both offline and online mobility scenarios. In the offline scenario with deterministic client mobility trajectories, we leverage alternating optimization theory to transform the client-edge association control problem into a weighted bipartite b-matching problem, and derive an efficient solution with linear relaxation and dependent rounding techniques. As for the online scenario, where each client arrives at different edge access points (APs) in an online manner, we design a fast and simple online subgradient projection algorithm with a bounded regret to make an online decision on client-edge association. Extensive experiment results on three public datasets and a real-world mobility trajectory dataset show that ALPHA has a superior learning performance with 1.40× – 2.89× convergence speedup compared to state-of-the-art solutions.