Fed-LyAHP: A Client Selection Method for Federated Learning based on Analytic Hierarchy Process and Lyapunov Optimization in Mobile Edge

Federated Learning (FL) is a distributed machine learning paradigm that trains global model parameters by aggregating model parameters from distributed clients, offering good privacy protection due to its distributed training characteristics. Additionally, due to distributed training, the heterogeneity between different clients can lead to instability in the global model training process. Therefore, it is necessary to select the right clients to participate in the global training. At the same time, since the local computing resources and battery power of most distributed clients are less than those of servers, from the perspective of reducing the long-term training cost of each client, it is beneficial for clients participating in the global model parameter update to be distributed across multiple clients. This paper proposes a new FL client selection method (Fed-LyAHP) based on the Analytic Hierarchy Process (AHP) and Lyapunov optimization, which can not only balance model accuracy and training time in the client selection process, but also introduce virtual queues to reduce the probability of a single client participating in the global update all the time. Experiments show that Fed-LyAHP is superior to Greedy methods, RBCS F and FedCS, in terms of training time, model accuracy, and queue length. Experiments on two different datasets show that Fed-LyAHP outperforms other algorithms by at least 10% in terms of queue length. In the fixed accuracy experiment, only 70% of the iterations of other algorithms were used to achieve the target accuracy. In experiments with fixed run-time, the accuracy can be improved by at least 10%.