Multicriteria Scoring for Cluster and Client Selection in Heterogeneous Hierarchical Federated Learning

In hierarchical federated learning, client selection is typically performed at edge servers and the central server. Edge servers perform intermediate model aggregation before transmitting updates to the central server, which aggregates the local models’ parameters. One of the challenges in federated learning is coping with the heterogeneity of clients’ resources and datasets, which can lead to slow global model convergence and poor performance. This paper introduces a client selection algorithm to address heterogeneity and scalability challenges in hierarchical scenarios. Our algorithm ranks client clusters based on a scoring function that incorporates resource availability, communication quality, and data-related attributes. Then, clients are selected from the top-ranked cluster. The proposed method is formulated as a combinatorial optimization problem and evaluated via emulation using the MNIST and CIFAR-10 datasets under heterogeneous scenarios. Experimental results show that, compared to existing clustering-based baselines, our method reduces CPU utilization by up to 42%, memory usage by 58%, and energy consumption by 48%. These improvements are obtained with negligible loss in model accuracy on MNIST and a slightly higher reduction in accuracy on CIFAR-10, consistent with its higher task complexity.