History-Aware Multi-Objective Client Selection for Federated Learning under Statistical and System Heterogeneity
Federated learning reduces raw-data movement, but partial participation under statistical and system heterogeneity makes the selected client cohort a major source of optimisation bias and delay. This study proposes a history-aware multi-objective selector that combines opt-in data-quality metadata, exponentially smoothed training utility, marginal label-balance gain, participation deficit, and an online latency penalty without querying every client's current private loss. A paired protocol compares Random FedAvg, Power-of-Choice, an Oort-inspired policy, a class-balance greedy policy, and the proposed method on Digits, MNIST, and CIFAR-10. The confirmatory benchmark uses 50 clients, three Dirichlet non-IID levels, ten fresh seeds, and Holm-adjusted Wilcoxon tests. Relative to Random FedAvg, the proposed method improves final-five-round accuracy in all six MNIST/CIFAR-10 conditions by 1.11–5.53 percentage points, with four adjusted-significant conditions. Component ablations show distinct roles. Label balance improves accuracy AUC in all six conditions, the adaptive penalty reduces simulated synchronous latency, and the fairness term improves participation uniformity without eliminating all dataset-dependent concentration. Accuracy gains remain positive when 6%, 10%, or 20% of clients participate per round. The evidence supports a practical accuracy–balance–latency trade-off rather than universal dominance, and the study defines clear boundaries for privacy and hardware claims.