Participation-Aware Search-Based Joint Client Selection and Adaptive Aggregation for Federated Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated learning is affected by differences in client data, behavior, and resource requirements. This paper develops and evaluates a search-based, client-participation-aware strategy for federated client selection and adaptive aggregation. It uses a genetic-search procedure with a multi-criteria objective. To avoid unstable aggregation, the searched weights are combined with the data-size weights. The experiment is performed on the non-IID MNIST dataset using a lightweight CNN. The proposed method is compared with standard federated learning baselines: Federated Averaging (FedAvg) and Federated Proximal (FedProx). The comparison also includes selection strategies such as loss-based and resource-aware selection. The method achieves the best final predictive performance, with an accuracy of 0.31, a macro-F1 of 0.22, and a loss of 2.14. However, it is limited by the final client disparity gap. Its selection entropy and stability variance also do not outperform all baselines. The proposed multi-criteria search improves predictive performance. Furthermore, ablation analysis shows performance deterioration when the participation component is removed. Similar degradation occurs when the searched weight coefficient in the aggregation varies. The fitness-weight and aggregation-ratio sensitivity analysis reports their influence on both predictive and participation behavior. However, the statistical analysis found that the search-based method is not significantly more predictive than the baselines. The results highlight the usefulness of participation-aware search for improving prediction-oriented federated training under non-IID data. Further, it is suggested that client-level balance requires more explicit disparity-aware design.