Adaptive entropy and robust aggregation for client selection in federated credit risk modeling

Federated learning enables collaborative modeling for financial risk prediction including credit default and transaction fraud without exposing raw local data. However, client heterogeneity, Non-IID data distribution, and dynamic participant involvement inevitably degrade model performance and incur excessive communication overhead. Existing client selection methods, either heuristic strategies or deep reinforcement learning solutions, suffer from inadequate exploration–exploitation tradeoff and overlook the inherent instability of local model updates. This paper proposes ERPPO-VA, an entropy-regularized proximal policy optimization framework for federated client selection. We formulate the client‑selection procedure using a heuristic‑augmented Markov‑decision‑process‑style framework, introduce an adaptive entropy coefficient together with prioritized experience replay to dynamically balance exploration and exploitation, and further design an exponential moving average based confidence-weighted aggregation mechanism to suppress the negative impact of unstable clients. Extensive experiments are conducted on three real-world financial datasets, namely CCFD, HCDR and IEEE-CIS, with over one million samples in total. The proposed approach is compared with Random selection, Greedy selection, DDQN, standard PPO, FedProx, MAPPO, and a faithful reimplementation of ERPPO. Under typical Non-IID settings, ERPPO-VA elevates the AUC metric by 4.5–8.7%, cuts total cumulative communication energy consumption by 52% over 200 federated rounds compared to ERPPO, accelerates the convergence speed by approximately 30%, and also achieves superior client fairness. Extensive empirical convergence analysis confirms the effectiveness of the proposed approach.