Enhancing credit scoring decisions with FedMOE: A Federated Mixture of Experts for imbalanced and heterogeneous data

Data silos, class imbalance, and data heterogeneity often hinder the effectiveness of credit scoring, which is fundamental to financial risk management. While federated learning (FL) addresses data privacy by enabling collaborative training without sharing raw data, existing FL methods struggle to simultaneously handle severe class imbalance and heterogeneous data distributions. To overcome these limitations, this paper introduces FedMOE, a novel Federated Mixture of Experts framework. The FedMOE architecture is composed of three specialised components: the Global Expert for learning shared knowledge, the Local Experts for capturing client-specific patterns, and the Imbalance Expert trained on a distribution-reversed dataset to enhance the detection of potential defaulters. A dynamic router adaptively assigns weights to these experts for each prediction. We conducted comprehensive experiments on four public credit scoring datasets under both IID and non-IID settings. The results demonstrate that FedMOE significantly outperforms state-of-the-art FL baselines, particularly in recall and the Kolmogorov–Smirnov (KS) statistic, two metrics crucial for risk control. Our findings indicate that FedMOE offers a robust and privacy-preserving solution that effectively tackles the dual challenges of data heterogeneity and class imbalance, thereby advancing the reliability of federated credit scoring systems.