FedLossMix: A Loss-Based Adaptive Federated Learning Approach for Credit Scoring

This study aims to address the challenge of systemic financial risk mitigation by improving credit scoring models trained under privacy constraints. It introduces Federated Loss-Based Mixing (FedLossMix) which is an adaptive Federated Learning (FL) aggregation method designed to handle data heterogeneity and class imbalance across financial institutions. The research develops an adaptive aggregation strategy that assigns dynamic weights to client model updates based on local validation loss. The method is evaluated using multiple heterogeneous credit risk datasets within a federated environment and compares its performance to traditional FL aggregation techniques such as FedAvg. FedLossMix demonstrates superior convergence stability and improved representational fairness across non-IID and imbalanced financial datasets. Experimental results show that the proposed approach consistently outperforms conventional FL aggregation methods in predicting borrower default probabilities. The proposed FedLossMix framework provides a robust and equitable way to collaboratively train Federated Learning (FL) models without sharing sensitive data. This method mitigates data-sharing and privacy risks by enabling better model performance under heterogeneous data conditions without requiring data centralization.

Federated Learning, Artificial Intelligence, Machine Learning, Distributed Systems, Credit Risk Management