머신러닝 기반 기업 신용평가 알고리즘 공정성 실증 연구

초록

Machine learning models are widely utilized in corporate credit scoring systems for automated risk assessment and decision support. However, group-level prediction bias arising from differences in data distribution and feature structures between corporate groups can undermine the consistency and reliability of these systems. This study empirically evaluates the algorithmic fairness of machine learning-based corporate credit scoring models using real-world corporate data, setting corporate type as a sensitive attribute. Under a stacking ensemble framework, we applied and compared the effects of three fairness intervention methods: pre-processing, in-processing, and post-processing. In the experiment, the Area Under the Receiver Operating Characteristic curve (AUC) was used as a metric for prediction performance, and various fairness metrics, such as Demographic Parity, Equal Opportunity, and Equalized Odds, were measured for group-level assessment. The analysis revealed a trade-off between fairness improvement and prediction performance loss. Post-processing methods significantly improved fairness but resulted in a decrease in AUC, whereas in-processing methods achieved fairness improvement with relatively minor performance degradation, offering a balance point between performance stability and fairness constraints. These findings provide practical implications for the application of fairness constraints and the selection of appropriate techniques in actual corporate credit scoring systems.

키워드