XAI-RAI-HITL 통합 구조를 활용한 신용평가 시스템의 공정성 개선과 설명가능성 확보에 관한 연구

초록

This study proposes an integrated XAI–RAI–HITL evaluation framework to mitigate bias and opacity in AI-based credit scoring systems. The framework operates as a unified process in which model transparency, fairness, and human oversight reinforce one another. Within the XAI component, SHAP (Shapley Additive Explanations) was employed to quantify feature contributions and visualize model decision logic, providing interpretable evidence to support downstream fairness and oversight decisions. Building on these explanations, the RAI component evaluated and improved fairness using Equal Opportunity Difference (EOD) and Disparate Impact (DI), applying group-specific threshold adjustments that substantially reduced EOD without compromising predictive performance. The HITL component operationalized human oversight by simulating reviewer intervention for “gray-zone” cases with predicted probabilities between 0.40 and 0.60. Under the assumption of human correction to ground truth, this mechanism improved Accuracy, Precision, and Recall, demonstrating how targeted human involvement can reduce decision errors while maintaining system stability. Overall, the findings empirically confirm that a integrated XAI–RAI–HITL structure can simultaneously enhance fairness and explainability while preserving model accuracy. The proposed framework provides a practical foundation for developing transparent, accountable, and trustworthy credit scoring systems suitable for high-stakes decision-making environments.

키워드