설명가능한 인공지능을 활용한 개인 신용카드 채무불이행 예측 연구

초록

This study aimed to empirically compare the performance of the logistic regression model and the XGBoost model for predicting individual credit card default, and to enhance the interpretability of the models through eXplainable Artificial Intelligence (XAI). Using American Express data from Kaggle, I analyzed feature importance through coefficient-based, gain-based, and SHAP (SHapley Additive exPlanations)–based methods with the following results. First, while both models demonstrated excellent predictive performance, XGBoost achieved zero False Negatives (Type II error) in the confusion matrix analysis, thereby perfectly identifying all actual default customers. This has important practical implications in risk management as it helps prevent financial losses for financial institutions. Second, the SHAP analysis confirmed that credit score, previous defaults, and credit limit usage are key variables in predicting default. SHAP, based on Shapley values, clearly decomposes the contribution of each variable to individual predictions while considering interactions among variables, leading to significantly improved interpretability compared with traditional methodologies such as coefficient-based and gain-based analysis. This provides important grounds for securing the transparency and credibility of models required in financial practice.

키워드