Comparative Analysis Research on Machine Learning Models in Credit Risk Assessment

Abstract

Credit risk assessment is crucial for the risk management and control of financial institutions, but it faces challenges such as sample imbalance, complex characteristics and the lack of model interpretability. In this study, two public datasets, "Give Me Some Credit" and "Loan Default", were used. The Synthetic Minority Over-Sampling Technique (SMOTE) was employed to balance the sample distribution and conduct feature engineering. Construct new features such as the income-debt ratio (Income_Debt_Ratio) to reduce variable redundancy. Meanwhile, by comparing the model's different performance among logistic regression, Random Forest (RF), the study improves the training efficiency. The experiment results depict that the integrated models (XGBoost, LightGBM) perform better on both datasets, with an average accuracy rate of 94% and an AUC value of 0.98 compared with the traditional models. Furthermore, SHapley Additive exPlanations (SHAP) values were used to develop the interpretability analysis. This study provides credit institutions with a high-precision and interpretable model construction scheme, and verifies the generalization ability of the model through cross-datasets, laying a theoretical and practical foundation for future credit risk control and the construction of an integrated system.