CT-GAN을 활용한 중소기업 채무불이행 예측 연구

In this paper, we study a machine learning-based prediction of SME defaults using large-scale SME data from Bank A in Korea. Based on SME loan information for five years from 2017 to 2022, a total of 616,007 data were collected and used to train machine learning models. In general, SME datasets for default prediction are highly imbalanced, with a significantly fewer default companies compared to normal companies. In this study, to compensate for insufficient number of data and imbalanced data, we generated synthetic data using CT-GAN(Conditional Tabular GAN), a generative artificial intelligence technique specialized in table data. By training machine learning models on datasets that included synthetic data, it was found that the predictive performance improved. Among the machine learning models, tree-based models such as XGBoost showed the best performance, while artificial neural network models performed relatively poorly. Additionally, this study used SHAP, a representative method of explainable artificial intelligence(XAI), to identify key variables in the machine learning models trained with synthetic data. Regardless of the synthetic data generation method used, the interest coverage ratio variable was found to be the most important. In this study, we provide important implications for the utilization of synthetic data and the interpretation of Explainable AI(XAI) in applying machine learning models to the credit evaluation field of SMEs in Korea.