범주 불균형 금융 문제의 해결을 위한 부스팅 학습

Abstract

The class imbalance problem is essential to be solved in order to improve the performance of classification and prediction models. This study proposes an AUCBooost technique that introduces a direct performance optimization technique to improve the performance of boosting algorithms applied to class imbalance problems in the financial sector. In this study, the performance of AUCBoost is verified on class imbalance problems such as corporate bankruptcy, card insolvency, and card fraud. For comparative analysis of performance, logistic, AdaBoost, GBM, and XGBoost are adopted as benchmark models. The results of repeated 10-fold cross-validation are as follows. First, unlike conventional algorithms only focusing on majority class in class imbalance problems, AUCBoost shows a balanced learning behavior that simultaneously considers the specificity of multiple class and the sensitivity of minority class. Second, compared to the benchmarking model, RMA shows that AUCBoost shows significantly improved performance in terms of AUC. Third, in class balanced data generated from data sampling, AUCBooost also shows the significantly improved performance compared to benchmarking models.