기업부실 예측모형의 성과개선을 위한 기하평균 최적화 기반의 부스팅 학습

The class imbalance problem is one of data quality issues that affect theperformance of classification models. This problem has drawn the attention ofresearchers and many approaches have been developed to address this issue. Thispaper proposed GMOPTBoost(Geometric Mean Optimization-based boostingalgorithm) to improve the performance of AdaBoost ensembles applied to solve theclass imbalance problem in bankruptcy prediction. GMOPTBoost derived the bestprediction by applying Gaussian gradient descent method to find the set of weightsassigned to base classifiers to optimize GM. The original dataset was collected fromKorean commercial bank. The original dataset was split into four datasets withdifferent IR values to determine whether GMOPTBoost contribute to theperformance improvement of AdaBoost ensembles. Three times of 10-fold crossvalidations are performed on each of imbalanced datasets. The performance of allclassification models were evaluated in terms of accuracy, GM, AUC. The mainfindings are as follows. First, the class imbalance problem has a negative effect onthe performance. As IR values increase, the performances of AdaBoost andGMOPTBoost ensembles decrease. Second, GMOPTBoost makes a significantcontribution to performance improvements of AdaBoost ensembles trained onimbalanced datasets at 1% significance level. Finally, GMOPTBoost ensemblesprovide more stable classification performance than AdaBoost ensembles.