Credit Rating Based on Hybrid Sampling and Dynamic Ensemble

Abstract

The core problem of the credit rating is how to build an efficient and accurate classifier on the imbalanced datasets. The ensemble learning and resampling technology have rich results in this field, but the efficiency of the classifier is limited when dealing with high imbalanced credit data. In this paper, we propose a credit rating model based on hybrid sampling and dynamic ensemble technique. Hybrid sampling can contribute to build a rich base classifier pool and improve the accuracy of the integrated learning model. The combination of hybrid sampling and dynamic ensemble can apply to various imbalanced data and obtain better classification results. In the resampling phase, synthetic minority over-sampling technique (SMOTE) and boundary-sensitive under-sampling techniques are used to process the training data set, and the clustering technique is used to improve the under-sampling and make it more adaptable to high imbalanced credit data, by generating more samples and more representative training subset to enhance the diversity of the basic classifier. A dynamic selection method is used to select one or more classifiers from the basic classifier pool for each test sample. Experiments on three credit data sets prove that the combination of hybrid sampling and dynamic ensemble can effectively improve the performance of the classification.