RMkNN and KNORA-IU: combining imbalanced dynamic selection techniques for credit scoring

Credit scoring has become a critical tool for financial institutions to discriminate "bad" applicants from "good" ones. One common characteristic of the credit datasets is the imbalance between good and bad applicants, with low defaults (no paid loans). Ensemble classification methodology is widely used in this field. However, dynamic ensemble selection approaches to imbalanced datasets have drawn little consideration. This study aims to measure the performance of the combination of two recent dynamic selection techniques for imbalanced credit scoring datasets, Reduced Minority k-NN (RMkNN) and KNORAImbalanced Union (KNORA-IU). We comprehensively evaluate the proposed combination against state-of-the-art competitors on six real-world public datasets and one private one. Experiments show that this combination improves the classification performance on the evaluated datasets in terms of AUC, balanced accuracy, H-measure, G-mean, F-measure, and Recall.