An Enhanced Dynamic Ensemble Selection Classifier for Imbalance Classification With Application to China Corporation Bond Default Prediction

Abstract

China corporation bond default prediction is important and can be formulated as an imbalance classification problem solved by static ensemble classifiers. However, dynamic ensemble selection (DES) classifiers have not been applied to this typical problem in the context of business research. DES classifiers are capable of selecting an ensemble classifier for each test sample, leading to better classification performance than static ensemble classifiers. Most existing DES classifiers can not address imbalance classification optimally and only use single criteria of competent for classifier selection, resulting in sub-optimal classification performance. In this paper, we propose an enhanced DES classifier, named META-DES-Diversity, that inherits strengths of data sampling, meta-learning, criteria of diversity, and dynamic weighting fusion scheme to alleviate such limitations. Specifically, the synthetic minority over-sampling technique is initially used to balance the training set before generating a candidate classifier pool. To select an ensemble classifier with a highly competent, the meta-learning framework META-DES is utilized to consider multiple criteria of competence. In complement with the meta-leaning framework, a two-phase selection strategy is utilized to perform competence and diverse ensemble classifier selection. Note that a competence-driven decision fusion scheme is employed to effectively fuse classification results from selected ensemble classifiers. Experiments on 14 two-class imbalanced data sets from the KEEL repository and one self-collected China corporation bond data set show the effectiveness and superiority of the proposed enhanced DES classifier.