Building Credit Scoring Systems Based on Support-based Support Vector Machine Ensemble

This paper proposes a new strategy - support-based SVM ensemble for building credit scoring systems. Different from the commonly used "one-member-one-vote" majority-ruled ensembles, our proposed new framework aggregates degrees of support, or confidence levels, of several SVM classifiers to generate the final classification results that represent the consensus of the SVM. Decision values of a member SVM classifier are a good measurement of its support to positive or negative classification of an unlabeled sample. Two publicly available credit dataset have been used to test the usefulness and predicting power of the new approach. Results of both tests indicated clearly that the new approach outperformed the other three commonly used approaches: single, single best, and majority-rule ensemble.