A loan default discrimination model using cost-sensitive support vector machine improved by PSO

• Published in: Information technology and management Vol. 14; no. 3; pp. 193 - 204
• Main Authors: Cao, Jie, Lu, Hongke, Wang, Weiwei, Wang, Jian
• Format: Journal Article
• Language: English
• Published: Boston Springer US 01.09.2013; Springer Nature B.V
• Subjects: Accuracy; Bank loans; Business and Management; Computer Communication Networks; Cost engineering; Costs; Credit risk; Credit scoring; Data Structures and Information Theory; Datasets; Default; Design; Discrimination; Genetic algorithms; Image retrieval; IT in Business; Loans; Machine learning; Marketing; Neural networks; Operations Research/Decision Theory; Pattern recognition; Portfolio management; Risk assessment; Studies; Support vector machines; Swarm intelligence; China; Cost sensitive learning; Support vector machine; Particle swarm optimization; Loan default prediction; Attribute reduction
• ISSN: 1385-951X; 1573-7667
• DOI: 10.1007/s10799-013-0161-1

This study proposes a novel PSO–CS-SVM model that hybridizes the particle swarm optimization (PSO) and cost sensitive support vector machine (CS-SVM) to deal with the problem of unbalanced data classification and asymmetry misclassification cost in loan default discrimination problem. Cost sensitive learning is applied to the standard SVM by integrating misclassification cost of each sample into standard SVM and PSO is employed for parameter determination of the CS-SVM. Meantime, the financial data are discretized by using the self-organizing mapping neural network. And the evaluation indices are reduced without information loss by genetic algorithm for decreasing the complexity of the model. The effectiveness of integrated model of CS-SVM and PSO is verified by three experiments comparing with traditional CS-SVM, PSO–SVM, SVM and BP neural network through real loan default data of companies in China. The corresponding results indicate that the accuracy rate, hit rate, covering rate and lift coefficient are improved dramatically by the developed approach. The proposed method can control the different types of errors distribution with various cost of misclassification accurately, reduce the total misclassification cost largely, and distinguish the loan default problems effectively.