A novel multi-stage ensemble model with enhanced outlier adaptation for credit scoring. (1st March 2021)

• Record Type: Journal Article
• Authors: Zhang, Wenyu; Yang, Dongqi; Zhang, Shuai; Ablanedo-Rosas, Jose H.; Wu, Xin; Lou, Yu
• Is Part Of: Expert systems with applications. Volume 165(2021)
• Publication Date: 2021-03-01
• Subjects: Machine learning -- Ensemble model -- Outlier adaptation -- Feature transformation -- Credit scoring; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.113872
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A novel multi-stage ensemble model is proposed for credit scoring. The local outlier factor approach is extended to generate outlier-adapted training sets. A new dimension-reduced feature transformation improves the feature interpretability. The stacking-based ensemble is enhanced through self-adaptive parameter optimization. The proposed model outperforms benchmark ensemble models. Abstract: Credit and credit-based transactions underlie the financial system. After decades of development, artificial intelligence and machine learning have brought new momentum to the credit scoring model. In this study, a novel multi-stage ensemble model with enhanced outlier adaptation is proposed to achieve good predictive power for credit scoring. To reduce the adverse effects of outliers existing in the noise-filled credit datasets, a local outlier factor algorithm is enhanced with the bagging strategy to effectively identify outliers and subsequently boost them back into the training set to construct an outlier-adapted training set that enhances the outlier adaptability of base classifiers. To improve the feature interpretability, a new dimension-reduced feature transformation method is proposed to hierarchically evolve features and extract salient features. To further strengthen the predictive power of the proposed model, a stacking-based ensemble learning method with self-adaptive parameter optimization is proposed to optimize the parameters of selected base classifiers automatically and then to construct a stacking-based multi-stage ensemble model. Ten datasets are tested with six evaluation indicators to evaluate the performance of the proposed model. The experimental results including statistical test results indicate the superior performance of the proposed model and prove its significance and effectiveness.