A focal-aware cost-sensitive boosted tree for imbalanced credit scoring. (1st December 2022)

• Record Type: Journal Article
• Authors: Liu, Wanan; Fan, Hong; Xia, Min; Xia, Meng
• Is Part Of: Expert systems with applications. Volume 208(2022)
• Publication Date: 2022-12-01
• Subjects: Credit scoring -- Cost-sensitive -- LightGBM -- Interpretability; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118158
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Credit scoring is an effective tool for banks or lending institutions to identify potential bad lenders and creditworthy applicants. Boosting ensemble approaches have made appealing progress for credit scoring. However, classical boosting ensemble models realize credit scoring to minimize the misclassification error of credit datasets, which implies that the misrecognition of good applicants and the misrecognition of bad applicants incur equal costs, while credit scoring is a typical cost-sensitive issue. In this paper, we propose a focal-aware cost-sensitive light gradient boosting machine (LightGBM-focal) for credit scoring. The proposed method considers a boosted tree as a base framework, which ensures the robustness of predicting risky and non-risky loans. Moreover, a cost-aware loss function-focal loss is introduced to address the skew-sensitivity issue in credit scoring. Besides, we investigate the effectiveness of LightGBM-focal by incorporating two interpretation methods, feature importance score and partial dependent plots. The experimental results on four accessible imbalanced credit datasets show that LightGBM-focal can effectively reduce the misrecognition rate of risky loans, thereby reducing the potential default risk of non-performing lenders. Besides, the performance on AUC, Type-I error, Type-II error and Gmean shows the superiority of LightGBM-focal. Finally, based on the good interpretability of the tree algorithm, the feature importance rankings and partial dependence plots further demonstrate that LightGBM-focal is a good solution for addressing the biased prediction problem in imbalanced credit scoring. Graphical abstract: Highlights: A cost-sensitive LightGBM is proposed for imbalanced credit scoring. Focal loss is embedded to transform LightGBM into a cost-sensitive version. We corroborate the validity of the algorithm by interpreting the results. Feature importance globally interprets the prediction results of LightGBM-focal. PDP tool locally interprets the credit scoring of LightGBM-focal.