Credit scoring is an important financial tool for banks to determine whether to issue the loan to potential borrowers. Ensemble algorithms, which mainly can be divided into bagging ensembles and boosting ensembles, have shown great promise for credit scoring. However, some problems need to be further addressed: (1) Bagging-type algorithms enrich the feature diversity while keep the training target unchanged. However, these methods acting as feature augmentation process that highly rely on the training targets may increase the statistical similarity of the prediction results. (2) Though boosting-type ensemble algorithms avoid the problem of high prediction similarity, boosting algorithms always work on the original credit features leading to the lack of feature diversity. (3) A more intelligent credit risk management system should well balance the accuracy and its interpretability. Based on the above considerations, in this study, a step-wise multi-grained augmented gradient boosting decision trees (mg-GBDT) is proposed for credit scoring. In the proposed method, a multi-grained scanning is introduced for feature augmentation, which enriches the input feature of GBDT; the GBDT-based step-wisely optimization mechanism ensures low-deviation of credit scoring; besides, the proposed method inherits the good interpretability of tree-based structure, which provides intuitive reference results for policy-makers. Experiments on 6 credit datasets show that the proposed method outperforms classic GBDT. Moreover, numerical results indicate that mg-GBDT provides an alternative to neural network-based feature enhancement. Finally, the global interpretation results and the visualized decision path demonstrate that mg-GBDT can be a good choice for accurate credit scoring interpretation.

信用评分是银行确定是否向潜在借款人发放贷款的重要财务工具。合奏算法主要可分为装袋合奏和增强合奏，对信用评分显示出了很大的希望。但是，还有一些问题需要进一步解决：（1）Bagging型算法丰富了特征多样性，同时保持训练目标不变。但是，这些高度依赖训练目标的充当特征增强过程的方法可能会增加预测结果的统计相似性。（2）尽管Boosting型集成算法避免了高预测相似性的问题，Boosting算法始终对原始信用特征起作用，导致缺乏特征多样性。（3）更智能的信用风险管理系统应在准确性和可解释性之间取得良好的平衡。基于上述考虑，在本研究中，提出了一种用于信用评分的逐步多粒度增强梯度提升决策树（mg-GBDT）。该方法引入了一种多粒度扫描进行特征增强，丰富了GBDT的输入特征。基于GBDT的逐步优化机制可确保信用评分的偏差小；此外，该方法继承了基于树的结构的良好解释性，为决策者提供了直观的参考结果。在6个信用数据集上进行的实验表明，该方法优于经典的GBDT。此外，数值结果表明mg-GBDT提供了基于神经网络的特征增强的替代方法。