主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (9): 21-28.doi: 10.16381/j.cnki.issn1003-207x.2023.0635

• • 上一篇    下一篇

面向不平衡数据的SMOTE-BO-XGBoost集成信用评分模型研究

李爱华1(), 刘婉昕1, 陈思帆1, 石勇2   

  1. 1.中央财经大学管理科学与工程学院,北京 102206
    2.中国科学院大学管理学院,中国科学院虚拟经济与数据科学研究中心、 大数据挖掘与知识管理重点实验室,北京 100080
  • 收稿日期:2023-04-17 修回日期:2024-01-14 出版日期:2026-09-25 发布日期:2026-09-01
  • 通讯作者: 李爱华 E-mail:neu_aihua@126.com
  • 基金资助:
    国家自然科学基金项目(71932008)

Research on SMOTE-BO-XGBoost Ensemble Credit Scoring Model for Unbalanced Data

Aihua Li1(), Wanxin Liu1, Sifan Chen1, Yong Shi2   

  1. 1.School of Management Science and Engineering,Central University of Finance and Economics,Beijing 102206,China
    2.School of Management,Center for Virtual Economy and Data Science Research,Key Laboratory of Big Data Mining and Knowledge Management,University of Chinese Academy of Sciences,Beijing 100080,China
  • Received:2023-04-17 Revised:2024-01-14 Online:2026-09-25 Published:2026-09-01
  • Contact: Aihua Li E-mail:neu_aihua@126.com

摘要:

信用评估研究中,有两类问题经常会遇到:一是数据不平衡会对模型的准确率有很大的影响;二是对非平衡数据处理较好但准确率不理想。为了同时解决这两方面的问题,本文构建了一种新的集成学习模型来解决数据类别不平衡下的信用风险评估问题——基于SMOTE-BO-XGBoost集成算法的类别非平衡信用评分模型。利用SMOTE方法处理样本不平衡问题,使用贝叶斯优化方法获取最优模型参数,进而构建优化后的集成分类模型。实验分析中,以准确率、AUC值、F1得分和第一类错误为评估指标,先后训练XGBoost、BO-XGBoost、SMOTE-XGBoost、SMOTE-BO-XGBoost四种集成模型,同时将SMOTE-BO-XGBoost模型与传统信用评分模型比较。实验结果表明:集成是有效的,其中SMOTE-BO-XGBoost模型具有最优的性能,优于一般的集成学习模型与传统分类算法,能够有效解决非平衡样本问题,为银行部门提供了一种有效的个人信用风险评估方法。

关键词: 信用评分, 集成学习, 非平衡数据, XGBoost

Abstract:

To handle the issues of some high-accuracy models neglecting data imbalance and other studies effectively handling imbalanced data but not achieving satisfactory accuracy, a novel ensemble learning model is constructed for credit risk assessment under imbalanced data categories—a class-imbalanced credit scoring model based on the SMOTE-BO-XGBoost ensemble algorithm. Firstly, the model tackles the sample imbalance problem through the Synthetic Minority Over-sampling Technique (SMOTE). Secondly, Bayesian optimization (BO) is employed to obtain optimal model parameters. Finally, an optimized XGBoost ensemble classification model is constructed. Based on the home-credit-default-risk dataset, four different ensemble models are sequentially trained, and the SMOTE-BO-XGBoost model is compared with traditional models. The experimental results demonstrate that: ① Ensembling is effective, and the multi-angle fused SMOTE-BO-XGBoost model exhibits the best model performance, outperforming general ensemble learning models and traditional classification algorithms; ② It can effectively overcome the problem of imbalanced samples. This model addresses individual credit risk issues of financial institution customers from two perspectives: data augmentation and model performance enhancement, providing an effective personal credit risk assessment method for banking sectors.

Key words: credit scoring, ensemble learning, class-imbalance, XGBoost

中图分类号: