管理评论

管理评论 ›› 2026, Vol. 38 ›› Issue (10): 79-90.

• 经济与金融管理 • 上一篇    

面向高维特征和类别分布不平衡的信用风险评估研究——基于复数值极限学习机集成视角

贾堰林1, 曾志豪1, 黄满翎2, 杨雁1, 王亚东3, 肖进4   

  1. 1. 西南石油大学理学院, 成都 610500;
    2. 中国联合网络通信有限公司眉山市分公司, 眉山 620020;
    3. 广州大学公共管理学院, 广州 510006;
    4. 四川大学商学院, 成都 610064
  • 收稿日期:2025-05-07 发布日期:2026-11-14
  • 作者简介:贾堰林，西南石油大学理学院讲师，博士；曾志豪，西南石油大学理学院硕士研究生；黄满翎，中国联合网络通信有限公司眉山市分公司助理工程师，硕士；杨雁，西南石油大学理学院教授，博士生导师，博士；王亚东，广州大学公共管理学院副教授，博士；肖进（通讯作者），四川大学商学院研究员，博士生导师，博士。
  • 基金资助:
    教育部人文社会科学研究项目（23YJCZH088）；国家社会科学基金项目(24VRC096)；国家自然科学基金项目(72401075)；四川省科技计划（2024NSFSC1957）。

A Credit Risk Assessment that Factors in High-dimensional Features and Imbalanced Class Distribution: From the Perspective of Complex-valued Extreme Learning Machine Ensemble

Jia Yanlin1, Zeng Zhihao1, Huang Manling2, Yang Yan1, Wang Yadong3, Xiao Jin4   

  1. 1. School of Sciences, Southwest Petroleum University, Chengdu 610500;
    2. China United Network Communications Co., Ltd. Meishan Branch, Meishan 620020;
    3. Public Administration School, Guangzhou University, Guangzhou 510006;
    4. Business School, Sichuan University, Chengdu 610064
  • Received:2025-05-07 Published:2026-11-14

PDF (PC)

1

摘要： 复数值神经网络已成功用于解决类别分布不平衡的信用风险评估问题，但这些研究仅在低维特征数据上进行建模，难以同时解决人工智能时代类别分布不平衡和高维特征数据问题。为此，本文引入具有更快学习速度的复数值极限学习机，并将其与重抽样技术、集成学习、多属性决策等方法有机结合，构建基于最优重抽样选择策略的信用风险评估复数值极限学习机集成模型，其中采用熵权TOPSIS方法选择最适合的重抽样技术。在两个高维不平衡的信用风险评估数据集上进行实证分析发现，在九种重抽样技术中，随机欠采样是最适合复数值极限学习机集成建模的；全特征建模的评估优于采用降维技术的评估；进一步，基于随机欠采样的复数值极限学习机集成模型的信用风险评估性能优于四种单一模型和五种集成模型。

关键词: 信用风险评估, 类别分布不平衡, 高维特征, 重抽样技术, 复数值极限学习机, 熵权TOPSIS

Abstract: Complex-valued neural networks have demonstrated success in addressing imbalanced credit risk assessment, but existing studies remain confined to low-dimensional feature spaces, failing to resolve the dual challenges of high-dimensional features and imbalanced class distribution prevalent in AI-driven financial environments. To this end, this study introduces a complex-valued extreme learning machine with faster learning speed and organically integrates it with resampling techniques, ensemble learning, multi-attribute decision-making and other methods to construct a credit risk assessment complex-valued extreme learning machine ensemble model based on the optimal resampling selection strategy, where the entropy-weighted TOPSIS method systematically identifies the most suitable resampling technique. Empirical evidence on two high-dimensional and imbalanced credit risk assessment data sets reveals that among nine resampling techniques, random undersampling emerges as the optimal technique for complex-valued extreme learning machine ensemble modelling. The evaluation based on full-feature modeling is superior to that using dimensionality reduction techniques. Furthermore, the proposed random undersampling-based complex-valued extreme learning machine ensemble model significantly outperforms four single models and five ensemble models.

Key words: credit risk assessment, imbalanced class distribution, high-dimensional features, resampling technique, complex-valued extreme learning machine, entropy-weight TOPSIS