Scholarlink

<p>Scarcity of credit samples and severe class imbalance pose significant challenges to reliable credit risk assessment (CRA), constraining effective decision making in financial institutions. To address these issues, an integrated CRA framework is developed based on an Improved Two-stage Filtering Feature Selection (ITFFS) algorithm enhanced with an uncertainty-driven data augmentation technique (UNN-Gauss). Specifically, the UNN-Gauss module uses information entropy to quantify sample uncertainty and synthesizes virtual default instances near the decision boundary to balance the credit category distribution. Subsequently, the ITFFS algorithm employs a two-stage credit feature selection strategy. The first stage performs fast filtering to eliminate irrelevant credit feature variables, while the second stage identifies a discriminative subset of credit features by capturing linear and nonlinear relationships. Experiments on two Chinese listed enterprises datasets demonstrate that the integrated CRA framework outperforms state-of-the-art baselines in terms of AUC, F1-score, and Kolmogorov–Smirnov (KS) statistics. The proposed ITFFS algorithm reduces the feature space from 52 to 18 features, showing its high level of dimensionality reduction. It also improves predictive performance with AUC gains of up to 0.018 and G-mean improvements of up to 0.039 across different classifiers. Furthermore, SHAP-based interpretability analysis confirms that supply chain linkages serve as critical leading indicators for credit risk transmission. Additionally, validation on two public credit datasets further confirms the accuracy of the integrated CRA model, which highlights its theoretical contribution to enterprise credit scoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An optimized two-stage feature selection algorithm for high-dimensional credit scoring enhanced with uncertainty-driven sample augmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Lu Bai,
  • Xuezhou Wen,
  • Longfei Xie

摘要

Scarcity of credit samples and severe class imbalance pose significant challenges to reliable credit risk assessment (CRA), constraining effective decision making in financial institutions. To address these issues, an integrated CRA framework is developed based on an Improved Two-stage Filtering Feature Selection (ITFFS) algorithm enhanced with an uncertainty-driven data augmentation technique (UNN-Gauss). Specifically, the UNN-Gauss module uses information entropy to quantify sample uncertainty and synthesizes virtual default instances near the decision boundary to balance the credit category distribution. Subsequently, the ITFFS algorithm employs a two-stage credit feature selection strategy. The first stage performs fast filtering to eliminate irrelevant credit feature variables, while the second stage identifies a discriminative subset of credit features by capturing linear and nonlinear relationships. Experiments on two Chinese listed enterprises datasets demonstrate that the integrated CRA framework outperforms state-of-the-art baselines in terms of AUC, F1-score, and Kolmogorov–Smirnov (KS) statistics. The proposed ITFFS algorithm reduces the feature space from 52 to 18 features, showing its high level of dimensionality reduction. It also improves predictive performance with AUC gains of up to 0.018 and G-mean improvements of up to 0.039 across different classifiers. Furthermore, SHAP-based interpretability analysis confirms that supply chain linkages serve as critical leading indicators for credit risk transmission. Additionally, validation on two public credit datasets further confirms the accuracy of the integrated CRA model, which highlights its theoretical contribution to enterprise credit scoring.