Scholarlink

<p>To improve the identification of medium-risk samples in credit risk assessment, this study integrates three-way decision with incremental learning and proposes an enhanced cost-sensitive Boosting model, extending traditional binary credit approval to ternary decision-making. To overcome the limitations of existing three-way decision (3WD) methods, which rely on static datasets, offline threshold optimization, and accuracy-driven threshold learning, the proposed model dynamically updates boundary thresholds as credit data arrive sequentially. It also constructs a dynamic cost matrix incorporating both misclassification and evaluation costs, enabling threshold learning to be guided by expected economic loss. In addition, incremental learning is embedded as an endogenous mechanism for boundary-sample re-evaluation and dynamic threshold optimization. By continuously integrating newly arrived, misclassified, and boundary-region samples, the model more effectively identifies uncertain cases, adapts to evolving data distributions, and strengthens decision support in dynamic credit environments. Interpretability analysis further reveals the key drivers of model decisions.</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

From Binary to Three-Way Decisions in Credit Scoring: An Enhanced Cost-Sensitive Boosting Model with Incremental Learning

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Feng Shen,
  • Jia Kuang,
  • Dao Lan,
  • Run Wang

摘要

To improve the identification of medium-risk samples in credit risk assessment, this study integrates three-way decision with incremental learning and proposes an enhanced cost-sensitive Boosting model, extending traditional binary credit approval to ternary decision-making. To overcome the limitations of existing three-way decision (3WD) methods, which rely on static datasets, offline threshold optimization, and accuracy-driven threshold learning, the proposed model dynamically updates boundary thresholds as credit data arrive sequentially. It also constructs a dynamic cost matrix incorporating both misclassification and evaluation costs, enabling threshold learning to be guided by expected economic loss. In addition, incremental learning is embedded as an endogenous mechanism for boundary-sample re-evaluation and dynamic threshold optimization. By continuously integrating newly arrived, misclassified, and boundary-region samples, the model more effectively identifies uncertain cases, adapts to evolving data distributions, and strengthens decision support in dynamic credit environments. Interpretability analysis further reveals the key drivers of model decisions.