Scholarlink

<p>Credit default prediction is crucial for financial institutions, as it enables more informed lending decisions. However, current methods face notable Limitations, especially in Handling extreme class imbalance and the scarcity of default instances. Additionally, research into uncovering hidden patterns in non-linear borrower data and improving model interpretability remains relatively underexplored. To mitigate these issues, we propose a hybrid approach that combines a conditional tabular generative adversarial network with a graph neural network, specifically the graph sample and aggregate, for inductive learning. This allows for generalization to new borrowers in large datasets and applies to real-world, dynamic financial environments. The generated default instances are added to the original dataset, creating a more balanced distribution between default and non-default classes. The data are then structured as a graph to capture the relationships among borrowers. This approach directly reduces class imbalance while maintaining non-linear relational context. Experimental results demonstrate that our approach achieves a 7.78% improvement in accuracy and a 13.35% increase in the area under the receiver operating characteristic curve over the non-augmented version, while outperforming baselines that use class balancing techniques. Moreover, Shapley additive explanations provide interpretability by highlighting key features influencing credit risk. These findings underscore the effectiveness of combining synthetic data augmentation with graph learning for robust credit risk prediction in imbalanced settings.</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

Graph-Based Inductive Learning for Credit Risk Prediction with Imbalance Mitigation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月11日

查看全文
  • Sogand Pourkhoshgoftar,
  • Asadollah Shahbahrami,
  • Nima Esmi

摘要

Credit default prediction is crucial for financial institutions, as it enables more informed lending decisions. However, current methods face notable Limitations, especially in Handling extreme class imbalance and the scarcity of default instances. Additionally, research into uncovering hidden patterns in non-linear borrower data and improving model interpretability remains relatively underexplored. To mitigate these issues, we propose a hybrid approach that combines a conditional tabular generative adversarial network with a graph neural network, specifically the graph sample and aggregate, for inductive learning. This allows for generalization to new borrowers in large datasets and applies to real-world, dynamic financial environments. The generated default instances are added to the original dataset, creating a more balanced distribution between default and non-default classes. The data are then structured as a graph to capture the relationships among borrowers. This approach directly reduces class imbalance while maintaining non-linear relational context. Experimental results demonstrate that our approach achieves a 7.78% improvement in accuracy and a 13.35% increase in the area under the receiver operating characteristic curve over the non-augmented version, while outperforming baselines that use class balancing techniques. Moreover, Shapley additive explanations provide interpretability by highlighting key features influencing credit risk. These findings underscore the effectiveness of combining synthetic data augmentation with graph learning for robust credit risk prediction in imbalanced settings.