Scholarlink

<p>In an era of complex digital transactions, financial fraud detection remains a critical challenge for institutions worldwide. Financial datasets are often highly imbalanced, with fraudulent transactions constituting only a small fraction of the overall volume. This imbalance hinders machine learning models, leading to diminished accuracy in identifying minority-class instances. To address this issue, this paper provides a comprehensive comparative analysis of various class imbalance adjusting schemes—including resampling, cost-sensitive, and ensemble methods—and highlights their respective limitations in handling complex transactional data. Building on these insights, the paper proposes a novel approach that integrates Conditional Generative Adversarial Networks (CGAN) with Graph Attention Networks (GAT). The CGAN+GAT framework synthesizes high-fidelity minority-class samples while preserving the intricate relational structures inherent in financial transactions. Extensive experiments on the IBM Transactions for Anti-Money Laundering (AML) dataset demonstrate that CGAN+GAT outperforms traditional resampling methods, cost-sensitive approaches, and baseline graph-based models in both statistical fidelity and classification performance. Evaluation metrics, including accuracy, F1-score, and G-Mean, indicate that the CGAN+GAT approach achieves superior detection rates, significantly reducing false negatives. This research underscores the importance of combining generative modeling with graph-based attention mechanisms to enhance fraud detection in highly imbalanced, complex financial datasets.</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

Mitigating Class Imbalance in Banking Transactions: A Graph-Based GAN Solution for Fraud Detection

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Girish K K,
  • Anuraag B V,
  • Harshith P Reddy,
  • Nishchal R Mayur,
  • Biswajit Bhowmik

摘要

In an era of complex digital transactions, financial fraud detection remains a critical challenge for institutions worldwide. Financial datasets are often highly imbalanced, with fraudulent transactions constituting only a small fraction of the overall volume. This imbalance hinders machine learning models, leading to diminished accuracy in identifying minority-class instances. To address this issue, this paper provides a comprehensive comparative analysis of various class imbalance adjusting schemes—including resampling, cost-sensitive, and ensemble methods—and highlights their respective limitations in handling complex transactional data. Building on these insights, the paper proposes a novel approach that integrates Conditional Generative Adversarial Networks (CGAN) with Graph Attention Networks (GAT). The CGAN+GAT framework synthesizes high-fidelity minority-class samples while preserving the intricate relational structures inherent in financial transactions. Extensive experiments on the IBM Transactions for Anti-Money Laundering (AML) dataset demonstrate that CGAN+GAT outperforms traditional resampling methods, cost-sensitive approaches, and baseline graph-based models in both statistical fidelity and classification performance. Evaluation metrics, including accuracy, F1-score, and G-Mean, indicate that the CGAN+GAT approach achieves superior detection rates, significantly reducing false negatives. This research underscores the importance of combining generative modeling with graph-based attention mechanisms to enhance fraud detection in highly imbalanced, complex financial datasets.