Scholarlink

In medical insurance fraud, there is an imbalance between normal and fraud samples, which seriously hinders the process of fraud identification. This study proposes a new hybrid sampling algorithm, RIGAN, by combining generative adversarial networks and reinforcement learning. The algorithm first denoises and deduplicates the initial samples through the natural neighbor search method while preserving the basic data information. Then, it uses an improved generative adversarial network to generate fraudulent samples, which is beneficial for the training of subsequent classification models. Finally, the PPO algorithm(Proximal Policy Optimization) is introduced to train a sample filter to screen out high-quality samples, achieving mutual reinforcement between the sample filter and the classifier. Experiments on the Kaggle medical insurance fraud dataset show that RIGAN significantly outperforms advanced GAN(Generative Adversarial Networks) methods such as IGAN, CWGAN, SNGAN, and TRANSWGAN, as well as traditional methods such as ADASYN, Borderline SMOTE, and CS-SVM in terms of accuracy (92.26%) and AUC (92.98%), demonstrating outstanding performance and robustness. Further ablation experiments on RIGAN verify the effectiveness of each component in RIGAN, among which component FRL has a stronger sample screening ability compared to traditional filters.

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

A Hybrid Sampling Algorithm Based on Generative Adversarial Networks and Reinforcement Learning for Medical Insurance Fraud Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yaping Zhao,
  • Longfeng Wang,
  • Gustave Florentin Nkoulou Mvondo,
  • Xing Yang

摘要

In medical insurance fraud, there is an imbalance between normal and fraud samples, which seriously hinders the process of fraud identification. This study proposes a new hybrid sampling algorithm, RIGAN, by combining generative adversarial networks and reinforcement learning. The algorithm first denoises and deduplicates the initial samples through the natural neighbor search method while preserving the basic data information. Then, it uses an improved generative adversarial network to generate fraudulent samples, which is beneficial for the training of subsequent classification models. Finally, the PPO algorithm(Proximal Policy Optimization) is introduced to train a sample filter to screen out high-quality samples, achieving mutual reinforcement between the sample filter and the classifier. Experiments on the Kaggle medical insurance fraud dataset show that RIGAN significantly outperforms advanced GAN(Generative Adversarial Networks) methods such as IGAN, CWGAN, SNGAN, and TRANSWGAN, as well as traditional methods such as ADASYN, Borderline SMOTE, and CS-SVM in terms of accuracy (92.26%) and AUC (92.98%), demonstrating outstanding performance and robustness. Further ablation experiments on RIGAN verify the effectiveness of each component in RIGAN, among which component FRL has a stronger sample screening ability compared to traditional filters.