Scholarlink

<p>Fraud is becoming more prevalent in various fields, resulting in substantial financial losses. Fraud detection has become critical in detecting and reducing fraud instances, thereby safeguarding individuals and corporations. However, fraud detection is not trivial due to the issue of data imbalance, which results in suboptimal performance. Most previous studies mainly focus on reweighting or resampling-type methods, often neglecting the inherent semantic relationship between users. In this paper, we propose <i>GuardNet</i>, an imbalance-aware graph neural networks-based framework for fraud detection, which achieves balance by considering the significance of both the majority and minority classes, leveraging the semantic relationships between users inherent in the data. Specifically, for datasets with diverse structures from various areas, we construct graphs that consider semantic relationships. Then, we devise an imbalance-aware subgraph shuffle method to reconstruct graphs, while highlighting the significance of nodes with minority class, and lowering the significance of nodes with majority class. The use of a GNN encoder facilitates the acquisition of high-level representations and subsequent fraud detection results. Extensive experiments on seven fraud detection datasets in various domains verify the superiority and robustness of our <i>GuardNet</i>, which outperforms all state-of-the-art methods, improving Recall, F1-Score and G-Means by 2.53%, 2.45%, and 1.69%, respectively. All code and datasets are available at <i>https://github.com/ZJUDataIntelligence/GuardNet</i>.</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

Guardnet: an imbalance-aware graph neural network for fraud detection

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Yuhao Chen,
  • Fanwei Zhu,
  • Zengwei Zheng,
  • Jianhua Ma,
  • Binbin Zhou

摘要

Fraud is becoming more prevalent in various fields, resulting in substantial financial losses. Fraud detection has become critical in detecting and reducing fraud instances, thereby safeguarding individuals and corporations. However, fraud detection is not trivial due to the issue of data imbalance, which results in suboptimal performance. Most previous studies mainly focus on reweighting or resampling-type methods, often neglecting the inherent semantic relationship between users. In this paper, we propose GuardNet, an imbalance-aware graph neural networks-based framework for fraud detection, which achieves balance by considering the significance of both the majority and minority classes, leveraging the semantic relationships between users inherent in the data. Specifically, for datasets with diverse structures from various areas, we construct graphs that consider semantic relationships. Then, we devise an imbalance-aware subgraph shuffle method to reconstruct graphs, while highlighting the significance of nodes with minority class, and lowering the significance of nodes with majority class. The use of a GNN encoder facilitates the acquisition of high-level representations and subsequent fraud detection results. Extensive experiments on seven fraud detection datasets in various domains verify the superiority and robustness of our GuardNet, which outperforms all state-of-the-art methods, improving Recall, F1-Score and G-Means by 2.53%, 2.45%, and 1.69%, respectively. All code and datasets are available at https://github.com/ZJUDataIntelligence/GuardNet.