An imbalanced learning method based on graph tran-smote for fraud detection.

Area of Science:

• Computer Science
• Data Science

Background:

• Fraud poses significant risks to individuals and societal stability, necessitating advanced detection methods.
• In social media, fraudsters are a scarce minority, often forming small groups, which challenges conventional graph neural networks.
• Existing graph neural networks struggle with insufficient fraud characteristic representation due to the scarcity of fraudulent nodes.

Purpose of the Study:

• To propose a novel method, Graph-based Trans-SMOTE (GTS), for improved fraud detection in graph-structured data.
• To enhance the representation of fraud characteristics by integrating structural and attribute features.
• To mitigate the class imbalance problem inherent in fraud detection scenarios.

Main Methods:

• Employs a subgraph neural network extractor to deeply mine structural node features.
• Integrates structural and attribute features using transformer technology for enriched node representation.
• Utilizes a feature embedding space and edge generator to create synthetic minority class nodes, addressing class imbalance.

Main Results:

• The proposed GTS method demonstrates superior performance compared to state-of-the-art baselines.
• Experiments on two real-world datasets validate the effectiveness of GTS in fraud detection.
• GTS successfully addresses inadequate feature representation and class imbalance issues.

Conclusions:

• GTS offers a robust solution for fraud detection, particularly in imbalanced graph data.
• The integration of transformer technology and synthetic data generation significantly improves detection accuracy.
• This approach provides a promising direction for future research in graph-based fraud detection.