Dynamic heterogeneous graph contrastive learning for uncovering collusive financial fraud

Area of Science:

• Graph Neural Networks
• Machine Learning
• Financial Forensics

Background:

• Modern banking fraud detection requires analyzing complex, evolving relationships between entities.
• Existing graph-based methods often need extensive labeled data, which is scarce in anti-money laundering (AML).
• Delayed fraud labels in AML hinder timely detection of illicit activities.

Purpose of the Study:

• To develop a dynamic heterogeneous graph neural network framework, Audit-HCL, for detecting collusion rings under label scarcity.
• To effectively model evolving structural interactions among heterogeneous entities in banking transactions.
• To improve early detection of financial fraud by leveraging dual-view contrastive learning.

Main Methods:

• Representing the transaction ecosystem as a temporal sequence of heterogeneous graph snapshots.
• Employing a metapath-guided heterogeneous attention encoder for feature extraction.
• Utilizing a GRU-based temporal dynamics module to track evolving node behavior.
• Implementing a cross-view contrastive objective with anomaly-aware negative sampling.

Main Results:

• Audit-HCL significantly outperforms fourteen baseline methods on two public benchmarks, improving AUC-ROC by 3.2% and F1-score by 6.8%.
• The framework demonstrates strong performance even with zero confirmed fraud labels.
• It detects money laundering patterns an average of 7.4 weeks earlier than the best baseline on a synthetic AML benchmark by identifying gradual structural drift.

Conclusions:

• Audit-HCL offers an effective solution for detecting financial collusion rings in low-label environments.
• The dual-view contrastive learning approach enhances the model's ability to discern legitimate from anomalous activities.
• The framework shows promise for proactive fraud detection by capturing subtle, evolving structural changes in financial networks.