GNNExplainer

GNNExplainer operates as a post-hoc, perturbation-based interpreter that treats the underlying GNN as a black box. It requires no access to internal gradients or model weights, making it compatible with any graph neural network architecture—including GraphSAGE, GCN, and GAT—deployed in production fraud detection pipelines.

• Compatible with any GNN variant without modification
• Works on pre-trained, frozen models already in production
• No need to retrain or alter the original fraud classifier

The algorithm learns a continuous edge mask applied to the local computational graph of the target node, optimizing to select the minimal subgraph that maximizes mutual information with the original prediction. This produces a human-intelligible explanation by highlighting only the few most influential transactions or relationships.

• Identifies the 5-10 most critical edges driving a fraud alert
• Suppresses irrelevant transactional noise from the explanation
• Enables analysts to trace the exact relational path behind a decision

Simultaneously with subgraph selection, GNNExplainer learns a feature selector mask that identifies which node attributes—such as transaction amount, device type, or account age—were most decisive. This dual explanation reveals both who was influential and why their attributes mattered.

• Ranks features by their contribution to the prediction
• Distinguishes between structural and attribute-based signals
• Example: flags that transaction frequency and shared device ID drove a collusion alert

By identifying the minimal sufficient subgraph, GNNExplainer implicitly surfaces counterfactual insights: if the highlighted edges or features were removed, the model's fraud score would drop significantly. This helps compliance teams answer the critical question: What would need to change for this alert to be dismissed?

• Provides actionable feedback for alert adjudication
• Supports regulatory requirements for adverse action reasoning
• Helps distinguish true positives from model over-sensitivity

GNNExplainer supports three distinct explanation targets within a financial graph:

• Node-level explanations: Why was a specific account flagged as fraudulent?
• Edge-level explanations: Why did the model predict a hidden collusion link between two merchants?
• Graph-level explanations: What overall transaction pattern caused an entire subgraph to be classified as a money laundering ring?

This flexibility maps directly to the tiered alert structure used in financial crime investigations.

The optimization objective includes an entropy regularization term that penalizes diffuse, unfocused explanations. This forces the mask to converge toward a sharp, discrete selection rather than assigning partial importance to many edges—producing explanations that are sparse, crisp, and operationally actionable for fraud analysts.

• Avoids overwhelming investigators with dozens of weakly relevant connections
• Produces binary or near-binary edge selections for clear audit trails
• Balances fidelity (matching the original prediction) with conciseness