Technical Summary: CEGNN-Fraud

1. Problem Formulation

Financial fraud detection is framed not as a static tabular classification task, but as a relational and temporal inference problem within continuously evolving heterogeneous networks. Existing approaches face three primary limitations:

1. Neighborhood Contamination: Static or purely temporal models often fail to distinguish between reliable and unreliable neighbors. Fraudsters may deliberately connect to benign entities (camouflage) or use stale relations, causing standard message-passing mechanisms to dilute fraud signals.
2. Operational Misalignment: Many models are evaluated under random train-test splits that leak future information or use metrics (e.g., accuracy, ROC-AUC) that do not reflect the operational reality of low-prevalence fraud, false-positive control, and the need for early warning before fraud onset.
3. Explanation Governance: Current graph explainers identify influential edges but often conflate statistical association with causality. Furthermore, integrating Large Language Models (LLMs) for reporting introduces risks of hallucination, where models may generate unsupported entities, alter numerical values, or claim causal relationships not present in the evidence.

The proposed framework, CEGNN-Fraud, addresses these issues by formulating fraud detection as an event-level screening problem that strictly adheres to temporal validity, handles extreme class imbalance, and separates statistical prediction from evidence-governed reporting.

2. Methodology

CEGNN-Fraud is a scalable heterogeneous temporal graph neural network architecture comprising three distinct stages:

A. Heterogeneous Temporal Graph Encoder

The core innovation is an Event-Conditioned Relational Reliability Gate. Unlike static neighbor selection or node-level similarity filtering, this gate evaluates the reliability of a historical neighbor conditional on the specific focal transaction being evaluated.

• Inputs: The gate jointly processes the focal event representation, the neighbor's pre-event state, the relation type, and the elapsed time since the interaction.
• Mechanism: It computes a reliability score based on relation semantics, behavioral discrepancy (using a clipped Mahalanobis distance), and temporal freshness.
• Normalization: Reliability is normalized within each relation type before aggregation. This prevents high-frequency relations from mechanically dominating less frequent but more diagnostic connections.
• Dynamic Memory: The model maintains dynamic entity memory states updated via GRUs after prediction (to prevent temporal leakage), capturing continuously evolving behavioral states.
• Loss Function: The model employs Focal Loss to address extreme class imbalance and includes regularizers to prevent uniform reliability scores and excessive state variation on routine events.

B. Evidence Validation Pipeline

Post-prediction, a separate pipeline extracts and validates evidence before it is presented to an LLM. This pipeline ensures three types of evidence are distinct:

1. Factual Graph Evidence: A sparse subgraph is extracted that preserves the original prediction (using edge masking and KL-divergence objectives).
2. Diagnostic Counterfactuals: The system identifies minimal, semantically admissible perturbations (e.g., masking a specific edge or replacing a mutable attribute) that change the model's decision. These are strictly diagnostic and do not claim real-world reversibility.
3. Temporal Predictive Evidence: Candidate signals are screened offline to ensure they precede the focal event and provide incremental predictive information beyond the history. This verifies temporal ordering without claiming structural causality.

C. Evidence-Constrained LLM Generation

The LLM is restricted to a communication role, not a decision-making role.

• Input: It receives a structured object containing the fraud probability, validated factual evidence, diagnostic counterfactuals, and temporally screened signals. It does not access raw customer histories or the graph structure directly.
• Constraints: The generation policy enforces that the LLM cannot introduce new entities, alter numerical values, or make unsupported causal claims (e.g., "caused fraud" is prohibited; "preceded the alert" is permitted).
• Action Policy: Operational actions are selected by a deterministic policy layer based on risk tiers and evidence types, not by the LLM. The LLM merely verbalizes these authorized actions.
• Verification: A rule-based verifier checks entity consistency and causal wording. Invalid outputs trigger regeneration or a fallback to a deterministic template.

3. Key Contributions

The paper claims three principal contributions:

1. Event-Conditioned Reliability Gate: A mechanism that evaluates neighbor reliability relative to the focal transaction, relation semantics, and time, directly targeting adversarial neighborhood contamination and camouflage.
2. Temporally Valid Event-Level Screening: A formulation and evaluation protocol that strictly separates pre-event information from the prediction, addresses class imbalance, and evaluates early-warning capabilities (anticipating fraud onset) rather than just transaction classification.
3. Evidence-Governed Reporting Architecture: A framework that decouples statistical prediction from explanation and language generation. It validates factual, counterfactual, and temporal evidence before LLM ingestion, ensuring the LLM acts as a controlled interface for validated model evidence rather than an independent reasoning engine.

4. Experimental Results

Experiments were conducted on a controlled synthetic dataset comprising 3.84 million transaction events (1.0% fraud prevalence) over 12 months, forming a heterogeneous temporal graph with ~2 million nodes and ~10.68 million edges.

• Predictive Performance: CEGNN-Fraud achieved a PR-AUC of 0.781, outperforming the strongest baseline (CaT-GNN, 0.744) and other tabular, GNN, and temporal models. It demonstrated superior performance in recall, F1-score, and Matthews Correlation Coefficient (MCC) while maintaining a lower False Positive Rate (FPR).
• Robustness: The model showed greater resilience to:
  • Class Imbalance: Performance degradation was minimal as fraud prevalence dropped to 0.1%.
  • Camouflage: It maintained higher PR-AUC than baselines when fraud-normal edges were injected (up to 30% increase).
  • Temporal Drift: It outperformed baselines under severe distribution shifts.
• Early Warning: The model anticipated 37.4% of fraud episodes within 1 hour of onset (vs. 33.8% for CaT-GNN) and achieved a mean warning lead time of 10.26 hours.
• Inductive Generalization: It generalized effectively to unseen entities (new accounts, devices, merchants), achieving a PR-AUC of 0.754 for any-new-entity scenarios.
• Explanation Quality: The framework produced explanations with higher factual fidelity (0.903) and semantic admissibility (0.908) than established explainers (e.g., SubgraphX, CLEAR).
• LLM Effectiveness: When supplied with validated evidence, the LLM achieved 91.8% evidence coverage and reduced unsupported statements to 4.1%, significantly outperforming inputs with only raw predictions or unvalidated explanations.
• Scalability: At the largest scale (2M nodes), the model achieved a P95 prediction latency of 18.6 ms (under 20ms), supporting real-time inference, while the asynchronous explanation pipeline operated at ~281ms.

5. Significance and Claims

The paper positions CEGNN-Fraud as a methodological and computational feasibility study rather than a production deployment on real institutional data. Its significance lies in:

• Methodological Integration: Demonstrating that combining event-conditioned graph learning with evidence-governed LLM reporting is feasible and effective.
• Operational Realism: Shifting the evaluation focus from average classification accuracy to robustness under adversarial conditions (camouflage, imbalance), temporal validity, and early-warning capability.
• Governance of AI: Establishing a rigorous architecture where LLMs are constrained to communicate validated model evidence without introducing hallucinations or independent causal claims, thereby separating statistical prediction from operational decision-making.

The author explicitly states that the findings establish methodological and computational feasibility but do not constitute external validation on specific financial institutions, noting that real-world deployment would require further evaluation of delayed labels, strategic fraudster adaptation, and end-to-end system latency.