Introduction
Financial fraud represents a multi-billion dollar challenge for the global economy, yet identifying illicit activity remains a "needle in a haystack" problem. Traditional machine learning approaches in this domain are frequently stymied by two primary obstacles: the Label Scarcity Gap and the "Black Box" Regulatory Barrier.
The Label Scarcity Gap arises because high-quality fraud labels are expensive and labor-intensive to produce, requiring extensive Know Your Customer (KYC) and Anti-Money Laundering (AML) investigations. Consequently, models that rely on massive supervised datasets often fail when faced with evolving fraud patterns. Furthermore, the "Black Box" barrier refers to the regulatory requirement for transparency; financial institutions cannot simply flag transactions without providing a justifiable explanation. This paper introduces SAGE-FIN, a semi-supervised Graph Neural Network (GNN) framework designed to address these challenges by leveraging the inherent network structure of financial data.
Methodology: The SAGE-FIN Framework
SAGE-FIN is built upon a semi-supervised architecture that maximizes the utility of limited labeled data by incorporating unsupervised structural learning. The model treats financial data as a bipartite graph consisting of two distinct node types: wallets and transactions.
Multi-Task Learning
The core of SAGE-FIN is a multi-task loss function that optimizes for three simultaneous objectives:
- Feature Reconstruction: The model learns to reconstruct node features, ensuring it captures the underlying distribution of the data.
- Structural Learning: By predicting the existence of edges, the model learns the topological "norms" of the network.
- Semi-Supervised Classification: Using the small pool of available labels, the model classifies nodes as either "licit" or "illicit."
This multi-tasking approach acts as a regularizer, preventing the model from overfitting to the sparse labeled examples and allowing it to learn "what normal looks like" from the vast amount of unlabeled data.
Bipartite Message Passing
Unlike standard GNNs that treat all nodes as homogeneous, SAGE-FIN explicitly models the different nature of wallets and transactions. It aggregates information across these distinct partitions and the edges connecting them, reflecting the real-world logic of financial systems where money flows between entities via specific transaction events.
Granger-Causal Explanations
To overcome the "Black Box" barrier, SAGE-FIN introduces a novel explanation module inspired by Granger causality. When a node is flagged as fraudulent, the model performs a "distillation" process. It systematically removes edges within the neighborhood of the flagged node and measures the resulting change in the model's prediction certainty (loss). The edges that cause the most significant drop in certainty are identified as the "causal" drivers of the prediction. This allows the system to present auditors with a concise, 10-edge subgraph that visually explains the "smoking gun" connections.
Experimental Setup and Data
The framework was evaluated using the Elliptic++ dataset, a large-scale representation of the Bitcoin transaction graph. The dataset is structured as a bipartite graph containing:
- 203,769 transaction nodes (with 165 local and aggregated features).
- 822,942 wallet address nodes (with 56 features, such as fees and total transacted amounts).
The dataset is characterized by a high degree of unlabeled nodes, making it an ideal environment for testing semi-supervised methodologies that rely on topology as a proxy for labels.
Results and Discussion
SAGE-FIN demonstrated strong performance, achieving F1-scores of approximately 0.806 for wallets and 0.807 for transactions. While state-of-the-art Gradient Boosted Trees (such as XGBoost) maintain a slight edge in raw precision, SAGE-FIN significantly outperformed baseline models like Logistic Regression.
The primary value proposition of SAGE-FIN lies in its interpretability. Qualitative analysis showed that the Granger-causal subgraphs were highly effective at isolating the most relevant information. In tests, a distilled subgraph of just 10 edges could maintain a fraud probability of 0.89, compared to 0.84 for the full local graph. This suggests that the explainer successfully identifies the critical structural patterns associated with fraud, providing investigators with a manageable path of money to follow.
Limitations and Future Directions
Despite its strengths, the current iteration of SAGE-FIN has limitations. The model treats the financial network as a static snapshot, whereas fraud is inherently temporal. Future work involves integrating Temporal GNNs to capture the evolution of transaction patterns over time. Additionally, the "Granger causality" employed is a statistical proxy within the model's logic; it does not account for interventionist causality or potential hidden confounders. Finally, while 10-edge subgraphs are a significant improvement over black-box outputs, further work is needed to ensure the underlying features of those edges are easily interpretable by non-technical human auditors.
Summary updated