Abstract
Detecting financial crime in massive, dynamic transactional networks requires scalable and inductive representation learning. This paper explores the application of GraphSAGE to a non-bipartite network of over 10 million nodes at NatWest. By generating unsupervised embeddings that capture structural and behavioral patterns, the approach significantly improves the precision of money mule detection, achieving a 57.1% relative improvement in Precision@20 for high-risk accounts.
Introduction
In modern banking, transactional data forms a complex, heterogeneous, and non-bipartite graph where accounts interact with both merchants and other accounts. Traditional graph embedding methods, such as DeepWalk or Graph Convolutional Networks (GCNs), are typically transductive, requiring the entire graph to be present during training. This poses a significant challenge for real-time banking applications where new customers and transactions are constantly added. To address this, we implement an inductive framework capable of generating embeddings for unseen nodes without full model retraining.
Methods
We utilized GraphSAGE (SAmple and aggreGatE), an inductive Graph Neural Network (GNN) framework. Unlike transductive models that learn a fixed embedding for each node, GraphSAGE learns a function that generates embeddings by sampling and aggregating features from a node’s local neighborhood.
Graph Schema and Sampling
The network was modeled with four primary node types and seven edge types representing various financial flows, including domestic and international transfers, point-of-sale (POS) transactions, and refunds.
To maintain scalability across 100 million edges, we employed a neighborhood sampling strategy with a fixed "fanout." This prevents computationally expensive "super-nodes" (e.g., major retailers) from bottlenecking the embedding generation process.
Unsupervised Learning
The model was trained using an unsupervised loss function designed to minimize the distance between neighboring nodes in the embedding space while maximizing the distance between non-neighbors. This allows the model to learn the underlying topology of the banking ecosystem without requiring manual labels.
Key Insights: Structural and Behavioral Patterns
Analysis of the resulting 32-dimensional embeddings revealed that the model captured significant real-world features purely from transaction topology.
Geographical Clustering
When projected into 2D space using UMAP, distinct clusters emerged that aligned with geographical regions. For instance, accounts associated with Ulster Bank in Northern Ireland formed a separate, tight cluster, indicating that regional economic activity creates unique transactional "neighborhoods."
Demographic and Account Type Proxies
The embeddings also naturally clustered by age group and account type. Current accounts and savings accounts occupied different regions of the manifold due to their distinct transactional behaviors; savings accounts typically only interact with a linked current account, creating a unique structural signature.
Results: Money Mule Detection
The efficacy of the GraphSAGE embeddings was evaluated through a downstream classification task for money mule detection. We compared a baseline gradient-boosted model using standard tabular features against an enhanced model incorporating the graph embeddings.
- Precision@20: +57.1% relative improvement.
- Precision@50: +22.2% relative improvement.
- PR-AUC: +4.3% gain.
The substantial gain in Precision@20 is particularly noteworthy for operational efficiency, as it directly improves the accuracy of the highest-priority alerts reviewed by human investigators.
Conclusion and Future Work
This study demonstrates that GraphSAGE can effectively scale to massive, non-bipartite transactional networks to produce meaningful, inductive representations. Future research will focus on the temporal evolution of these embeddings, exploring how shifts in a node's position within the embedding manifold over time can serve as early warning signals for fraudulent activity.
Summary updated