Daemons-Q/qsh-finance-aml-gnn-causal
Multi-PNA+EU model for financial transaction analysis
2 min readupdated
An official implementation of the Multi-PNA+EU architecture (Graph Neural Networks for directed multigraphs) trained on the IBM AMLworld HI-Small dataset for financial transaction classification and anti-money laundering (AML) detection. The model uses reverse message passing, multigraph port numbers, node IDs, and edge updates, and is adapted for a causal protocol with modified timestamp inputs. The repository is exported to ONNX (opset 18) and built for research demos and running in onnxruntime-web.
Key features
- Based on the official IBM Multi-PNA+EU code (AAAI 2024, arXiv:2306.11586).
- Uses relative time inputs (lag) instead of absolute timestamps to prevent out-of-range issues beyond the training window.
- Evaluated across 5 seeds with BatchNorm statistics recomputed on the fitting window.
- Exported to
ONNXfor inference in web environments.
What you can do
- Integrate the model into web applications powered by
onnxruntime-webto assess financial transfer risks. - Use the repository's weights and patches to reproduce financial fraud detection experiments.
- Study how causal time protocols affect GNN accuracy on time-series tasks.
Example
The model loads via the onnxruntime-web runtime to score transactions in plaintext mode using a pre-sampled neighborhood (eval_sample.json).
Read more
DetailsMetrics (seed 1, exported method): accuracy — 0.9989, auroc — 0.9816, auprc — 0.5789, minority_class_f1 — 0.6252.
Trained on the IBM AMLworld HI-Small dataset (5M transactions, CDLA-Sharing-1.0 license).
Requirements: ONNX Runtime, WebAssembly / web environment support for in-browser execution.
Limitations: research prototype, not intended for production systems without further work.










Comments