AML detection on the Elliptic Bitcoin dataset
Graph neural networks vs. tabular baselines for anti-money-laundering classification. PyTorch Geometric + cuML. Best RF F1: 0.80.
Problem
The Elliptic dataset is a labeled transaction graph: licit, illicit, unknown. Goal — classify the illicit nodes from graph structure plus tabular features, and see whether GNNs beat boring baselines.
Approach
Two tracks in parallel:
- Tabular baselines — Logistic Regression and Random Forest on the feature matrix, accelerated with RAPIDS cuML.
- GNN track — GCN and SkipGCN variants in PyTorch Geometric, using the temporal split (train: steps 1–34, test: steps 35–49).
- Hybrid — GCN embeddings fed as features into Random Forest.
Five orchestrated phases: preprocessing → training → evaluation → embedding extraction → hybrid eval.
Results
| Model | Precision | Recall | F1 (illicit) |
|---|---|---|---|
| Random Forest | 0.8992 | 0.7248 | 0.8027 |
| RF + GCN Embeddings | 0.9929 | 0.6445 | 0.7816 |
| GCN | — | — | 0.69 |
Tabular won on this dataset — the engineered features carry a lot. The hybrid approach traded recall for very high precision.
Takeaways
- “Use a GNN” is not free; the feature engineering it implicitly replaces may already be doing the heavy lifting.
- Reproducibility hygiene matters more than model choice for write-ups like this.
- High precision / lower recall in the hybrid is a meaningful trade-off in AML: false positives are costly to investigate.