Project
Elliptic Graph ML for Illicit Transaction Detection
Leakage-safe graph ML for illicit transaction detection with operational metrics and robustness checks.
Python PyTorch Geometric XGBoost Logistic regression TensorBoard Streamlit Calibration analysis
Read article Private repo
Source code is private. The public case study and article focus on the methodology, evaluation design, and operational tradeoffs.
Business context
Fraud and compliance teams need to identify risky transaction flows early, but graph-based financial datasets are easy to evaluate incorrectly because of time leakage and unrealistic thresholds. This project focused on building a defensible detection workflow, not just chasing a strong offline score.
Outcome
- Compares feature-only baselines against GCN, GraphSAGE, and GAT models.
- Uses temporal splits, calibration, robustness checks, Precision@K, PR-AUC by timestep, and workload curves.
- Documents SAGE-ResBN configurations as the strongest-performing setup in the repo’s experiments.
- Includes interpretability and ensemble-analysis paths for both baseline and graph models.
Key decisions
- Prioritized leakage-safe temporal splits before model tuning.
- Benchmarked simpler baselines before claiming graph-model gains.
- Used precision-at-investigation-budget style metrics instead of relying only on ROC-AUC.
- Added calibration, robustness, and hub-ablation checks to test operational stability.
System design
Raw Elliptic CSVs are transformed into a processed graph artifact, then passed through baseline and GNN training pipelines. Analysis modules handle by-time drift, calibration, workload curves, bootstrap comparison, robustness checks, explanations, and optional ensembling.
Stack
- Python, PyTorch Geometric, XGBoost, and logistic-regression baselines
- TensorBoard and Streamlit for experiment presentation
- Graph preprocessing, temporal evaluation, calibration, and explainability tooling
