This repository implements a hybrid, explainable fraud/anomaly detection pipeline that combines:
- Graph Neural Networks (GNNs) over a transaction graph
- Tabular Gradient Boosting (baseline) for calibrated probabilities
- Autoencoder for unsupervised anomaly scoring
- Post‑hoc explainability via SHAP, Permutation Importance, and GNNExplainer-like node/edge importance
- Drift tracking (population stability index, PSI) and threshold optimization (F1/PR-AUC maximizing)
It includes a configurable training CLI, reproducible experiments, and a lightweight unit test suite.
pip install -r requirements.txt
python -m src.experiments.main --config configs/default_config.json
# visualize
python -m src.experiments.evaluate --artifacts results/run_000explainable-ai-fraud-detection/
├── README.md
├── requirements.txt
├── configs/
│ └── default_config.json
├── data/
│ └── transactions.csv # sample synthetic data
├── src/
│ ├── data/
│ │ ├── build_graph.py
│ │ └── tabular_features.py
│ ├── models/
│ │ ├── gnn_model.py
│ │ ├── autoencoder.py
│ │ └── tabular_baseline.py
│ ├── explainability/
│ │ ├── shap_explainer.py
│ │ └── gnn_explainer_like.py
│ ├── experiments/
│ │ ├── main.py
│ │ └── evaluate.py
│ ├── utils.py
│ └── metrics.py
├── docs/
│ └── project_report.md
├── results/
│ └── (artifacts saved here)
└── tests/
└── test_shapes.py
- All hyperparameters and seeds are in
configs/default_config.json. - Artifacts (metrics, confusion matrix, feature importances, explainability plots) are saved under
results/run_xxx.
torch-geometricinstallation depends on your local CUDA/CPU stack. See https://pytorch-geometric.readthedocs.io for wheels.- If
torch-geometricis unavailable, you can run tabular + autoencoder only by setting"use_gnn": falsein the config.
