🛡️ Real-Time Credit Card Fraud Detection
An end-to-end fraud system on the Sparkov dataset (1.85M transactions, ~0.5% fraud).
- Model — LightGBM (PR-AUC 0.97) · plus a GraphSAGE GNN and an autoencoder baseline
- Honest evaluation — PR-AUC, cost-optimal thresholds, an imbalance study + real-data validation
- Production-shaped — leakage-safe features, SHAP explanations, drift monitoring, ~10 ms scoring
👉 Try it: open the Live Scoring tab, set an amount / time / distance, and click Score.
Score a transaction in real time
Set the transaction details and click Score. Fraud rises with a high amount, a late hour, and a large distance from home. Score several in a row to build up the card's velocity history and watch the risk climb.
Built by Muhammad Fikri Wahidin · Sparkov dataset · LightGBM · PyTorch Geometric · SHAP · FastAPI