Interpretable fraud detection for payment transactions — every flag comes with a reason.
DATANOVA 2026 · Cybersecurity & Fraud Detection track · All figures below are real outputs of the model run in run.py on a 100% synthetic, seeded dataset (no real customer data).
Model vs. rules-only baseline (held-out test set)
Try it: move the decision threshold
Recomputed live in your browser from the actual test-set scores embedded in this page — drag to see the precision/recall trade-off analysts tune every day.
Why was it flagged? — feature drivers
Standardised logistic-regression coefficients. Positive = pushes toward fraud. This is the explanation an analyst (and a customer) gets with every alert.