Modern card networks score every transaction against a gradient-boosted classifier
(an XGBoost-style ensemble of decision trees) in single-digit milliseconds.
The model turns raw features — transaction amount, recent transaction velocity, and
device/IP risk — into a single fraud probability. This lab visualizes that scoring
process as a 3D feature space: each historical transaction is a point plotted by its
three features, colored blue (legitimate) or red (fraud), and a translucent
decision plane shows exactly where the model currently draws the line.
Because fraud is rare (often well under 1% of transactions), a model that predicts "not fraud" for everything can still score 99%+ accuracy — which is why real fraud systems are tuned on precision and recall, not raw accuracy, and why the threshold is a live business dial, not a fixed constant.
A 3D feature-space view of a gradient-boosted fraud classifier: historical transactions cluster as legitimate or fraudulent points, a movable decision plane marks the current probability threshold, and live transactions stream through, get scored, and route to an approved or blocked gate.
Moving the decision threshold slides the boundary plane along the axis between the legitimate and fraud clusters, directly trading recall for precision — exactly the tuning fraud teams do in production.
Adjust the threshold, stream rate, base fraud rate and model separability, then watch precision/recall update and transactions get routed to the APPROVED or BLOCKED gate in real time. Toggle explainability bars to see each transaction's feature contributions.
Because genuine fraud is rare, a model that flags nothing can still be "99% accurate" — real systems are judged on precision, recall and F1, not raw accuracy, and the threshold is a live business dial tuned against cost of fraud versus customer friction.