Card Fraud Detection
Models that score every card payment for fraud risk in milliseconds, combining supervised classifiers such as gradient boosting with anomaly detection on spending patterns.
Also payment fraud detection · Visa Advanced Authorization · transaction scoring · fraud model
Card networks and banks; Visa says it began neural-network risk scoring in 1993
Your bank knows your habits: coffee in the morning, groceries on Saturday. A sudden purchase of five laptops abroad at 3 a.m. looks very unlike you, so it asks you to confirm.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
How it works
Every time you tap your card, a model decides in a few milliseconds whether the payment looks legitimate. Card networks and banks have used machine learning for this for decades; Visa says it started scoring transactions with neural networks in 1993.
What makes it hard:
- Extreme imbalance: only a tiny share of payments is fraud. A model that says “never fraud” is 99.9% accurate and useless, so teams look at precision and recall and ROC/PR curves.
- Delayed labels: you only learn a payment was fraud when the cardholder disputes it, weeks later.
- Adversaries: fraud rings change tactics as soon as they get blocked (data drift).
- Speed: the decision must fit inside the payment’s time budget (model serving).
Typical systems combine:
- Gradient boosting or neural networks trained on past fraud, using features such as “amount compared with this customer’s average” or “distance from the last purchase”.
- Anomaly detection to catch new patterns no labeled example covers yet.
- Rules and human analysts for the uncertain middle.
The threshold is a business decision: block more fraud, or annoy fewer honest customers with declined cards.
Everyday use cases
- Case 01
How you've probably met it
A text from your bank asking "Did you just spend 900 euros in another country?", or a declined card on holiday.
Strengths, weaknesses and when to use it
Strengths
- Stops large amounts of fraud while letting the vast majority of payments through instantly.
Weaknesses
- False declines annoy honest customers; fraudsters adapt quickly.
When to use itAs the everyday example of imbalanced classification, anomaly detection and low-latency serving.