My Model Hit 99.72% Accuracy on Fraud Detection. It Was Completely Useless.

A Hands On KNN Fraud Detection Project on Imbalanced Data: The Mistakes, The Fixes, and What Actually Matters. Continue reading on Medium »

Source: medium.com

Expert opinions (1)

• Anton Erpulev, CEO & Co-Founder at PowerSun Market: Stripe's first fraud detection model famously had 98% accuracy and would have cost them millions if they'd shipped it. The problem? Their transaction data was 99% legitimate, so the model just learned to label everything as "not fraud" and call it a day. This article captures something I see…