Target Leakage

noun

say it: “TAR-git LEE-kij” /ˈtɑrɡɪt ˈlikɪdʒ/

When a model learns from a feature that is only known after the result it is meant to predict, so it seems accurate in testing but fails in real use.

What is target leakage?

Target leakage is data leakage through a feature that's only filled in after the answer is known, such as refund_issued in a fraud model. The model learns to read the answer off that column, so its test scores look great, and it fails in real use, where the column is still empty.

In a sentence

  • “Removing the target leakage dropped the test score from 0.99 to a believable 0.81.”
  • “The chargeback column was target leakage: it was only filled in after a fraud was confirmed.”

Level: Freelancer · More AI Basics terms at this level

Updated