Data Leakage
Leakage is any information flow from outside the training fold into model fitting — scaling with full-dataset statistics, target-derived features, duplicate rows across splits, or future data in temporal problems — and it inflates validation scores that collapse in production. The rule: every fitted transformation must see only the training fold, which is why preprocessing lives inside the Pipeline, not before the split.
This Concept is waiting for its first lesson!
Leakage is any information flow from outside the training fold into model fitting — scaling with full-dataset statistics, target-derived features, duplicate rows across splits, or future data in temporal problems — and it inflates validation scores that collapse in production. The rule: every fitted transformation must see only the training fold, which is why preprocessing lives inside the Pipeline, not before the split.
Are you a teacher? Sign in to start contributing.
Sign In