Data Leakage (Machine Learning)

Updated 17 Sep 2026

•
What it is
- •
  information unavailable at the intended prediction time improperly influences model development or evaluation
•
Examples
- •
  fitting a scaler on the entire dataset before splitting
- •
  placing near-duplicate recordings in both training and test sets
- •
  using future information to predict an earlier outcome
- •
  tuning repeatedly against the test set
•
Prevention
- •
  split before fitting learned preprocessing
- •
  fit preprocessing within each training fold
- •
  group related samples and respect time order
- •
  keep final evaluation separate from model selection.
•
Distinguish from
- •
  a confidentiality breach — the term here concerns invalid evaluation, even when no private information is disclosed
•
Related
- •
  Dataset Splitting
- •
  Unit of Analysis

Linked references 2