Label Leakage
Definition
Label leakage occurs when information about the target illegitimately enters training features or preprocessing.
This can cause a model to achieve unrealistically good offline results because it uses signals that would not be available at the actual prediction time.
Simply explained
Like accidentally including parts of the solution in the exam question.
Why it matters in practice
Highly relevant for any model evaluation, especially with time series and operational data.
Example from business
A churn model accidentally uses a field that is only set after contract cancellation.
Limits and pitfalls
- Fitting preprocessing on all data before splitting; using information after the prediction time; overlooking indirect leakage via derived features.
Origin & history
Data leakage is a well-known ML problem; label/target leakage specifically refers to target information leaking into features or pipeline.
Use in mid-sized companies
- Possible application in SMEs: explicitly document feature availability at the real decision time.
Frequently asked questions
Why is leakage dangerous?
It overestimates model quality. Can leakage be indirect? Yes, via derived fields or preprocessing.
Sources
- Label leakage · Google ML Glossary, Label leakage
- Common pitfalls, Abschnitte 12.2 und 12.2.1 · scikit-learn, Common pitfalls, Abschnitte 12.2 und 12.2.1
Sources checked: 10./12.10.2026.
Editorially reviewed on: