the.com/leakage
When your model aces the test because it secretly saw the answers.
means The accidental seepage of information from training or test data into a machine learning model's evaluation, making its performance look better than it really is.
from A plain old word for liquid escaping through cracks — from Middle English 'leken,' related to Old Norse 'leka,' to drip or leak. Data scientists borrowed the image directly: just as water finds hidden seams in a hull, answers find hidden seams between training and test sets, seeping through where they shouldn't be.
- common cause
- future data sneaking into training sets
- classic tell
- suspiciously perfect validation accuracy
- sneaky source
- preprocessing done before splitting data
- real cost
- great demo, catastrophic production failure