Train/validation/test split
Dividing data three ways: one part to fit on, one to make decisions against, and one untouched until the end to report an honest number.
What people get wrong
The split must respect the structure of the data. Random splitting of time series, of repeated measurements on the same user, or of near-duplicate documents leaks the answer and inflates every score.
A definition is not an understanding. This is taught properly in Memorising instead of learning, part of Overfitting & regularisation.
Learn this properlyTerms from the same topic