SMOTE
Also: synthetic minority oversampling
Creating extra minority-class rows by interpolating between real ones and their neighbours, rather than duplicating them.
What people get wrong
It must be applied inside the training fold only. Oversampling before the split puts synthetic relatives of validation rows into training, which is leakage and produces excellent, meaningless scores.
A definition is not an understanding. This is taught properly in When one class is rare, part of Imbalanced and messy data.
Learn this properlyTerms from the same topic