SMOTE addresses the classic imbalanced data problem where minority classes get ignored by models. Instead of duplicating rare samples—which causes overfitting—it generates new synthetic examples by interpolating between existing minority neighbors. This creates diversity, helping models generalize rather than memorize.

The technique works by finding a minority sample's nearest neighbors, drawing lines between them, and placing new points randomly along those lines. This linear interpolation creates plausible, new data points. For instance, a dragon classifier with just 47 examples saw accuracy jump from 12% to 94.7% after SMOTE generated thousands of synthetic dragons.

SMOTE isn't a single algorithm but a family. Borderline-SMOTE focuses on decision boundaries, ADASYN adapts to harder regions, and SMOTE-NC handles mixed data types. However, it fails with noisy labels or extreme dimensionality. Best practice is to apply SMOTE only after train-test splits and within cross-validation pipelines to avoid data leakage.