SMOTE (Synthetic Minority Over-sampling Technique)
What it is
A method that creates new synthetic examples of a rare class instead of just copying existing ones.
What It’s Used for
Imbalanced classification such as fraud, medical screening or rare faults like arc events.
Why and how it Helps
It interpolates between a minority sample and its nearest neighbours, so the classifier sees a fuller decision region. Apply it only to the training split, never before splitting.
When to Use it
Use it on tabular or feature-vector data when a classifier keeps ignoring a rare class, after trying class weights first. Skip it for raw waveforms or images (augment the signal instead, e.g. shifts, noise, mixing in load currents) and when minority samples are very few or noisy, since it amplifies the noise.