SMOTE is a data augmentation technique that makes extra minority-class training examples by combining existing ones. It is used when one class, like fraud, appears too rarely in the data. section 1.27
The basic move is simple: take the small pile of underrepresented examples and synthesize more of them from what is already there. The warning is just as simple: it does not create new information, so it may be a band-aid for a poorly framed task. section 1.27
What it does and does not do:
- Generates more examples of an underrepresented category. section 1.27
- Combines existing examples instead of collecting new ones. section 1.27
- Adds no new information to the training data. section 1.27
Fraud detection
A model for credit card fraud may have very few fraud-labeled transactions. SMOTE creates additional fraudulent examples by combining existing ones so the minority class is less sparse. section 1.27
