Handling Imbalanced Datasets

The trap

Fraud, disease, and defects are rare. A model can score 99% accuracy by always predicting "normal" — and catch nothing useful.

What to do

• Pick the right metric — precision/recall/F1, not accuracy.
• Resample — oversample the minority (SMOTE) or undersample the majority.
• Class weights — tell the model to penalize minority errors more.
• Adjust the threshold to favor recall where misses are costly.

Mindset

Define which error is more expensive before training, then optimize for it. Imbalance is a framing problem as much as a modeling one.