Imbalanced Data
When one class vastly outnumbers another — fraud among transactions, disease among patients — and naive accuracy becomes a liar that rewards models for ignoring the class you care about.
Reviewed July 16, 2026Stable
Reach for something else instead
- Threshold tuning — move the decision threshold instead of touching the data.
- Cost-sensitive learning — reweight the loss rather than resample.
- Anomaly detection when the minority class is so rare it's better framed as "unusual."
Sources & further reading
- Chawla et al. (2002), SMOTE: Synthetic Minority Over-sampling Technique — the classic resampling method.
- He & Garcia (2009), Learning from Imbalanced Data — the standard survey.
- Saito & Rehmsmeier (2015), The Precision-Recall Plot Is More Informative than the ROC Plot on Imbalanced Datasets.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reporting accuracy on imbalanced data — a trivial majority-class model scores high and catches nothing.
- Resampling the whole dataset before splitting, leaking test information into training.
- Using ROC-AUC uncritically under heavy imbalance, where precision-recall is more honest.
At a glance
FieldMachine Learning
Problemone class vastly rarer than another
Metric fixprecision, recall, F1 (not accuracy)
Data fixesSMOTE, reweighting, threshold tuning
DifficultyIntermediate