Think ROC-AUC Fails on Imbalanced Data? Think Again

Why the myth that “ROC-AUC fails with imbalanced data” persists, even among experts.

Data science is full of myths. One that is surprisingly persistent is that the ROC-AUC metric (Area Under the ROC Curve) is not reliable when applied to imbalanced datasets (that is, those where positive cases make up a small fraction of the total, like 1% or 5%).

Since many real-world datasets are indeed imbalanced, if this myth were true, we’d rarely b…