A different curve
The precision recall curve plots precision against recall as the threshold sweeps. The area under it, PR AUC or average precision, summarizes the whole curve.
Why it suits imbalance
Both precision and recall focus on the positive class and ignore true negatives entirely. When negatives vastly outnumber positives, that is exactly what you want.
- ROC AUC can look impressive while precision is terrible
- PR AUC exposes the cost of false positives because precision sits in the picture
The baseline shifts
A crucial difference from ROC: the no skill baseline for PR AUC is not 0.5. It equals the positive prevalence.
- If positives are 1 percent of data, random PR AUC is about 0.01
- A PR AUC of 0.3 there is actually strong, not weak
Always compare PR AUC against the prevalence baseline, not against 0.5.
When to use which
Use ROC AUC for balanced data or when you care about ranking across both classes. Use PR AUC when positives are rare and false positives are expensive.
Key idea
PR AUC ignores true negatives and centers the rare positive class, making it the better summary under heavy imbalance. Judge it against prevalence, not 0.5.