A curve focused on the positive class
The precision recall curve plots precision against recall as the threshold varies. Unlike the ROC curve, it ignores true negatives entirely, which makes it the right tool when the positive class is rare.
Why it beats ROC on imbalance
When negatives vastly outnumber positives, even many false positives barely move the false positive rate, so the ROC curve and AUC can look flattering. The precision recall curve has no such blind spot because precision falls sharply the moment false positives creep in.
Summarizing the curve
- Average precision summarizes the curve as a weighted mean of precision values across recall levels.
- A useful baseline is the fraction of positives in the data, which is what a random model achieves.
- A model is only impressive if its average precision sits well above that baseline.
Key idea
The precision recall curve ignores the abundant true negatives, so on rare positive problems it reveals weaknesses that an optimistic ROC curve would hide.