What is a precision-recall curve?
A precision-recall curve is built by sweeping the model's confidence threshold from high to low and, at each step, recording the resulting precision and recall. Plotting those pairs traces a curve that shows how precision changes as you demand more recall. A model that keeps precision high even at high recall pushes the curve toward the top right, which is the ideal.
Because it captures behavior across all thresholds, the curve tells you far more than a single precision or recall value measured at one arbitrary cutoff.
Key takeaways
- It shows the precision-recall trade-off across every threshold, not just one.
- A curve hugging the top right indicates strong performance.
- The area under it is summarized by average precision.
How it works
Predictions are ranked by confidence, then the threshold is lowered one prediction at a time. Each threshold yields a precision and a recall, and those points form the curve. The area under the curve is commonly reduced to a single number called average precision, which makes curves easy to compare across models or classes.
Why it matters
Real systems have to pick an operating threshold, and the curve is how you choose one deliberately rather than by default. It also exposes models that look similar at one cutoff but behave very differently as you push for higher recall, which is exactly the situation where a single-number score can mislead.
Frequently asked questions
What does the area under a precision-recall curve mean?
It is summarized as average precision, a single number where higher is better and which reflects performance across all thresholds.
When is a precision-recall curve better than an ROC curve?
It is generally more informative than an ROC curve when the positive class is rare, because it is not flattered by a large number of true negatives.
Related terms