Diagnostics › Diagnostic accuracy

Precision-Recall curve

A precision-recall curve and its average precision (AUC-PR) describe classifier performance, and are often more informative than ROC when the outcome is heavily imbalanced.

What is Precision-Recall curve?

When the positive class is rare (< 5%), high specificity is easy and ROC AUC inflates. PR curves use precision (= PPV) on the y-axis instead of (1 − specificity), so they're prevalence-sensitive in the way ROC isn't.

Average precision = area under the PR curve, computed as the trapezoidal integral over the (recall, precision) pairs at each unique threshold.

When should I use Precision-Recall curve?

  • Imbalanced classification (rare disease detection, fraud, IR retrieval).
  • Where false positives are costly relative to total negatives.

What data does it need?

Continuous predictor + binary outcome.

What does it report?

PR curve plot + AUC-PR.

How do I interpret the result?

Baseline AUC-PR = prevalence (not 0.5). Compare improvement over baseline, not over a fixed 0.5 reference.

See also

References

  • Saito & Rehmsmeier (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One 10(3).