precisionrecallcurves

Precisionrecallcurves, commonly referred to as precision-recall curves, are diagnostic plots used to evaluate binary classifiers by illustrating the trade-off between precision and recall across different score thresholds. For a given threshold, instances with predicted scores above the threshold are labeled positive. Precision is computed as TP divided by TP plus FP, and recall as TP divided by TP plus FN. The curve is drawn with recall on the x-axis and precision on the y-axis, tracing points corresponding to a range of thresholds from high-confidence positives to low-confidence positives.

PR curves are particularly informative when the positive class is rare or imbalanced, because they foreground performance on detecting positives rather than overall discrimination. Unlike ROC curves, which can present an optimistic view on imbalanced data, PR curves emphasize the classifier’s ability to retrieve relevant instances without bystander counts of true negatives. A common scalar summary is the area under the curve (AUPRC) or, in some contexts, average precision, which approximates the integral of precision with respect to recall. Higher AUPRC indicates better performance, though the baseline depends on the prevalence of the positive class.

In practice, PR curves are used for model comparison and threshold selection in domains where identifying positive cases is critical, such as medical screening or information retrieval. In multiclass settings, curves are typically constructed in a one-vs-rest fashion or averaged across classes. Limitations include sensitivity to class prevalence and the need for well-calibrated probability scores, as poorly calibrated scores can distort the curve and the derived AUPRC.