PR AUC, which stands for “Area Under the Precision–Recall Curve,” is a performance metric used to evaluate machine learning models, particularly for classification tasks. The PR curve plots precision (the ratio of true positives to all positive predictions) against recall (the ratio of true positives to all actual positives) at various threshold settings. The area under this curve—PR AUC—summarizes the trade-off between precision and recall in a single value, which ranges from 0 to 1. A higher PR AUC means the model does a better job distinguishing between the positive and negative classes.
PR AUC is especially useful when working with imbalanced datasets, where the number of positive examples is much smaller than the number of negative examples. In such cases, traditional metrics like accuracy or even the ROC AUC (Area Under the Receiver Operating Characteristic Curve) can be misleading. That’s because a model could achieve high accuracy by simply predicting the majority class, while missing almost all of the minority (often more important) class. PR AUC, on the other hand, focuses on how well the model identifies positive instances, making it a preferred metric for tasks like fraud detection, medical diagnosis, and information retrieval.
To compute the PR AUC, you first calculate precision and recall at many different thresholds. As you change the threshold for what constitutes a “positive” prediction, both precision and recall will change. By plotting these pairs and measuring the area under the resulting curve, you get a single number that reflects the model‘s ability to balance identifying positives (recall) without introducing too many false positives (precision).
Unlike ROC AUC, which evaluates the trade-off between true positive rate and false positive rate, PR AUC is more sensitive to the performance of the model on the positive class. This makes it more informative when positive events are rare or especially important. The interpretation is straightforward: a PR AUC of 1.0 means perfect precision and recall at all thresholds; a value close to 0.0 indicates poor performance.
When comparing models, a higher PR AUC generally means better performance in terms of identifying true positives without generating excessive false alarms. However, it’s important to compare PR AUC scores only on the same dataset, since the metric is sensitive to class imbalance. Practitioners often visualize both the PR curve and the ROC curve to get a full picture of model performance, but for rare-event classification, PR AUC is usually the gold standard.
