Precision-Recall curves provide a way to evaluate how well a model can classify categorical responses. They can also be used to select a threshold for classification. A precision-recall curve plots the precision values against the recall values at a variety of thresholds. Precision is defined as the proportion of predicted responses that are correctly classified as positives. Recall is defined as the proportion of actual positive responses that are correctly classified as positives. Next to each precision-recall curve plot, there is a table that contains the area under the curve for each response. The area under the precision-recall curve is a single metric that summarizes the overall performance of the model. A higher area under the curve indicates better model performance, with a value of one indicating a perfect fit.
Precision-recall curves are especially useful when classes are imbalanced. Class imbalance occurs when the number of observations at one response level is much smaller than the number of observations at the other response level. These are referred to as the minority and majority classes, respectively. In many instances, such as rare disease identification, fraud detection, or equipment failure, the minority class is the class of interest. Classification measurements that rely on accuracy of predictions can be misleading in these scenarios. This can occur if a model predicts the majority class for every observation, it could have a low overall misclassification rate. However, it would have no practical use if you are interested in classifying observations in the minority class. Precision-recall curves can help focus on the performance of predicting the observations that are in the minority class. Precision-recall curves are also useful in cases where incorrectly classifying a positive response as negative (a false negative) has severe consequences.