Precision and Recall

Also known as · precision-recall

The paired classification metrics: precision is the share of predicted positives that are really positive, TP/(TP+FP)TP/(TP+FP), and recall is the share of actual positives that were caught, TP/(TP+FN)TP/(TP+FN).

When to use

Use them instead of plain accuracy on imbalanced classes such as churn. Precision matters when each alarm has a cost (ad spend, calls); recall matters when missing a case is costly (fraud). Lowering the threshold raises recall and lowers precision, so pick the threshold from the business cost of each error; F1 gives a single balanced number.

See DS Lec 3.

Appears in