Precision and Recall

Precision

true positives / all positive predictions

$$\dfrac{\text{TP}}{\text{TP} + \text{FP}}$$

“What fraction of positive predictions were actually correct?” — focus on the correctness of the positive predictions
or: “Did the model predict positive too often?”
e.g. precision = 0.5, model is correct 50% of the time

true positives / all actual positives

$$\dfrac{\text{TP}}{\text{TP} + \text{FN}}$$

“What fraction of actual positives were identified correctly?” — focus on the identification of the actual positives
or: “How many misses?”
e.g. recall = 0.11, model correctly identifies 11% of the actual positive cases

Increased precision: predict positive only when we are confident in the positive prediction
- –> increasing the classification threshold / casting a smaller net
- reduces false positives
Increased recall: predict more positive so as to prevent the number of misses
- –> lowering the classification threshold / casting a wider net to catch all actual positive cases
- reduces false negatives
- increases false positives
Always look at both precision and recall!

Helpful graphic

Source: https://en.wikipedia.org/wiki/Precision_and_recall