Recall asks a simple question: among the examples that really are positive, how many did the classifier find? It is especially useful when missing a positive case—a false negative—is costly, such as failing to identify a sick patient. section 7.7 section 7.18
Count the true positives, then divide by every actual positive: recall = true positives / (true positives + false negatives). A low score means the model missed many positive examples; false positives do not directly lower recall. section 7.7 section 7.17
Two coronavirus classifiers
Suppose 10 people are sick. A model that identifies none has recall 0/10 = 0. A second model identifies 8 sick people and misses 2, so its recall is 8/10 = 0.8, or 80%. Although the second model makes more total mistakes, it is better at finding sick people. section 7.7
Table 7.3 The confusion matrix of our second coronavirus model
|
Coronavirus model 2
|
Diagnosed sick
|
Diagnosed healthy
|
|
Sick
|
8 (true positives)
|
2 (false negatives)
|
|
Healthy
|
48 (false positives)
|
942 (true negatives)
|
Use the confusion matrix to locate the 8 true positives and 2 false negatives, then calculate recall as 8 divided by 8 + 2. section 7.7
Recall is also called sensitivity or the true positive rate. It is not precision: precision asks whether positive predictions were actually positive, while recall asks whether actual positives were found. section 7.13 section 7.17
Choosing recall means choosing which error matters most:
- Favor high recall when false negatives are more serious than false positives, as in a model meant to find sick people. section 7.11 section 7.18
- Balance recall with precision when both kinds of error matter; an Fβ-score gives recall more weight when β is larger. section 7.18 section 7.11
- Adjust a probability threshold to trade sensitivity against specificity: lowering the threshold raises sensitivity while lowering specificity. section 7.14 section 7.16


































