Recall (Sensitivity)
When you build a classifier, you’re usually trying to “catch” something: fraudulent transactions, sick patients, churn-risk customers. Recall (also called sensitivity or true positive rate) measures how many of the real cases your model actually catches.
What recall measures
Recall focuses on the positive class (the thing you care about detecting). It answers: “Out of all the truly positive examples, what fraction did the model label as positive?” Using the confusion matrix terms:
- True Positives (TP): positives correctly predicted as positive
- False Negatives (FN): positives incorrectly predicted as negative (misses)
The formula is:
Recall = TP / (TP + FN)High recall means few false negatives; low recall means you’re missing many real positives.
Why it matters in practice
Recall becomes the headline metric when missing a positive case is costly. For example:
- Disease screening: low recall means sick patients are told they’re fine.
- Fraud detection: low recall means fraudulent charges slip through.
- Safety or compliance: low recall means dangerous content or events aren’t flagged.
In these settings, you typically accept more false alarms to avoid misses—so recall is evaluated alongside precision.
How recall is controlled and reported
Recall depends on the decision threshold. Lowering the threshold (predicting “positive” more easily) usually increases recall while hurting precision. In scikit-learn, you’ll see it via sklearn.metrics.recall_score, and you can study the trade-off with a precision-recall curve.
Recall (Sensitivity) is the fraction of actual positive cases that a classifier correctly identifies: TP / (TP + FN), where TP is true positives and FN is false negatives. It measures how completely the model captures the positive class (e.g., detecting diseased patients among all diseased patients). It matters when missing positives is costly, and it is central to threshold selection and trade-offs with precision.
Imagine a smoke alarm. A “good” alarm is one that catches as many real fires as possible, even if it sometimes beeps when you burn toast. Recall (also called Sensitivity) is that same idea for AI classifiers: out of all the cases that truly are positive, how many did the model successfully catch?
For example, in medical screening, high recall means the system finds most people who really have the disease (few missed cases). In spam detection, high recall means it catches most spam emails. It matters most when missing a true positive is costly or dangerous.