Recall

Recall measures how completely a model finds the positive cases that actually exist. Of all the true positives in the data, recall is the fraction the model correctly identified. It is defined as true positives divided by the sum of true positives and false negatives.

Recall says nothing about how many false alarms a model raised along the way, only about how much of the real signal it captured. It is the natural counterpart to precision, and the two are almost always read together.

You take every actual positive in the ground truth, then check how many the model successfully predicted as positive. Recall is that count divided by the total number of real positives. Lowering a model's confidence threshold usually raises recall, because the model makes more positive predictions and catches more of the true cases, though this tends to reduce precision at the same time.

Recall is the metric to prioritize whenever a missed positive is costly or dangerous. In medical screening, a hazard detector for a moving vehicle, or fraud detection, failing to catch a real case matters far more than an occasional false alarm. Because recall and precision pull against each other, teams tune the decision threshold to the tolerance for misses versus false alarms, and often summarize the balance with an F1 score or a precision-recall curve.

Recall measures how many actual positives were found. Precision measures how many predicted positives were correct. Recall punishes false negatives, precision punishes false positives.

When a missed positive is expensive or unsafe, such as screening for disease or detecting hazards, where catching every real case matters most.

It depends on the task and the cost of a missed case. Recall should be read alongside precision, since a model can reach perfect recall simply by predicting positive for everything.

Precision, F1 score, Accuracy, Confusion matrix, Precision-recall curve.