Recall is the proportion of actual positive cases that a model correctly identifies, calculated as true positives divided by the sum of true positives and false negatives. In model evaluation, recall answers one question: of everything the model should have found, how much did it find? It is also called sensitivity or the true positive rate.
Recall is the metric of choice when missing a positive case is costly, as in disease screening, fraud detection, safety filtering or document retrieval for legal review. It is always read together with precision, because a model can reach perfect recall trivially by labeling everything as positive.
How recall is calculated
Recall is computed from the confusion matrix, which counts how a classifier’s predictions relate to the true labels. Only the cases that are actually positive enter the formula: Recall = TP / (TP + FN).
| Term | Meaning | Role in recall |
|---|---|---|
| True positive (TP) | Positive case predicted as positive | Numerator and part of denominator |
| False negative (FN) | Positive case predicted as negative | Part of denominator, lowers recall |
| False positive (FP) | Negative case predicted as positive | Not used in recall, lowers precision |
| True negative (TN) | Negative case predicted as negative | Not used in recall |
Recall ranges from 0 to 1, often expressed as a percentage. A recall of 0.92 means the model found 92 of every 100 real positives and missed 8. Because false positives do not appear in the formula, recall says nothing about how many false alarms the model raises.
Recall, precision and the decision threshold
Most classifiers output a score or probability, and a decision threshold turns it into a yes or no. Lowering the threshold flags more cases as positive, which raises recall and usually lowers precision. This is the central trade-off in classification.
| Threshold change | Effect on recall | Effect on precision |
|---|---|---|
| Lower threshold | Increases, fewer misses | Usually decreases, more false alarms |
| Higher threshold | Decreases, more misses | Usually increases, fewer false alarms |
| Threshold of zero | Reaches 1.0, every positive flagged | Falls to the positive base rate |
The precision-recall curve plots both metrics across all thresholds, and its area, often reported as average precision, summarizes a model independently of any single threshold. Practitioners often fix a minimum recall requirement, for example 0.95, and then choose the threshold that gives the best precision at that level.
Recall compared with related metrics
Recall belongs to a family of metrics that look at the confusion matrix from different angles. Choosing among them depends on which error matters more and how balanced the classes are.
| Metric | Formula | Question it answers |
|---|---|---|
| Recall | TP / (TP + FN) | How many real positives were found |
| Precision | TP / (TP + FP) | How many flagged items are truly positive |
| F1 score | Harmonic mean of precision and recall | How balanced are both |
| Specificity | TN / (TN + FP) | How many real negatives were correctly rejected |
| Accuracy | (TP + TN) / all cases | How many predictions were correct overall |
Accuracy can be badly misleading with imbalanced data. If 1% of transactions are fraudulent, a model that never flags fraud has 99% accuracy and zero recall. The F-beta score generalizes F1: with beta greater than 1, such as F2, recall is weighted more heavily than precision.
Recall in retrieval and generative AI evaluation
In search and retrieval-augmented generation, recall is measured over a ranked list. Recall at k is the share of all relevant documents that appear among the top k results, for example recall at 10. It shows whether the retriever brings the needed evidence into the model’s context at all.
For generative models, recall also appears in summary and extraction evaluation. ROUGE, a family of summarization metrics, is recall-oriented by design: it measures how much of a reference text’s content appears in the generated summary. In information extraction, recall is the share of entities or facts in the source that the model successfully extracted.
With multiple classes, recall is computed per class and then averaged. Macro averaging gives every class equal weight and exposes weak performance on rare classes, while micro averaging pools all cases and is dominated by frequent classes.
Common mistakes when reporting recall
Recall is simple to compute but easy to misreport. The most frequent problems concern incomplete ground truth, missing context and small samples.
| Mistake | Why it misleads | Better practice |
|---|---|---|
| Reporting recall alone | A model that flags everything scores 1.0 | Report precision or F1 alongside |
| Incomplete labels | Unlabeled positives inflate apparent recall | Audit samples of negatives for hidden positives |
| Tuning on the test set | Threshold choice overfits the evaluation data | Choose thresholds on a validation set |
| Tiny positive class | A few cases make recall swing widely | Report confidence intervals or counts |
| Mixing averaging methods | Macro and micro recall can differ greatly | State the averaging method explicitly |
The evaluation dataset must also reflect the deployment population. Recall measured on a curated benchmark can drop sharply when real inputs contain new variants of the positive class, so monitoring recall after deployment is part of responsible evaluation.
In simple terms
Recall tells you how much a model catches out of everything it should catch. Imagine searching a beach for 100 lost coins. If you find 90 of them, your recall is 90%, no matter how many bottle caps you also picked up by mistake.
That last part is the catch. Recall ignores false alarms, so a searcher who scoops up the entire beach finds every coin. This is why recall is always paired with precision, which measures how much of what you picked up was actually a coin.
A practical example
A hospital evaluates an image model that flags chest X-rays for possible pneumonia before radiologists review them. On a validation set of 2,000 images with 200 confirmed cases, the model at its default threshold of 0.5 finds 170 cases, a recall of 0.85, with a precision of 0.80.
Because a missed case means a delayed diagnosis, clinicians require recall of at least 0.95. The team lowers the threshold to 0.25, where the model finds 191 of the 200 cases. Precision falls to about 0.55, so radiologists review more false alarms, but only 9 cases are missed instead of 30.
The hospital accepts the extra review workload as the price of safety and schedules a monthly check of recall on newly labeled images to catch any drift.