Recall is a metric that answers the question: of all the cases that really are positive, how many did the model find? It is the number of true positives divided by the number of actual positives, which is true positives plus false negatives [1]. If 100 patients have a disease and a screening model flags 90 of them, recall is 90 percent, and ten sick patients are missed. Recall is also called sensitivity or the true positive rate. It is a core part of model evaluation, the companion of precision, and is read from the confusion matrix.
The formula
Recall = TP / (TP + FN)
where TP is the number of true positives and FN is the number of false negatives, the real positives that the model failed to flag. Recall ranges from 0 to 1. A model that flags everything achieves perfect recall, which shows why recall must be read together with precision: calling every email spam catches all spam but also blocks every real message.
Example
A security system must detect intrusions in network traffic. Over a week there are 40 real intrusions. The model flags 34 of them (TP = 34) and misses 6 (FN = 6). Recall is 34 divided by 40, or 85 percent. The six missed intrusions are the cost. If the team lowers the model’s threshold to be more sensitive, recall might rise to 95 percent, but the model will flag more harmless traffic, and precision will usually fall.
What recall tells you
Recall measures the cost of misses. High recall means few real cases slip through. It is the main concern whenever failing to detect something is serious or irreversible. It is blind to false alarms, which is why a complete picture needs precision, the F1 score or curves that show both. Unlike accuracy, recall focuses on the positive class and is not fooled by large numbers of easy negatives.
Recall versus precision
The two measures pull against each other. As the decision threshold falls, a model flags more items, so recall rises while precision usually drops. As the threshold rises, the reverse happens [1]. The right balance depends on the costs involved.
- Favor recall when a miss is dangerous or expensive: cancer screening, fraud detection where losses are large, safety monitoring, search for evidence in legal discovery, detection of harmful content involving children, or finding every relevant document in a systematic review.
- Favor precision when false alarms are costly: automated penalties, spam filtering of legitimate mail, or costly manual follow-up.
Systems often combine the two with a staged design. A high-recall first model casts a wide net, and a more precise second stage or a human reviewer filters the candidates.
Sensitivity and specificity
In medicine and statistics, recall is called sensitivity, the proportion of people with a condition whom a test identifies [2]. Its counterpart is specificity, the proportion of healthy people the test correctly clears, which equals one minus the false positive rate. Together they describe a diagnostic test. The ROC curve plots sensitivity against one minus specificity across thresholds.
Variants
- Macro, micro and weighted recall. Ways of averaging per-class recall in multi-class problems.
- Recall at k. The share of all relevant items that appear in the top k results, used in search, recommendation and retrieval for retrieval-augmented generation.
- Mean average recall. Used in some object detection evaluations; see object detection.
- Recall for generated text. In summarization metrics such as ROUGE, recall counts how much of a reference summary’s content appears in the generated one.
Improving recall
- Lower the decision threshold.
- Collect more examples of the under-detected cases, perhaps with data augmentation.
- Rebalance training with class weights or sampling, since models trained on rare positives tend to under-predict them.
- Add features that separate hard cases.
- Ensemble several models, or use a second model that rescues missed items.
- Check labels, since positives labeled as negatives hide true recall; see data labeling.
Always measure the effect on precision and on downstream workload, since catching more cases generates more items to review.
Pitfalls
- Reporting recall alone. It is trivially maximized by flagging everything.
- Unrepresentative positives. If the test positives are easier than real ones, recall is overstated.
- Unknown true positives. In search and discovery tasks, the total number of relevant items may not be known, so recall must be estimated by sampling.
- Ignoring subgroups. Recall can differ across groups, so a model may miss more cases for some populations; see algorithmic bias.
- Threshold drift. Changes in data can move recall over time; see data drift.
Recall in practice: three scenarios
Medical screening. A mammography screening model is tuned for high recall because a missed cancer can be fatal, while a false alarm leads to a follow-up scan. The program accepts that many flagged cases will turn out to be benign, and plans radiologist capacity accordingly. Recall is reported by age group and breast density, since a model may find tumors less reliably in some groups.
E-discovery. In legal disputes, lawyers must find nearly all relevant documents among millions. Courts have accepted statistical estimates of recall, for example by reviewing a random sample of documents the system set aside to estimate how many relevant ones were missed. Teams often aim for recall of 70 to 80 percent or more, depending on the matter, and measure it carefully because it is the legal standard of reasonable effort.
Retrieval for language models. A system that answers questions from company documents first retrieves passages and then asks a model to write an answer. If the retriever misses the passage that holds the answer, no amount of skill in the language model can recover it. Recall at k of the retriever is therefore a leading indicator of answer quality, and teams tune chunking, embeddings and search settings to raise it before worrying about the wording of answers.
Estimating recall when positives are unknown
In many real tasks nobody has a complete list of true positives. To estimate recall, analysts draw a random sample from the items the model did not flag, have experts check it, and estimate the share of hidden positives. Another approach is to plant known positives, such as test messages or synthetic fraud patterns, and see how many the system catches. Both methods come with uncertainty, so reports should give a range and describe how it was obtained.
FAQs
What is recall in machine learning? It is the share of all real positive cases that the model correctly finds.
What is the formula for recall? True positives divided by the sum of true positives and false negatives.
Is recall the same as sensitivity? Yes. Sensitivity and true positive rate are other names for recall.
What is the difference between recall and precision? Recall asks how many real positives were found. Precision asks how many flagged items were truly positive.
When is high recall most important? When missing a positive case is costly or dangerous, as in disease screening or safety monitoring.
More in AI Evaluation and Reliability
Key facts
- Recall is the number of true positives divided by the number of true positives plus false negatives, and is also called sensitivity or the true positive rate1.
- Lowering a classifier's decision threshold raises recall and usually lowers precision1.
- In medical testing, sensitivity is the proportion of people with a disease whom the test correctly identifies2.
Related concepts
Quick checkWhat is recall in machine learning?Show answer
Recall is the share of all actual positive cases that a model correctly identifies, calculated as true positives divided by the sum of true positives and false negatives. It measures how completely a model finds what it should find.
Sources
- Tom Fawcett, An Introduction to ROC Analysis, Pattern Recognition Letters, https://doi.org/10.1016/j.patrec.2005.10.010
- Douglas Altman and J. Martin Bland, Diagnostic tests 1: Sensitivity and specificity, BMJ, https://doi.org/10.1136/bmj.308.6943.1552
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
