Category: Incident & Remediation
Recall
Also known as: Sensitivity, True Positive Rate
Simply put
In the context of AI and machine learning, recall measures how well a model finds the items it is supposed to find. It answers the question: of all the cases that were actually positive, what fraction did the model correctly identify? A high recall means the model misses few of the relevant cases, though it does not indicate how many of its positive predictions were wrong.
Formal definition
Recall is a classification performance metric commonly defined as the proportion of actual positive instances that a model correctly classifies as positive, typically expressed as true positives divided by the sum of true positives and false negatives. It is frequently reported alongside precision, since the two capture different error types and often trade off against one another; recall alone does not account for false positives. Note: the evidence packet provided describes 'recall' only in unrelated senses (product/vehicle safety recalls and the psychological process of memory retrieval) and contains no source defining recall as a machine learning evaluation metric. This entry therefore reflects the term as commonly used in model evaluation, but the supplied evidence does not substantiate it; the definition should be treated as unsourced pending appropriate references.
Why it matters
Recall is central to evaluating AI systems in settings where failing to catch a positive case carries serious consequences. In domains such as fraud detection, medical screening, or anti-money-laundering monitoring, a missed positive (a false negative) can be far costlier than a false alarm. Recall quantifies exactly this exposure: it tells you the fraction of genuinely relevant cases the model actually surfaced. For model risk managers and compliance officers, tracking recall helps make the cost of missed detections visible and auditable rather than buried in an aggregate accuracy figure.
Who it's relevant to
Model Risk Managers and Validators
Those responsible for measuring and monitoring model risk use recall to quantify exposure to missed positive cases (false negatives), particularly where such misses carry regulatory or safety consequences. They should insist that recall be evaluated alongside precision and the underlying base rate rather than in isolation.
Data Scientists and ML Engineers
Practitioners who build and tune classification models adjust decision thresholds knowing that recall and precision often trade off against one another. They rely on recall to understand how many relevant cases a model captures and to communicate that error profile to reviewers.
Compliance Officers and Auditors
Professionals reviewing AI systems in higher-stakes settings should be cautious about accepting a single recall figure as evidence of performance. They should confirm that both error types are documented and that the metric's meaning is not confused with unrelated senses of the word 'recall.'
Domain Specialists in Detection-Oriented Applications
Experts in areas such as fraud monitoring, screening, or safety-critical review, where a missed positive is especially costly, use recall to make the rate of missed detections explicit and to weigh it against the operational burden of false positives.
Inside Recall
True positives
The count of instances the model correctly identified as belonging to the positive (target) class. Recall measures how completely the model captures these among all actual positive cases.
False negatives
Actual positive instances the model failed to identify. Recall is sensitive to false negatives; as they rise, recall falls. This makes recall a common focus where missed detections carry high cost, such as fraud, disease screening, or safety-critical classification.
Recall formula
Commonly defined as true positives divided by the sum of true positives and false negatives (i.e., true positives over all actual positives). It answers the question: of all real positive cases, what proportion did the model detect?
Relationship to precision
Recall is typically evaluated alongside precision, which measures the proportion of predicted positives that are correct. The two often trade off against one another as the classification threshold changes, which is why they are frequently reported together rather than in isolation.
Threshold dependence
For scoring or probability-based classifiers, recall is a function of the decision threshold applied to convert scores into class labels. Lowering the threshold generally increases recall while typically reducing precision.
Class-conditional interpretation
Recall is computed per class and is conditioned on the actual (ground-truth) label. In multiclass settings, recall is often reported per class and may be aggregated (for example, by macro or weighted averaging) depending on the evaluation goal.
Common questions
Answers to the questions practitioners most commonly ask about Recall.
Does a high recall value mean a model is performing well overall?
No. Recall measures only the proportion of actual positive cases the model correctly identifies; it says nothing about how many of the model's positive predictions are correct. A model can achieve very high recall by labeling nearly everything as positive, which typically produces many false positives and low precision. Recall should be interpreted alongside precision and other metrics rather than treated as a standalone indicator of overall performance.
Are recall and accuracy the same thing?
No. Accuracy reflects the proportion of all predictions that are correct across both positive and negative classes, whereas recall focuses specifically on the model's ability to capture actual positive cases. In imbalanced datasets, accuracy can appear high while recall for the minority class remains low, so the two metrics are not interchangeable and can move in opposite directions.
When should recall be prioritized over precision in practice?
Recall is commonly prioritized when the cost of missing a true positive is high relative to the cost of a false positive, such as in certain screening or detection contexts. The appropriate emphasis depends on the specific use case, error costs, and risk tolerance, and this trade-off should typically be documented and justified rather than assumed. The determination of acceptable thresholds is generally a business and risk decision, not a purely technical one.
How does the choice of classification threshold affect recall?
For models that output probabilities or scores, lowering the decision threshold generally increases recall by classifying more cases as positive, while often reducing precision. Adjusting the threshold is a common way to tune the recall-precision balance. Any threshold selected for production use should typically be validated and monitored, since the balance that holds during development may shift as data changes over time.
How should recall be monitored after a model is deployed?
Recall can degrade over time as input data or the underlying population shifts, which may reflect model performance degradation distinct from the model's inherent design. Ongoing monitoring typically involves tracking recall against defined thresholds using labeled outcomes as they become available, though delays or gaps in obtaining ground-truth labels can complicate this and should be accounted for in the monitoring approach.
How should recall be reported when there are multiple classes?
In multiclass settings, recall is typically calculated per class and can be aggregated using approaches such as macro or micro averaging, which can produce different results, particularly under class imbalance. To avoid ambiguity, documentation should generally specify which class or averaging method a reported recall value refers to, since a single headline figure can obscure poor performance on individual classes.
Common misconceptions
A high recall score means the model is performing well overall.
Recall alone can be trivially maximized by predicting the positive class for nearly everything, which inflates false positives and typically collapses precision. Recall should be interpreted together with precision and other measures appropriate to the use case rather than treated as a standalone indicator of quality.
Recall and precision are interchangeable or measure the same thing.
They answer different questions. Recall measures completeness of detection among actual positives (sensitivity to false negatives), while precision measures correctness among predicted positives (sensitivity to false positives). Experts keep these distinct because optimizing for one commonly degrades the other.
Recall is a fixed property of the model.
For scoring models, recall depends on the chosen decision threshold and on the underlying data distribution. A reported recall figure is meaningful only in reference to the threshold, dataset, and class definition used to compute it, and may shift as data or thresholds change.
Best practices
Report recall alongside precision (and, where relevant, threshold-independent summaries) rather than in isolation, so the completeness-versus-correctness trade-off is visible to decision-makers.
State the decision threshold, the dataset, and the class definition used whenever a recall figure is documented, since recall is threshold- and data-dependent.
Set target recall levels based on the operational cost of false negatives for the specific use case, and document the rationale for the chosen threshold.
For multiclass problems, report per-class recall and specify the aggregation method (for example, macro or weighted) so results are not obscured by class imbalance.
Monitor recall over time on representative data to detect changes in model behavior as input distributions shift, and treat unexplained movement as a trigger for review.
Avoid tuning solely to maximize recall; validate that gains in recall do not produce unacceptable increases in false positives for the intended application.