Recall is a performance metric used to evaluate classification models, measuring the proportion of actual positive instances that are correctly identified by the model. Also known as sensitivity or the true positive rate, it quantifies the model’s ability to capture all relevant cases within a dataset. A high recall value indicates that the algorithm successfully minimizes false negatives, ensuring that very few instances of the target class are overlooked.
How it works
Recall is calculated by dividing the number of true positives by the total number of actual positive instances. In this context, a true positive occurs when the model correctly predicts the positive class for an instance that truly belongs to that class. The denominator includes all instances that actually belong to the positive class, regardless of whether the model identified them correctly or missed them. This calculation yields a value between zero and one, where one represents perfect recall—meaning every single actual positive instance was detected.
The metric focuses exclusively on the actual positive cases, ignoring the true negatives. This means that recall does not penalize the model for incorrectly labeling negative instances as positive (false positives). Instead, it answers the question: of all the items that are actually positive, how many did the model find? By focusing on the actual positives, recall highlights the model’s sensitivity to the target class. If a model has a recall of zero, it means it failed to identify any of the actual positive instances, even if it correctly identified all the negative ones.
In practice, recall is often considered alongside precision to understand the full picture of model performance. While recall measures the completeness of detection among actual positives, it does not indicate the purity of the positive predictions. A model can achieve perfect recall by simply predicting that every single instance in the dataset is positive. In such a scenario, the model captures all actual positives, but it also captures many false positives, resulting in low precision. Therefore, recall is a measure of coverage rather than accuracy in isolation.
Where it is used
Recall is particularly valuable in classification tasks involving binary outcomes or imbalanced datasets, where the cost of missing a positive case is high. It is widely used in scenarios where the primary goal is to ensure comprehensive coverage of the target class, even if it means accepting a higher number of false positives. This trade-off is critical in domains where the consequence of a false negative is more severe than that of a false positive.
Common applications include medical diagnosis, where failing to detect a disease (a false negative) can lead to serious health consequences, and fraud detection, where missing a fraudulent transaction can result in financial loss. In these settings, it is often preferable to err on the side of caution by flagging more instances as positive, ensuring that few actual cases are overlooked. The metric reinforces the idea that comprehensive detection is prioritized over strict precision in these high-stakes environments.
Recall is also essential in information retrieval and search systems, where the objective is to retrieve all relevant documents for a query. If a search engine misses relevant documents, the user experience is degraded, even if the documents it does return are highly relevant. Similarly, in anomaly detection systems, recall measures the system’s ability to catch all anomalies, which is crucial for maintaining security or operational integrity.
Limitations and trade-offs
The primary limitation of recall is that it does not account for false positives. A model can achieve high recall by being overly aggressive in its positive predictions, which may lead to a high rate of false alarms. This can be problematic in contexts where the cost of investigating or acting on a false positive is significant. For example, in spam filtering, a high recall might mean that all spam is caught, but if legitimate emails are also frequently flagged as spam, the user experience suffers due to the loss of valid communications.
Another trade-off is that recall alone does not provide a complete picture of model performance. It must be interpreted in conjunction with other metrics, such as precision or the F-score, which combines recall and precision into a single value. Without considering precision, a high recall value might mask a model that is essentially guessing that everything is positive. Additionally, in highly imbalanced datasets where positive cases are rare, a model can achieve high recall by simply predicting the positive class more often, but this might not reflect true predictive power if the precision is very low.
Related terms
- False Negative – Recall is directly determined by the number of false negatives, as it measures the proportion of actual positives that were not missed.
- False Positive – Recall ignores false positives, focusing only on the detection of actual positives, which creates a trade-off with precision.
- F-score – The F-score combines recall and precision into a single metric, providing a balanced measure of a model’s performance.
- Type II Error – A Type II error corresponds to a false negative, which is the event that recall aims to minimize.
- Supervised Learning – Recall is a standard evaluation metric used in supervised classification tasks to assess model performance.
- Ground Truth – Recall is calculated by comparing the model’s predictions against the ground truth labels of the actual positive instances.
