glossary terms
Recall
- Category
- Evaluation, Safety & Governance
- Difficulty
- Intermediate
Definition
Recall is a classification performance metric that calculates the proportion of actual positive instances that were correctly identified by the model.
How It Works and Context
In machine learning, recall (also known as sensitivity or the true positive rate) is a fundamental metric used to evaluate the performance of classification models. It is calculated by dividing the number of true positives by the sum of true positives and false negatives. A high recall score indicates that the model is effective at capturing most of the positive instances, minimizing the number of missed detections. This metric is particularly vital in scenarios where missing a positive case carries significant consequences, such as medical diagnostics, fraud detection, or safety-critical systems. However, recall must be balanced against precision; a model that predicts 'positive' for every single input will achieve perfect recall but suffer from extremely low precision, leading to an overwhelming number of false alarms.
Why It Matters
Recall is essential for assessing the reliability of AI systems in high-stakes environments. In fields like healthcare or cybersecurity, a model with low recall is dangerous because it fails to flag critical issues. By monitoring recall, developers can ensure their systems are not missing important signals, which is a prerequisite for building safe, trustworthy, and effective AI applications that users can depend on for accurate identification.
Real-world Example
In a cancer screening AI, recall is the most important metric. If the model has 95% recall, it correctly identifies 95 out of 100 patients who actually have the disease. The remaining 5 patients are 'false negatives'—they have the disease but were missed by the AI. In this context, maximizing recall is prioritized over precision to ensure that as few sick patients as possible are sent home with a false sense of security.
Common Mistakes
- Confusing recall with precision: Recall focuses on missing positive cases, while precision focuses on the accuracy of the positive predictions made.
- Ignoring the tradeoff: Assuming a model can have both perfect recall and perfect precision simultaneously without adjusting thresholds.
- Over-optimizing for recall: Creating a model that flags everything as positive, which renders the system useless due to excessive false positives.
Frequently Asked Questions
How does recall relate to the F1-score?
The F1-score is the harmonic mean of precision and recall. It provides a single metric that balances both, which is useful when you need to compare models that have different strengths in precision and recall.
Can recall be 100%?
Yes, recall can be 100% if the model correctly identifies every single positive instance in the dataset. However, this often comes at the cost of very low precision, as the model may also label many negative instances as positive.
When should I prioritize recall over precision?
You should prioritize recall when the cost of a false negative (missing a positive case) is much higher than the cost of a false positive (a false alarm). Examples include life-threatening medical conditions or critical security threats.