Definition
Recall in AI is a metric measuring the proportion of relevant items successfully retrieved from a total set of relevant items. It calculates the completeness of a model's predictions by dividing the number of correctly identified positive instances by the total number of actual positive instances. Higher recall indicates the model's ability to capture more true positive results in tasks like information retrieval, classification, and recommendation systems.
When we evaluate how well an AI model performs, especially in classification and information retrieval tasks, we need more than just overall accuracy. One important measure is recall. Recall tells us how good a model is at finding all the relevant cases.
In simple terms, recall measures completeness. It answers the question, of all the real positive cases out there, how many did the model actually catch?
What recall means
Recall is the fraction of relevant instances that were successfully retrieved. It focuses only on the positive class, the cases we care most about detecting. The formula is:
Recall = true positives divided by all actual positives or Recall = TP / (TP + FN)
Here:
- True positives are items correctly identified as positive
- False negatives are items that were positive but the model missed
A higher recall means the model is missing fewer real positive cases.
A simple example
Imagine a system that recognises dogs in photos. In one picture, there are twelve dogs and ten cats. The model says it sees eight dogs. Out of those eight:
- Five really are dogs, these are true positives
- Three are actually cats, these are false positives
However, there are seven dogs the model failed to detect. These are false negatives. So the recall is:
5 / 12
The model found five of the twelve real dogs. Even though it detected some correctly, it still missed many, so recall is not very high.
Recall versus precision
Recall and precision are closely linked but focus on different things.
- Precision measures how many of the detected positives were actually correct
- Recall measures how many of the real positives were detected at all
It is possible to get perfect recall by labelling everything as positive. You would not miss any real positives, but you would also produce many false positives. That would make precision very low.
Because of this trade off, recall is rarely used on its own.
When recall matters most
Recall is especially important when missing a positive case is very costly. Examples include:
- Medical diagnosis, where missing a disease case can be dangerous
- Fraud detection, where failing to flag a fraudulent transaction can cause financial loss
- Safety and security systems that must detect all potential threats
In such situations, it is often better to accept more false alarms than to miss real problems.
Limits of recall
Recall does not consider false positives. A model with very high recall might still be impractical if it flags too many normal cases as positive. This is why recall is usually looked at together with precision, or combined into a single measure such as the F1 score.
Key takeaways
- Recall measures how many real positive cases a model successfully finds
- It is calculated as true positives divided by all actual positives
- High recall means few important cases are missed
- Recall is crucial when false negatives are costly
- It should be balanced with precision to get a full picture of model performance