Recall (TPR, Sensitivity)

Recall, also known as true positive rate (TPR) and sensitivity, measures the proportion of all positive ground truths that a model correctly predicts, ranging from 0 to 1 (where 1 is best).

As shown in this diagram, recall is the fraction of all positive ground truths that are correctly predicted:

\[\text{Recall} = \frac{\text{TP}}{\text{TP} + \text{FN}}\]

In the above formula, \(\text{TP}\) is the number of true positive inferences and \(\text{FN}\) is the number of false negative ground truths.

Guide: True Positive / False Negative

Read the TP / FP / FN / TN guide if you're not familiar with "TP" and "FN" terminology.

• API Reference: recall ↗

Implementation Details

Recall is used across a wide range of workflows, including classification, object detection, instance segmentation, semantic segmentation, and information retrieval. It is especially useful when the objective is to measure and reduce false negative ground truths, i.e. model misses.

For most tasks, recall is the ratio of the number of correct positive inferences to the total number of positive ground truths.

\[ \text{Recall} = \frac {\text{# True Positives}} {\text{# True Positives} + \text{# False Negatives}} \]

For workflows with a localization component, such as object detection and instance segmentation, see the Geometry Matching guide to learn how to compute true positive and false negative counts.

Examples

Perfect model inferences, where every ground truth is recalled by an inference:

Metric	Value
TP	20
FN	0

\[ \begin{align} \text{Recall} &= \frac{20}{20 + 0} \\[1em] &= 1.0 \end{align} \]

Partially correct inferences, where some ground truths are correctly recalled (TP) and others are missed (FN):

Metric	Value
TP	85
FN	15

\[ \begin{align} \text{Recall} &= \frac{85}{85 + 15} \\[1em] &= 0.85 \end{align} \]

Zero correct inferences — no positive ground truths are recalled:

Metric	Value
TP	0
FN	20

\[ \begin{align} \text{Recall} &= \frac{0}{0 + 20} \\[1em] &= 0.0 \end{align} \]

Multiple Classes

So far, we have only looked at binary classification/object detection cases, but in multiclass or multi-label cases, recall is computed per class. In the TP / FP / FN / TN guide, we went over multiple-class cases and how these metrics are computed. Once you have these four metrics computed per class, you can compute recall for each class by treating each as a single-class problem.

Aggregating Per-class Metrics

If you are looking for a single recall score that summarizes model performance across all classes, there are different ways to aggregate per-class recall scores: macro, micro, and weighted. Read more about these methods in the Averaging Methods guide.

Limitations and Biases

As seen in its formula, recall only takes positive ground truths (TP and FN) into account; negative ground truths (TN and FP) are not considered. Thus, recall only provides one half of the picture, and should always be used in tandem with precision: precision penalizes false positives (FP), whereas recall does not.

For a single metric that takes both precision and recall into account, use F₁-score, which is the harmonic mean between precision and recall.