Keras · Class

metrics.AUC

keras.metrics.AUC estimates the area under a ROC or precision-recall curve for binary or multi-label classification. It accumulates thresholded confusion statistics and returns the approximated area as a scalar metric.

How it works

AUC accumulates classification statistics across predefined prediction thresholds and estimates the area under either a ROC or precision-recall curve. Call update_state() with labels and predictions, then result() to obtain the current scalar; the approximation uses bucketed values and a selected summation method.

Syntax

keras.metrics.AUC(num_thresholds=200, curve="ROC", summation_method="interpolation", name=None, dtype=None, thresholds=None, multi_label=False, num_labels=None, label_weights=None, from_logits=False)

Parameters

Name	Type	Description
`num_thresholds`	`int`	Number of thresholds used to discretize the ROC curve; must be greater than 1 and defaults to 200.
`curve`	`str`	Curve to compute: "ROC" or "PR"; defaults to "ROC".
`summation_method`	`str`	Method used to sum the discretized curve.
`name`	`str or None`	Optional metric name.
`dtype`	`str or None`	Optional metric dtype.
`thresholds`	`list or None`	Explicit threshold values instead of the default linearly spaced thresholds.
`multi_label`	`bool`	Whether inputs represent multiple labels.
`num_labels`	`int or None`	Number of labels for multi-label data.
`label_weights`	`list or None`	Optional weights for labels.
`from_logits`	`bool`	Whether predictions are logits rather than probabilities.

Returns

A scalar tensor containing the estimated AUC, returned by result().

Key points

AUC estimates the area under either the ROC curve or the precision-recall curve.
Predictions are grouped into value buckets, and per-bucket averages are interpolated to estimate the curve area.
Unlike accuracy, the metric summarizes model behavior across all operating points.
num_thresholds controls how finely the curve is discretized; its default is 200.
Use curve="ROC" for ROC-AUC or curve="PR" for PR-AUC.
The implementation tracks true positives, true negatives, false positives, and false negatives to construct the estimate.

Common mistakes

num_thresholds must be greater than one; otherwise the metric cannot discretize the curve correctly.

Example: Compute ROC-AUC from predictions

import keras

x = [0.1, 0.9, 0.8, 0.2]
y = [0, 1, 1, 0]

auc = keras.metrics.AUC(curve="ROC")
auc.update_state(y, x)
print(float(auc.result()))

Creates a ROC-AUC metric, accumulates labels and predicted probabilities, then reads the scalar result.

When to use it

Monitor ROC-AUC during training

Binary classification with probability outputs.

Track PR-AUC

Imbalanced binary classification where precision and recall matter.

Evaluate logits directly

Models producing raw scores before sigmoid or softmax conversion.

Measure multi-label performance

Multiple independent labels per example.

Interview questions

QWhat does keras.metrics.AUC measure?

AUC estimates the area under a ROC or precision-recall curve. It accumulates predictions into threshold buckets and computes the area from the resulting precision, recall, or false-positive-rate values.

Follow-up: Why can the value differ from an exact ranking-based AUC? Because Keras uses a finite threshold grid and a Riemann-sum approximation, so increasing num_thresholds can improve resolution.

QHow do ROC and PR modes differ?

curve selects the curve type: ROC is the default, while PR requests the precision-recall curve. ROC-AUC and PR-AUC evaluate model behavior across operational points rather than one classification threshold.

Follow-up: When might PR-AUC be preferable? It is often more informative when positive examples are uncommon because it focuses on precision and recall.

QHow is AUC accumulated during evaluation?

The metric is stateful: update_state() accumulates the batch statistics, and result() returns the current scalar AUC. A reset_state() call is used to clear accumulated state before a new evaluation period.

Follow-up: What happens if state is not reset? Results combine observations from multiple periods, so the reported metric no longer represents only the intended dataset.

keras.metrics.Precision
keras.metrics.Recall
keras.metrics.TruePositives
keras.metrics.TrueNegatives
keras.metrics.FalsePositives
keras.metrics.FalseNegatives
keras.metrics.PrecisionAtRecall

Source: Classification metrics based on True/False positives & negatives. Checked 2026-10-15.