AUC score
The area under the ROC curve, abbreviated AUC or AUROC, summarizes the overall performance of a binary classifier in a single number ranging from zero to o...
A Single Number Summary of Classifier Quality
The area under the ROC curve, abbreviated AUC or AUROC, summarizes the overall performance of a binary classifier in a single number ranging from zero to one. An AUC of one corresponds to a perfect classifier. An AUC of 0.5 corresponds to a random classifier with no discriminative power. In practice, useful classifiers typically fall between 0.7 and 0.99 depending on the difficulty of the task.
The AUC has an elegant probabilistic interpretation: it equals the probability that the model assigns a higher predicted score to a randomly chosen positive example than to a randomly chosen negative example. This interpretation is independent of any threshold decision and makes AUC directly meaningful even when the optimal operating threshold is not yet known.
Because AUC aggregates performance across all thresholds, it is particularly useful for comparing two classifiers without committing to a specific threshold. If classifier A has a higher AUC than classifier B, it tends to rank positive examples above negative examples more reliably regardless of where you draw the threshold.
AUC is also relatively robust to class imbalance compared with accuracy, because it measures ranking quality rather than counting correct predictions at a single threshold. However, for very imbalanced problems where the positive class is extremely rare, the precision-recall AUC, sometimes called average precision, is an even more informative metric because it focuses attention on the positive class.
A concrete example: a credit card fraud detector trained on heavily imbalanced data achieves an accuracy of 99.5 percent by predicting no fraud ever. Its AUC is only 0.61, revealing that the model can barely distinguish fraudulent from legitimate transactions. A better model achieves 0.94 AUC, indicating strong discriminative ability.
Exam tip: when asked how to evaluate a classifier on imbalanced data, first mention AUC as a threshold-independent metric, then add that precision-recall AUC may be preferable when the positive class is extremely rare.