ROC-AUC

In one sentence ROC-AUC measures how well a model ranks positive cases above negative ones, across every possible decision threshold at once.

ROC-AUC is the probability that your model scores a randomly chosen positive example higher than a randomly chosen negative one.

Picture a talent judge scoring 100 singers, of whom 20 can genuinely sing. You do not ask whether each score is "correct". You ask: if I line everyone up by score, do the real singers rise to the top? A judge whose top 20 scores belong to the 20 real singers is a perfect ranker. ROC-AUC measures that ranking skill.

The name unpacks like this. Most classifiers output a score, and you pick a threshold to turn scores into yes or no. The ROC curve (receiver operating characteristic) plots the trade-off between catching positives and raising false alarms, at every threshold. AUC (area under the curve) squashes that whole curve into one number.

1.0         perfect ranking
0.8 - 1.0   strong
0.6 - 0.8   modest — some signal, plenty of mixing
0.5         coin flip: the scores carry no signal
below 0.5   rankings are inverted — check your labels

Because it ignores the threshold, ROC-AUC is useful for comparing models before you choose an operating point. Its known weakness: under heavy class-imbalance it can look flattering, and a precision-recall curve gives the more honest view.

Where to go next

Full lesson: Model evaluation
Related terms: precision, recall, accuracy, logistic-regression

Where to go next#

Where to go next