Why is the cutoff for AUC (area under ROC curve) to be considered an acceptable classfier at .7 and what is it acceptable for?

Question

From reading about interpreting the area under a ROC curve, it seems the general consensus is that anything from .7 - .8 is considered an acceptable classifier. What is it considered "acceptable" for? It seems like a quite vague interpretation. I was also wondering why anything above .7 is considered good. For example, some classifiers need high accuracy to be used (like using face recognition as a password on phones) while others don't need nearly that high of accuracy.

Answer 1

The area under an ROC curve

is equal to the probability that a classifier will rank a randomly chosen positive instance higher than a randomly chosen negative one (assuming 'positive' ranks higher than 'negative').

So if a 70% or 80% probability of getting the case ordering correct is good enough, then those AUC values are "acceptable." As you note, whether that's actually good enough depends on the field of application. My guess is that whatever you read that led you to your sense of a "consensus" about the quality of those AUC values might have been based more on the difficulty of getting models in practice to do much better than that, at least in fields like clinical medicine.