AUC is simply the portion of the ROC chart "square" that is underneath the ROC curve of the ML. Since it is a proportion, it will always range from 0 to 1. An AUC of 1 represents a curve that covers the entire graph, and therefore it is perfectly aligned to the upper-left corner. This would be the aforementioned rarely observed, yeti-like perfect model.

The random model represented by the dashed line in the ROC chart has an AUC of 0.5. So, in real world situations, you will be looking at AUC numbers from 0.5 to 1.0. How to interpret an AUC number depends a lot on the industry and business situation. An AUC of 0.6, indicating better than random, may be a fantastic model for stock picking in the finance industry; while the same AUC measure would be an utter failure for cancer detection in the medical industry.

AUC numbers are a useful way to compare models using a single number instead of chart. This comparison could be between different models, or on a model against itself after it has been tweaked or retrained. However, there is a danger when only looking at the AUC value as you are losing some valuable information.

An AUC number gives you a rough idea of performance but tells you little about the actual shape of the curve. You lose the ability to determine the performance of a model at different threshold settings. At more conservative settings, an ML model with a lower AUC value may actually perform better than another model with a higher AUC.

AUC values are useful shorthand for doing quick comparisons and for boiling down an ROC curve to one number. But you should always review the full ROC curve before making any business decisions, especially if the AUC values are close.

AUC values can be quickly generated in R using the pROC package also. Building on the prior ROC code example, the AUC values can be obtained by using the auc function:

auc(rocCurve)