At the very first glance, AUC ROC can look pretty ugly. But it turns out that is quite simple and intuitive! Let’s take a look.
Firstly, True Positive Rate (TPR) describes how many of the actual positives are predicted as positives. This is also called as Recall or Sensitivity.
False Positive Rate (FPR) describes how many of the actual negatives are predicted as positives.
The classes are predicted based on the probability threshold, which by default is 0.5. Increasing this threshold would mean more number of instances are predicted as negative, which implies a lower TPR.
In essence, TPR and FPR depend on this threshold (domain dependent), which can be any value between 0 and 1. So, if we plot all the TPR and FPR for every value of the threshold, we get the Receiver Operator Characteristics ROC curve!
And the area under this curve, AUC ROC, gives us the measure of separability of classes by the model.
If AUC=1, it means that the model is predicting every TP and TN correctly. The ideal model!
If AUC=0.5, it means that the model is giving purely random predictions.
And if AUC=0, it means that the model is predicting completely opposite of what is desired. The worst model… Or another ideal model…??
#datascience #machinelearning