Receiver Operating Characteristics: ROC
#Machine Learning #Performance #Basics
ROC space is the twodimensional space spanned by True Positive Rate and False Positive Rate.
AUC: Area under Curve
 TPR = TP Rate
 FPR = FP Rate
The ROC curve is defined by the relation $f(TPR, FPR)$. Area under the ROC curve is
$$ \int TPR(FPR) d(FPR) \sim \sum_i TPR_i *\Delta FPR. $$
If AUC = 1, we have TP Rate = 1 for all FP Rate. This is the best performance a model could have.
How to Calculate ROC Curve
Not every model has an AUC. To get the AUC curve, we need a hyperparameter to be tuned to get different TP Rate and FP Rate.
In logistic regression, a threshold $T$ is predetermined to decide which label to use in classifications.
By tuning the threshold $T$, we get different TP Rate $TPR$ and FP Rate $FPR$, i.e., $TPR(T)$ and $FPR(T)$. The parametric relations between $TPR(T)$ and $FPR(T)$ forms the ROC curve.
L Ma (2020). 'Receiver Operating Characteristics: ROC', Datumorphism, 05 April. Available at: https://datumorphism.leima.is/wiki/machinelearning/performance/roc/.
Table of Contents
Current Ref:

wiki/machinelearning/performance/roc.md