2

Does someone have a program to compute the Area Under the Roc Curve, a well-known statistics to assess the performance of a binary classifier ? The result is given for some Procs (like the Proc logistic), but it would be also interesting to get it for a.o. “expert models” based on human expertise.

Thank you, CC

flag

1 Answer

3

Toloc,

How are you obtaining the ROC Curve to begin with? As you indicated, some SAS procedures produce the estimate for the area under the ROC Curve as part of their procedures. If you estimate the equation of the ROC curve (perhaps via regression), then you could use some numerical intergration procedure to simply calculate the area under the curve of the regression function. The following SUGI article uses the elementary "trapazoid" rule from calculus to obtain and estimate for the area under a curve: http://www2.sas.com/proceedings/sugi27/p229-27.pdf. I could imagine you could adapt your program to make use of this?

link|flag
Hello MarkLamias, Thank you, this article helps. Indeed, to compute this statistics you need first to fit the Roc Curve (way to compute the Roc Curve: kddcup-orange.com/evaluation.php). I don't have tried to do it as this curve is per se only used to compute this statitics (another curve, the Lorenz Curve gives more visual insights). Does someone already have a way to fit this curve ? Regards – Toloc Nov 18 at 9:02

Your Answer

Not the answer you're looking for? Browse other questions tagged or ask your own question.