Hi

ROC curves are thought for binary problems by definition, since you have to plot the True Positive Fraction (TPF) or also called sensibility versus the False Positive Fraction (FPF) or 1-specificity

In the case of knn, for example, if you have only two classes and you use 62 neighbours (62-nn) the output of your classifier is the number of postiive samples among the 62 nearest neighbours.

In Matlab for example you can find information here:

http://es.mathworks.com/help/stats/perfcurve.html

The function perfcurve must receive:

A vector called labels that contains the label of every sample (values 0 or 1 because it is a binary classification problem).

A vector called scores that is the output of the 62-nn (the number of positive samples among the 62 nearest neighbours).

thank you for your answer.