E = ROC(A,W,C,N)
Computes N points on the receiver-operator curve of the classifier W for class C in the labeled dataset B, which is typically the result of B = A*W; or for the dataset A labelled by applying the (cell array of) trained classifiers W.
Note that a Receiver-Operator Curve is related to a specific class (class C) for which the errors are plotted horizontally. The total error on all other classes is plotted vertically. The class index C refers to its position in the label list of the dataset (A or B). It can be found by GETCLASSI.
The curve is computed for N thresholds of the posteriori probabilities stored in B. The resulting error frequencies for the two classes are stored in the structure E. E.XVALUES contains the errors in the first class, E.ERROR contains the errors in the second class. In multi-class problems these are the mean values in a single class, respectively the mean values in all other classes. This may not be very useful, but not much more can be done as for multi-class cases the ROC is equivalent to a multi-dimensional surface.
Use PLOTE(E) for plotting the result. In the plot the two types of error are annotated as 'Error I' (error of the first kind) and 'Error II' (error of the second kind). All error estimates are weighted according the class prior probabilities. Remove the priors in A or B (by setprior(A,)) to produce a vanilla ROC.
Train set A and test set T:
1. R.O. Duda, P.E. Hart, and D.G. Stork, Pattern classification, 2nd edition, John Wiley and Sons, New York, 2001.