Plotting ROC for fitcecoc svm classifier

Hi I have created a 4 level SVM classifier by fitcecoc. I need to generate ROC curve for each class. This is the code: template = templateSVM('KernelFunction', 'gaussian', 'PolynomialOrder', [], ... 'KernelScale', 1, 'BoxConstraint', 1, 'Standardize', true);

[classificationSVM,HyperparameterOptimizationResults] = fitcecoc(predictors... , response, 'Learners', template, 'Coding', 'onevsone', 'OptimizeHyperparameters',... {'BoxConstraint','KernelScale'},'HyperparameterOptimizationOptions',... struct('AcquisitionFunctionName', 'expected-improvement-plus',... 'Repartition',true,'MaxObjectiveEvaluations',10,'kfold',5),'Verbose',0,... 'FitPosterior',1);

[label,NegLoss,PBScore,Posterior] = resubPredict(classificationSVM);

Apparently, I need to use perfcurve function to get the ROC. Yet, the help instructions are for fitcsvm that does not work for fitcecoc. I am trying to use the following code for class 1: [Xsvm,Ysvm,Tsvm,AUCsvm] = perfcurve(response,Posterior(:,1),'true'); Yet, I receive the below error message: Error using perfcurve>membership (line 693) Positive class is not found in the input data.

Error in perfcurve (line 437) [W,subYnames] = membership(cls(sorted),weights(sorted),... I was wondering how I can proceed from here.

Thanks, Roohollah

[SI LIU]

Undefined function 'Posterior' for input arguments of type 'ClassificationECOC'.

Answer 1

[Prashant Arora]

Hi Roohollah,

The reason you are getting this error because you defined the positive class to be "true" in the perfcurve function. This only work if the labels are logical values, i.e. can be defined as "true" or "False". If your labels are not logical values, you must define the Positive class as one of the member of the labels. You can find more information about this at the following link:

https://www.mathworks.com/help/stats/perfcurve.html?s_tid=srchtitle#inputarg_posclass

Hope this helps!

Prashant

[Roohollah Milimonfared]

Thanks Prashant. I was wondering if there is any standard procedure to generate ROC plots for multi-level classifiers. I am using one-vs-one coding design for the binary learners. Does that mean I have to use purcurve function for each binary learner separately? Thanks.

[satinder nagra]

may be, you are correct