I am training a multiclass classification model based on SVM using the function fitcecoc with coding design 'allpairs', meaning that binary models are trained for all possible combinations of class pairs.
You can cross-validate this multiclass (ECOC) classifier and estimate its generalization error by for example doing:
Mdl = fitcecoc(X,Y,'Learners',t,...
CVMdl = crossval(Mdl);
oosLoss = kfoldLoss(CVMdl)
In addition to this, would it also be possible to cross-validate and estimate the generalization error for the single binary models?