fitcecoc with specified Box Contraint and Kernel Scale

Hello,

I want to create an SVM model for multiclass classification, with specified hyperparameters: BoxContraint = 500 and KernelScale = 0.01

I triend using:

cvpt = cvpartition(Y_train, "kFold", 5);
fitcecoc(X_train,"class",  "CVPartition",cvpt,  "Coding","onevsone",  "BoxContraint", 500,  "KernelScale", 0.01)

I get the following error:

Error in classreg.learning.FitTemplate.make (line 124)

temp = fillIfNeeded(temp,type);

Error in classreg.learning.FitTemplate/fillIfNeeded (line 448)

classreg.learning.FitTemplate.make(this.Method,'type',this.Type,...

Error in classreg.learning.FitTemplate.make (line 124)

temp = fillIfNeeded(temp,type);

Error in ClassificationECOC.template (line 111)

temp = classreg.learning.FitTemplate.make('ECOC','type','classification',varargin{:});

Error in ClassificationECOC.fit (line 115)

temp = ClassificationECOC.template(varargin{:});

Error in fitcecoc (line 339)

obj = ClassificationECOC.fit(X,Y,ecocArgs{:});

Answer 1

[Drew]

The error occurs because BoxConstraint and KernelScale are NOT name-value pairs of fitcecoc. They are properties of the binary learners (in this case, SVMs) that fitcecoc uses under the hood.

"fitcecoc" itself only orchestrates the ECOC (error-correcting output codes) reduction of a multiclass problem to a collection of binary (two-class) classification problems. For example, one SVM per binary problem in the default one-vs-one scheme.

So, BoxConstraint and KernelScale have to be set in the SVM learner via templateSVM, and that templateSVM is then handed to fitcecoc through the "Learners" name-value argument. Here is an example:

% The templateSVM captures every SVM-specific setting (kernel function, box
% constraint, kernel scale, standardization, etc.). fitcecoc then clones
% this template for each binary sub-problem.
t = templateSVM( ...
    KernelFunction = "linear", ...
    BoxConstraint  = 0.5, ...       
    KernelScale    = "auto", ...   
    Standardize    = false);

mdl = fitcecoc(X, Y, Learners = t, ClassNames = unique(Y));

See the documentation page https://www.mathworks.com/help/stats/fitcecoc.html for more information. Note that fitcecoc also supports other types of binary learners.