'naivebayes' learner option fails when optimizing hyper-parameters for 'fitcecoc' function.

I am attempting to optimize a multi-class classifier. The classifier uses a (187 x 20) predictor matrix and a (187 x 1) categorical label vector (6 possible categories labeld 1 through 6). I am trying to run the optimization as follows:

Mdl = fitcecoc(predictorMat, labelVec, 'Learners', 'naivebayes',  'OptimizeHyperparameters','all',...
                       'HyperparameterOptimizationOptions',struct('AcquisitionFunctionName',...
                        'expected-improvement-plus', 'MaxObjectiveEvaluations', 10, 'UseParallel', ...
                        true, 'Kfold', 5));

This returns the following error:

Error using classreg.learning.paramoptim.BayesoptInfoCECOC/templateFromLearnersArg (line 127)
Optimizing hyperparameters for fitcecoc with learner type 'naivebayes' is not supported.
Error in classreg.learning.paramoptim.BayesoptInfoCECOC/getWeakLearnerTemplate (line 63)
            Template = templateFromLearnersArg(this, LearnersArg);
Error in classreg.learning.paramoptim.BayesoptInfoCECOC (line 29)
            this.WeakLearnerTemplate = getWeakLearnerTemplate(this, FitFunctionArgs);
Error in classreg.learning.paramoptim.BayesoptInfo.makeBayesoptInfo (line 127)
            Obj = ConstructorFcn(Predictors, Response, FitFunctionArgs);
Error in classreg.learning.paramoptim.fitoptimizing (line 17)
BOInfo = classreg.learning.paramoptim.BayesoptInfo.makeBayesoptInfo(FitFunctionName, Predictors, Response, FitFunctionArgs);
Error in fitcecoc (line 283)
    [obj, OptimResults] = classreg.learning.paramoptim.fitoptimizing('fitcecoc',X,Y,varargin{:}); 

This is odd since the expected hyper-parameter optimization behavior of 'fitceoc' with 'naivebayes' learners is described in the function's docs (see the Hyperparameter Optimization section about 3/4 of the way down that doc: https://www.mathworks.com/help/stats/fitcecoc.html). Moreover, in the code above changing the leaerner to 'svm' (or any of the other learner types such as 'knn' or 'kernel') goes through the optimization as expected. I am trying this on Matlab 2021a. Thanks for any help.

Answer 1

[Don Mathis]

It's true that Naive Bayes is not supported for optimization via fitcecoc. But since Naive Bayes is already a multiclass classifier, you can optimize it by itself:

Mdl = fitcnb(X,Y,  'OptimizeHyperparameters','all',...
'HyperparameterOptimizationOptions',struct('AcquisitionFunctionName',...
'expected-improvement-plus', 'MaxObjectiveEvaluations', 10, 'UseParallel', ...
true, 'Kfold', 5));

[Andres G.]

Thanks for this answer - this solves my problem. Perhaps the documentation for fitcecoc could be updated as it really does seem wrong (or at least confusing) as is. Anyway, thank you!

[Alan Weiss]

Indeed, we are now aware of the documentation problem and will fix it. Sorry for misleading you, and causing you to waste your time.

Alan Weiss

MATLAB mathematical toolbox documentation