Help req. in using fitcsvm()

Hi!

I am trying to use fitcsvm() to implement SVM. Previously, I was using LibSVM. I know from the results obtained using LibSVM that the best kernel for my problem is RBF. Now, I want to find the kernel parameters. For this, I am using the following code:

opts=struct('Optimizer','bayesopt','ShowPlots',true, 'Repartition',1);
svmmod=fitcsvm(ftTrn,CLTrn,'KernelFunction','rbf','OutlierFraction',0.05,...
  'OptimizeHyperparameters','auto','HyperparameterOptimizationOptions',opts);
                     % ftTrn: Training data, %CLTrn: corresponding classlabels

1) Is this code right for my purpose?

2) svmmod contains the SVM trained on the entire training data or on a subset (on a fold used for determining the best values for the kernel parameters)?

3) Are there any other parameters I can tweak for improving the classification performance?

Answer 1

[Don Mathis]

(1) Yes that's right. In that case it will optimize BoxConstraint and KernelScale.

(2) svmmod contains the SVM trained on the entire training data, using the best hyperparameters found. 5-fold crossvalidated misclassification rate was used as the objective function during optimization.

(3) You can optimize more variables. You can find out what hyperparameters are eligible like this:

>> h = hyperparameters('fitcsvm',ftTrn,CLTrn)
>> h.Name
h = 
  5×1 optimizableVariable array with properties:

      Name
      Range
      Type
      Transform
      Optimize

ans =
    'BoxConstraint'
ans =
    'KernelScale'
ans =
    'KernelFunction'
ans =
    'PolynomialOrder'
ans =
    'Standardize'

And then you can optimize additional hyperparameters like this:

svmmod=fitcsvm(ftTrn,CLTrn,'KernelFunction','rbf','OutlierFraction',0.05,...
  'OptimizeHyperparameters',{'BoxConstraint','KernelScale','Standardize'},'HyperparameterOptimizationOptions',opts)

Because you're fixing the kernel function, the 'PolynomialOrder' hyperparameter is not relevant. So 'Standardize' ends up being the only additional hyperparameter.

One more note: Since you're now optimizing 3 variables, you might want to run the optimization longer, say 60 evaluations:

opts=struct('Optimizer','bayesopt','ShowPlots',true, 'Repartition',1, 'MaxObjectiveEvaluations',60);
svmmod=fitcsvm(ftTrn,CLTrn,'KernelFunction','rbf','OutlierFraction',0.05,...
  'OptimizeHyperparameters',{'BoxConstraint','KernelScale','Standardize'},'HyperparameterOptimizationOptions',opts)

[Don Mathis]

Yet one more note: If you've got some time on your hands, why not let it try other kernel functions, too?

opts=struct('Optimizer','bayesopt','ShowPlots',true, 'Repartition',1, 'MaxObjectiveEvaluations',60);
svmmod=fitcsvm(ftTrn,CLTrn,'OutlierFraction',0.05,...
  'OptimizeHyperparameters','all','HyperparameterOptimizationOptions',opts)

[UserJ]

Thanks a lot Don