A simple question about cross-validation

Question

I want to use SVM using 10-fold cross-validation (CV). My code is as follows:

svm.model <- svm(train.d, y=train.l,
                 type='C-classification',
                 cross=kfold,
                 kernel=kernel)

I know that in k-fold CV, k different models are trained. Which of those k models is the final svm.model? Is it the trained model with maximum accuracy?

Kirill Dubovikov · Accepted Answer · 2016-02-22T21:56:07.310

2

Which library do you use? Cross-validation is not a model selection method by itself, it's rather a way to estimate mean square error (for regression) or accuracy rate (for classification). It is likely that the model you get is the SVM fitted for the data specified in the parameters, no specific selection should be done on behalf of CV.

Cross-validation allows you to fit the model to all available data and still be informed about it's performance.

edited Feb 22 '16 at 21:56

answered Feb 22 '16 at 21:38

Kirill Dubovikov

164
1
3

1

I am using 'e1071' R-package. You mean the output model is an independent model trained by all of training data? – WhoAmI Feb 22 '16 at 21:54
1

Yes, I mean exactly that. – Kirill Dubovikov Feb 22 '16 at 21:56
1

CV can estimate any out-of-sample metric, not just accuracy and RMSE (even though these are often default choices). – Sycorax Feb 22 '16 at 22:37

A simple question about cross-validation

1 Answers1