R - SVM (radial) regression using tune() (e1071) - How to uncover influential features?

Question

I have some questions regarding SVM and regression.

1) How to interpret SVM (regression) results.

2) How to make a proper plot (containing decent information)

e1071's tune() is used to uncover the best cost (C) and gamma (y) parameters. The model is fitted against a training set (train2).

svm.radial = tune( svm , dep_sev_fu ~ . , data = train2 , kernel = "radial" , type = "eps-regression" , ranges = list( cost = c(0.001 , 0.01 , 0.1 , 1 , 5 , 10 , 50 ) , gamma = c( .0001 , .001 , .01 , .1 , 1 , 5 , 10 ) ) )

Optimal model is fitted against test set:

svm.tuned <- svm( dep_sev_fu ~ . , data = test , kernel = "radial" , type = "eps-regression" , ranges = list( cost = 1 , gamma = .01 , epsilon = .1 ) )

In order to extract (significant) regression weights, there's a function called 'rfe' within caret that applies backward selection. But it is unclear how to specify a model when using SVR.

I looked at a kernlab example, where they mention the W vector. After running w = t(svm.tuned$coefs) %*% svm.tuned$SV a vector of coefficients emerge. But how do I know if they are significant weights?

Note: Do I need to predict() of some sort? Not really sure how validation should work at this step

Bonus

Expanding on the validation approach, how can I fit using 10-fold CV?

* Example of the first 10 cases & 8 variables *

disTypecomorbid.disorder disTypedepressive.disorder Sexemale Age aedu IDS BAI FQ 1 0 0 50 10 6 10 4 0 1 0 35 11 7 5 2 0 0 0 51 15 4 3 14 0 1 0 43 15 11 7 3 0 0 0 38 10 7 8 15 1 0 0 38 10 15 15 32 0 0 0 45 9 12 12 2 0 0 0 57 9 9 14 4 1 0 0 43 10 12 3 0 1 0 1 49 11 14 11 3

Answer 1

You can't really interpret SVM models unless a linear kernel is used (you have the RBF in your code above). The matrix multiplication is only for the linear case too (I believe).

There is an excessive amount of documentation on rfe at the package website. Look there for syntax.