interpretation of svm models with the e1071 package

Dear all,

after having calibrated a svm model through the svm() command of the e1071
package, is there a way to
i) represent the modeled relationships between the y and X variables
(response variable vs. predictors)?

ii) rank the influence of the predictors used in the model?

Hi,

On Fri, Jul 9, 2010 at 12:15 PM, manuel.martin
<manuel.martin at orleans.inra.fr>  wrote:
   

Dear all,

after having calibrated a svm model through the svm() command of the e1071
package, is there a way to
i) represent the modeled relationships between the y and X variables
(response variable vs. predictors)?
     

Can you explain a bit more ... how do you want them represented?

   

ii) rank the influence of the predictors used in the model?
     

One technique that's often/sometimes used is to calculate the SVM's W
vector by using the support vectors along with their learned
weights/alphas.

This comes up every now and again. Here's an older post explaining how
you might do that with the svm model from e1071:

http://article.gmane.org/gmane.comp.lang.r.general/158272/match=w+b+vector+svr

Hope that helps.

Hi,

On Fri, Jul 9, 2010 at 12:15 PM, manuel.martin
<manuel.martin at orleans.inra.fr>  wrote:
   

Dear all,

after having calibrated a svm model through the svm() command of the e1071
package, is there a way to
i) represent the modeled relationships between the y and X variables
(response variable vs. predictors)?
     

Can you explain a bit more ... how do you want them represented?
   

ii) rank the influence of the predictors used in the model?
     

One technique that's often/sometimes used is to calculate the SVM's W
vector by using the support vectors along with their learned
weights/alphas.

This comes up every now and again. Here's an older post explaining how
you might do that with the svm model from e1071:

http://article.gmane.org/gmane.comp.lang.r.general/158272/match=w+b+vector+svr

Hope that helps.

   

Steve,

Couldn't he also just use the decision.value property to see the equivilent
of t(x) %*% b for each row?

On 07/10/2010 04:11 AM, Steve Lianoglou wrote:

On Fri, Jul 9, 2010 at 12:15 PM, manuel.martin
<manuel.martin at orleans.inra.fr> ?wrote:

Dear all,

after having calibrated a svm model through the svm() command of the
e1071
package, is there a way to
i) represent the modeled relationships between the y and X variables
(response variable vs. predictors)?

Can you explain a bit more ... how do you want them represented?

I was thinking to a simple ? = fi(Xi) plot, fi resulting from the fitted svm
model. Xi is the predictor, among the whole set of predictors, X, one wish
to see the relationship with the response.
For boosted regression trees, which I am more familiar with, this is fi
function is estimated by averaging the effects of all predictors but Xi, and
plotting how ? varies as Xi does.