Just early this month, an R package "Partykit" was released.  It is an implementation of the random forest and bagging ensemble algorithms utilizing conditional inference trees (ctree) as base learners.

Also, I posted the link to an article.

Good luck.

http://partykit.r-forge.r-project.org/partykit/

http://jmlr.org/papers/volume16/hothorn15a/hothorn15a.pdf

The short answer is yes, it is possible. You can do exactly as you suggest and run the algorithms sequentially, first performing feature selection and then running an additional algorithm to look for interactions, or you can use a method developed more specifically for that purpose.

I don't know many papers off the top of my head, but Gradient Boosted Feature Selection (link follows) is a method of doing what you are asking about, as part of a single method.

I know that the randomForestSRC package for R also offers a method for finding interactions after performing feature selection.

http://alicezheng.org/papers/gbfs.pdf

Dear Manhal,

I understand your question in term of is it possible to check Multicollinearity, which is certainly different than variable section. If it is variable section boosting or RF will do the the job.

Now if you are with Multicollinearity that is on other question as if you will not know what variable to remove, which one really has effect, with RF you do not know as the variable are select randomly during the split... with boosting the effect will be different the shrinkage will happen, so you can leave then !

In few words RF not so good for Multicollinearity boosting behave better, if RF you should remove all the Multicollinearity variables. 

If you are speaking about non Multicollinearity variables and only reduction then forget what I wrote !!

Still hope it give you few ideas about some problem you could face :)

Alain