Unable to use "var.select" method of randomForestSRC

[littleisland8]

Hello.

I would like to use Random forest minimal depth (method="md" of var.select method of randomForestSRC ) and random forest variable hunting (method="vh" of var.select method of randomForestSRC) approaches to perform feature selection on a survTask.
I load the learner with:
learner = lrn("surv.rfsrc", importance="permute")
When I try to run flt("importance", learner = learner) I am able to get the feature score based on permutation score as the learner got importance="permute" as parameter. However, I cannot use var.select method to perform minimal depth and variable hunting because in the mlr_filters dictionary I cannot find something related to var.select of randomForestSRC.

How can I use var.select into the mlr3verse?

[bblodfon]

Hi @littleisland8,

As you've probably seen, for the flt("importance", learner = learner) to work, learner$importance() needs to return the importance scores of the features (usually from a default method from the package that implements the learner). randomForestSRC::var.select seems more like a more elaborate/extra analysis variable selection method with a lot of parameters, and maybe more suited to variable selection.

If you really want to and can get some sort of importance scores from such a function, then I would suggest extending the respective learner (code line) and override the importance() method by using inside var.select. Then using that new augmented learner with mlr3filters::flt("importance", ...) would work.