Forum: help

RE: Using pre-defined predictor importance in cforest/mob fit? [ Reply ]
By: Torsten Hothorn on 2017-12-01 10:20

Dear Kohleth,

We are about to add this functionality. Pls check out the current partykit version from R-forge and look at the file "extree.R". In line 56, we sample predictor variables to be evaluated with possible weights in variable "partyvars". This feature is not yet exported (ie you can't change partyvars from ctree or mob) but you could tweak ctree if you want to play around with it a little.

Best,

Torsten

Using pre-defined predictor importance in cforest/mob fit? [ Reply ]
By: Kohleth Chia on 2017-10-13 22:51

[forum:45386]

Dear all,

I recently came across of a class of random forest called 'Enriched Random Forest' which assigns predictor importance prior to fitting, and use that to aid the model training process. (http://www.rci.rutgers.edu/~cabrera/DNAMR/)

This is a very useful feature for building interpretable model. For example, I have a clinical application where I am building a random forest to predict the pre/absence of certain impairments. I used ~140 predictors, but most clinician will only recognize ~20 as being clinically relevant. I still want to use all 140 predictors, but would like the split in the trees to prioritize the 20 that are pre-marked as important variables.

I am guessing the partykit/mob framework is my best bet at implementing this. Can anyone here provide some guidance as to whether this is possible at with partykit? if so, where to start?

Thanks.