Variable importance between classes with randomforest

I'm using the randomforest package to classify my dataset of continuous variables grouped into four categories (a, b, c, d). Using the varImp() and importance() functions I can get the class-specific and overall permutation variable importance. I would like to find out which of my variables provide the greatest separation between two and two groups - i.e. which variables have the highest variable importance for separating between group a and group c. I've come across the varImpGroup() function in RFgroove, but this seems to be directed at variable importance for a subset of variables instead of groups.

Is there a way to produce the variables importance for pairwise comparisons of groups, without having to run the classification separately for only two groups at a time?

Update: For anyone wondering, I contacted Andy Liaw who manages the randomforest package. He suggested using the localImp = TRUE argument to produce a matrix of casewise variable importance, and average the values across all the individuals composing the selected groups to determine between-group variable importance.

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.

If you have a query related to it or one of the replies, start a new topic and refer back with a link.