Feature Importance in (Caret) package

Solution 1:

Per the varImp() documentation, the scale argument in the caret::varImp() function scales the variable importance values from 0 to 100.

Absent a reproducible example, we'll use the vowel data from the Elements of Statistical Learning book to generate a random forest, and rescale the variable importance data so the sum is equal to 1 by dividing each variable importance number by the sum of all importance numbers.

library(readr)
vowel.train <- subset(read_csv("https://web.stanford.edu/~hastie/ElemStatLearn/datasets/vowel.train"),
                      select = -row.names) 
vowel.test <- subset(read_csv("https://web.stanford.edu/~hastie/ElemStatLearn/datasets/vowel.test"),
                     select = -row.names)
library(caret)
library(randomForest)
vowel.train$y <- as.factor(vowel.train$y)
vowel.test$y <- as.factor(vowel.test$y)
set.seed(33833)
tr1Control <- trainControl(method="boot")
modFit <- train(y ~ .,method="rf",trControl=tr1Control,data=vowel.train)
# Variable Importance: caret function, extract importance data frame & rescale
v <- varImp(modFit,scale = TRUE)[["importance"]]
v$Overall <- v$Overall / sum(v$Overall)
v

..and the output:

> v
         Overall
x.1  0.318660495
x.2  0.327734091
x.3  0.018931795
x.4  0.021533916
x.5  0.126744531
x.6  0.089627688
x.7  0.000000000
x.8  0.067066743
x.9  0.027072197
x.10 0.002628545

...and to demonstrate that sum(v$Overall) is now 1: