(Conditional) Permutation Importance for Random Forests.

This R-package computes the Conditional Permutation Importance (CPI; Strobl, 2008) using an alternative implementation that is both faster and more stable (Debeer & Strobl 2020). The (C)PI can be computed for random forest fit using (a) the original impurity reduction method ( randomForest-package), and (b) using the Conditional Inference framework (party-package). In addition, a plotting method for the resulting VarImp-object is included.

Installation

The package can be installed using using the devtools-package:

install.packages("devtools")
devtools::install_github("ddebeer/permimp")

Documentation

The workhorse is the permimp-function.

?permimp

For documentation about the plotting function:

?plot.VarImp

Example

library(party)
library(randomForest)
library(permimp)

### set seed
set.seed(542863)

### get example data
airq <- subset(airquality, !(is.na(Ozone) | is.na(Solar.R)))

### fit a random forest 
### ... using the party package
cfAirq5 <- cforest(Ozone ~ ., data = airq,
                    control = cforest_unbiased(mtry = 3, ntree = 1000,
                                               minbucket = 5, 
                                               minsplit = 10))
                                               
### compute the conditional permutation importance
permimp_cf <- permimp(cfAirq5, conditional = TRUE)
plot(permimp_cf, type = "box", interval = "quantile")


### fit a random forest ...
### ... using the randomForest package         
rfAirq5 <- randomForest(Ozone ~ ., data = airq, 
                        mtry = 3, ntree = 1000, importance = TRUE, 
                        keep.forest = TRUE, keep.inbag = TRUE)
                        
### compute the conditional permutation importance                      
permimp_rf <- permimp(rfAirq5, conditional = TRUE)
plot(permimp_rf, horizontal = TRUE)

Parallel Processing

For forests with large trees parallel processing may speed up the computations. Parallel processing is possible via the cl argument. Under the hood, the pblapply function from the pbapply-package.

Tip: when using parallel processing set progressBar = FALSE. The additional communication between the nodes for updating the progress bar will slow down the computations.