R: Caret rfe with a customized variable importance permimp - r

Question

I want to run a recursive feature elimination with caret rfe() with the alternative variable importance algorithm permimp. The permimp permutation importance uses cforest with cforest_unbiased(). Which other caret functions do I need to customize in order run rfe with permimp() and cforest?
This is my code so far:
library(caret)
permimpRFE <- list(summary = defaultSummary,
fit = function(x, y, first, last, ...){
library(party)
tmp <- as.data.frame(x, stringsAsFactors = TRUE)
tmp$y <- y
party::cforest(y ~ ., data = tmp,
control = party::cforest_unbiased(ntree = 50))
},
pred = function(object, x) predict(object, x),
rank = function(object, x, y) {
library(permimp)
vimp <- permimp::permimp(object, conditional = TRUE, threshold = .95, do_check = FALSE)
vimp <- as.data.frame(vimp$values)
colnames(vimp) <- "Overall"
vimp <- vimp[order(vimp$Overall, decreasing = TRUE),, drop = FALSE]
if (ncol(x) == 1) {
vimp$var <- colnames(x)
} else vimp$var <- rownames(vimp)
vimp
},
selectSize = pickSizeBest,
selectVar = pickVars)
# specify rfeControl
contr <- caret::rfeControl(functions=permimpRFE, method="repeatedcv", number=3, repeats=2,
saveDetails = TRUE)
dat <- as.data.frame(ChickWeight)[1:50,]
preds <- dat[,2:4]
response <- dat[,1]
# recursive feature elimination caret (Algorithm 2)
set.seed(43, kind = "Mersenne-Twister", normal.kind = "Inversion")
results <- caret::rfe(x = preds,
y = response,
sizes=c(1:3),
metric= "RMSE",
rfeControl=contr)
I get the error Error in { : task 1 failed - "invalid 'x' type in 'x && y'"
How can I get the rfe running with permimp and cforest?

Answer 1

When trying to customize the variable importance function for rfe(), it is important to implement the exact syntax of the desired function as well as its dependent functions.
In my case, I had to change predict(object, x) to predict(object, newdata = x).
Now the code snippet is working:
library(caret)
permimpRFE <- list(summary = defaultSummary,
fit = function(x, y, first, last, ...){
library(party)
tmp <- as.data.frame(x, stringsAsFactors = TRUE)
tmp$y <- y
party::cforest(y ~ ., data = tmp,
control = party::cforest_unbiased(ntree = 50))
},
pred = function(object, x) {
x <- as.data.frame(x, stringsAsFactors = TRUE)
predict(object, newdata = x)
},
rank = function(object, x, y) {
library(permimp)
vimp <- permimp::permimp(object, conditional = TRUE, threshold = .95, do_check = FALSE)
vimp <- as.data.frame(vimp$values)
colnames(vimp) <- "Overall"
vimp <- vimp[order(vimp$Overall, decreasing = TRUE),, drop = FALSE]
if (ncol(x) == 1) {
vimp$var <- colnames(x)
} else vimp$var <- rownames(vimp)
vimp
},
selectSize = pickSizeBest,
selectVar = pickVars)
# specify rfeControl
contr <- caret::rfeControl(functions=permimpRFE, method="repeatedcv", number=3, repeats=2,
saveDetails = TRUE)
dat <- as.data.frame(ChickWeight)[1:50,]
preds <- dat[,2:4]
response <- dat[,1]
# recursive feature elimination caret (Algorithm 2)
set.seed(43, kind = "Mersenne-Twister", normal.kind = "Inversion")
results <- caret::rfe(x = preds,
y = response,
sizes=c(1:3),
metric= "RMSE",
rfeControl=contr)