Example/srbct_cv.R

In robingenuer/VSURF: Variable Selection Using Random Forests

# This script is an example of the use of an external cross-validation around VSURF procedure, to estimate prediction errors of random forests involving only variables selected from VSURF.
# WARNING: lots of calculations are done. It takes almost 20 minutes on a linux computing server with 40 cores
library(VSURF)
data(srbct, package = "mixOmics")
set.seed(322, "L'Ecuyer-CMRG")

K <- 5
x <- srbct$gene
y <- srbct$class
n <- length(y)
folds <- replicate(ceiling(n / K), sample(1:K))[1:n]

errtest.mat <- matrix(nrow = K, ncol = 2)
colnames(errtest.mat) <- c("interp", "pred")
res.cv <- vector("list", K)

for (k in 1:K) {
  xtrain <- x[-which(folds == k),]
  ytrain <- y[-which(folds == k)]
  xtest <- x[which(folds == k),]
  ytest <- y[which(folds == k)]
  
  vsurf.fold <-
    VSURF(
      xtrain, ytrain, parallel = TRUE, clusterType = "FORK", ncores = 40
    )
  errtest.mat[k, 1] <-
    sum(ytest != predict(vsurf.fold, newdata = xtest, step = "interp"))
  errtest.mat[k, 2] <-
    sum(ytest != predict(vsurf.fold, newdata = xtest, step = "pred"))
  res.cv[[k]] <- vsurf.fold
}

errtest <- colSums(errtest.mat) / n
errtest