Random Forest all data vs subsamples

rf_subsamples.R

library(randomForest)
library(parallel)

genr_data <- function(n,p) { 
  X <- matrix(rnorm(n*p),n,p)
  y <- as.factor(apply(X,1, function(x) 
    ifelse(sum(x^2)>qchisq(0.5,p),"+","-")))
    ## Hastie etal 10.2
  data.frame(X,y)
}

err_rate <- function(md, d_test) {
  yp <- predict(md, newdata=d_test)
  sum(d_test$y!=yp)/nrow(d_test)
}

n <- 3000
p <- 300

d_train <- genr_data(n,p)
d_test <- genr_data(10000,p)

n_trees <- 500

rf_par <- function(frm, data, n_trees) {
  nproc <- 10  ##detectCores()
  mds <- mclapply(rep(floor(n_trees/nproc),nproc),
            function(nn) randomForest(frm, data=data, ntree = nn),
            mc.cores = nproc)
  md <- do.call("combine", mds)  
}


system.time({
##md <- randomForest(y~., d_train, ntree=n_trees)
md <- rf_par(y~., d_train, n_trees)
})

err_rate(md, d_test)

m <- 10

system.time({
mds_split <- lapply(0:(m-1), function(k) {
  idx <- (1:nrow(d_train)) %% m == k
  ##randomForest(y~., d_train[idx,], ntree=n_trees/m)
  rf_par(y~., d_train[idx,], n_trees/m)
})
})

md_split <- do.call(combine, mds_split)

err_rate(md_split, d_test)