Random forest bootstrap training and forest generation

问题

I have a huge training data for random forest (dim: 47600811*9). I want to take multiple (let's say 1000) bootstrapped sample of dimension 10000*9 (taking 9000 negative class and 1000 positive class datapoints in each run) and iteratively generate trees for all of them and then combine all those trees into 1 forest. A rough idea of required code is given below. Can anbody guide me how can I generate random sample with replacement from my actual trainData and optimally generate trees for them iteratively? It will be great help. Thanks

library(doSNOW)
library(randomForest)
cl <- makeCluster(8)
registerDoSNOW(cl)

for (i=1:1000){
B <- 1000 
U <- 9000 
dataB <- trainData[sample(which(trainData$class == "B"), B,replace=TRUE),] 
dataU <- trainData[sample(which(trainData$class == "U"), U,replace=TRUE),] 
subset <- rbind(dataB, dataU)

I am not sure if it is the optimal way of producing a subset again and again (1000 times) from actual trainData.

rf <- foreach(ntree=rep(125, 8), .packages='randomForest') %dopar% {
  randomForest(subset[,-1], subset$class, ntree=ntree)
}
}
crf <- do.call('combine', rf)
print(crf)
stopCluster(cl)

回答1:

Although your example parallelizes the inner rather than the outer loop, it may work reasonably well as long as the inner foreach loop takes more than a few seconds to execute, which it almost certainly does. However, your program does have a bug: it is throwing away the first 999 foreach results and only processing the last result. To fix this, you could preallocate a list of length 1000*8 and assign the results from foreach into it on each iteration of the outer for loop. For example:

library(doSNOW)
library(randomForest)
trainData <- data.frame(a=rnorm(20), b=rnorm(20),
                        class=c(rep("U", 10), rep("B", 10)))
n <- 1000         # outer loop count
chunksize <- 125  # value of ntree used in inner loop
nw <- 8           # number of cluster workers
cl <- makeCluster(nw)
registerDoSNOW(cl)
rf <- vector('list', n * nw)
for (i in 1:n) {
  B <- 1000
  U <- 9000
  dataB <- trainData[sample(which(trainData$class == "B"), B,replace=TRUE),]
  dataU <- trainData[sample(which(trainData$class == "U"), U,replace=TRUE),]
  subset <- rbind(dataB, dataU)
  ix <- seq((i-1) * nw + 1, i * nw)
  rf[ix] <- foreach(ntree=rep(chunksize, nw),
                    .packages='randomForest') %dopar% {
    randomForest(subset[,-1], subset$class, ntree=ntree)
  }
}
cat(sprintf("# models: %d; expected # models: %d\n", length(rf), n * nw))
cat(sprintf("expected total # trees: %d\n", n * nw * chunksize))
crf <- do.call('combine', rf)
print(crf)

This should fix the problem that you mention in the comment that you directed to me.

回答2:

Something like this would work

# Replicate expression 1000 times, store output of each replication in a list
# Find indices of class B and sample 9000 times with replacement
# Do the same 1000 times for class U. Combine the two vectors of indices

i = replicate(1000, {c(sample(which(trainData$class == "B"), 9000, replace = T), sample(which(trainData$class == "U"), 1000, replace = T))})

Then feed i into a parallel version of lapply

mclapply(i, function(i, ntree) randomForest(trainData[i,-1], trainData[i,]$class, ntree=ntree)

来源：https://stackoverflow.com/questions/39494010/random-forest-bootstrap-training-and-forest-generation