增加样本数据的大小 - R.

Question

One of my colleagues indicated that randomForest() does not perform well with very large data sets. Now, I am just trying to figure out if that really is the case, but since the data set cannot be shared (sensitive information), I thought I might as well try to come up with a large data set. I have tried following, but cannot make sense of the error message:

library(randomForest)
data(iris)
dataFile <- iris
newdataFile <- dataFile[sample(dataFile, size= 1:1000000000, replace=T),]

Error message:

Error in xj[i] : invalid subscript type 'list'

Can anyone please guide me here ?

Answer 1

sample accepts a vector. When sampling from a data.frame, one usually samples the rows by referring to them as a number, much akin to subsetting but in this case, with replacement.

newdataFile <- iris[sample(nrow(iris),100000,replace=T),]

Answer 2

The assertion that Random Forests does not perform well with large datasets is absurd. It is notably well suited to high dimensional problems both from a sample size and multivariate standpoint. The primary issues with RF and very large problems are: 1) tractability and 2) sample balance.

If you have a problem where one class is proportionally larger (>30%) then the bootstrap can be biased and the OOB validation, and possibly the estimate, is incorrect. The result, of say a binary problem with [0=10000,1=200], would be a very high prediction rate to 0 and very low to 1 resulting in a very good, but quite inflated, OOB error rate for the model but very poor performance for class 1.

This is obviously not representative of the model performance and you will have very low prediction prevalence for class 1. If you have a class balance issue I would follow the methodologies in either Chen et. al., (2004) or Evans & Cushman (2009).

Chen C, Liaw A, Breiman L (2004) Using random forest to learn imbalanced data. http://www.stat.berkeley.edu/tech-reports/666.pdf

Evans, JS and SA Cushman (2009) Gradient Modeling of Conifer Species Using Random Forests. Landscape Ecology 5:673-683.