CODEWITHSUNDEEP
rrandom-forest
Zelazny7
Zelazny7

Reputation: 40628

nodesize parameter ignored in randomForest package

Does the randomForest package ignore the nodesize parameter? When I predict the terminal nodes for a dataset and check the counts, I see values that are less than the nodesize. I would submit a fix for this myself but the underlying code was written in Fortran. If someone can confirm this behavior I will reach out to the package maintainer and hopefully start a fix.

> library(randomForest)
> set.seed(1)
> rf <- randomForest(mtcars[,-1], mtcars[,1], nodesize = 5)
> nodes <- attr(predict(rf, mtcars[,-1], nodes = TRUE), 'nodes')

# node counts of first tree
> table(nodes[,1])

# first row is the terminal node ID#, second row is the count
 2  6  9 10 11 14 15 16 18 19 
 5  3  3  6  4  2  3  1  3  2

Adding system info:

Session info----------------------------------------------------------------
 setting  value                       
 version  R version 3.1.1 (2014-07-10)
 system   x86_64, mingw32             
 ui       RStudio (0.98.1049)         
 language (EN)                        
 collate  English_United States.1252  
 tz       America/Chicago             

Packages--------------------------------------------------------------------
 package      * version date       source        
 randomForest * 4.6.10  2014-07-17 CRAN (R 3.1.1)

Upvotes: 4

Views: 729

Answers (1)

Zelazny7
Zelazny7

Reputation: 40628

Response from package maintainer:

That parameter behaves as the way that Leo Breiman intended. The bug is in how the parameter was described. It’s the same as minsplit in the rpart:::rpart.control() function:

the minimum number of observations that must exist in a node in order for a split to be attempted.

I will change the description in the help file in the next version to resolve this confusion.

Best, Andy

Upvotes: 1

Related Questions