2

Other random forest tools have the "dial" that limits max depth of splits on a particular branch.

h2o.randomForest has "max_depth", for example.

What is the version of that for "ranger"?

Dennis Jaheruddin
  • 21,208
  • 8
  • 66
  • 122
EngrStudent
  • 1,924
  • 31
  • 46

1 Answers1

1

I'm not familiar with the h2o.randomForest package, but my general understanding of random forests is that each tree will be grown until a certain minimum number of data points fit into each leaf of the tree. In other words, a tree will keep splitting until a certain level of classification of each data point has been achieved. In the standard randomForest package, there is a parameter called nodesize which controls this:

https://stats.stackexchange.com/questions/158583/what-does-node-size-refer-to-in-the-random-forest

The analogous parameter in the ranger package seems to be min.node.size. You can compare the information in the link above with the documentation to convince yourself that they are both discussing the same thing.

Tim Biegeleisen
  • 502,043
  • 27
  • 286
  • 360
  • max depth is about how many splits between stump and leaf. It is a control parameter that is used to avoid over-fitting. – EngrStudent May 31 '17 at 18:28
  • @EngrStudent I don't believe that over fitting is a big concern with random forests. And in any case the node size gives you the same control over splitting. – Tim Biegeleisen May 31 '17 at 23:35
  • Anything can over-fit. It is not the largest issue, but if you want tuned performance, then you must tune. If you limit the number of splits made for a binary tree, then you put an upper bound on compute time. – EngrStudent Jun 01 '17 at 14:46
  • My understanding of random forests is that they are relatively not prone to overfitting, but maybe I'm wrong. In any case, controlling the node size should give you the same control that controlling the number of splits would give you. – Tim Biegeleisen Jun 01 '17 at 14:51
  • Consider page 24 of this: http://www2.ims.nus.edu.sg/Programs/014swclass/files/mark.pdf – EngrStudent Jun 01 '17 at 15:47