total leaf count in h2o

Question

I made a random forest in h2o using r. It has trees and leaves.

I want to know how many leaves. I like to compare my total number of rows to leaves.

I have 200 trees, 8 layers deep, require 5 rows per terminal leaf. Am I traversing much of my data? I have 20k rows.

Is there a clean way to count the number of leaves in the h2o randomForest?

A random forest model is actually a collection of trees. Why do you need to compute the number of leaves? — Tim Biegeleisen
I want to use the ratio of leaves to rows to think about over-fitting or under-fitting in terms of variable importance. If I use each row 12 times, then there is likely a problem. If I use less than 0.5% of rows, then I could be under-sampling, even if it does look bootstrappy. Perhaps my trees use the same row 12 times, or 0.1 times. I want to think in terms of coverage, and don't know how to access the actual 'h2o' leaves. — EngrStudent
This approach is not right AFAIK. The variables which are used to split early on usually dominate the behavior of a given tree. If you want variable importance, you should be using importance() or for a simpler option try the best constant model. — Tim Biegeleisen
@TimBiegeleisen - I use "importance", aggressively. I have enjoyed getting into the fundamentals of the 'Boruta' package including the difference between z and gini. That is not the problem. My problem is that in some cases I have sparse data, and the tree looks like a subset of rows are dominating the fit. In some cases I have huge data and I want to be sure it is adequately sampled. All the "importance" in the world is going to be meaningless if I don't make a healthy forest. I want some better measure than an eyeball norm on a convergence plot. — EngrStudent

Darren Cook Darren Cook · Accepted Answer · 2016-05-19T09:57:05

You'll be glad to know H2O stores that information! E.g. (this is for the Iris data set)

m <- h2o.randomForest(1:4, 5, data)

When printing m I see:

number_of_trees model_size_in_bytes min_depth max_depth mean_depth min_leaves  max_leaves mean_leaves
            150               20217         1         9    3.72667          21         15     6.17333

So there are 926 leaves (6.17333 * 150).

To get it with code:

ms = m@model$model_summary
ms$number_of_trees * ms$mean_leaves

total leaf count in h2o

1 Answers