Ensure that min.node.size is strictly enforced.

[jtibshirani]

Currently, we don't prevent splits from occurring that could result in nodes with size less than min.node.size. Instead, the algorithm simply stops splitting if a node's size is less than or equal to min.node.size.

Although this is also the current behavior in both ranger and the original randomForest package, it's not intuitive and can be quite misleading. We should look into updating the splitting algorithm so that min.node.size is enforced as a true minimum.

[swager]

Just to add a little more context, the behavior of the randomForest package is discussed here.

We still intend to update the behavior of our implementation to properly enforce min.node.size as a true minimum node size.

[ras44]

Hi @jtibshirani I'd be happy to continue contributing via PRs to grf. I was thinking either this one or any of the below:

• 420 cobalt style balance plots
• 291 classification forests

... or any other suggestions you or the team may have. Thanks!

[swager]

Thanks for reaching out, @ras44! Let's start with #420 (and move the conversation there). If you're interested in doing this, could you please prepare a short document to start the conversation about how you envision the feature (in terms of scope, what kinds of functionality we should have, what types of forests the function can be applied to, if / how we should set up a cobalt dependency, etc.)?

[halflearned]

The implementation of min.node.size in grf is now documented in the REFERENCE. See min.node.size and Selecting Balanced Splits.