1
$\begingroup$

I am considering to implement the Quantile Regression Forest proposed here http://www.jmlr.org/papers/volume7/meinshausen06a/meinshausen06a.pdf

By using the normal Random Forest Regression framework on the data the splitting is very conservative i.e. in most of the times the right node gets 5% of samples whereas the left node retains the remaining 95%. After fixing the tree depth I end up with a bunch of leaves having several samples and one leaf having 10 times more for instance. In the context of the Quantile Regression Forest, where we build the forecast distribution from the leaves of the trees this imbalance doesn't seem right. The author of the paper doesn't address this particular aspect. He just mentions that the minimal number of observation in a node is set to greater than 10. I'm curious if anyone has already dealt with that problem and can elaborate more.

$\endgroup$
0
$\begingroup$

So actually the way they deal with it in the paper is basically by weighting the samples in each leaf by 1 over the number of the samples in a leaf i.e. the bigger the sample in the terminal node the less weight the samples from that particular node are having in the forecast distribution.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.