Github user manishamde commented on the pull request:
https://github.com/apache/spark/pull/886#issuecomment-44388751
I fully agree.
I will give others a day or two to raise any concerns if they have any and
then proceed to implement the two-step solution for multiclass classification
that I mentioned above. The second step will be the O(k) algorithm (k is the
number of categorical feature values) that will come up with k sorted
categorical feature splits using the target variable entropy for ordering.
The O(n^2) algorithm looked promising at first but I think it might end up
dominating the tree computation time.
In general, getting 0(k) splits is more important than ensuring that they
are sorted since we now have a way of dealing with unsorted splits with this
PR. I currently don't have a good intuition on what makes a good subset of
splits but we could keep adding more heuristics later.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
