Github user sethah commented on the issue:
https://github.com/apache/spark/pull/14547
One questions I had - this PR creates an inherent coupling between the
impurity used to train the tree and the loss used for boosting. This is not how
I understood tree boost. My impression was that, regardless of how the tree was
trained (i.e. what impurity was used), that tree boost would simply modify the
leaf node predictions to minimize the *boosting loss*.
In fact, there is no real coupling done in this PR, but the framework is
there. In scikit, there is no implied coupling. They simply train the tree, and
modify the leaf node predictions after training. It may be hard to do this in a
performant way here, so I'm not sure what is best. Just wanted to get some
clarification on the design.
---
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.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
