Github user jkbradley commented on the pull request:
https://github.com/apache/spark/pull/5055#issuecomment-84204731
Apologies for my late response too. I feel like what we really need to do
is clarify the intended behavior of feature scaling. Currently, feature
scaling changes the optimal solution vector for regularized models since it
changes each feature's relative amount of regularization. I see 2 options:
* Keep the current behavior.
* If we go with this behavior, then we should expose it as an option
since it changes the optimal solution vector.
* Adjust the regularization parameter for each feature such that the
optimal solution vector is identical (after rescaling) to the solution for the
original problem (before scaling).
* I believe this is what libsvm (or maybe liblinear) does.
* If we do feature scaling under the hood (and do not expose it as an
option), then we should use this behavior. Otherwise, users will be confused
when the optimal solution is not what they would expect.
I strongly vote for the 2nd option: It has the intended benefit of
improving optimization behavior, and it is better for users since it gives them
the solution they expect.
---
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]
