[
https://issues.apache.org/jira/browse/SPARK-17163?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15435497#comment-15435497
]
Seth Hendrickson commented on SPARK-17163:
------------------------------------------
I completely agree about the fact that there would not be much need for LOR if
they are kept separate. It seems we will have only a single interface, whether
we force it explicitly by merging them, or let it happen as one falls out of
use. If that's the case, I prefer to have a single interface called just
{{LogisticRegression}} which handles all cases.
As a clarification on the pivoting subject - we elected explicitly NOT to
implement pivoting for the multinomial case as it matches glmnet and it doesn't
subjectively select one class as a pivot. Implementing pivoting could be as
simple as locking one set of coefficients at zero during training (i.e. we
don't have to hard code a modified gradient update formula like in mllib). What
do you think about exposing a merged API which has a {{family}} parameter. When
the family is set to {{"binomial"}} we produce normal logistic regression with
pivoting and when it is set to {{"multinomial"}} (default) it produces logistic
regression with pivoting. This is basically how glmnet exposes it. The concern
would be that users get different results between 2.0 and 2.1, but we could fix
that by making the default "binomial" for binary classification.
I vote to merge them into a single interface {{LogisticRegression}}. What a fun
PR that would be :)
> Decide on unified multinomial and binary logistic regression interfaces
> -----------------------------------------------------------------------
>
> Key: SPARK-17163
> URL: https://issues.apache.org/jira/browse/SPARK-17163
> Project: Spark
> Issue Type: Sub-task
> Components: ML, MLlib
> Reporter: Seth Hendrickson
>
> Before the 2.1 release, we should finalize the API for logistic regression.
> After SPARK-7159, we have both LogisticRegression and
> MultinomialLogisticRegression models. This may be confusing to users and, is
> a bit superfluous since MLOR can do basically all of what BLOR does. We
> should decide if it needs to be changed and implement those changes before 2.1
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
