[
https://issues.apache.org/jira/browse/SPARK-7372?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14529592#comment-14529592
]
Renat Bekbolatov commented on SPARK-7372:
-----------------------------------------
The main difference is probably an answer to this question: at which
abstraction level is this algorithm available?
If a user just wants to use at least some (say, one-vs-all SVM) straight out of
the box, without having to go to the higher abstraction of ML pipeline (meta
algorithms in package "ml"), it would be convenient to find it in the base
library along with other classification algorithms (package "mllib"). This is
similar to how scikit-learn makes these algorithm effortlessly available.
It is possible that in future ML pipeline will be the main way users interact
with MLlib, in which case maybe this is not a big usability issue. However,
even in that case, redundant lower level (package "mllib") implementations
efficiently reproducible with ML pipeline could be phased out and replaced with
those versions.
Basically, as a user, I would want to just run the classifier. Maybe ML
pipeline already makes this simple - I'll have to check. What do you think?
> Multiclass SVM - One vs All wrapper
> -----------------------------------
>
> Key: SPARK-7372
> URL: https://issues.apache.org/jira/browse/SPARK-7372
> Project: Spark
> Issue Type: Question
> Components: MLlib
> Reporter: Renat Bekbolatov
> Priority: Trivial
>
> I was wondering if we want to have a some support for multiclass SVM in
> MLlib, for example, through a simple wrapper over binary SVM classifiers with
> OVA.
> There is already WIP for ML pipeline generalization: SparkSPARK-7015,
> Multiclass to Binary Reduction
> However, if users prefer to just have basic OVA version that runs against
> SVMWithSGD, they might be able to use it.
> Here is a code sketch:
> https://github.com/Bekbolatov/spark/commit/463d73323d5f08669d5ae85dc9791b036637c966
> Maybe this could live in a 3rd party utility library (outside Spark MLlib).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
