[
https://issues.apache.org/jira/browse/SPARK-7770?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14805343#comment-14805343
]
Yanbo Liang commented on SPARK-7770:
------------------------------------
[~josephkb] After investigate of convergenceTol in GradientDescent, I'm also
agree to compare ValidateError with tolerance switching between relative and
absolute ones. I think users can understand this parameter more clear and it
also consistent with the semantic of GradientDescent.
One thing need to discuss is that it will break the semantic of current
validationTol. I have updated the PR, please feel free to comment there.
> Should GBT validationTol be relative tolerance?
> -----------------------------------------------
>
> Key: SPARK-7770
> URL: https://issues.apache.org/jira/browse/SPARK-7770
> Project: Spark
> Issue Type: Improvement
> Components: ML, MLlib
> Reporter: Joseph K. Bradley
> Assignee: Yanbo Liang
> Priority: Minor
>
> In spark.mllib, GBT validationTol uses absolute tolerance. Relative
> tolerance is arguably easier to set in a meaningful way. Questions:
> * Should we change spark.mllib's validationTol meaning?
> * Should we use relative tolerance in spark.ml's GBT (once we add validation
> support)?
> I would vote for changing both to relative tolerance, where the tolerance is
> relative to the current loss on the training set.
--
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
