[
https://issues.apache.org/jira/browse/SPARK-9963?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14720824#comment-14720824
]
Trevor Cai commented on SPARK-9963:
-----------------------------------
[~lkhamsurenl]: Are you still working on this?
I took a quick look at the code, and it looks to me like it's not trivial to
convert the binnedFeatures and splits arrays into a single vector mapping
feature to threshold. The below section of code seems to indicate to me that
some features don't have a threshold and should always move right, and using
Double.MAX_VALUE as the threshold in those cases seems like it could
potentially cause issues.
{code}
override private[tree] def shouldGoLeft(binnedFeature: Int, splits:
Array[Split]): Boolean = {
if (binnedFeature == splits.length) {
// > last split, so split right
false
} else {
val featureValueUpperBound =
splits(binnedFeature).asInstanceOf[ContinuousSplit].threshold
featureValueUpperBound <= threshold
}
}
{code}
As a result, if you're still working on this, your second proposal makes more
sense. If not, I can pick up the issue and implement the second option.
[~josephkb] Does this seem reasonable to you?
> ML RandomForest cleanup: replace predictNodeIndex with predictImpl
> ------------------------------------------------------------------
>
> Key: SPARK-9963
> URL: https://issues.apache.org/jira/browse/SPARK-9963
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Reporter: Joseph K. Bradley
> Priority: Trivial
> Labels: starter
>
> Replace ml.tree.impl.RandomForest.predictNodeIndex with Node.predictImpl.
> This should be straightforward, but please ping me if anything is unclear.
--
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
