[
https://issues.apache.org/jira/browse/SPARK-18678?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15714960#comment-15714960
]
Sean Owen commented on SPARK-18678:
-----------------------------------
It's helpful if you point out specifically what you mean in the code, but I
think it's this:
{code}
var l = i.toLong
val rand = new XORShiftRandom(seed)
while (input.hasNext) {
val item = input.next()
val replacementIndex = (rand.nextDouble() * l).toLong
if (replacementIndex < k) {
reservoir(replacementIndex.toInt) = item
}
l += 1
}
{code}
l needs to be incremented before the random choice, right?
> Skewed feature subsampling in Random forest
> -------------------------------------------
>
> Key: SPARK-18678
> URL: https://issues.apache.org/jira/browse/SPARK-18678
> Project: Spark
> Issue Type: Bug
> Components: ML
> Affects Versions: 2.0.2
> Reporter: Bjoern Toldbod
>
> The feature subsampling performed in the RandomForest-implementation from
> org.apache.spark.ml.tree.impl.RandomForest
> is performed using SamplingUtils.reservoirSampleAndCount
> The implementation of the sampling skews feature selection in favor of
> features with a higher index.
> The skewness is smaller for a large number of features, but completely
> dominates the feature selection for a small number of features. The extreme
> case is when the number of features is 2 and number of features to select is
> 1.
> In this case the feature sampling will always pick feature 1 and ignore
> feature 0.
> Of course this produces low quality models for few features when using
> subsampling.
--
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
