[
https://issues.apache.org/jira/browse/SPARK-12815?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Robert Dodier updated SPARK-12815:
----------------------------------
Description:
The Wilcoxon-Mann-Whitney rank sum statistic (also known by other permutations
of those names) is a useful assessment of relevance of an input field for a
classification problem. As such it would nice to have in ML or MLlib (I don't
know what's a more suitable package for it).
I have created a Spark package,
[spark-wilcoxon|http://spark-packages.org/package/robert-dodier/spark-wilcoxon],
to demonstrate an implementation. If there is interest in this issue, I'll
create a pull request. spark-wilcoxon computes the scaled rank sum statistic
{{U/(n0*n1)}}, where {{U}} is the rank sum statistic and {{n0}} and {{n1}} are
the numbers of data in class 0 and class 1, respectively.
There exists already the Spearman rank correlation statistic in MLlib (in
...mllib.stat.correlation.SpearmanCorrelation) but that is not equivalent to
the WMW statistic -- the one cannot be derived from the other because the
Spearman correlation contains squares of rank differences and the WMW contains
only first-order terms.
See the Wikipedia article [Mann-Whitney U
test|https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test] for formulas
and background information. At this point, I am proposing only to compute the
rank sum statistic, not to implement the significance test.
was:
The Wilcoxon-Mann-Whitney rank sum statistic (also known by other permutations
of those names) is a useful assessment of relevance of an input field for a
classification problem. As such it would nice to have in ML or MLlib (I don't
know what's a more suitable package for it).
I have created a Spark package,
[spark-wilcoxon|http://spark-packages.org/package/robert-dodier/spark-wilcoxon],
to demonstrate an implementation. If there is interest in this issue, I'll
create a pull request. spark-wilcoxon computes the scaled rank sum statistic
`U/(n0*n1)`, where `U` is the rank sum statistic and `n0` and `n1` are the
numbers of data in class 0 and class 1, respectively.
There exists already the Spearman rank correlation statistic in MLlib (in
...mllib.stat.correlation.SpearmanCorrelation) but that is not equivalent to
the WMW statistic -- the one cannot be derived from the other because the
Spearman correlation contains squares of rank differences and the WMW contains
only first-order terms.
See the Wikipedia article [Mann-Whitney U
test|https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test] for formulas
and background information. At this point, I am proposing only to compute the
rank sum statistic, not to implement the significance test.
> Compute Wilcoxon-Mann-Whitney rank sum statistic
> ------------------------------------------------
>
> Key: SPARK-12815
> URL: https://issues.apache.org/jira/browse/SPARK-12815
> Project: Spark
> Issue Type: New Feature
> Components: ML, MLlib
> Reporter: Robert Dodier
> Priority: Minor
>
> The Wilcoxon-Mann-Whitney rank sum statistic (also known by other
> permutations of those names) is a useful assessment of relevance of an input
> field for a classification problem. As such it would nice to have in ML or
> MLlib (I don't know what's a more suitable package for it).
> I have created a Spark package,
> [spark-wilcoxon|http://spark-packages.org/package/robert-dodier/spark-wilcoxon],
> to demonstrate an implementation. If there is interest in this issue, I'll
> create a pull request. spark-wilcoxon computes the scaled rank sum statistic
> {{U/(n0*n1)}}, where {{U}} is the rank sum statistic and {{n0}} and {{n1}}
> are the numbers of data in class 0 and class 1, respectively.
> There exists already the Spearman rank correlation statistic in MLlib (in
> ...mllib.stat.correlation.SpearmanCorrelation) but that is not equivalent to
> the WMW statistic -- the one cannot be derived from the other because the
> Spearman correlation contains squares of rank differences and the WMW
> contains only first-order terms.
> See the Wikipedia article [Mann-Whitney U
> test|https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test] for formulas
> and background information. At this point, I am proposing only to compute the
> rank sum statistic, not to implement the significance test.
--
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
