[
https://issues.apache.org/jira/browse/SPARK-14478?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Joseph K. Bradley updated SPARK-14478:
--------------------------------------
Description:
Currently, MLlib's StandardScaler scales columns using the corrected standard
deviation (sqrt of unbiased variance). This matches what R's scale package
does.
However, it is a bit odd for 2 reasons:
* Optimization/ML algorithms which require scaled columns generally assume unit
variance (for mathematical convenience). That requires using biased variance.
* scikit-learn, MLlib's GLMs, and R's glmnet package all use biased variance.
*Question*: Should we switch to unbiased?
*Decision*: No. Document what we do, and possibly add support for unbiased
later on.
was:
Currently, MLlib's StandardScaler scales columns using the unbiased standard
deviation. This matches what R's scale package does.
However, it is a bit odd for 2 reasons:
* Optimization/ML algorithms which require scaled columns generally assume unit
variance (for mathematical convenience). That requires using biased variance.
* scikit-learn, MLlib's GLMs, and R's glmnet package all use biased variance.
*Question*: Should we switch to unbiased?
*Decision*: No. Document what we do, and possibly add support for unbiased
later on.
> Should StandardScaler use biased variance to scale?
> ---------------------------------------------------
>
> Key: SPARK-14478
> URL: https://issues.apache.org/jira/browse/SPARK-14478
> Project: Spark
> Issue Type: Question
> Components: ML, MLlib
> Reporter: Joseph K. Bradley
> Assignee: Joseph K. Bradley
>
> Currently, MLlib's StandardScaler scales columns using the corrected standard
> deviation (sqrt of unbiased variance). This matches what R's scale package
> does.
> However, it is a bit odd for 2 reasons:
> * Optimization/ML algorithms which require scaled columns generally assume
> unit variance (for mathematical convenience). That requires using biased
> variance.
> * scikit-learn, MLlib's GLMs, and R's glmnet package all use biased variance.
> *Question*: Should we switch to unbiased?
> *Decision*: No. Document what we do, and possibly add support for unbiased
> later on.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
