[jira] [Commented] (SPARK-5127) Fixed overflow when there are outliers in data in Logistic Regression

["Apache Spark (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-5127?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14267381#comment-14267381 ] 

Apache Spark commented on SPARK-5127:
-------------------------------------

User 'dbtsai' has created a pull request for this issue:
https://github.com/apache/spark/pull/3928

> Fixed overflow when there are outliers in data in Logistic Regression
> ---------------------------------------------------------------------
>
>                 Key: SPARK-5127
>                 URL: https://issues.apache.org/jira/browse/SPARK-5127
>             Project: Spark
>          Issue Type: Bug
>          Components: MLlib
>            Reporter: DB Tsai
>
> gradientMultiplier = (1.0 / (1.0 + math.exp(margin))) - label
> However, the first part of gradientMultiplier will be suffered from overflow if there are samples far away from hyperplane, and this happens when there are outliers in data. As a result, we use the equivalent formula but more numerically stable.
>     val gradientMultiplier =
>       if (margin > 0.0) {
>         val temp = math.exp(-margin)
>         temp / (1.0 + temp) - label
>       } else {
>         1.0 / (1.0 + math.exp(margin)) - label
>       }



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org