[ 
https://issues.apache.org/jira/browse/SPARK-18845?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15746385#comment-15746385
 ] 

Andrew Ray commented on SPARK-18845:
------------------------------------

[~srowen] No that's a different thing just whether the result sums to 1 or n. 
But to expand on that, since we are using the version that sums to n our 
initial ranks need to sum to n or it takes a lot longer to converge. 

> PageRank has incorrect initialization value that leads to slow convergence
> --------------------------------------------------------------------------
>
>                 Key: SPARK-18845
>                 URL: https://issues.apache.org/jira/browse/SPARK-18845
>             Project: Spark
>          Issue Type: Bug
>          Components: GraphX
>    Affects Versions: 1.2.2, 1.3.1, 1.4.1, 1.5.2, 1.6.3, 2.0.2
>            Reporter: Andrew Ray
>
> All variants of PageRank in GraphX have incorrect initialization value that 
> leads to slow convergence. In the current implementations ranks are seeded 
> with the reset probability when it should be 1. This appears to have been 
> introduced a long time ago in 
> https://github.com/apache/spark/commit/15a564598fe63003652b1e24527c432080b5976c#diff-b2bf3f97dcd2f19d61c921836159cda9L90
> This also hides the fact that source vertices (vertices with no incoming 
> edges) are not updated. This is because source vertices generally* have 
> pagerank equal to the reset probability. Therefore both need to be fixed at 
> once.
> PR will be added shortly
> *when there are no sinks -- but that's a separate bug



--
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

Reply via email to