[ 
https://issues.apache.org/jira/browse/MAHOUT-796?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13092179#comment-13092179
 ] 

Nathan Halko commented on MAHOUT-796:
-------------------------------------

The reorthogonalizations aren't essential and if its a barrier to power 
iterations we should forego them at the moment.  A quick and dirty trick to 
avoid even sweeping through A again is to neglect the cross terms in the 
product (AA')^qA\Omega and just use (A_iA_i')^qA_i\Omega. This could be 
extremely naive but I've been getting some good results with it.  The accuracy 
typically falls about half way between single pass and full power iterations so 
it could be useful (although it could be dangerous as well).  

sig_51  <-  optimal
   60.6531

full power iters 1
   81.2668

full power iters 4
   67.5545

row-wise power iters 1
   89.4983

row-wise power iters 4
   82.2247

single pass
   92.8736

norm A
  100.0000 

The 'row-wise power iters' being (A_iA_i')^qA_i\Omega.  
  

> Modified power iterations in existing SSVD code
> -----------------------------------------------
>
>                 Key: MAHOUT-796
>                 URL: https://issues.apache.org/jira/browse/MAHOUT-796
>             Project: Mahout
>          Issue Type: Improvement
>          Components: Math
>    Affects Versions: 0.5
>            Reporter: Dmitriy Lyubimov
>            Assignee: Dmitriy Lyubimov
>              Labels: SSVD
>             Fix For: 0.6
>
>
> Nathan Halko contacted me and pointed out importance of availability of power 
> iterations and their significant effect on accuracy of smaller eigenvalues 
> and noise attenuation. 
> Essentially, we would like to introduce yet another job parameter, q, that 
> governs amount of optional power iterations. The suggestion how to modify the 
> algorithm is outlined here : 
> https://github.com/dlyubimov/ssvd-lsi/wiki/Power-iterations-scratchpad .
> Note that it is different from original power iterations formula in the paper 
> in the sense that additional orthogonalization performed after each 
> iteration. Nathan points out that that improves errors in smaller eigenvalues 
> a lot (If i interpret it right). 

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to