[ 
https://issues.apache.org/jira/browse/MAHOUT-796?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13091601#comment-13091601
 ] 

Ted Dunning commented on MAHOUT-796:
------------------------------------

One problem here is that the Q's are large and potentially dense.  Thus, 
accumulating them is not a great idea.  That can be worked around in the single 
iteration because we can keep R in memory and can reconstruct chunks of Y given 
chunks of A.

That trick becomes a bit more involved if we want to keep all of the Q's in 
such an implicit form.  Computing (AA')^q A\Omega is relatively simple as you 
point out if A' is available as a linear operator, but I thought I understood 
that reorthogonalizing was a good idea.  What I don't see is how to 
re-orthogonalize without keeping very large matrices in memory or doing a 
dangerously dense operation.  Yet.





> 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