[
https://issues.apache.org/jira/browse/MAHOUT-376?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12919567#action_12919567
]
Dmitriy Lyubimov edited comment on MAHOUT-376 at 10/10/10 12:59 PM:
--------------------------------------------------------------------
so i got to singular values now. I run a unit test so that k+p=n. When i
parameterize algorithm so that only one Q-block is produced , the eigenvalues
match the stock result at least as good as 10E-5. Which is expected under the
circumstances. however as soon as i increase number of Q-blocks >1, the
singular values go astray as much as 10%. Not good. In both cases, the entire Q
passes the orthonormality test. I guess it means that as i thought before,
doing block orthonormalization this way does result in a subspace different
from original span by Y.I need to research on doing orthonormalization with
blocks. I think that's the only showstopper here that is still left. It may
result in a rewrite that splits one job producing both Q and Bt, into several
though.
was (Author: dlyubimov2):
ok, i got to BBt normalized for rank deficiences in Q blocks.
What eigensolver should i use in context of Mahout? I know there's something in
that Cern library but it is all deprecateed. There's also a solver in apache
commons, but it is not readily imported into the mahout core.
Thanks.
PS.
so i got to singular values now. I run a unit test so that k+p=n. When i
parameterize algorithm so that only one Q-block is produced , the eigenvalues
match the stock result at least as good as 10E-5. Which is expected under the
circumstances. however as soon as i increase number of Q-blocks >1, the
singular values go astray as much as 10%. Not good. In both cases, the entire Q
passes the orthonormality test. I guess it means that as i thought before,
doing block orthonormalization this way does result in a subspace different
from original span by Y.I need to research on doing orthonormalization with
blocks. I think that's the only showstopper here that is still left. It may
result in a rewrite that splits one job producing both Q and Bt, into several
though.
> Implement Map-reduce version of stochastic SVD
> ----------------------------------------------
>
> Key: MAHOUT-376
> URL: https://issues.apache.org/jira/browse/MAHOUT-376
> Project: Mahout
> Issue Type: Improvement
> Components: Math
> Reporter: Ted Dunning
> Assignee: Ted Dunning
> Fix For: 0.5
>
> Attachments: MAHOUT-376.patch, sd-bib.bib, sd.pdf, sd.tex, Stochastic
> SVD using eigensolver trick.pdf
>
>
> See attached pdf for outline of proposed method.
> All comments are welcome.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
