Trying to do dimensionality reduction with SSVD then running the new doc matrix 
through kmeans. 

The Lanczos + ClusterDump test of SVD + kmeans uses A-hat = A^t V^t. 
Unfortunately this results in anonymous vectors in clusteredPoints after A-hat 
is run through kmeans. The doc ids are lost due to the transpose I assume?

In any case Dmitriy pointed out that this might have been done because Lanczos 
does not produce U. 

So I need to do US^-1? This would avoid the transpose and should preserve 
doc/row ids for kmeans? And doing the PCA in SSVD will weight things properly 
so I don't need the --halfSigma?

Please correct me if I'm wrong.


On Sep 5, 2012, at 4:59 PM, Dmitriy Lyubimov <[email protected]> wrote:

Yes i have an option to output U * Sigma^0.5 already.

But strictly speaking the way PCA space is defined would require just
U*Sigma output. Or it is not worth it?


On Wed, Sep 5, 2012 at 4:56 PM, Ted Dunning <[email protected]> wrote:
> Yes.  (A-M)V is U \Sigma.  You may actually want something like U \sqrt
> \Sigma instead, though.
> 
> 
> On Wed, Sep 5, 2012 at 4:10 PM, Dmitriy Lyubimov <[email protected]> wrote:
> 
>> Hello,
>> 
>> I have a question w.r.t what to advise people in the SSVD manual for PCA.
>> 
>> So we have
>> 
>> (A-M) \approx U \Sigma V^t
>> 
>> and strictly speaking since svd is reduced rank, we need to re-project
>> original data points as
>> 
>> Y= (A-M)V
>> 
>> However we can assume (A-M)V \approx U \Sigma, can't we? I.e. instead of
>> recomputing tough job of (A-M)V we can just advise to use U\Sigma or just U
>> in some cases, can't we?
>> 
>> Thanks.
>> -d
>> 

Reply via email to