Thread Subject: Can't perform PCA

My data set is too large for me to use the built-in Matlab
functions such as "processpca" to compute principal
components due to memory limitations.

Does anyone have an m-file that performs PCA interatively,
returning one component at a time, or any algorithm that
allows PCA to be performed in some way without computing
the entire covariance matrix?

Thanks

"Dave Krebs" <djk37@mathworks.com> wrote in message
<fjck47$91t$1@fred.mathworks.com>...
> My data set is too large for me to use the built-in Matlab
> functions such as "processpca" to compute principal
> components due to memory limitations.
>
> Does anyone have an m-file that performs PCA interatively,
> returning one component at a time, or any algorithm that
> allows PCA to be performed in some way without computing
> the entire covariance matrix?
>

I don't (not in MATLAB), but you are better chance to find
what you are looking for with keywords such as "Lanczos" and
"Krylov's space".

Bruno

On Dec 7, 6:14 pm, "Dave Krebs" <dj...@mathworks.com> wrote:
> My data set is too large for me to use the built-in Matlab
> functions such as "processpca" to compute principal
> components due to memory limitations.
>
> Does anyone have an m-file that performs PCA interatively,
> returning one component at a time, or any algorithm that
> allows PCA to be performed in some way without computing
> the entire covariance matrix?

Go to Google Groups and search on

nipals -oliver

Hope this helps.

Greg

Dave Krebs wrote:
> My data set is too large for me to use the built-in Matlab
> functions such as "processpca" to compute principal
> components due to memory limitations.
>
> Does anyone have an m-file that performs PCA interatively,
> returning one component at a time, or any algorithm that
> allows PCA to be performed in some way without computing
> the entire covariance matrix?

Dave, do you have access to the Statistics Toolbox? There is no good reason to
compute the cov matrix, and at least two reasons not to. The PRINCOMP function
computes a PCA directly from the data. If your data are wide, you'll want to
use the 'econ' flag:

 >> help princomp
  PRINCOMP Principal Components Analysis.
[snip]
     [...] = PRINCOMP(X,'econ') returns only the elements of LATENT that are
     not necessarily zero, i.e., when N <= P, only the first N-1, and the
     corresponding columns of COEFF and SCORE. This can be significantly
     faster when P >> N.

Since you mention computing the cov matrix as the problem, I'm guessing your
data _are_ wide.

That being said, there are any number of implementations of the NIPALS algorithm
in MATLAB out there; that may be what you're looking for, but I'd look at
PRINCOMP first.

Hope this helps.

- Peter Perkins
   The MathWorks, Inc.

I never use it, but svds() can potentially solve your problem.

Bruno