[Numpy-discussion] very large matrices.

[Dave P. Novakovic]

There are definitely elements of spectral graph theory in my research
too. I'll summarise

We are interested in seeing the each eigenvector from svd can
represent in a semantic space
In addition to this we'll be testing it against some algorithms like
concept indexing (uses a bipartitional k-meansish method for dim
reduction)
also testing against Orthogonal Locality Preserving indexing, which
uses the laplacian of a similarity matrix to calculate projections of
a document (or term) into a manifold.

These methods have been implemented and tested for document
classification, I'm interested in seeing their applicability to
modelling semantics with a system known as Hyperspace to analog
language.

I was hoping to do svd to my HAL built out of reuters, but that was
way too big. instead i'm trying with the traces idea i mentioned
before (ie contextually grepping a keyword out of the docs to build a
space around it.)

Cheers

Dave

On 5/14/07, Charles R Harris <charlesr.harris at gmail.com> wrote:
>
>
> On 5/13/07, Dave P. Novakovic <davidnovakovic at gmail.com> wrote:
> > > Are you trying some sort of principal components analysis?
> >
> > PCA is indeed one part of the research I'm doing.
>
> I had the impression you were trying to build a linear space in which to
> embed a model, like atmospheric folk do when they try to invert spectra to
> obtain thermal profiles. Model based compression would be another aspect of
> this. I wonder if there aren't some algorithms out there for this sort of
> thing.
>
> Chuck
>
>
> _______________________________________________
> Numpy-discussion mailing list
> Numpy-discussion at scipy.org
> http://projects.scipy.org/mailman/listinfo/numpy-discussion
>
>