SVD for a large sparse matrix?

[byungki]

Hi, can the current SVD routine in the mlpack handle large sparse
matrices? When I used the SVD routine in the mlpack, I got "out of
memory" error. The dimension of the matrix is about 15000x5000, but
it's very sparse. Is there an easy way that I don't need to rely on
other SVD tools?

[Dongryeol Lee]

The SVD routine defined in fastlib/la/la.h is a dense-matrix based method, and will
require 8 bytes for each of the 15000 * 5000 entries. You would have to rely on
an external package that is not included in the current release of MLPACK on
mloss.org.

--
Dongryeol Lee
Ph.D. Student, College of Computing
Georgia Institute of Technology

[Nikolaos Vasiloglou II]

The internal version of MLPACK uses trillinos, although it hasn't been maintained. If you really need it we could do something for it although we are busy  until NIPS

[byungki]

Well, if so, I'll just use other tools such as svdpack and try to
integrate it to fastlib as much as I can. It's just a suggestion, but
it would be nicer if mlpack could handle larger matrices because then
dimensionality reduction techniques, for example, become more useful.

On Dec 4, 11:32 pm, "Nikolaos Vasiloglou II" <nva...@ieee.org> wrote:
> The internal version of MLPACK uses trillinos, although it hasn't been
> maintained. If you really need it we could do something for it although we
> are busy  until NIPS
>