Large-scale SVD and subspace-based methods for Information retrieval

O.  Marques*, H. Simon* and H. Zha

*NERSC,
Lawrence Berkeley National Laboratory
1 Cyclotron Road
Berkeley, CA 94720

307 Pond Laboratory
Department of Computer Science and Engineering
The Pennsylvania State University
University Park, PA 16802-6103


Abstract.

We present a theoretical foundation based on subspaces for latent
semantic indexing (LSI) in information retrieval. We show that our
model leads to a low-rank-plus-shift structure that is approximately
satisfied by the cross-product of the term-document matrices. This
structure can be exploited for the compution of  the partial singular
value decomposition (SVD) of a large sparse term-document matrix
used in LSI. We also discuss several parallel implementation issues and
present emperical numerical results on Cray T3E using text collections
with millions of documents.