Folding-up

Latent semantic indexing (LSI) is an information retrievalmethod that represents a dataset as a term-document matrix. LSIuses a matrix factorization method known as the partial singularvalue decomposition (PSVD) to reduce noise in the data byprojecting the term-document matrix into a lower-dimensional vectorspace. Calculating the PSVD of a large term-document matrix iscomputationally expensive. In a rapidly expanding environment, aterm-document matrix is altered often as new documents and termsare added. Recomputing the PSVD of the term-document matrix eachtime changes occur can be too expensive. Folding-in is one methodof adding new documents or terms to an LSI database; updating thePSVD of the database is another method. Folding-in iscomputationally inexpensive, but may cause a loss of accuracy inthe PSVD. PSVD-updating is more expensive, but maintains the PSVDaccuracy. We introduce folding-up, a new method that combinesfolding-in and PSVD-updating. Folding-up offers a significantimprovement in computation time compared with recomputing the PSVDor PSVD-updating, while avoiding the degradation in the PSVD thatcan occur when the folding-in method alone is used.