Document Representation Using Nonnegative Matrix Factorization

Non-negative matrix factorization (NMF) is an emerging technique of latent semantic analysis from the given document corpus. The existing NMF algorithms don not use the intrinsic structure information of original document corpus. In order to preserve intrinsic structure information in latent semantic space extracted by NMF, a NMF algorithm with intrinsic structure information properties is presented. The primary ideal is to extend the original NMF through incorporating the intrinsic structure information constraints inside the NMF decomposition. Our experimental results performed on the RCV1 and SECTOR data sets show that the proposed method is superior to NMF for document latent semantic analysis.