Document Clustering using NMF based on Spectral Information

摘要

Nonnegative matrices abound and their effective dimensional reduction is an important goal in data analysis. In this paper we seek meaningful nonnegative low rank factorizations using the truncated SVD. The starting point is the recently proposed NNDSVD method, for which we propose block as well bisecting versions. Numerical experiments indicate that the proposed methodology can provide fast and effective approximations for clustering tasks. Moreover, depending on the task at hand, the resulting approximations can be used directly or as initializations for other NMF algorithms.