Subtractive Initialization of Nonnegative Matrix Factorizations for Document Clustering

Abstract

Nonnegative matrix factorizations (NMF) have recently assumed an important role in several fields, such as pattern recognition, automated image exploitation, data clustering and so on. They represent a peculiar tool adopted to obtain a reduced representation of multivariate data by using additive components only, in order to learn parts-based representations of data. All algorithms for computing the NMF are iterative, therefore particular emphasis must be placed on a proper initialization of NMF because of its local convergence. The problem of selecting appropriate starting initialization matrices becomes more complex when data possess special meaning, and this is the case of document clustering. In this paper, we present a new initialization method which is based on the fuzzy subtractive scheme and used to generate initial matrices for NMF algorithms. A preliminary comparison of the proposed initialization with other commonly adopted initializations is presented by considering the application of NMF algorithms in the context of document clustering.