online read us now
Paper details
Number 2 - June 2014
Volume 24 - 2014
Regularized nonnegative matrix factorization: Geometrical interpretation and application to spectral unmixing
Rafał Zdunek
Abstract
Nonnegative Matrix Factorization (NMF) is an important tool in data spectral analysis. However, when a mixing matrix
or sources are not sufficiently sparse, NMF of an observation matrix is not unique. Many numerical optimization algorithms,
which assure fast convergence for specific problems, may easily get stuck into unfavorable local minima of an
objective function, resulting in very low performance. In this paper, we discuss the Tikhonov regularized version of the
Fast Combinatorial NonNegative Least Squares (FC-NNLS) algorithm (proposed by Benthem and Keenan in 2004), where
the regularization parameter starts from a large value and decreases gradually with iterations. A geometrical analysis and
justification of this approach are presented. The numerical experiments, carried out for various benchmarks of spectral
signals, demonstrate that this kind of regularization, when applied to the FC-NNLS algorithm, is essential to obtain good
performance.
Keywords
blind source separation, nonnegative matrix factorization, active-set algorithm, regularized NMF, polytope approximation