English

A convex model for non-negative matrix factorization and dimensionality reduction on physical space

Machine Learning 2015-05-27 v1

Abstract

A collaborative convex framework for factoring a data matrix XX into a non-negative product ASAS, with a sparse coefficient matrix SS, is proposed. We restrict the columns of the dictionary matrix AA to coincide with certain columns of the data matrix XX, thereby guaranteeing a physically meaningful dictionary and dimensionality reduction. We use l1,l_{1,\infty} regularization to select the dictionary from the data and show this leads to an exact convex relaxation of l0l_0 in the case of distinct noise free data. We also show how to relax the restriction-to-XX constraint by initializing an alternating minimization approach with the solution of the convex model, obtaining a dictionary close to but not necessarily in XX. We focus on applications of the proposed framework to hyperspectral endmember and abundances identification and also show an application to blind source separation of NMR data.

Keywords

Cite

@article{arxiv.1102.0844,
  title  = {A convex model for non-negative matrix factorization and dimensionality reduction on physical space},
  author = {Ernie Esser and Michael Möller and Stanley Osher and Guillermo Sapiro and Jack Xin},
  journal= {arXiv preprint arXiv:1102.0844},
  year   = {2015}
}

Comments

14 pages, 9 figures. EE and JX were supported by NSF grants {DMS-0911277}, {PRISM-0948247}, MM by the German Academic Exchange Service (DAAD), SO and MM by NSF grants {DMS-0835863}, {DMS-0914561}, {DMS-0914856} and ONR grant {N00014-08-1119}, and GS was supported by NSF, NGA, ONR, ARO, DARPA, and {NSSEFF.}

R2 v1 2026-06-21T17:21:29.429Z