Revising algorithm for nonnegative matrix factorization based on minimizing quasi-L1 norm

Motoaki Mouri, Ichi Takumi, Hiroshi Yasukawa, Andrzej Cichocki

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

4 Citations (Scopus)

Abstract

Previously, we developed a nonnegative matrix factorization (NMF) algorithm named QL1-NMF that is based on minimizing the quasi-L1 norm of an error matrix. When the data includes many outliers, the QL1-NMF algorithm returns better results than ISRA, which is one of the basic NMF algorithms. However, the update functions in the QL1-NMF algorithm are based on a differential function with distortion. Moreover, the solutions it provides sometimes diverge to infinity. The method therefore required improvement to enable it to produce more accurate analysis. In the work described in this paper, we replaced its update functions with others that were based on a simple differential function without distortion. We also contrived ways to implement adjustment factors into the update functions. Computer simulation results confirm the revised algorithm works better than the previous one.

Original languageEnglish
Title of host publication2014 IEEE Asia Pacific Conference on Circuits and Systems, APCCAS 2014
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages767-770
Number of pages4
EditionFebruary
ISBN (Electronic)9781479952304
DOIs
Publication statusPublished - 5 Feb 2015
Externally publishedYes
Event2014 IEEE Asia Pacific Conference on Circuits and Systems, APCCAS 2014 - Ishigaki Island, Okinawa, Japan
Duration: 17 Nov 201420 Nov 2014

Publication series

NameIEEE Asia-Pacific Conference on Circuits and Systems, Proceedings, APCCAS
NumberFebruary
Volume2015-February

Conference

Conference2014 IEEE Asia Pacific Conference on Circuits and Systems, APCCAS 2014
Country/TerritoryJapan
CityIshigaki Island, Okinawa
Period17/11/1420/11/14

Keywords

  • blind source separation (BSS)
  • L1 norm
  • nonnegative matrix factorization (NMF)
  • outlier

Fingerprint

Dive into the research topics of 'Revising algorithm for nonnegative matrix factorization based on minimizing quasi-L1 norm'. Together they form a unique fingerprint.

Cite this