Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization

Non-negative matrix factorization is a relatively new method of matrix decomposition which factors an <i>m</i> × <i>n</i> data matrix <i>X</i> into an <i>m</i> × <i>k</i> matrix <i>W</i> and a <i>k</i> × <i>n&l...

Descripción completa

Guardado en:
Detalles Bibliográficos
Autores principales: José M. Maisog, Andrew T. DeMarco, Karthik Devarajan, Stanley Young, Paul Fogel, George Luta
Formato: article
Lenguaje:EN
Publicado: MDPI AG 2021
Materias:
PCA
Acceso en línea:https://doaj.org/article/17df6d63695d43868d77feb0c575755d
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
id oai:doaj.org-article:17df6d63695d43868d77feb0c575755d
record_format dspace
spelling oai:doaj.org-article:17df6d63695d43868d77feb0c575755d2021-11-25T18:16:26ZAssessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization10.3390/math92228402227-7390https://doaj.org/article/17df6d63695d43868d77feb0c575755d2021-11-01T00:00:00Zhttps://www.mdpi.com/2227-7390/9/22/2840https://doaj.org/toc/2227-7390Non-negative matrix factorization is a relatively new method of matrix decomposition which factors an <i>m</i> × <i>n</i> data matrix <i>X</i> into an <i>m</i> × <i>k</i> matrix <i>W</i> and a <i>k</i> × <i>n</i> matrix <i>H</i>, so that <i>X</i> ≈ <i>W</i> × <i>H</i>. Importantly, all values in <i>X</i>, <i>W</i>, and <i>H</i> are constrained to be non-negative. NMF can be used for dimensionality reduction, since the <i>k</i> columns of <i>W</i> can be considered components into which <i>X</i> has been decomposed. The question arises: how does one choose <i>k</i>? In this paper, we first assess methods for estimating <i>k</i> in the context of NMF in synthetic data. Second, we examine the effect of normalization on this estimate’s accuracy in empirical data. In synthetic data with orthogonal underlying components, methods based on PCA and Brunet’s Cophenetic Correlation Coefficient achieved the highest accuracy. When evaluated on a well-known real dataset, normalization had an unpredictable effect on the estimate. For any given normalization method, the methods for estimating <i>k</i> gave widely varying results. We conclude that when estimating <i>k</i>, it is best not to apply normalization. If the underlying components are known to be orthogonal, then Velicer’s MAP or Minka’s Laplace-PCA method might be best. However, when the orthogonality of the underlying components is unknown, none of the methods seemed preferable.José M. MaisogAndrew T. DeMarcoKarthik DevarajanStanley YoungPaul FogelGeorge LutaMDPI AGarticlenon-negative matrix factorizationnormalizationPCAfactorization ranknumber of factored componentshigh-dimensional dataMathematicsQA1-939ENMathematics, Vol 9, Iss 2840, p 2840 (2021)
institution DOAJ
collection DOAJ
language EN
topic non-negative matrix factorization
normalization
PCA
factorization rank
number of factored components
high-dimensional data
Mathematics
QA1-939
spellingShingle non-negative matrix factorization
normalization
PCA
factorization rank
number of factored components
high-dimensional data
Mathematics
QA1-939
José M. Maisog
Andrew T. DeMarco
Karthik Devarajan
Stanley Young
Paul Fogel
George Luta
Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
description Non-negative matrix factorization is a relatively new method of matrix decomposition which factors an <i>m</i> × <i>n</i> data matrix <i>X</i> into an <i>m</i> × <i>k</i> matrix <i>W</i> and a <i>k</i> × <i>n</i> matrix <i>H</i>, so that <i>X</i> ≈ <i>W</i> × <i>H</i>. Importantly, all values in <i>X</i>, <i>W</i>, and <i>H</i> are constrained to be non-negative. NMF can be used for dimensionality reduction, since the <i>k</i> columns of <i>W</i> can be considered components into which <i>X</i> has been decomposed. The question arises: how does one choose <i>k</i>? In this paper, we first assess methods for estimating <i>k</i> in the context of NMF in synthetic data. Second, we examine the effect of normalization on this estimate’s accuracy in empirical data. In synthetic data with orthogonal underlying components, methods based on PCA and Brunet’s Cophenetic Correlation Coefficient achieved the highest accuracy. When evaluated on a well-known real dataset, normalization had an unpredictable effect on the estimate. For any given normalization method, the methods for estimating <i>k</i> gave widely varying results. We conclude that when estimating <i>k</i>, it is best not to apply normalization. If the underlying components are known to be orthogonal, then Velicer’s MAP or Minka’s Laplace-PCA method might be best. However, when the orthogonality of the underlying components is unknown, none of the methods seemed preferable.
format article
author José M. Maisog
Andrew T. DeMarco
Karthik Devarajan
Stanley Young
Paul Fogel
George Luta
author_facet José M. Maisog
Andrew T. DeMarco
Karthik Devarajan
Stanley Young
Paul Fogel
George Luta
author_sort José M. Maisog
title Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
title_short Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
title_full Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
title_fullStr Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
title_full_unstemmed Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization
title_sort assessing methods for evaluating the number of components in non-negative matrix factorization
publisher MDPI AG
publishDate 2021
url https://doaj.org/article/17df6d63695d43868d77feb0c575755d
work_keys_str_mv AT josemmaisog assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
AT andrewtdemarco assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
AT karthikdevarajan assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
AT stanleyyoung assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
AT paulfogel assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
AT georgeluta assessingmethodsforevaluatingthenumberofcomponentsinnonnegativematrixfactorization
_version_ 1718411365190729728