Maximum likelihood estimation of Gaussian mixture models with different class-specific covariance matrices is known to be problematic. This is due to the unboundedness of the likelihood, together with the presence of spurious maximizers. Existing methods to bypass this obstacle are based on the fact that unboundedness is avoided if the eigenvalues of the covariance matrices are bounded away from zero. This can be done imposing some constraints on the covariance matrices, i.e. by incorporating a priori information on the covariance structure of the mixture components. The present work introduces a constrained approach, where the class conditional covariance matrices are shrunk towards a pre-specified target matrix Ψ . Data-driven choices of the matrix Ψ , when a priori information is not available, and the optimal amount of shrinkage are investigated. Then, constraints based on a data-driven Ψ are shown to be equivariant with respect to linear affine transformations, provided that the method used to select the target matrix be also equivariant. The effectiveness of the proposal is evaluated on the basis of a simulation study and an empirical example.

A data driven equivariant approach to constrained Gaussian mixture modeling

GATTONE, Stefano Antonio;
2017-01-01

Abstract

Maximum likelihood estimation of Gaussian mixture models with different class-specific covariance matrices is known to be problematic. This is due to the unboundedness of the likelihood, together with the presence of spurious maximizers. Existing methods to bypass this obstacle are based on the fact that unboundedness is avoided if the eigenvalues of the covariance matrices are bounded away from zero. This can be done imposing some constraints on the covariance matrices, i.e. by incorporating a priori information on the covariance structure of the mixture components. The present work introduces a constrained approach, where the class conditional covariance matrices are shrunk towards a pre-specified target matrix Ψ . Data-driven choices of the matrix Ψ , when a priori information is not available, and the optimal amount of shrinkage are investigated. Then, constraints based on a data-driven Ψ are shown to be equivariant with respect to linear affine transformations, provided that the method used to select the target matrix be also equivariant. The effectiveness of the proposal is evaluated on the basis of a simulation study and an empirical example.
File in questo prodotto:
File Dimensione Formato  
ADAC.pdf

Solo gestori archivio

Tipologia: PDF editoriale
Dimensione 768.8 kB
Formato Adobe PDF
768.8 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
draft_3rdrevision.pdf

accesso aperto

Tipologia: Documento in Pre-print
Dimensione 399.03 kB
Formato Adobe PDF
399.03 kB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11564/663407
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 10
  • ???jsp.display-item.citation.isi??? 8
social impact