This article is cited in 3 papers

Selecting the dimensionality for mixture of probabilistic principal component analyzers

M. P. Krivenko

Institute of Informatics Problems, Federal Research Center "Computer Science and Control" of the Russian Academy of Sciences, 44-2 Vavilov Str., Moscow 119333, Russian Federation

Abstract: The article considers the problems of choosing structural parameters characterizing the model of a mixture of probabilistic principal component analyzers, namely, the number of elements of the mixture and the dimensions of these elements. Among the set of approaches used in practice for the task of classifying data, only sampling management methods are actually available. To implement the choice of dimensions, it is proposed to use a combination of the known methods for model selecting. The mixture of probabilistic principal component analysis allows one to model bulk data using a relatively small number of free parameters. The number of free parameters can be controlled by selecting the latent dimension of the data.

Keywords: probabilistic principal component analysis (PPCA), mixtures of PPCA, model selection criterion, bootstrap, cross-validation.

Received: 16.07.2019

DOI: 10.14357/08696527190301