Work

Assessment of the number of components in a mixture. Application to classification?

Year: 1994
Type: dissertation
Abstract: This thesis is concerned with Mixture Distributions and Classification. Basing Cluster Analysis within the context of multivariate Mixture Models provides a fruitful line of approach for finding the right number of clusters in a given data set. Estimating the number of components (k) in a mixture of distributions is a difficult problem for which there several competing approaches, since the traditional likelihood ratio test can not be used under the mixture context. We propose a new procedure based on the SEM algorithm for assessing k. We have also introduced a new prior, called empirical prior, for the approach of Aitkin-Rubin. Then we performed a simulation study for comparing all the approaches under the same underlying model. In the cluster analysis point of view, we propose and entropy criterion (NEC) which aims to choose the mixture model from the which a clustering structure arises with the greatest evidence. This criterion allows us to analyze the relation between the mixture model and cluster analysis. We derive some asymptotic calculations describing the limit behavior of NEC in the context of simple univariate gaussian mixtures. The numerical Monte-Carlo experiments display also good performance of NEC
Author: Pereira, Gilda Maria C. F. Soromenho
Language: English
Open Access status: green
OpenAlex ID: W7017239241
Cites:
Cited by:
Related to: 10