Model choice for binned-EM algorithms of fourteen parsimonious Gaussian mixture models by BIC and ICL criteria

Abstract

Choosing the right model is an important step in model-based clustering approaches. In this framework, BIC and ICL criteria were proposed to choose a model for clustering of standard data. On the other hand, in order to accelerate the data processing when using EM algorithm, this algorithm was adapted to binned data (binned-EM algorithm). Then fourteen binned-EM algorithms of fourteen parsimonious Gaussian mixture models were developed to replace the binned-EM algorithm of the most general Gaussian mixture model when data have a simple structure. So this paper studies the application of BIC and ICL criteria to select a good model which better fits binned data, when clustering is based on these fourteen binned-EM algorithms. Numerical experiments on simulated and real data are performed, and the experimental results are analyzed.