Assessing Robust Clustering for Gaussian Mixtures in the Presence of Outliers and Noise Variables with Component Overlap-A Simulation Study

DOI: 10.5176/2251-1938_ORS17.23

Authors: Fikriye Kabakci and Umashanger Thayasivam

Abstract:

Model based clustering based on finite mixture of Gaussian distributions [1,2,3,4] is being used frequently in recent times. This article presents a comparison study of robustness for clustering with L2E [5,6,7,8], Expectation Maximization Algorithm (EM) [9,10] and robust model based clustering method which is known as Robust Improper Maximum Likelihood Estimation (RIMLE) [11,12,13]. In order to make comparison of their robustness, we employ MixSim [14] package in R to simulate mixtures of Gaussian distributions. We examine the results for different type of noise variables, when presence of outliers with different levels of overlap between clusters.

Keywords: component; Gaussian mixture model; L2E estimation; Expectation Maximization; Uniform Noise; Maximum Likelihood Estimation (MLE);MixSim;R;RIMLE