A Robust EM Clustering Approach: ROBEM

Abstract

Cluster analysis is defined as a group of multivariate statistical methods that are used to classify identical, or similar units. As is the case with all other classical statistical methods, classical clustering analysis gives misleading results when there is an outlier in the multivariate data set. To solve this problem many approaches have been proposed. This study focuses on developing a new approach, aiming to make the expectation maximization (EM) clustering algorithm resistant to outliers. We proposed a new robust hybrid clustering algorithm called robust EM (ROBEM) to reach our aim. This algorithm combines the EM clustering algorithm with robust principal component analysis (ROBPCA) algorithm. Spatial EM algorithm was proposed as a robust EM algorithm in the literature, but our simulation results and sample data applications showed that the ROBEM algorithm was more successful than the spatial EM algorithm in terms of outlier detection rate and faulty classification rate. Moreover, the proposed algorithm ROBEM provides similar results to the other well known robust clustering algorithms, such as TCLUST and Trimmed k-Means. © 2020 Taylor & Francis Group, LLC.