Robust model-based clustering via cellwise approach
TCLUST is a robust model-based methodology for obtaining unsupervised classification. It corresponds to a Maximum Likelihood Estimator, which is modified for getting robust behavior via the joint application of trimming and constraints. TCLUST generalizes to multiple populations the Minimum Determinant of the Covariance Matrix Estimator (MCD), a robust procedure to estimate multivariate location and scatter. The Robustness of TCLUST and MCD are related to avoid the influence of casewise outliers, which are atypical individuals. However, they are not expected to work under the presence of cellwise contamination corresponding to a relatively low percentage of cells. CellMCD is able to avoid the effect of cellwise contamination when estimating multivariate location and scatter. By generalizing cellMCD to the k populations case, we propose a version of TCLUST for coping cellwise contamination, cellGMM. The presentation will provide details about cellGMM and its robust performance.
Keywords: clustering robustness cellwise contamination trimming