Robust Constrained Clustering in Presence of Entry-Wise Outliers

We propose a robust heteroscedastic model-based clustering method based on snipping. An observation is snipped when some of its dimensions are discarded, but the remaining are used for estimation. An expectation-maximization algorithm augmented with a stochastic optimization step is used to derive inference, and its convergence properties are studied. We show global robustness of our resulting sclust procedure also when outliers arise entry-wise. The method is robust to contamination, even when most or even all of the observations contain outliers. Simulations and two real data applications illustrate and compare the approach with existing methods.