Robust model-based clustering with covariance matrix constraints

The Optimally Tuned Robust Improper Maximum Likelihood Estima- tor (OTRIMLE) for robust model-based clustering is introduced. It is based on a ML-type procedure for a pseudo model in which clusters are represented by a finite mixture of Gaussian distributions, while noise is represented with the addition of an improper constant density (ICD). The OTRIMLE requires constraints on the underly- ing covariance matrices that prevent spurious solutions. These constraints may have strong impact on the final clustering and alternative algorithms are provided with the OTRIMLE software.