Christian Hennig - (University College London)
Abstract
This presentation introduces the OTRIMLE estimator for robust clustering. OTRIMLE stands for "optimally tuned robust improper maximum likelihood". The idea is to fit a Gaussian mixture model to the data that includes a "noise component" to catch observations that should not be assigned to any cluster. This idea goes back to Banfield and Raftery (1993), who proposed a uniform distribution over the convex hull of the data as noise component. We use a constant pseudo-density over the whole Euclidean space. This has the advantage of improved robustness.
I will present consistency and robustness theory and the computation of OTRIMLE, and I will discuss in some depth a simulation study to compare OTRIMLE with other methods for robust clustering, which involves some remarks on a proper problem definition and on unification of the methods to make them more comparable.
Contact person
Cinzia Viroli