|
Activity Number:
|
66
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Sunday, August 2, 2009 : 4:00 PM to 5:50 PM
|
|
Sponsor:
|
IMS
|
| Abstract - #304639 |
|
Title:
|
Robust model-based clustering using an improper constant density selected by a distance optimization for non-outliers
|
|
Author(s):
|
Pietro Coretto*+ and Christian Hennig
|
|
Companies:
|
University of Salerno and University College London
|
|
Address:
|
Via Ponte Don Melillo, Fisciano, International, 84084, Italy
|
|
Keywords:
|
Model-based clustering ; robust improper maximum likelihood estimator ; improper density ; outliers, noise
|
|
Abstract:
|
In this paper we present a robust model-based clustering method. Gaussian mixtures are a popular tool used to represent elliptical-shaped clusters. The parameters of a Gaussian mixture can be estimated by maximum likelihood (ML). However, a small proportion of outliers (or noisy points), i.e. points not consistent with the Gaussian mixture model, can affect the ML estimator and the corresponding clustering severely. Hennig (2004) and Coretto (2008) have introduced and studied a robust alternative where the presence of outliers and points not belonging to any cluster is treated in a model-based fashion. We add a component with an improper constant density to teh mixture, and this plays the role to represent ``noise'. The present paper addresses the crucial issue of how to fix the improper constant density in a multidimensional setup.
|
