Multiple Scaled Contaminated Normal Distribution and Its Application in Clustering

TL;DR

This paper introduces the multiple scaled contaminated normal (MSCN) distribution, a flexible model for robust multivariate clustering that accounts for dimension-specific outliers and improves detection and estimation in contaminated data.

Contribution

It proposes the MSCN distribution with dimension-specific contamination parameters and an EM algorithm extension for robust clustering, advancing outlier detection and parameter estimation.

Findings

MSCN effectively detects bad points in each dimension.

The model improves clustering robustness over existing heavy-tailed distributions.

Simulation and real data demonstrate superior performance in contaminated scenarios.

Abstract

The multivariate contaminated normal (MCN) distribution represents a simple heavy-tailed generalization of the multivariate normal (MN) distribution to model elliptical contoured scatters in the presence of mild outliers, referred to as "bad" points. The MCN can also automatically detect bad points. The price of these advantages is two additional parameters, both with specific and useful interpretations: proportion of good observations and degree of contamination. However, points may be bad in some dimensions but good in others. The use of an overall proportion of good observations and of an overall degree of contamination is limiting. To overcome this limitation, we propose a multiple scaled contaminated normal (MSCN) distribution with a proportion of good observations and a degree of contamination for each dimension. Once the model is fitted, each observation has a posterior probability of being good with respect to each dimension. Thanks to this probability, we have a method for simultaneous directional robust estimation of the parameters of the MN distribution based on down-weighting and for the automatic directional detection of bad points by means of maximum a posteriori probabilities. The term "directional" is added to specify that the method works separately for each dimension. Mixtures of MSCN distributions are also proposed as an application of the proposed model for robust clustering. An extension of the EM algorithm is used for parameter estimation based on the maximum likelihood approach. Real and simulated data are used to show the usefulness of our mixture with respect to well-established mixtures of symmetric distributions with heavy tails.