A Bayesian Approach for Clustering Skewed Data Using Mixtures of Multivariate Normal-Inverse Gaussian Distributions

• Activity Number: 281 - Clustering with Mixtures: Towards Emerging Data Types
• Type: Invited
• Date/Time: Tuesday, July 31, 2018 : 8:30 AM to 10:20 AM
• Sponsor: The Classification Society
• Author(s): Sanjeena Dang*
• Companies: Binghamton University (SUNY)

Abstract

Multivariate normal inverse Gaussian (MNIG) distributions possess an appealing property that they can represent symmetric as well as skewed populations with computational simplicity. This makes MNIG very attractive for model-based clustering. The MNIG model arises from a mean-variance mixture of a multivariate normal distribution with the inverse Gaussian distribution. A Bayesian approach using Gibbs sampler which is an alternate approach to the traditional EM algorithm, is extended here for mixtures of MNIG models. A novel approach to simulating from matrix generalized-inverse Gaussian (MGIG) distribution is also discussed. Application on simulated data sets with symmetric and skewed subpopulations as well as a real data set is presented.