Merging Mixture Components for Model-Based Clustering

• Activity Number: 245
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Volodymyr Melnykov*+
• Companies: The University of Alabama
• Keywords: finite mixture model ; model-based clustering ; misclassification probability

Abstract

Model-based clustering is a popular technique based on finite mixture models. It is traditionally assumed that each group of data points can be adequately modeled by a single mixture component. In this case, there exists an appealing correspondence between components and clusters. Unfortunately, more than one component might be needed for modeling a group of data. If this happens, model-based clustering loses its attractive interpretation. One possible remedy is to merge mixture components for clustering. A new merging approach based on misclassification probabilities is proposed and carefully illustrated.