Mixture Model Component Trees: A Visualization Tool for Merging Clusters

• Activity Number: 308
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Rebecca Nugent*+ and Nema Dean
• Companies: Carnegie Mellon University and University of Glasgow
• Keywords: Cluster Analysis ; Finite MIxture Models ; Combining Components ; Curvilinear Clusters ; Model-based Clustering

Abstract

Model-based clustering is a popular parametric cluster analysis method based on fitting mixture models of (often) Gaussian densities. Decisions about the number of mixture components can be made automatically using information criteria. Usually each component is assumed to represent a distinct cluster, an estimate of a population group. If a group is skewed or not normally distributed, more than one component may be fit to that group. To summarize the hierarchical structure of the mixture components (and perhaps identify a combination of components that better represent a group) the density estimate from the mixture model is used to create a (dis)similarity measure between the components. This measure is then used as input to an agglomerative hierarchical clustering algorithm; the resulting dendrogram, the mixture model component tree, can be used as a tool to merge clusters as needed.