To Merge or Not to Merge: An Interactive Visualization Tool for Local Merges of Mixture Model Components

• Activity Number: 79
• Type: Contributed
• Date/Time: Sunday, August 3, 2014 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Graphics
• Author(s): Elizabeth Lorenzi*+ and Rebecca Nugent and Nema Dean
• Companies: Carnegie Mellon and Carnegie Mellon and University of Glasgow
• Keywords: model-based clustering ; component tree ; merge criteria ; interactive visualization

Abstract

Model-based clustering (MBC) is a common clustering technique that fits mixture models to data, identifying and characterizing the subpopulations in the underlying population (Fraley, Raftery 2002). Typically, multivariate Gaussian distributions are used to represent each subpopulation. Sometimes MBC can overestimate the number of clusters, fitting more components than needed in an effort to better approximate the underlying density. Recent merge criteria, such as ridgeline and entropy (Ray, Lindsay 2005; Hennig 2010; Baudry et al, 2010), can provide insights into whether to merge components to better match the true cluster structure. We present mixture model component trees, a tool that calculates inter-component similarity and displays the corresponding hierarchical structure in a dendrogram. This approach provides an interactive visualization tool for the cluster and component structure of data of any dimensionality. Clicking on the tree returns the summary statistics and locally-estimated merge criteria of the corresponding components. With our tool, users can explore subsets of the data (rather than the entire data set), providing insights into how to merge locally.