Cluster Pruning: Finding a Better Cluster Representative Object by Dimension Reduction

• Activity Number: 44
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Amy Wagaman*+
• Companies: Amherst College
• Keywords: cluster analysis ; dimension reduction ; representative object ; multidimensional scaling

Abstract

Cluster analysis is a significant research area with many applications. While new clustering methods and cluster validation have been a focus of substantial research work, extracting a good representative cluster object appears to have received less attention. In this article, we propose a method to prune clusters obtained from any non-fuzzy clustering method, eliminating unusual cluster members in order to obtain a better representative object from the cluster. The method uses dimension reduction to identify the unusual cluster objects which are then removed. We demonstrate the method via simulations and with applications. One application is extracting protein potential native structures after clustering frames from a molecular dynamics simulation. In this application, averaging frames in a cluster to produce a representative frame could result in a nonsensical protein structure, so extracting a representative frame is important.