Assessing Clusterability in a Model-Based Framework

• Activity Number: 683
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and Marketing
• Author(s): Ewa Nowakowska*+
• Companies: GfK Polonia
• Keywords: clusterability ; cluster analysis ; model-based approach ; Gaussian mixture models

Abstract

The notion of clusterability captures the ability of a given data set to be clustered, facilitating the assessment of chances for a stable and actionable clustering solution. It is particularly vital when some flexibility in the feature space selection is granted. In the theoretical framework of multivariate Gaussian heteroscedastic mixture model, we propose to characterize and quantify the level of clusterability in an intrinsic way, which does not take into account a specific clustering method choice. We then evaluate the efficiency of the corresponding measure introduced in a data set setup and show its consistency with the theory. The applications in market research are also discussed and illustrated with typical examples.