Feature Selection in Segmentation Studies

• Activity Number: 586
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and Marketing
• Author(s): Ewa Nowakowska*+
• Companies: GfK Polonia
• Keywords: segmentation ; feature selection ; clusterability ; Fisher discriminant ; spectral dimensionality reduction

Abstract

The process of attribute selection is essential for any segmentation study as it determines the nature of the resulting segments to a great extent. The problem is particularly vital for marketing research where the segmentation solution is frequently the basis for strategic decisions and marketing actions. There exist many different approaches to the topic with the efficiency varied according to the data structure. In this work the main ones are briefly summarized and compared. However, the focus is on a new method that seeks for the set of variables providing the highest distinctness of the clustering structure (clusterability). It is shown, that in the theoretical framework of multivariate Gaussian heteroscedastic mixture model, spectral dimensionality reduction preceded by appropriate data transformation, leads to an easy estimate of the distinctness of the clustering structure. It does not require actual clustering solution which makes it computationally feasible to test and compare clusterability in many different feature subspaces. For illustrative purposes empirical examples are presented and method's applications for marketing research are also discussed.