Modifying Cluster-Based Segments to Enhance Agreement with an Exogenous Response Variable

Abstract

K-means clustering programs are frequently used to group buyers into market segments. Segment members exhibit similar background profiles, based on such characteristics as psychographics, benefits seeking, conjoint-based partworths, and so on. In addition, researchers also may use data on one or more exogenous variables for the same respondents. The authors describe and apply an algorithm for systematically modifying the original K-means segmentation to enhance prediction of an exogenous variable (either continuous or categorical). The modification is designed to respect a user-specified constraint on the permissible percentage decrease in the initial partition's ratio of among-groups sums of squares to total-groups sums of squares. The authors compare this approach with models drawn from the highly popular clusterwise regression literature.

Emergent clustering methods for empirical OM research

“…Krieger and Green (1996) recognized that a serious drawback of clusterwise regression is a failure to produce clusters that are homogeneous with respect to important descriptor variables. In other words, although excellent prediction of the response variable is achieved via the clustering process, the clusters are not identifiable based on salient descriptive measures.…”

“…Brusco et al (2008) demonstrated the propensity for overfitting the data with clusterwise regression and the challenges of assigning new cases to clusters within the context of these models. In light of some of the challenges in establishing meaningful cluster profiles, Krieger and Green (1996), DeSarbo and Grisaffe (1998), and Brusco et al (2002, 2003) have proposed multicriterion clusterwise regression procedures that attempt to provide a tradeoff between traditional K ‐means criteria for clustering variables and explained variation for response variables. Unfortunately, these models provide an additional difficult decision for the cluster analysis, namely the appropriate weights for the relevant criteria.…”

Scatter tabu search for multiobjective clustering problems

“…The market segmentation problem described by Krieger and Green (1996) is a typical case of partitioning using multiple dissimilarity matrices. The problem consists of finding homogenous clusters in order to plan targeted marketing efforts.…”

Recommending a platform leveraging strategy based on the homogeneous or heterogeneous nature of a product line

“…In this domain, conjoint analysis (Green and Rao 1971) has been very helpful in measuring buyer preferences using part-worths for attribute levels to specify product lines. For market segmentation, cluster (and conjoint) analysis techniques have been used to identify market segments (Krieger and Green 1996). Techniques usually pursue maximisation of the buyer preferences via part-worth and maximisation of the company profit via commonality.…”