Segmentation-based competitive analysis with MULTICLUS and topology representing networks

Abstract

Two neural network approaches, Kohonen's self-organizing (feature) map (SOM) and the topology representing network (TRN) of Martinetz and Schulten are employed in the context of competitive market structuring and segmentation analysis. In an empirical study using brands preferences derived from household panel data, we compare the SOM and TRN approach to MULTICLUS, a parametric latent vector multi-dimensional scaling (MDS) model approach which also simultaneously solves the market structuring and segmentation problem. Our empirical analysis shows several benefits and shortcomings of the three methodologies under investigation. As compared to MULTICLUS, we find that the non-parametric neural network approaches show a higher robustness against any kind of data preprocessing and a higher stability of partitioning results. As compared to SOM, we find advantages of TRN which uses a more flexible concept of adjacency structure. In TRN, no rigid grid of units must be prespecified. A further advantage of TRN lies in the possibility to exploit the information of the neighborhood graph for adjacent prototypes which supports ex-post decisions about the segment configuration at both the micro and the macro level. However, SOM and TRN also have some drawbacks as compared to MULTICLUS. The network approaches are, for instance, not directly accessible to inferential statistics. Our empirical study indicates that especially TRN may represent a useful expansion of the marketing analyst's tool box. Scope and purpose Determination of competitive market structure among rival brands and market segmentation represent well-known concepts in strategic marketing planning. During the last decade, approaches that combine the two interrelated tasks into one single model have been introduced into marketing literature. Most of them respect consumer heterogeneity by including ‘fixed’ parameters (e.g., demographic or past purchase behavior variables) for each individual or by assuming consumer parameters to be distributed according to a (mixture of) probability distribution(s). However, the key to the success of these statistical modeling approaches seems to lie in the proper choice of parametric model assumptions and/or heterogeneity distributions. Due to its non-parametric nature, the neuro-computing methodology presented in this article imposes less rigorous assumptions on data properties and derives segment-specific patterns of competitive relationships between brands in a purely data-driven way.