Bayesian network clustering and self-organizing maps under the test of Indian districts. A comparison

Abstract

This paper compares Hierarchical Clustering, Bayesian Networks and Self-Organizing Map Neural Networks (SOM and superSOM) approaches used for clustering purposes in geographic space. The same dataset, covering the Republic of India and made of 55 indicators for 640 spatial units (administrative districts), is used in the three analyses. Indicators descry the several aspects of urban, economic and socio-demographic development in India. Bayesian Networks use a likelihood function while SOM/SuperSOM and Hierarchical Clustering minimize variance of Euclidean distance in variable space, the former by preserving the topological properties within the output space and the latter by successively combining similar items. Relatively similar multi-step protocols have been implemented for the three techniques, to take into account variable redundancy. Methods as well as clustering results are compared. From this perspective, the aim of the paper is to highlight the similarities between the protocols and to evaluate the differences between the segmentation approaches (geographical and variable space comparisons). A few key points are also discussed such as the data pre-processing steps, the conception of latent factors and the choice of the number of clusters.