A methodology for the characterization of business-to-consumer E-commerce
This thesis concerns the field of business-to-consumer electronic commerce. Research on Internet consumer behaviour is still in its infancy, and a quantitative framework to characterize user profiles for e-commerce is not yet established. This study proposes a quantitative framework that uses latent variable analysis to identify the underlying traits of Internet users' opinions. Predictive models are then built to select the factors that are most predictive of the propensity to buy on-line and classify Internet users according to that propensity. This is followed by a segmentation of the online market based on that selection of factors and the deployment of segment-specific graphical models to map the interactions between factors and between these and the propensity to buy online. The novel aspects of this work can be summarised as follows: the definition of a fully quantitative methodology for the segmentation and analysis of large data sets; the description of the latent dimensions underlying consumers' opinions using quantitative methods; the definition of a principled method of marginalisation to the empirical prior, for Bayesian neural networks, to deal with the use of class-unbalanced data sets; a study of the Generative Topographic Mapping (GTM) as a principled method for market segmentation, including some developments of the model, namely: a) an entropy-based measure to compare the class-discriminatory capabilities of maps of equal dimensions; b) a Cumulative Responsibility measure to provide information on the mapping distortion and define data clusters; c) Selective Smoothing as an extended model for the regularization of the GTM training.