choice models ; data mining ; econometric models ; hierarchical Bayes analysis ; marketing tools ; regression and other statistical techniques;
The authors present a framework within which robust models of preferences are computationally efficiently estimated using quadratic optimization methods. Within this framework, general highly non-linear models can be computationally efficiently estimated, while at the same time avoiding overfitting problems that such models typically have.They compare these models with standard logistic regression and recently proposed polyhedral conjoint methods. The robustness of the methodology enables the estimation of preference models when the data is noisy and the number of attributes describing the choices is very large while the amount of example choices - past information - is small.The approach can therefore be useful also for analyzing data from very small questionnaires regarding multi-attribute choices with large numbers of attributes, or data that are noisy like, for example, that describing users' choices or clicks on the Internet. The methods discussed in this paper are in the spirit of recently suggested polyhedral algorithms for conjoint estimation, and they bring together well founded tools of statistical learning theory as well as polyhedral optimization theory to the field of preference modeling.