Neural Modelling of Ranking Data with an application to stated preference data Catherine KRIER

Although neural networks are commonly encountered to solve classification problems, ranking data present specificities which require adapting the model. Based on a latent utility function defined on the characteristics of the objects to be ranked, the approach suggested in this paper leads to a perceptron-based algorithm for a highly non linear model. Data on stated preferences obtained through a survey by face-to-face interviews, in the field of freight transport, are used to illustrate the method. Numerical difficulties are pinpointed and a Pocket type algorithm is shown to provide an efficient heuristic to minimize the discrete error criterion. A substantial merit of this approach is to provide a workable estimation of contextually interpretable parameters along with a statistical evaluation of the goodness of fit. The research underlying this paper is performed within the framework of the second Scientific Support Plan for a sustainable Development Policy (SPSP II) for the account of the Belgian State, Prime Minister's Office – Federal Science Policy Office (contract: CP/17/361). Financial support from the IAP research network no P5/24 of the Belgian Government (Federal Office for Scientific, Technical and Cultural Affairs) is also acknowledged. Comments by an anonymous referee and by several participants of seminars where this paper has been presented are gratefully acknowledged.