Applying random forest in discrete choice modeling: a case study of household car allocation

Abstract

This paper proposes a novel approach to discrete choice model estimation, which we term a methodological-iterative (MI) approach. This approach utilizes the results of an ensemble learner, in this case the random forest classifier, to incorporate an algorithm of recursive feature elimination to utility function specification and model estimation. The method is applied to a case study of car allocation among household members.

The MI method reduces the model specification efforts required considerably, can reveal significant explanatory variables that may be overlooked, and may prove helpful when the modeler lacks the behavioral insight of the dataset or the phenomenon being investigated or when the dataset is very large, as it is estimated by an algorithm rather than by the conventional trial and error approach. The method is a hybrid tool of data-driven machine learning and a theory-driven discrete choice model to obtain an improved model and forecast. As such, it harnesses some of the advantages of a powerful data-driven machine learning classifier while obtaining the interpretability of discrete choice models.

The MI method is implemented in this paper to investigate car allocation choice behavior. The MI car allocation choice model produced a forecast that is similar to the data-driven machine learning classifiers prediction wise, but with full interpretability.