Beyond i.i.d: complex Random Utility Model specifications with gradient boosting

Abstract

This paper extends RUMBoost, a novel discrete choice modelling approach that combines the interpretability and behavioural robustness of Random Utility Models with the generalisation and predictive ability of deep learning methods, to complex RUM specifications. With RUMBoost, we obtain non-linear pseudo-utilities in the form of piece-wise constants by replacing each linear parameter in the utility functions of a RUM with an ensemble of gradient boosted regression trees. We further use an optimisation-based smoothing technique to identify non-linear utility functions with defined gradients from the piece-wise constants. This allows for the estimation of behavioural indicators such as the Value of Time (VoT) or the willingness to pay. Finally, we demonstrate how RUMBoost can mimic the estimation of complex model specifications with a case study on a mode choice dataset. This is achieved by adapting the probability function to account for alternative correlations in the error term.