A new flexible and interpretable choice model with monotonicity constraints, non-linearity, and taste heterogeneity

Abstract

This study proposes a flexible and interpretable discrete choice model (DCM) capturing key behavioural mechanisms simultaneously: (i) interactions between alternative-specific and individual-specific attributes (e.g., taste heterogeneity), (ii) interactions between alternative-specific attributes, (iii) inherent non-linear utility of alternative-specific attributes (e.g., diminishing marginal utility of travel cost). Deep neural networks (DNNs) have been considered as candidates to flexibly capture these mechanisms, but they fail to provide trustworthy and explainable economic information (i.e., interpretability) obeying domain-specific knowledge (e.g., decrease in utility of travel mode due to an increase in its travel cost). We propose a DCM based on a lattice network (LN) that efficiently imposes attribute-specific monotonicity constraints in the utility specification while ensuring the trustworthy interpretation of DNNs. The proposed LN-based DCM is benchmarked against DNN in a Monte Carlo study. The results show that it outperforms even the parametric DCM in terms of interpretability while slightly underperforming the DNN in terms of predictability.