Discrete Choice Model with Generalized Additive Utility Network for Interpretable Policy Evaluation

Discrete choice models (DCMs) are widely used for prediction and policy analysis, but standard multinomial logit (MNL) models with linear utility may fail to capture nonlinear response patterns in travel behavior. Neural extensions, such as deep neural network with alternative-specific utility functions (ASU-DNN), improve predictive flexibility, but their learned utility representations can be difficult to interpret and may be less stable under counterfactual shifts. This study proposes the generalized additive utility network (GAUNet) and its interaction variant, generalized additive and interactive utility network (GAIUNet), which represent systematic utility as a low-dimensional additive decomposition of learned neural shape functions. The proposed models are evaluated using synthetic transport mode choice data and real-world probe-person data collected in Tokyo, Japan, from 2018 to 2021. In synthetic experiments, GAUNet achieves predictive performance comparable to ASU-DNN and better than linear MNL, while yielding more stable policy evaluation under counterfactual changes in travel cost and access time. In real-world data, GAUNet and GAIUNet outperform linear MNL and remain competitive with ASU-DNN, while producing utility shapes that are easier to inspect at the attribute level. The results suggest that the proposed framework can recover interpretable nonlinear utility patterns without requiring cutoff locations to be specified in advance, thereby providing a transparent and flexible alternative to more entangled neural DCMs.