Integrating neural network-based habit learning into random utility models for transport mode choice

Abstract

Discrete choice models are very powerful models able to incorporate a large range of behaviours. However, the theoretical interpretability that is a key feature of these models often limits the possibility of adequately modelling the behavioural complexity behind the decision process, despite the availability of a rich dataset. An exemplary case of such complexity is the inertia effect. Several functional forms are used in the literature, each based on a specific type of information. This study exploits the potential of Neural Networks to autonomously extract the temporal dynamics of habit, leading to inertia. Specifically, the network extracts a continuous ’history-support’ index that adapts dynamically to the decision environment. These effects are subsequently integrated and tested within a standard, interpretable random utility model, and the results are compared with the same effects measured in the traditional way. In an empirical application using the sixweek Mobidrive panel from Karlsruhe, Germany, the results confirm that the habit component exhibits a large individual effect and significantly improves the explanation of observed choices. More interestingly, it reveals that the learned habit influence exhibits temporal decay and varies significantly by trip purpose, capturing rich behavioural dynamics that traditional static frequency measures fundamentally fail to represent.