Mixture Model for Contextual Route Choice in Multimodal Transportation Systems

Abstract

The development of personalized travel recommendation systems (TRSs) is crucial in assisting travelers to navigate intricate transportation networks and providing affordable, reliable multimodal mobility options. However, predicting route behavior based on trip data becomes a challenging task because of the diversification of intermodal options and latent factors such as trust in the recommended routes. In this paper, we propose an Inverse Mixture Model (IMM) composed of two interconnected components: (1) a set of dynamic route choice models with arc-to-arc transition functions contingent on the contextual traffic conditions, and (2) mixture probabilities determining the selection of route choice models. There mixture probabilities are influenced by the expected value of following recommended routes, thus emulating travel behavior in dynamic and convoluted traffic networks. Two EM algorithms are tailored for estimating IMMs in the following scenarios: (1) when mixture probabilities are predetermined upon arriving at the origin, and (2) when mixture probabilities vary at each intersection. The effectiveness of IMM and estimation methods are validated through our numerical results, using both synthesized and real-world trip data. The expansion to a mixture of route choice enables the identification of heterogeneous routing behavior at a granular level, thereby facilitating the development of data-driven travel behavior prediction models.

Keywords: Mixture of experts, dynamic route choice models, inverse optimization