Abstract Manager

Travel decision modeling plays a critical role in understanding and predicting people’s travel behavior to support sustainable mobility solutions. Traditional Discrete Choice Models (DCMs), such as the Multinomial Logit (MNL) and Mixed Logit models, have provided valuable insights through their theoretical foundation and interpretability. However, these models often struggle to capture complex, nonlinear relationships within large-scale datasets, limiting their predictive accuracy and applicability to evolving multimodal networks. In contrast, data-driven approaches like machine learning can efficiently uncover complex patterns and nonlinear interactions in data with high predictive power. However, these methods also face some challenges, and the most significant is a lack of interpretability. This paper introduces a new hybrid neural-utility framework to address these challenges that bridge the gap between theory-driven and data-driven approaches. The framework integrates the discrete choice utility function into the neural network architecture, allowing it to retain the interpretability of utility-based models while leveraging the nonlinear modeling power of machine learning. The model is optimized using evolutionary algorithms, regularization techniques, and bootstrap methods to ensure stable parameter estimation and robust performance across diverse datasets. Evaluated by two distinct transportation datasets, the model demonstrates significant improvements in predictive accuracy, computational efficiency, and explanatory power compared to traditional DCMs. Moreover, an experiment is conducted to compare the performance of the full proposed model with and without the non-linear part. The results show that the intricate cross-effects among travel attributes significantly influence decision-making. That would provide actionable insights for planners and policymakers. Besides, the proposed model also exhibits strong generalization capabilities for unseen data, outperforming traditional models that rely heavily on in-sample accuracy. When generalization is not prioritized, the model also showcases exceptional performance on training datasets, highlighting its adaptability and robustness. Therefore, this work represents a step forward in integrating artificial intelligence into travel decision modeling, providing an innovative tool to address the complexity of network modeling for multimodal mobility systems.