Incorporating domain knowledge in deep neural networks for discrete choice models

This paper explores the integration of domain knowledge into deep neural network (DNN) models to support the interpretability of travel demand predictions in the context of discrete choice models (DCMs). Traditional DCMs, formed as random utility models (RUM), are widely employed in travel demand analysis as a powerful theoretical econometric framework. But, they are often limited by subjective and simplified utility function specifications, which may not capture complex behaviors. This led to a growing interest in data-driven approaches. Due to their flexible architecture, DNNs offer a promising alternative for learning unobserved non-linear relationships in DCMs. But they are often criticized for their “black box” nature and potential deviations from established economic theory. This paper proposes a framework that incorporates domain knowledge constraints into DNNs, guiding the models toward behaviorally realistic outcomes while retaining predictive flexibility. The framework's effectiveness is demonstrated through a synthetic dataset and an empirical study using the Swissmetro dataset. The synthetic study confirms that domain knowledge constraints enhance consistency and economic plausibility, while the Swissmetro application shows that constrained models avoid implausible outcomes, such as negative values of time, and provide stable market share predictions. The proposed approach is independent of the model structure, making it applicable on different model architectures. The methodology was applied on both standard DNN and an alternative-specific utility DNN (ASU-DNN). Although constrained models exhibit a slight reduction in predictive fit, they generalize better to unseen data and produce interpretable results. This study offers a pathway for combining the flexibility of machine learning with domain expertise for DCMs, across diverse model architectures and datasets.