Enhanced utility estimation algorithm for discrete choice models in travel demand forecasting

Recent data-driven discrete choice models in travel demand forecasting have achieved improved predictability. However, such prediction improvements come at the cost of black-box models and lost transparency in travel demand forecasting, which makes scenario testing and transportation planning difficult (if not impossible). Furthermore, these predictability gains have often been modest compared to handcrafted parsimonious models, which benefit from enhanced behavioural interpretability and transparency. This paper introduces a novel bi-level model and estimation framework (DUET) to enhance predictability in traditional utility-based discrete choice models. The proposed model improves the specification process by identifying effective variable transformations and interactions in utility functions. Utilising a genetic algorithm, the upper level of our framework selects feasible functional forms from an extensive array, while the lower level applies iterative singular value decomposition and maximum likelihood estimation to optimise model parameters and prevent overfitting. This approach ensures superior predictability through a general utility functional form that considers extensive variable interactions. Case studies on both synthetic data and the Swissmetro dataset highlight the framework’s effectiveness in improving model performance and uncovering critical behavioural patterns and latent trends. Notably, incorporating interactions among variables in Swissmetro data, our model demonstrates a 6.5% improvement in the Brier score (probabilistic prediction accuracy) compared to the state-of-the-art deep neural network-based discrete choice model.Lastly, our results on transport mode choice data align with existing literature, indicating that younger individuals are less sensitive to travel costs. This confirms the need for targeted pricing policies to encourage public transit use among different age groups.