Neural-Bayesian Structure Learning Enhances Discrete Choice Modeling.

Researchers have introduced Neural-Bayesian Structure Learning (Neural-BSL), a novel framework designed to improve discrete choice modeling by integrating differentiable structure learning with traditional random-utility-based estimation. Unlike conventional models that treat explanatory variables as independent inputs, Neural-BSL internally determines how related attributes should adjust when one is intentionally changed. The framework learns the underlying dependency structure among attributes and jointly estimates random-utility parameters. This learned structure is then used to propagate the effects of interventions: when an attribute is modified, its model-implied downstream changes are computed in topological order, leading to updated utilities and choice probabilities. This provides a more nuanced understanding of how choices shift. Evaluated using real-world stated-preference data from Seoul and revealed-preference data from London, Neural-BSL demonstrated predictive performance on par with established benchmarks. Crucially, it also recovered behaviorally coherent dependency structures, offering insights into the mechanisms behind predicted mode-share responses and the associated adjustments in traveler or trip attributes.