A new computational travel model produced different simulated policy outcomes from a conventional multinomial logit (MNL) benchmark while keeping its predictive performance close to benchmark levels. Its central move is to let a change in one attribute flow through a learned network of dependencies before recalculating choice probabilities.
What the model adds
Called Neural-BSL, the method jointly learns a directed acyclic graph, or DAG, that maps dependencies among attributes without loops, and a random-utility choice model. The choice outcome is excluded from the graph. Instead, the learned structure enters the utility calculation through interactions, and a modeled change is propagated through downstream nodes in topological order.
The evaluation used two separate, non-pooled datasets. The stated-preference set contained 863 respondents and 5,178 scenario-level observations. Respondent-level splitting assigned 604 respondents to training, 129 to validation and 130 to testing, with 3,624, 774 and 780 observations in those partitions. Every scenario from a respondent stayed in the same partition. The revealed-preference LPMC set contained 81,086 trips from 31,954 individuals in 17,616 households. Its choice set was walking, cycling, public transport and driving.
Prediction stayed close to the benchmarks
In five-fold cross-validation, Neural-BSL reached 0.584 accuracy, a 0.551 weighted F1 score and a mean log-likelihood of -0.916 in the stated-preference data. The corresponding revealed-preference figures were 0.742, 0.729 and -0.675. A feedforward multilayer perceptron, or MLP, had the highest accuracy in both datasets, at 0.593 and 0.746. The model was therefore competitive, but not the accuracy leader.
Every reported level-of-service, or LOS, own-effect in both datasets had the theoretically expected non-positive sign, and the mean utility coefficients agreed with the expected signs in every case. The reported high-support dependencies included license to car in the stated-preference data, with an edge probability above 0.999 and agreement across all 10 particles. In the revealed-preference data, age to license and license to car availability had standardized weights of 0.465 and 0.321. Age to commute purpose and weekend to commute purpose had negative weights of -0.296 and -0.293.
The reported downstream-to-direct contribution ratios were 28.7% for license possession, 23.4% for the male indicator, 21.1% for monthly income and 17.9% for high-school education or below in the stated-preference data. In the revealed-preference data, they were 13.9% for age, 8.5% for weekend travel, 4.5% for license possession, 3.1% for work or education travel and 2.4% for the age-based fare concession. Direct contributions were still larger than the summed downstream contributions for every source attribute in the revealed-preference data.
Simulated policies diverged
The differences appeared in modeled policy scenarios, not just in diagnostic scores. For license holders aged 50 or above in the stated-preference license-surrender scenario, DRT share fell by 0.99 percentage points under Neural-BSL, compared with 0.64 points under MNL. Bus share rose 0.47 points versus 0.87, while subway share rose 0.53 points under Neural-BSL but fell 0.23 points under MNL.
Other stated-preference scenarios also separated the models. One fewer habitual transfer increased bus share by 1.50 percentage points with Neural-BSL and 1.68 with MNL; subway share fell by 0.64 and 0.24 points, and DRT share fell by 0.86 and 1.43 points, respectively. Moving the age category from the 40s to 50 or older reduced bus share by 1.16 points under Neural-BSL and 1.10 under MNL; subway rose 1.85 versus 2.15 points, and DRT fell 0.69 versus 1.05 points. A 20% DRT-fare reduction increased DRT share by 9.79 points under Neural-BSL and 9.51 under MNL.
The revealed-preference scenarios showed the same kind of model dependence. Changing commute purpose reduced public-transport share by 2.96 percentage points under Neural-BSL and 2.06 under MNL, while driving share rose 2.31 and 3.06 points. Setting public-transport fare to zero increased public transport by 1.84 points versus 0.70 and reduced driving by 0.93 versus 0.16. Zero public-transport fare was already present in 32.5% of the trips.
A model, not a settled explanation
Propagation changed the other inputs as well. In the stated-preference license-surrender scenario, car ownership fell by 1.215 standardized units, while one fewer transfer reduced the long-commute indicator by 0.362. The stated-preference age shift increased license possession by 0.336, higher-income status by 0.321 and car ownership by 0.211. In the revealed-preference age scenario, license possession rose by 0.191 and car availability by 0.031, while work or education travel fell by 0.125 and rush-hour travel by 0.070.
The figures are simulated outputs, and the learned arrows are model-based dependencies rather than established causal findings. The authors identify dependence on structural assumptions and tuning choices, unresolved directions in cross-sectional data and fewer distinct respondent profiles for structure learning in the repeated-task stated-preference survey.
The LPMC data and Neural-BSL source code are publicly available, while the Seoul stated-preference data are not publicly redistributable. The work reports support from the National Research Foundation of Korea, funded by Korea's MSIT, under grant No. RS-2025-00520515.
Paper data and sources
Original title: Neural-Bayesian Structure Learning for Discrete Choice Modeling
Authors: Hyunsoo Yun, Eun Hak Lee, Jiaru Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text