🤖 machine learning

Neural-Bayesian Structure Learning for Discrete Choice Modeling

This paper introduces Neural-Bayesian Structure Learning (Neural-BSL), a novel framework that jointly learns differentiable attribute dependency structures and random-utility parameters to enable more behaviorally coherent discrete choice modeling and accurate prediction of downstream attribute adjustments under policy interventions.

Original authors: Hyunsoo Yun, Eun Hak Lee, Jiaru Zhang, Ziran Wang, Eui-Jin Kim

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hyunsoo Yun, Eun Hak Lee, Jiaru Zhang, Ziran Wang, Eui-Jin Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, millions of people make a simple decision: how to get from home to work. They weigh time, cost, and comfort, choosing between a bus, a train, a car, or a ride-share. Transportation planners rely on models to understand these choices, hoping to predict how a new subway line or a change in bus fares will shift the crowd. For decades, these models have treated a traveler's life as a flat list of facts: their age, income, and whether they own a car are all fed into a computer as separate, unrelated inputs. The computer then calculates which option looks best. This approach works well for guessing what people did in the past, but it struggles when asked to imagine the future. It cannot easily answer what happens if a policy changes one part of a person's life, because in reality, changing one thing often changes the rest. If a city raises the price of gas, people might not just drive less; they might also stop buying a second car, or they might move to a neighborhood closer to work. Traditional models often miss these ripple effects, treating each piece of information as if it exists in a vacuum.

A team of researchers has developed a new way to look at these decisions, one that maps the hidden connections between a traveler's life details before predicting their choice. They call their method Neural-Bayesian Structure Learning. Instead of just asking a computer to guess the best mode of transport, they first ask it to learn the story of how a person's life is put together. The researchers discovered that attributes like income, education, and car ownership are not just random facts sitting side by side; they are linked in a chain. A person's age might influence whether they get a driver's license, and having a license might influence whether they buy a car. By uncovering these hidden chains, the new model can simulate what happens when a policy changes one link. If a policy makes it harder to get a license, the model understands that this will likely lead to fewer people owning cars, which in turn changes how they choose to travel. This allows planners to see not just the final choice, but the entire chain of adjustments a person might make along the way.

The researchers tested this idea using two very different sets of data. The first came from a survey of commuters in Seoul, South Korea, where people were asked to imagine their travel choices under different scenarios. The second came from real-world records of over 80,000 trips made by people in London. In both cases, the new model performed just as well as the best existing methods at predicting what people actually chose. However, when the researchers used the model to simulate policy changes, the results were strikingly different. In the Seoul data, they simulated a scenario where older adults gave up their driver's licenses. A traditional model predicted that these people would simply switch to buses. The new model, however, saw that giving up a license was part of a larger shift in their mobility resources. Because the model understood that losing a license often meant losing access to a private car, it predicted that these travelers would not just take the bus, but would also shift significantly toward the subway, a mode that offers higher capacity and reliability. The traditional model missed this nuance because it treated the license and the car as separate, unrelated facts.

In the London data, the differences were even more pronounced when looking at age. When the researchers simulated a ten-year increase in age for a group of travelers, the traditional model predicted a very small change in how they traveled. The new model, which understood that older age is linked to having a driver's license and owning a car, predicted a much larger shift. It foresaw that as people age, they are more likely to have these mobility resources, which in turn makes them much more likely to drive rather than take public transit. The model also showed that when a trip's purpose changes from work to something else, the timing of the trip and the distance traveled adjust automatically within the simulation, leading to a different choice of transport. These adjustments happen because the model follows the logical flow of a person's life, rather than just swapping out a single number in a list.

The researchers found that for many important attributes, such as having a driver's license or being in a certain age group, a significant portion of their influence on travel choices comes through these downstream connections. For example, in the Seoul study, the influence of having a driver's license on travel choice was about 29% stronger when the model accounted for how that license led to car ownership. In the London study, age influenced choices through a similar chain of effects involving licenses and car availability. This means that the "downstream" effects are not just small details; they are a major part of how people make decisions. The model successfully captured these relationships without the researchers having to guess the connections beforehand. Instead, the computer learned the structure of these relationships directly from the data, creating a map of how a person's life details are connected.

This approach offers a clearer picture for policymakers who need to plan for the future. When a city considers a new fare policy or a change in service, the old way of thinking might suggest a simple shift in ridership. The new method suggests that the change will trigger a series of adjustments in a traveler's life that ultimately lead to a different outcome. It does not claim to know the absolute truth about every individual's mind, but it provides a much more realistic simulation of how changes in one part of a system can ripple through the rest. By treating a traveler's life as an interconnected web rather than a list of isolated facts, the researchers have built a tool that can see further into the consequences of policy decisions. This allows planners to anticipate not just the immediate reaction to a change, but the deeper, structural adjustments that people will make in response.

Technical Summary: Neural-Bayesian Structure Learning for Discrete Choice Modeling

1. Problem Statement

Travel mode choice modeling is central to transportation planning, yet conventional approaches face a fundamental limitation regarding policy simulation. Standard discrete choice models (e.g., Multinomial Logit, MNL) and modern machine learning benchmarks (e.g., Neural Networks, Random Forests) are primarily estimated from observational data. They treat explanatory covariates (e.g., income, vehicle ownership, age) as parallel, independent inputs. Consequently, these models lack an internal mechanism to determine how related attributes should adjust when one is deliberately changed.

This creates a disconnect between prediction and explanation. While a model may accurately predict mode shares under observed conditions, it fails to represent the causal response to intervention. For instance, a policy simulation that changes income while holding vehicle ownership fixed ignores the downstream adjustment in vehicle ownership that would naturally occur if income were the intervention target. Existing methods either assume a pre-specified structural ordering (as in Structural Equation Modeling) or treat variables as a flat input vector, failing to learn the directed dependency structure among traveler attributes from the data itself. Furthermore, integrating structure learning with discrete choice is methodologically challenging because the observed choice is a mutually exclusive categorical outcome driven by latent utility comparisons, not a standard variable in a dependency graph. Treating the choice as a graph node can distort the learned attribute structure.

2. Methodology: Neural-BSL

The authors propose Neural-Bayesian Structure Learning (Neural-BSL), a framework that couples differentiable structure learning with random-utility-based discrete choice estimation in a single, end-to-end differentiable procedure. The framework consists of three core components:

A. Differentiable Bayesian Structure Learning
Instead of learning a single point-estimate Directed Acyclic Graph (DAG), Neural-BSL learns a posterior distribution over DAGs using $K$ parallel particles.

Representation: Each particle parameterizes a soft adjacency matrix $\mathbf{G}_k$ via low-dimensional latent embeddings ( $\mathbf{U}_k, \mathbf{V}_k$ ). The entry $G_k[i, r]$ represents the probability of a directed edge $r \to i$ .
Acyclicity: The acyclicity constraint is enforced via a differentiable penalty function $h(\mathbf{G}_k) = \text{tr}((\mathbf{I} + \mathbf{G}_k/D_X)^{D_X}) - D_X$ , incorporated into the loss via an augmented Lagrangian schedule.
Likelihood: A mixed-type generalized linear score models each attribute conditional on its parents (Bernoulli for binary, Gaussian for continuous). The structural loss evaluates how well the graph explains the observed attribute data.
Uncertainty: The framework uses Stein Variational Gradient Descent (SVGD) to update particles, approximating the posterior distribution over graph structures and avoiding overcommitment to a single fragile topology.

B. Utility Specification with Structure-Weighted Interactions
The framework integrates the learned structure directly into the utility function without including the observed choice in the DAG.

Structure-Weighted Interaction: For each traveler $n$ , a structure-weighted interaction vector $\mathbf{s}_{n,k}$ is computed as the elementwise product of the traveler's attributes $\mathbf{X}_n$ and the weighted sum of their parents in the learned DAG. This captures the "causal context" of each attribute.
Utility Construction: The systematic utility for alternative $j$ $j$ combines:
1. Linear Terms: Standard coefficients on observed attributes and Level-of-Service (LOS) attributes.
2. Structure-Weighted Terms: Coefficients on the interaction vector $\mathbf{s}_{n,k}$ , allowing the learned DAG topology to modulate sensitivity to LOS attributes.
3. Nonlinear Residuals: Small Multi-Layer Perceptrons (MLPs) capture additional nonlinearities, initialized to ensure the model starts as a specification-matched MNL baseline.
LOS Handling: LOS attributes (e.g., travel time, fare) are treated as exogenous supply variables entering the utility directly, bypassing the attribute DAG to preserve the distinction between traveler characteristics and alternative-specific conditions.

C. Joint Training and Policy Simulation

Joint Optimization: The model minimizes a combined loss function comprising the choice negative log-likelihood, the structural likelihood, sparsity penalties, and an economic sign penalty (to ensure LOS coefficients remain non-positive). The structural and choice objectives inform each other through the shared structural parameters.
Intervention Propagation: For policy analysis, the framework executes a Pearl-style do-operation. When an attribute is intervened upon, the change is propagated through the learned DAG in topological order. Downstream attributes are updated based on the fitted structural relationships (preserving residuals for continuous variables and sampling from conditional probabilities for binary variables). Choice probabilities are then recomputed using the updated attribute vector.

3. Key Contributions

The paper identifies three primary contributions:

Unified Framework: It is the first framework to combine data-driven structure learning and discrete choice modeling within a single end-to-end differentiable estimation procedure. This allows the discovered dependency structure and the choice model to inform one another, preserving the theoretical grounding of random utility theory.
Integrated Utility Specification: The framework introduces a utility specification where individual-level LOS coefficients and Alternative-Specific Constants (ASCs) depend on both observed attributes and structure-weighted interactions derived from the learned DAG. The structure is not merely a descriptive output but an active mechanism in the choice model.
Structural Policy Simulation: The framework supports simulations that propagate interventions through learned downstream relationships before recomputing mode choice. This yields not only predicted mode-share changes but also the accompanying adjustments in downstream traveler or trip attributes, exposing the structural basis of the forecast.

4. Results

The framework was evaluated using two datasets: a Stated-Preference (SP) survey from the Seoul Metropolitan Area and a Revealed-Preference (RP) dataset (London Passenger Mode Choice).

Predictive Performance: Neural-BSL achieved predictive performance (accuracy, weighted F1, log-likelihood) comparable to both specification-matched MNL and feedforward Neural Network (MLP) benchmarks. On the SP dataset, Neural-BSL matched the MNL (0.584 vs. 0.583 accuracy) and was slightly below the MLP (0.593). On the RP dataset, it performed similarly to the MNL (0.742 vs. 0.740) and slightly below the MLP (0.746).
Discovered Structures: The model successfully recovered behaviorally coherent dependency structures.
- In the SP data, it identified a "mobility-resource pathway" linking socioeconomic status (income, education, gender) to license possession and car ownership.
- In the RP data, it linked age to license possession and car availability, and connected trip context (commute purpose, rush hour) with demographic factors.
Policy Simulation Divergence: The simulations revealed significant differences between Neural-BSL and the static MNL when interventions were applied to structurally connected attributes.
- License Surrender: In the SP scenario, Neural-BSL predicted a larger shift from Demand-Responsive Transit (DRT) to subway compared to MNL. This was because the learned structure propagated the loss of a license to a loss of car ownership, shifting car-deprived travelers toward high-capacity rail, whereas MNL treated the license loss in isolation.
- Age Intervention: In the RP scenario, a 10-year age increase led to a much larger shift from public transit to driving under Neural-BSL (+2.69 pp) compared to MNL (+0.43 pp). This was driven by the propagation of age to increased license possession and car availability.
- Downstream Effects: The framework explicitly quantified how interventions alter downstream attributes (e.g., license surrender reducing car ownership by 1.215 standardized units), providing a mechanistic explanation for the mode-share shifts.

5. Significance and Claims

The paper claims that Neural-BSL bridges the gap between predictive machine learning and explanatory causal modeling in transportation. By learning the dependency structure among traveler attributes directly from data and integrating it into the random utility framework, the model moves beyond the "ceteris paribus" assumption of standard choice models.

The authors emphasize that the framework does not claim to recover the "true" data-generating mechanism in the presence of unmeasured confounders; rather, it recovers the structural conditional dependencies supported by the observed data under specific modeling assumptions. The significance lies in the ability to simulate policy interventions that respect the interdependencies of traveler characteristics, revealing how upstream changes cascade through the system to affect both traveler attributes and final mode choices. This provides planners with a more behaviorally grounded tool for anticipating the full impact of policies, rather than just the direct effect on a single variable.

Limitations Acknowledged:

The causal interpretation relies on the assumption of causal sufficiency (no unmeasured confounders).
The learned structure depends on regularization settings and prior masks.
Edge directions in cross-sectional data cannot always be fully resolved without temporal ordering.
The SP dataset had a limited number of unique respondent profiles for structural learning compared to the total number of choice observations.