PaperlayerView exact revision on arXiv

Source-linked AI summary

LEBGen: An LLM-Enhanced Bayesian Network Framework for Few-Shot Travel Survey Data Generation

Zijian Shen, Bin Zhou, Jiguang Wang, Ya Zhao, Jintao Ke

arXiv:2609.08288v1cs.AI

TL;DR

Few-shot travel survey samples provide insufficient coverage for reliably learning dependencies among heterogeneous traveler characteristics and behavior. LEBGen combines LLM-derived personas with BN structure refinement while parameterizing the refined model from observed data. On the 2022 Hong Kong TCS at a 2% sampling rate, it improves distributional and dependency fidelity over representative baselines.

  • Problem

    Few-shot travel survey samples provide incomplete coverage of heterogeneous traveler groups and insufficient evidence for recovering reliable demographic–behavioral dependencies.

  • Method

    LEBGen uses LLM-derived traveler personas and structure refinement to enhance BN construction, while parameterizing the refined BN exclusively from observed data.

  • Results

    LEBGen achieves the lowest overall mean JSD and smallest mean absolute pairwise Cramér’s V error on the 2022 Hong Kong TCS under a 2% few-shot setting, outperforming representative baselines.

  • Takeaways & Limitations

    Persona discovery and LLM-guided structure refinement provide complementary benefits, with advantages most pronounced at small sample sizes.

  • Takeaways & Limitations

    Evaluation uses a single survey, and detailed origin–destination relationships remain challenging to reconstruct.

Abstract

from arXiv · show

Travel survey data are essential for transportation planning and travel behavior analysis, yet collecting large-scale representative samples is costly and time-consuming. A practical alternative is to generate synthetic survey records from a few-shot sample. However, such samples provide incomplete coverage of heterogeneous traveler groups and insufficient evidence for recovering the complex dependencies between demographic characteristics and travel behavior. Existing approaches have complementary limitations. Probabilistic generative models such as Bayesian networks (BNs) offer explicit distributional control, but structures learned from few-shot samples may omit meaningful dependencies or retain spurious ones. Large language models (LLMs) can help address these difficulties in BN structure learning by providing behavioral knowledge that complements the limited statistical evidence. We therefore propose LEBGen, an LLM-enhanced BN framework that uses this knowledge to refine network structure for few-shot travel survey data generation. Specifically, the LLM first identifies traveler personas from demographic attribute and travel behavior statistics, then recovers dependencies missed by the persona-augmented BN structure and prune spurious ones. The refined BN is parameterized exclusively from the observed data to generate synthetic records. Under a 2% few-shot setting on the 2022 Hong Kong Travel Characteristics Survey, LEBGen reduces the mean marginal Jensen-Shannon divergence from 0.0671 to 0.0091 and the mean absolute Cramer's V error by 14.3% over the best-performing baseline, substantially improving both distributional and dependency fidelity.

1 Introduction

LEBGen addresses few-shot travel survey generation, where limited observations make dependency estimation unreliable and existing models may miss meaningful behavioral relationships. It combines LLM-derived traveler personas with BN structure refinement and improves distributional and dependency fidelity on the 2022 Hong Kong TCS.

  • Travel survey data support transportation planning, demand forecasting, behavior analysis, and calibration of activity-based and agent-based models.
  • Few-shot samples can omit meaningful demographic–behavioral dependencies or retain spurious relationships because many traveler combinations are sparsely observed.
  • LEBGen uses LLM-derived behavioral knowledge to enhance BN construction at both node and edge levels.
  • The persona discovery agent summarizes demographic combinations and travel statistics into interpretable traveler personas represented as states of an additional BN node.
  • The structure refinement agent reviews few-shot-learned dependencies to identify missing or superfluous relationships before synthetic generation.
  • LEBGen outperforms representative baselines in distributional fidelity, including marginal, temporal, and spatial distributions, and in dependency fidelity on the 2022 Hong Kong TCS.

2 Literature Review

Prior transportation synthesis methods include reconstruction pipelines, probabilistic models, deep generative models, and LLM-based approaches, but few-shot settings still challenge dependency and heterogeneity preservation. BN structure learning and LLM semantic reasoning motivate the integrated approach studied by LEBGen.

  • Synthetic population and survey data generation in transportation: Synthetic transportation data support activity-based and agent-based modeling by representing households, individuals, activities, and travel behavior.
  • Synthetic population and survey data generation in transportation: Existing approaches span microsimulation pipelines, probabilistic graphical models, VAEs, and other deep generative methods.
  • Synthetic population and survey data generation in transportation: Few-shot synthesis remains underexplored because limited evidence can misrepresent heterogeneity, unconstrained relationships, spatial allocation, and temporal patterns.
  • Bayesian networks for travel behavior modeling and data synthesis: BNs provide interpretable joint-distribution modeling, but limited observations can cause structure-learning algorithms to omit meaningful dependencies or retain sampling-induced relationships.
  • LLMs in transportation and synthetic data generation: LLMs have been used for structural constraints, causal-order priors, structured reasoning, and transportation prediction and generation.
  • LLMs in transportation and synthetic data generation: LLM-based survey and mobility models can improve behavioral coherence but may show reduced variation, prompt sensitivity, and distorted relationships among survey variables.

3 Problem Formulation

The paper formulates generation of individual-level travel survey records from a few-shot sample over demographic and travel-behavior variables. The generator must match admissible schemas while approximating both individual distributions and cross-variable dependencies without additional target-population observations.

  • The survey variables are partitioned into traveler and household characteristics and travel behavior attributes.
  • Each observed record is represented as a vector of values in the joint variable space.
  • The few-shot sample is independently drawn from an unknown target distribution but lacks dense coverage of heterogeneous demographic–behavioral combinations.
  • The task is to construct a substantially larger synthetic dataset using the few-shot sample and survey metadata, with no additional target-population observations.
  • Generated records must follow the observed survey’s variable schema and admissible domains while approximating marginal distributions and dependencies between traveler characteristics and travel behavior.

4 Methodology

LEBGen builds a persona-augmented Bayesian network from few-shot travel-survey observations, then uses two LLM agents to refine its structure while preserving data-grounded parameter estimation. Personas group demographic profiles with similar travel behavior, supporting shared conditioning and assignment of unseen profiles.

  • Framework representation: LEBGen discretizes survey variables for structure learning while retaining original continuous values for later generation.The representation distinguishes demographic and travel-behavior variables and discrete and continuous variables.
  • Few-shot BN initialization: The framework first learns an initial Bayesian network from the few-shot structural representation, using score-based BIC optimization with greedy edge operations.The search considers valid edge additions, deletions, and reversals and stops when no single-edge operation improves the score.
  • Motivation: LEBGen addresses sparse demographic-behavior combinations by sharing behavioral information without removing demographic conditioning altogether.Separate distributions can be unreliable for sparsely observed profiles, while removing conditioning may discard relevant behavioral differences.
  • Traveler persona discovery: The persona discovery agent groups demographic profiles by associated travel-behavior summaries and assigns interpretable personas as states of an auxiliary network node.Profile summaries include state probabilities and robust continuous-variable statistics, while profile frequency records empirical support.
  • Traveler persona discovery: Persona assignment provides rule-based extrapolation for demographic combinations absent from the few-shot sample and shares conditioning across profiles with similar behavior.The persona node can replace multiple demographic parents, reducing separately estimated local configurations.
  • Edge structure refinement: The structure refinement agent reviews the persona-augmented graph and proposes dependency modifications that preserve graph validity while admitting semantically plausible weakly observed relationships.The two agents have distinct sequential roles: persona construction and graph-operation proposal.

5 Experiments

The experiments evaluate LEBGen on the 2022 Hong Kong Travel Characteristics Survey under controlled few-shot conditions. Models receive identical inputs and output sizes, while complete-survey data are reserved for evaluation.

  • Dataset: The 2022 TCS includes household-, individual-, and trip-level variables covering demographics, trip characteristics, and travel behavior.Variables include age, car availability, trip purpose, departure time, journey time, districts, and travel mode.
  • Experimental setting: The main experiment uses 2% of the TCS training records as the few-shot sample and evaluates generated data against a held-out reference.The complete survey is not used for model fitting or calibration.
  • Experimental setting: Each method receives the same few-shot observations and generates M = 80,000 synthetic travel survey records.Fixing the training subset and output size makes reconstruction differences attributable to the generative mechanisms.
  • LEBGen configuration: LEBGen discovers personas and refines the BN structure with GPT-4o, then estimates discrete and continuous local distributions from augmented few-shot data.Synthetic records are generated by ancestral sampling, after which the auxiliary persona variable is removed.
  • Sensitivity analysis: Few-shot sensitivity analysis varies the available training share from 1%−100% while holding the subset and synthetic output size consistent across methods.The analysis examines performance as progressively more direct statistical evidence becomes available.
  • Baseline methods: The comparison includes Gaussian Copula, CTGAN, TVAE, and MTabGen, representing statistical, adversarial, variational, and diffusion-based generators.All baselines use the identical few-shot subset and output size.

5.3 Evaluation Metrics

The evaluation measures synthetic travel-survey quality through distributional fidelity and dependency fidelity. It uses JSD for distribution agreement and Cramér’s V-based statistics for pairwise association agreement.

  • Distributional fidelity measures agreement between synthetic and reference marginal, temporal, spatial, and selected joint travel distributions.
  • JSD compares reference and synthetic variable distributions, with lower values indicating closer agreement.Categorical frequencies are compared directly, while numerical variables use common discretization for quantitative JSD calculation.
  • Cramér’s V measures association strength between categorical variable pairs, with numerical variables discretized for association analysis.
  • Dependency fidelity evaluates associations between traveler-characteristic and travel-behavior variables, as well as among travel-behavior variables.The evaluated pairs form the set Q.
  • Five statistics summarize dependency agreement, including synthetic association strength, mean absolute error, systematic bias, correlation, and inflation share.Inflation share is desirable near 0.5; lower mean absolute error and bias closer to zero indicate better agreement.

5.4 Few-Shot Distributional Fidelity

Under the 2% few-shot setting, LEBGen most closely reproduces marginal and selected spatial distributions overall, while dependency comparisons show stronger association fidelity than the baselines. Its main spatial exception is OD-pair reconstruction, where MTabGen performs better.

  • Distributional fidelity: LEBGen more closely preserves the principal shapes and relative frequencies of categorical and continuous distributions than the baseline methods.The advantage is especially visible for departure time and journey time, where several deep generative baselines show substantial shifts.
  • Distributional fidelity: 0.0091 mean marginal JSD is achieved by LEBGen, versus 0.0671 for the strongest baseline, MTabGen.LEBGen also records JSD values of 0.0045 for journey time and 0.0048 for car availability.
  • Spatial fidelity: 0.0724 OD-pair JSD for LEBGen exceeds MTabGen’s 0.0536, although LEBGen provides the strongest overall distributional reconstruction.The framework nevertheless follows observed spatial heterogeneity and reproduces many district-level trip-share differences.
  • Dependency fidelity: 0.054 mean absolute deviation is achieved by LEBGen, compared with 0.063 for MTabGen, while correlation with the real Cramér’s V vector reaches 0.868 versus 0.850.
  • Dependency fidelity: 0.005 signed association error indicates little systematic bias, while LEBGen’s 0.425 inflation share is close to the balanced reference value of 0.5.Gaussian Copula, CTGAN, and MTabGen generally attenuate dependencies, whereas TVAE shows association inflation.

5.6 Sensitivity to the Few-Shot Data Budget

LEBGen maintains a data-efficiency advantage when only a small survey fraction is available, especially from 1% to 10%. As more observations become available, its JSD decreases smoothly toward the full-data reference.

  • 1%–10% survey shares show the clearest advantage for LEBGen, which generally achieves the lowest or among the lowest JSD values across evaluated variables.The evaluated variables include mode, purpose, departure time, age, journey time, car availability, origin district, and destination district.
  • Baseline performance generally improves with more observations, but substantial differences remain for temporal and continuous attributes.Gaussian Copula retains errors for departure and journey time, while TVAE is unstable for several spatial variables.
  • CTGAN is especially sensitive to extremely small training samples and produces large JSD values at several low-budget points before improving.
  • LEBGen follows a substantially smoother convergence pattern as the available few-shot share increases.Its JSD decreases rapidly and approaches the full-data reference across most variables.

5.7 Ablation Study

The ablation study compares six configurations under the 2% few-shot setting and shows that LLM-guided structure refinement and persona augmentation contribute complementary improvements. The complete LEBGen framework achieves the strongest distributional and dependency fidelity.

  • Each configuration generates 80,000 synthetic records to isolate marginal modeling, BN dependencies, LLM-derived marginal knowledge, persona augmentation, and structure refinement.
  • Mean JSD is evaluated for distributional fidelity, while absolute Cramér’s V discrepancy measures dependency fidelity; lower values are better.
  • The empirical marginal sampler and initial few-shot BN produce comparable distributional errors, whereas LLM Marginal performs worse because it samples variables independently without explicit dependencies.
  • Adding personas to the unrefined BN yields only a modest distributional improvement, while LLM-guided refinement substantially reduces mean JSD.
  • The complete framework achieves the lowest distributional error, indicating complementary benefits from persona augmentation and BN structure refinement.
  • The complete framework produces the smallest and most concentrated absolute Cramér’s V discrepancies, with personas further reducing median error and dispersion after refinement.

6 Conclusion

LEBGen addresses few-shot travel survey generation by using semantic behavioral knowledge to improve Bayesian-network representation and dependency structure. Experiments show improved distributional and dependency fidelity, while broader validation remains necessary because evaluation uses one survey.

  • Few-shot observations provide insufficient evidence for reliably learning dependencies linking demographic characteristics, household conditions, and travel behavior.
  • LEBGen incorporates LLM-derived knowledge at both BN node and edge levels through persona discovery and structure refinement.
  • LEBGen achieves the lowest overall mean JSD and smallest mean absolute pairwise Cramér’s V error on the 2022 Hong Kong TCS under a 2% few-shot setting.
  • Persona discovery and LLM-guided structure refinement provide complementary benefits, with advantages most pronounced at small sample sizes.
  • Evaluation is based on a single survey, while local distribution estimation and detailed origin–destination reconstruction remain constrained by few-shot coverage.

A LLM Prompt Implementation

The prompt implementation constrains persona discovery and structure refinement through fixed roles, structured inputs, explicit evidence boundaries, and deterministic validation. These controls keep persona rules executable and graph updates valid while allowing LLM-derived semantic judgments.

  • A.1 Prompt Architecture and Output Control: Both LLM agents use fixed role-specific prompts and JSON-formatted runtime inputs tailored to persona discovery or graph refinement.
  • A.1 Prompt Architecture and Output Control: Persona discovery defines a persona as a semantic grouping of demographic profiles with similar travel-behavior patterns, not an individual respondent or population estimate.
  • A.1 Prompt Architecture and Output Control: The prompt evidence boundary prohibits invented probabilities, category codes, observations, and quantitative claims, and treats runtime input as data rather than instructions.
  • A.3 structure refinement agent: Structure-refinement operations are validated sequentially, and invalid graph operations are rejected without changing the current graph.
  • A.1 Prompt Architecture and Output Control: The LLM Marginal ablation constructs only one univariate distribution per variable and does not construct personas, model dependencies, or refine the graph.
  • A.2 persona discovery agent: The persona task selects the number of personas adaptively and requires mutually exclusive, collectively exhaustive demographic rules that exclude travel outcomes.
  • A.2 persona discovery agent: Membership rules use explicit Boolean and leaf operators with nesting-defined evaluation order, excluding ranges, wildcards, free-form expressions, and default branches.
  • A.2 persona discovery agent: Persona outputs require sequential identifiers, observed supporting profiles, and a complete persona set, while deterministic validators check coverage, exclusivity, and support.

B TCS Persona Set and Bayesian Network Structure

The 2% few-shot TCS sample yields eleven traveler personas that define the auxiliary persona node. The corresponding network comparison contrasts the initial data-driven graph with the refined persona-augmented structure.

  • The Persona Discovery Agent identifies eleven traveler personas from demographic profiles and aggregate travel-behavior summaries in the 2% few-shot TCS sample.
  • The persona labels and demographic characterizations describe traveler groups, while behavioral signatures summarize associated aggregate travel patterns.
  • The personas constitute the state space of the auxiliary persona node used in the refined Bayesian network.
  • Figure 8 compares the initial graph G0 learned by BIC-based hill climbing with the refined persona-augmented graph G∗.

A. BIC-initialized BN graph B. Refined persona-augmented BN graph

The TCS few-shot experiment compares a BIC-initialized Bayesian network with a refined persona-augmented graph. LEBGen preserves data-supported dependencies while adding, deleting, and reversing relationships using demographic, persona, and behavioral semantics.

  • A. BIC-initialized BN graph: The BIC-initialized graph captures dependencies among demographic and travel-behavior variables directly from few-shot observations.Its travel-behavior layer connects trip purpose, main mode, departure time, and journey time.
  • B. Refined persona-augmented BN graph: The refined graph strengthens connections between traveler characteristics and travel behavior, including mobility resources, activity roles, and age-related differences.Car availability and licence status inform mode choice, while employment, student status, and age group relate to trip purpose and temporal patterns.
  • B. Refined persona-augmented BN graph: Persona nodes represent demographic–behavioral heterogeneity through deterministic demographic assignments and links to selected travel-behavior variables.Personas are derived from demographic combinations and associated trip-purpose, mode, temporal, and other travel summaries.
  • B. Refined persona-augmented BN graph: Deleted edges remove less informative dependencies after additional demographic and persona-mediated relationships are introduced.The refinement replaces some few-shot statistical relationships with more coherent representations through other demographic variables or the expanded network.
  • B. Refined persona-augmented BN graph: Reversed edges preserve initial associations while changing their direction in the generative factorization.This affects tightly connected attributes such as trip purpose, main mode, departure time, and journey time.
  • B. Refined persona-augmented BN graph: The final graph combines BIC-supported structure with persona augmentation and LLM-guided refinement to introduce additional direct and persona-mediated dependencies.Figure 8 distinguishes retained, added, deleted, reversed, and fixed demographic-to-persona relationships by edge color and style.
Loading 2609.08288v1…