Source-linked AI summary
Feasible and Novel Synthetic Population Generation with Tabular and Sequential Travel Attributes
Farbod Abbasi, Zachary Patterson, Bilal Farooq
TL;DR
Limited survey samples omit valid attribute combinations and may yield infeasible profiles, while realistic synthetic populations must also represent sequential travel behavior. The paper proposes a regularized two-stage framework combining tabular WGAN-GP synthesis with conditional sequence generation and expanded evaluation. Regularization improves tabular feasibility, diversity, and novelty over vanilla WGAN-GP, while LSTM-Attention best matches trip lengths and Transformer achieves higher sequential F1 with strong cross-stage consistency.
Problem
Limited samples create sampling zeros and generative models may produce structural zeros, while synthetic populations must represent both static attributes and sequential travel behavior.
Method
The framework regularizes WGAN-GP tabular synthesis and conditions Transformer and LSTM-Attention sequence generators on synthesized tabular profiles, using novelty and count-aware metrics.
Results
All regularized models outperform vanilla WGAN-GP across feasibility, diversity, novelty, and count-aware scores; IGP has the strongest count-aware novelty F1, while Transformer achieves higher sequential F1 than LSTM-Attention.
Takeaways & Limitations
The framework jointly produces more feasible and novel tabular profiles, behaviorally realistic trip chains, and strong compatibility between mobility status and generated sequences.
Takeaways & Limitations
Soft regularization cannot fully prevent structural violations or guarantee recovery of all sampling-zero combinations, and performance depends on hyperparameters and training-sample quality.
Abstract
from arXiv · showhide
Synthetic populations are critical inputs for activity-based travel demand models, yet generating realistic populations from limited survey data remains challenging. Small samples miss valid attribute combinations, known as sampling zeros, and generative models may also produce infeasible structural zeros. Moreover, realistic synthetic populations must capture both static socio-demographic attributes and sequential travel behaviour, such as trip chains. This paper proposes a regularized two-stage generative framework to address these challenges, where regularization refers to additional loss terms that guide the generator toward broader valid coverage and fewer infeasible samples. In Stage 1, a Wasserstein GAN with gradient penalty is augmented with three regularization terms, IGP, LDR, and CLAP, to improve feasibility, diversity, and novelty in tabular population synthesis. In Stage 2, Transformer and LSTM-Attention models generate sequential travel attributes, including departure time, trip purpose, and travel mode, conditioned on the synthesized tabular profiles. We also introduce novelty and count-aware metrics to evaluate whether valid unseen combinations are recovered and generated in realistic proportions. Results show that regularized models outperform the vanilla WGAN-GP across feasibility, diversity, and novelty. Regularization increases feasibility by 2.1 to 3.7 percentage points and novelty by 6.6 to 10.0 percentage points, improving sampling-zero recovery without sacrificing feasibility. The F1 score improves by 6.3 to 8.6 percentage points. For sequential attributes, LSTM-Attention best matches the trip-length distribution, while Transformer achieves higher overall sequential F1, 90.6\% versus 89.1\%. Cross-stage validation confirms strong consistency between generated mobility status and generated trip chains.
1. Introduction
The paper targets reliable synthetic populations for activity-based models by addressing limited-sample coverage, infeasible combinations, sequential travel behavior, and incomplete evaluation. It proposes a unified regularized two-stage framework spanning tabular synthesis and conditional trip-chain generation.
- Synthetic populations support disaggregate activity-based models whose downstream realism depends on statistically representative individual attributes and behaviors.
- Limited surveys make joint attribute-distribution modeling difficult, because samples typically cover only one to five percent of the population.
- Sampling zeros are valid unseen combinations, whereas structural zeros are logically impossible combinations; pursuing novelty and diversity can therefore conflict with feasibility.
- Sequential travel attributes add variable-length temporal dependencies to static socio-demographic profiles, including trip purposes, departure times, and travel modes.
- Novelty and count-aware diversity and novelty metrics extend evaluation beyond feasibility and coverage by testing unseen-combination recovery and realistic proportions.
- The framework uses a regularized WGAN-GP for tabular synthesis and Transformer or LSTM-Attention models for conditional sequential generation.
2. Literature Review
Prior work advances marginal, probabilistic, Bayesian, deep-generative, and sequential synthesis, but existing approaches rarely address sampling zeros, structural zeros, sequential behavior, and evaluation completeness together. The paper positions its framework as a unified response to these gaps.
- Traditional marginal-fitting methods are efficient but cannot capture joint dependencies essential for behavioral realism.
- Probabilistic, Bayesian, and hierarchical methods model dependencies more flexibly but rely on structural assumptions or face scalability limits as attributes increase.
- Deep-generative studies vary in how they address sampling and structural zeros, with existing evaluation often failing to distinguish observed from valid unseen combinations.
- Count-aware metrics assess whether recovered combinations follow realistic population proportions, complementing novelty measurement.
- Existing sequential-generation studies model trip chains but generally do not explicitly address sampling-zero recovery or structural-zero reduction.
- The proposed framework combines regularized tabular synthesis, sequential generation, and novelty- and count-aware evaluation.
3. Data and Case study
The case study uses the 2018 Montreal Origin-Destination survey to examine synthesis from limited samples, with static attributes paired with ordered daily trip chains. A one-percent training sample exposes substantial missing combination coverage that the generator must recover.
- The 2018 Montreal OD survey contains travel information for 162,588 individuals and approximately four percent of the total population.
- The dataset includes seven individual attributes with 35 categorical classes plus sequential departure-time, trip-purpose, and trip-mode features.
- Daily travel behavior is represented as an ordered trip chain whose sequence length equals the individual’s number of trips.
- The study treats the full OD dataset as representative and evaluates whether a subset-trained generator can reconstruct its individuals and missing combinations.
- As sample size increases, a larger proportion of the full dataset’s unique combinations is recovered.
- At one-percent sampling, 14.7% of unique feature combinations are observed, while they represent 78.2% of dataset individuals.
4. Methodology
The paper uses a two-stage framework that first synthesizes tabular socio-demographic profiles with regularized WGAN-GP and then generates conditioned travel sequences. Its regularization targets broader valid coverage and novelty while preserving feasibility, with sequential quality assessed through trip-length and activity-pattern comparisons.
- Framework overview: The framework decomposes synthetic individuals into tabular attributes and sequential travel behaviour, then models these components in two stages.Stage 1 learns tabular attributes; Stage 2 generates sequences conditioned on tabular profiles.
- Stage 1: Tabular synthesis: WGAN-GP maps latent noise to categorical individual profiles, while its critic uses a gradient penalty to improve training stability and enforce the Lipschitz constraint.The generated profiles include household size, vehicle ownership, age group, employment, license, gender, and mobility status.
- Stage 1: Regularization: Regularized generator losses guide broader latent-space exploration to recover rare valid combinations while preserving feasibility, rather than guaranteeing removal of every invalid profile.The regularizers are soft penalties, and each is tested separately for effects on diversity, novelty, and feasibility.
- Stage 1: Regularization: IGP encourages different latent vectors to produce meaningfully different outputs, increasing effective latent-space use and supporting more diverse and novel samples.Its negative loss encourages the output-to-latent distance ratio to increase up to threshold τ.
- Evaluation: Tabular feasibility requires generated attribute combinations to appear at least once in the full OD dataset, while sequential evaluation compares trip-length distributions and frequent purpose-sequence shares with real data.These evaluations address both valid attribute combinations and interpretable activity-sequence patterns.
5. Results and discussion
Regularized models improve tabular synthesis across distributional similarity, feasibility, diversity, novelty, and count-aware evaluation, while sequence-model performance varies by metric. IGP leads weighted novelty F1, whereas LSTM-Attention better matches trip-length distributions and Transformer achieves the highest overall sequential F1.
- Experimental setup: A 1% Montreal OD sample contains 628 of 4,282 full-dataset tabular combinations, leaving 3,654 sampling-zero combinations for recovery.The experiment tests whether models can generate valid profiles absent from the limited training sample.
- Distributional similarity: Regularized models consistently outperform vanilla WGAN-GP in SRMSE, with CLAP achieving the lowest error across univariate, bivariate, and trivariate distributions.CLAP-generated marginal distributions closely follow the real data without a clear pattern of category over- or under-representation.
- Interpretation: Distributional similarity is only an initial evaluation, so further assessment is needed for feasibility, novelty, diversity, and downstream usefulness.Low SRMSE alone does not establish overall synthetic-population quality.
- Feasibility: 0.935 is CLAP’s feasibility score versus 0.898 for Vanilla, while IGP and LDR achieve 0.919 and 0.934, respectively.All regularized variants improve validity, indicating that broader exploration does not sacrifice feasibility.
- Diversity: 0.741 is IGP’s diversity score versus 0.638 for Vanilla, while CLAP and LDR achieve 0.678 and 0.714, respectively.When weighted by population frequencies, IGP and CLAP both score 0.684, narrowing IGP’s standard-diversity advantage.
- Novelty: 0.706 is IGP’s novelty score versus 0.600 for Vanilla, while CLAP and LDR achieve 0.637 and 0.676, respectively.For weighted novelty, IGP reaches 0.579, ahead of CLAP at 0.548 and LDR at 0.545, showing that realistic frequency recovery is stricter than unseen-combination recovery alone.
- Model comparison: 0.711 is IGP’s weighted novelty F1 score and the highest model value, supporting the ranking IGP, CLAP, LDR, and Vanilla.Novelty is prioritized over diversity because it directly measures recovery of valid combinations absent from the training sample.
- Sequential attributes: LSTM-Attention yields lower JSD for trip-length distributions, while CLAP combined with LSTM-Attention achieves the best W1 score and Transformer-based combinations achieve the highest overall sequential scores.The best overall sequential model is CLAP + Transformer, with mean F1 of 0.906.
6. Conclusions
The paper presents a two-stage framework that synthesizes feasible, diverse, and novel tabular populations, generates sequential travel behaviour, and evaluates both with expanded metrics. Regularization improves tabular synthesis, while sequential models reproduce trip-chain characteristics and remain consistent with synthesized mobility profiles.
- The framework combines tabular synthesis, sequential travel-behaviour generation, and feasibility, diversity, novelty, and count-aware evaluation.
- Tabular synthesis: Regularized WGAN-GP models outperform vanilla WGAN-GP across tabular feasibility, diversity, novelty, and count-aware scores.IGP achieves the strongest overall count-aware novelty F1 balance between feasible individuals and recovery of combinations absent from the 1% training sample.
- Sequential generation: Transformer and LSTM-Attention models reproduce trip-length, activity, and sequence n-gram characteristics when conditioned on synthesized tabular profiles.LSTM-Attention is competitive for some trip-length distributions, whereas Transformer generally provides stronger sequential validity for higher-order trip-chain patterns.
- Cross-stage validation: Mobility-status consistency exceeds 98% for LSTM-Attention and 99% for Transformer-generated sequences.The validation indicates that generated behaviour remains logically compatible with synthesized socio-demographic profiles.
- Limitations: Soft regularization cannot fully prevent structural violations or guarantee recovery of all sampling-zero combinations.Performance depends on hyperparameter choices and the quality of the limited training sample.
- Future work: Future work should test transferability across regions and datasets and investigate household-level constraints and interactions.
- Future work: Future extensions include richer spatial information, dynamic population synthesis, and evaluation within full activity-based modelling pipelines.
Declaration of generative AI
The authors used ChatGPT for language editing, clarity improvement, manuscript formatting, and preparation of tables and figures, then reviewed and edited the content.
- ChatGPT assisted with language editing, clarity improvement, manuscript formatting, and preparation of tables and figures.
- The authors reviewed and edited the content after using ChatGPT and retain full responsibility for the published article.
Data availability
The study uses data from the 2018 Montreal Origin-Destination survey, which are licensed and not publicly available from the authors.
- The findings rely on data from the 2018 Montreal Origin-Destination survey.
- The survey data are used under license and are not publicly available from the authors.