SemaPop-GAN: Persona-Conditioned Population Synthesis with Marginal Consistency

Abstract

Population synthesis is a core component of agent-based transport modeling, requiring realistic individual-level populations that preserve both statistical consistency and behavioral heterogeneity. Existing methods primarily rely on structured attributes and marginal constraints, limiting their ability to capture high-level behavioral semantics. We propose SemaPop-GAN, a semanticconditioned generative framework that integrates large language model (LLM)-derived persona representations into population synthesis. To ensure statistical validity, we introduce marginal regularization during training and a post-hoc marginal calibration procedure for deployment. Extensive experiments demonstrate that SemaPop-GAN substantially improves generative performance, achieving closer alignment with both target marginal and joint distributions while preserving sample-level feasibility and diversity under semantic conditioning. Furthermore, through post-hoc marginal calibration, we show that SemaPop-GAN remains robust under additional calibration constraints. Overall, our framework provides a principled approach to bridging semantic expressiveness and statistical consistency in population synthesis, enabling more realistic and scenario-aware transport modeling.