Generation of Tabular and Sequential Attributes in Synthetic Population with Enhanced Diversity and Feasibility

Abstract

Synthetic population generation is essential for activity-based modeling, where individual level data are needed to simulate travel behavior. However, such data are often limited due to high costs and privacy concerns. A realistic synthetic population should capture both tabular and sequential attributes while preserving diversity (the ability to represent a wide range of plausible attribute combinations) and feasibility (the consistency of generated attributes with logically valid combinations). Despite recent advances in generative modeling, most existing approaches focus primarily on tabular attribute synthesis, and do not model sequential attributes. Moreover, diversity and feasibility of generated attribute combinations are often not explicitly addressed within the modeling framework. In this paper, we propose a unified generative framework for synthetic population generation that integrates tabular and sequential attributes. In the first stage, a Generative Adversarial Network (GAN) is used to learn the distribution of tabular attributes, augmented with an Inverse Gradient Penalty (IGP) regularization term to improve diversity while maintaining feasible attribute combinations. Since tabular attributes such as socio-demographic characteristics influence mobility patterns, the second stage uses a Transformer-based model to generates sequential attributes conditioned on the synthesized tabular attributes. The proposed framework is evaluated using distributional similarity, diversity, and feasibility metrics for tabular attributes, and unigram, bigram, and trigram pattern comparisons for sequential features. Results show that the regularized GAN significantly improves diversity coverage while maintaining high feasibility, and the Transformer model successfully reproduces dominant behavioral patterns observed in the real data.