Generating Synthetic Population Using Transformer Based Networks

Microsimulation models are widely used in various fields. However, accessing individual-level data can be challenging due to confidentiality concerns and high costs. To address this, researchers often generate synthetic populations based on available data of the target population. In this paper, we present a novel Transformer-based approach for generating both numerical and categorical demographic attributes in synthetic populations. Our model achieves good fit with minor errors on the test set and preserves population heterogeneity, as indicated by the number of unique individuals in the real and synthetic populations (< 0.23% difference). Compared to the actual data, the synthetic population is moderately diverse, with an average distance of 0.00636 and standard deviation of 0.13836 between synthetic agents and real data. Our approach has potential applications in various domains, and we believe it can contribute to the development of more accurate and representative microsimulation models.

I. Introduction

When creating microsimulation models, which aim to simulate the individual behavior or activity patterns of agents in fields such as economics, transportation, ecology, and epidemiology, researchers and policy makers often face the challenge of accessing individual-level data. Confidentiality concerns and high acquisition costs can limit or prevent the collection of the necessary sample data, which is required for a realistic representation of the entire population under study. As the agent-based microsimulation models must reflect the realistic social dynamics of the population, it is essential to ensure that the agents' attributes accurately represent those of the entities they represent. To achieve this, one common approach is to generate synthetic populations of agents based on sociodemographic data available for the target population.