Population Synthesis with Deep Generative Models - is it worth it? Exploring new models and metrics

Abstract

Population synthesis is a fundamental clement for activity-based simulation models, consisting in generating a full synthetic population from a restricted dataset. Even if many methods have been studied, diffusion models — the current state-of-the-art Deep Generative Models (DGM) for tabular data — have not yet been thoroughly investigated. This paper addresses this gap and conducts a benchmark comparison with previous models, including investigations on the impact of encoding choices. To build a robust benchmark, metrics to assess the distribution, originality and realism of a generated population are redefined. Furthermore, this benchmark is conducted using a large French census dataset allowing an accurate evaluation. By considering two scenarios of training set size —representing typical use cases — this study offers generalizable guidelines. Notably, with appropriate data encoding, diffusion models outperform other DGMs. However, Bayesian Networks, a probabilistic model, demonstrate comparable or superior performances while requiring fewer computational resources, making them attractive in similar applications.