Adaptive synthetic generation using one-step Gibbs Sampler

Most existing state-of-the-art synthetic generation methods produce static snapshots of data that fail to adapt to demographic changes over time, which makes them quickly outdated. This paper introduces an adaptive approach to synthetic population generation using a one-step Gibbs Sampler that allows for the maintenance of synthetic data by integrating new information adaptively, rather than requiring a complete regeneration of datasets each time an update is necessary. We compare existing independent regeneration methods with the proposed adaptive generation and demonstrate that our approach creates a synthetic population of the same level of accuracy, but more efficiently. Also, we show that when the initial data is scarce or biased, the adaptive generator is particularly effective in enhancing dataset quality by adaptively enriching the population sample. To account for updates, we introduce a new Gibbs-resampling technique as an intermediate step that uses information from the most recent disaggregated real data to correct the errors and improve the representativeness and heterogeneity of the synthetic data. Furthermore, our results indicate that the adaptive approach is robust to unforeseen events, which helps mitigate the lack of representativeness in real data during such occurrences. • We propose an adaptive population synthesis method using a one-step Gibbs sampler. • Method updates synthetic data without full regeneration, saving computation time. • New Gibbs-resampling method improves long-term synthetic data representativeness. • The results show improved robustness to unforeseen events compared to other methods. • Integrating new data enhances heterogeneity compared to random resampling.