Using Generative Adversarial Networks to Generate Synthetic Population

• Activity Number: 344 - Expanding Data Utility - Issues in Disclosure and Modeling
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Government Statistics Section
• Author(s): Yijun Wei* and Luca Sartore and NELL Sedransk
• Companies: NISS and National Institute of Statistical Sciences and NISS
• Keywords: US Census of Agriculture; GAN; synthetic population

Abstract

The US Census of Agriculture is conducted by the USDA National Agriculture Statistics Service (NASS) every five years, in years ending in 2 and 7. The Census provides comprehensive information of the US farms, ranches, and the people who operate them. NASS wants to make the Census data widely available. However, to avoid disclosing confidential information provided by respondents, the record-level dataset must be modified. Generative Adversarial Networks (GANs) are a type of deep learning models that generate synthetic data having the same properties as the original data. GANs have been successfully applied in a number of areas, such as images creation and transfer learning. A GAN consists of two networks: a generative network and a discriminative network. These two networks work together to create new different objects that resemble the original reference data. A GAN methodology is proposed in this study to generate a synthetic population for 2012 Census of Agriculture dataset. The performance of the model is evaluated by comparing the distributions of the synthetic population and the original one.