generative adversarial nets , synthetic data , differential privacy , R package , multiple imputation
Abstract:
Generative Adversarial Nets (GANs) are a robust framework for learning complex data distributions and sampling new examples. This dissertation introduces GANs to the social science community by enhancing accessibility, exploring potential use cases, and identifying limitations.
The study introduces GANs, making the framework accessible to social science researchers. The dissertation presents the RGAN package, explicitly designed for GAN modeling in R, to facilitate implementation.
As a specific use case, the dissertation explores privacy-preserving synthetic data as a potential data-sharing means. It introduces a new method to produce better output from GANs trained to satisfy differential privacy.
Lastly, the dissertation critically examines the limitations and challenges of GANs in multiple imputation of missing data, providing insights into boundaries and considerations for their application. To that end, it also introduces a novel benchmark to evaluate multiple imputation algorithms.
Dieser Eintrag ist Teil der Universitätsbibliographie.
Das Dokument wird vom Publikationsserver der Universitätsbibliothek Mannheim bereitgestellt.
Search Authors in
