Dissertation
Generative Adversarial Networks for Multi-Objective Synthetic Data Generation
Washington State University
Doctor of Philosophy (PhD), Washington State University
05/2024
DOI:
https://doi.org/10.7273/000006581
Abstract
Synthetic data has become increasingly accessible due to remarkable advancements in machine learning. This data is extremely useful to researchers due to its wide range of applications. Synthetic data may be used to robust populations that are under-sampled, or to create permutations of some existing data, generating combinations not seen in the original data. Synthetic data may also be used in place of the original data completely when sensitive aspects limit the distribution.
Previously, research in synthetic data generation has been primarily focused on generating data that is maximally realistic. Significantly less attention has been paid to assurances of other components of the data, such as privacy concerns or data diversity. This has left a gap in the field of synthetic data generation. We address this through the investigation of multi-agent synthetic data generation.
In this dissertation, we expand the scope of data generation by introducing agents that optimize various facets of synthetic data, such as privacy, class diversity, and training utility. We propose a novel, multi-objective synthetic generation framework to allow all of these objectives to be optimized. We finally demonstrate this framework can generate high quality data across multiple domains for an arbitrary number of objectives.
Metrics
10 File views/ downloads
21 Record Views
Details
- Title
- Generative Adversarial Networks for Multi-Objective Synthetic Data Generation
- Creators
- Chance Nicholas DeSmet
- Contributors
- Diane J Cook (Chair)Hassan Ghasemzadeh (Committee Member)Lawrence B Holder (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 182
- Identifiers
- 99901122438501842
- Language
- English
- Resource Type
- Dissertation