EAGER: Multi-objective generation of synthetic time series data to boost model robustness and data privacy

Machine learning models require a sufficient amount and diversity of training data to ensure robustness and minimize bias. A dearth of data can negatively impact the predictive performance of machine learning algorithms. Because researchers recognize the surrogate role offered by synthetic data generators, they have been creating methods to generate increasingly realistic data proxies. Synthetic data can be automatically generated that balance the need to maximize multiple objectives. The goal of this project is to design a synthetic data generator that creates realistic individual and time series data to optimize cooperating or competing objectives in an adversarial manner. Building on this contribution, the algorithm will also be evaluated as a mechanism to increase model robustness, improve privacy preservation, and decrease model bias. The outcomes of this project include the design of a novel multi-agent generative adversarial network (GAN) architecture, called HydraGAN, that balances multiple, possibly competing, data goals. While researchers have investigated constrained versions of multi-criteria GANs, this project will introduce a novel method that facilitates balancing any number of data criteria using a multi-agent GAN. A second outcome of the project will be a formal proof that the system will reach a Nash equilibrium during training. Furthermore, the HydraGAN algorithm will be enhanced to not only explore a traditional i.i.d. data generation format but also handle the more complex nature of time-series data, representing one of the first efforts to create multi-agent GANs for time series data. Multiple discriminator agents will be defined for the multi-agent system. Exploring beyond the traditional criteria of sample realism, cooperating or competing agents will address the underexplored constraints of privacy preservation, distribution realism, and meeting of diversity constraints. The project will demonstrate the utility of the created synthetic data for tackling machine learning challenges including data sparsity and representation bias. While traditional performance metrics focus on the realism of individual data samples, this work will introduce an Area Under the Radar Chart metric, or AURC, to evaluate the data based on an arbitrary number of data quality criteria. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.