Generative adversarial networks for multivariate time series : design and evaluation

Abstract

The objectives of this research include the development of a synthetic data pipeline and the enhancement of dataset utility using machine learning methods. Especially in scenarios where original datasets suffer from underrepresentation of certain data points, synthesizing data appears as a solution to achieve balance and enhance dataset utility.

This work takes a look into the methods of generating synthetic data and the recent advancements in synthetic data generation. Focus is particularly in the generative adversarial network systems but this thesis also explores matters of data privacy and ethics, and the constraints these pose to data availability, usage and synthetic data generation.

As the results of this work, a pipeline for generating synthetic data is developed along with rich set of methods for evaluating the generated data. The findings suggest that the synthetic data holds a potent promise in augmenting datasets proven by the fidelity of the generated data. There is room for improvement regarding utility of the synthetic data but this work gives ideas for improving the framework and lays base for future development.