Can Generative Adversarial Networks Compete against Diffusion Models at Generating Time Series?

: Diffusion models have become the state-of-the-art at synthetic data generation, widely replacing generative adversarial networks (GANs) in recent years. GANs often suffer from issues like training instabilities and mode collapse. However, they also offer advantages. They are computationally efficient, as they only require a single forward pass for generating a synthetic data sample. Thus, we ask the question whether GANs can still compete with diffusion models. Specifically, we tackle the task of generating conditional time series—one of the most common data modalities in industrial applications. Making a model follow conditions appropriately is not always trivial. In this work, we leverage a structured noise space (SNS) to implicitly condition GANs for generating time series data. The SNS allows to use an unconditional generator architecture for conditional data generation by training it on a certain latent space structure. We extensively evaluate the method on time series datasets from the UCR classification archive, comparing to other GAN, variational autoencoder and diffusion approaches. We find advantages over explicit conditioning regarding fidelity of generated samples and training stability. With our improved architecture, we show that SNS-GAN is able to compete with diffusion at a fraction of the sampling time. Additionally, we examine limitations and propose alternatives of structuring the noise space so that the model can effectively learn the condition in varying scenarios.