SeriesGAN: Time Series Generation via Adversarial and Autoregressive Learning

TL;DR

This work introduces an advanced framework that integrates the advantages of an autoencoder-generated embedding space with the adversarial training dynamics of GANs, and employs two discriminators: one to specifically guide the generator and another to refine both the autoencoder and generator’s output.

Abstract

Current Generative Adversarial Network (GAN)-based approaches for time series generation face challenges such as suboptimal convergence, information loss in embedding spaces, and instability. To overcome these challenges, we introduce an advanced framework that integrates the advantages of an autoencoder-generated embedding space with the adversarial training dynamics of GANs. This method employs two discriminators: one to specifically guide the generator and another to refine both the autoencoder's and generator's output. Additionally, our framework incorporates a novel autoencoder-based loss function and supervision from a teacher-forcing supervisor network, which captures the stepwise conditional distributions of the data. The generator operates within the latent space, while the two discriminators work on latent and feature spaces separately, providing crucial feedback to both the generator and the autoencoder. By leveraging this dual-discriminator approach, we minimize information loss in the embedding space. Through joint training, our framework excels at generating high-fidelity time series data, consistently outperforming existing state-of-the-art benchmarks both qualitatively and quantitatively across a range of real and synthetic multivariate time series datasets.

Figures (5)

• Figure 1: The figure illustrates the architecture of five distinct time series generation techniques: TimeGAN, RCGAN, GAN, Teacher Forcing, and Professor Forcing. Each method presents a unique approach to generating time series data, offering various strengths and applications depending on the specific requirements of the task.
• Figure 2: The figure showcases the architecture of SeriesGAN for time series generation. It includes two autoencoders, which play a crucial role in loss function calculation and facilitate lower-dimensionality training of the GAN network. Additionally, it incorporates two discriminators that enhance the data reconstruction capabilities of the autoencoder and improve the data generation quality of the generator network.
• Figure 3: PCA visualizations illustrate how the distributions of original and synthetic data align. The top row shows SeriesGAN results, with TimeGAN, Standard GAN, Teacher Forcing, and Professor Forcing visualizations displayed sequentially underneath. From left to right, the plots correspond to the Stocks, Sines, ECG, and SWAN-SF datasets.
• Figure 4: t-SNE visualizations show distribution alignment between original and synthetic data. The top row presents SeriesGAN results, followed by TimeGAN, Standard GAN, Teacher Forcing, and Professor Forcing from top to bottom. Left to right, plots represent Stocks, Sines, ECG, and SWAN-SF datasets.
• Figure 5: This illustration compares the original dataset samples (red) with their synthetic counterparts generated by the SeriesGAN algorithm (green) for both Sines and ECG samples.