Multivariate Time Series Synthesis Using Generative Adversarial Networks

Abstract

Figure 1: From left to right, top to bottom, the training process of our GAN network is depicted. The generated data is plotted as a density chart throughout the training process, showing how the network learns to reflect the fidelity of the original data.

[SoK] The Great GAN Bake Off, An Extensive Systematic Evaluation of Generative Adversarial Network Architectures for Time Series Synthesis

“…This measurement is only suitable for GAN models, which can generate time series with various lengths. To account for spatial correlations, Leznik et al [24] also compared the correlation coefficient of the generated data. The authors synthesised a Content Delivery Network (CDN) data set from a production environment, which contains time series with high temporal and spatial correlations.…”

“…The authors synthesised a Content Delivery Network (CDN) data set from a production environment, which contains time series with high temporal and spatial correlations. Comparing descriptive measurements of generated time series as can also provide insights [24]. Leznik et al [24] utilized different entropy measurements to compare the generated time series with regard to information and noise.…”

“…Comparing descriptive measurements of generated time series as can also provide insights [24]. Leznik et al [24] utilized different entropy measurements to compare the generated time series with regard to information and noise. Using downstream tasks can also provide insight if the GAN was able to learn the underlying data distribution correctly.…”

“…Network Architectures Based on the state of the art, we consider RNN, TCN and transformer architectures for the comparison task. The RNN architectures we use are based on Esteban, Hyland, and Rätsch [10], Lin et al [27] and Leznik et al [24]. These authors utilize an LSTM block to process the time series and MLP layers to achieve the desired output dimensionality.…”

“…In line with Leznik et al [24] this work focuses on synthesising time series with a fixed length. According to Lin et al [27], RNN-based GAN models struggle to synthesise time series with more than 500 time steps.…”