Abstract:Generative Adversarial Networks (GANs) have proven to be a powerful tool for generating realistic synthetic data. However, traditional GANs often struggle to capture complex relationships between features which results in generation of unrealistic multivariate time-series data. In this paper, we propose a Graph-Attention-based Generative Adversarial Network (GAT-GAN) that explicitly includes two graph-attention layers, one that learns temporal dependencies while the other captures spatial relationships. Unlike RNN-based GANs that struggle with modeling long sequences of data points, GAT-GAN generates long time-series data of high fidelity using an adversarially trained autoencoder architecture. Our empirical evaluations, using a variety of real-time-series datasets, show that our framework consistently outperforms state-of-the-art benchmarks based on \emph{Frechet Transformer distance} and \emph{Predictive score}, that characterizes (\emph{Fidelity, Diversity}) and \emph{predictive performance} respectively. Moreover, we introduce a Frechet Inception distance-like (FID) metric for time-series data called Frechet Transformer distance (FTD) score (lower is better), to evaluate the quality and variety of generated data. We also found that low FTD scores correspond to the best-performing downstream predictive experiments. Hence, FTD scores can be used as a standardized metric to evaluate synthetic time-series data.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to generate high - quality multivariate time - series data. Traditional time - series generation methods, such as the generative adversarial network (GAN) based on the recurrent neural network (RNN), have difficulties in capturing the complex relationships between features, which results in unauthentic generated data. Especially when dealing with long - sequence data, the RNN is difficult to model long - time dependencies and is easily influenced by data points close to the current time step. In addition, existing generative models often overlook the extraction of spatial correlations, which is an important limiting factor for many practical applications, such as traffic prediction, social science research, criminology, and nonlinear system analysis. To solve the above problems, the paper proposes a time - series generative adversarial network based on the graph attention mechanism (GAT - GAN). This model can generate long - sequence time - series data with high fidelity by introducing two graph attention layers, which are respectively used to learn time dependencies and capture spatial relationships. GAT - GAN uses an auto - encoder architecture for adversarial training, in which the encoder (generator) maps the posterior distribution of latent - space variables to the prior distribution, and the decoder maps it back to the original distribution. In this way, GAT - GAN can not only effectively capture the dynamic changes over time, but also handle the spatial correlations simultaneously, thus generating more authentic multivariate time - series data. In addition, the paper also proposes a new evaluation metric - the Frechet Transformer Distance (FTD) - for evaluating the quality and diversity of generated samples. The experimental results show that GAT - GAN outperforms existing methods in a series of evaluation metrics on multiple time - series datasets, especially in generating long - sequence data. These results prove the effectiveness and superiority of GAT - GAN in generating high - quality synthetic time - series data.

GAT-GAN : A Graph-Attention-based Time-Series Generative Adversarial Network

Fully Embedded Time-Series Generative Adversarial Networks

Time-series Generative Adversarial Networks

Time-series Transformer Generative Adversarial Networks

TTS-GAN: A Transformer-based Time-Series Generative Adversarial Network

GT-GAN: General Purpose Time Series Synthesis with Generative Adversarial Networks

Generative adversarial networks in time series: A survey and taxonomy

Generative Adversarial Networks in Time Series: A Systematic Literature Review

Dyngraphgan: Dynamic Graph Embedding Via Generative Adversarial Networks

Representation Learning of Multivariate Time Series using Attention and Adversarial Training

SeriesGAN: Time Series Generation via Adversarial and Autoregressive Learning

Towards Better Long-range Time Series Forecasting using Generative Adversarial Networks

Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs

TCGAN: Convolutional Generative Adversarial Network for Time Series Classification and Clustering.

AEC-GAN: Adversarial Error Correction GANs for Auto-Regressive Long Time-Series Generation

TSI-GAN: Unsupervised Time Series Anomaly Detection using Convolutional Cycle-Consistent Generative Adversarial Networks

Directed Chain Generative Adversarial Networks

GraphGAN: Graph Representation Learning with Generative Adversarial Nets

Ten Years of Generative Adversarial Nets (GANs): A survey of the state-of-the-art

ChronoGAN: Supervised and Embedded Generative Adversarial Networks for Time Series Generation