TimeGAE: A Multivariate Time-Series Generation Method via Graph Auto Encoder

Zhao Bai, Fangda Guo, Yuxin Xi, Zhuoming Zhu, Yu Guo, Rongfang Bie

Abstract: In numerous practical scenarios, the scarcity of data presents a challenge for the implementation of large and complex deep learning models. To address this issue, time series generation methods have emerged as a promising approach to alleviate data scarcity. However, most existing methods do not explicitly consider multivariate time series, thereby failing to fully exploit the potential spatial dependencies among different variables. The ability to capture temporal dynamics also plays a critical role in generating diverse and feasible synthetic data. In this paper, we propose TimeGAE, a novel multivariate time series generation model based on graph auto encoder. Our approach explicitly models multivariate time series by leveraging graph learning layers to learn the internal relationships among variables. This enables the generation of realistic time series based on the learned internal relationships. We also incorporate transformer-encoder or RNN modules to enhance the ability to retain temporal dynamics for time series feature extraction. Several comparisons and ablation experiments on three multivariate time series datasets have been conducted. Our results demonstrate that TimeGAE-generated synthetic data is significantly more realistic and that TimeGAE outperforms other state-of-the-art time series generation methods on multivariate time series datasets.