IIT-GAN: Irregular and Intermittent Time-series Synthesis with Generative Adversarial Networks

Jinsung Jeon, Jeonghak Kim, Haryong Song, Noseong Park

Keywords: time-series synthesis, GANs, differential equations

Abstract: Time-series data is one of the most popular data types in the field of machine learning. For various reasons, there is a strong motivation to synthesize fake time-series data. Several disparate settings for time-series synthesis have been previously solved, ranging from synthesizing time-series without any missing values to time-series of multiple signals with different frequencies. In this paper, we solve the problem of synthesizing irregular and intermittent time-series where values can be missing and may not have specific frequencies, which is far more challenging than existing settings. To this end, we adopt various state-of-the-art deep learning concepts, such as autoencoders (AEs), generative adversarial networks (GANs), neural ordinary differential equations (NODEs), neural controlled differential equations (NCDEs), and so on. Our contribution lies in carefully re-designing those heterogeneous technologies and proposing our unified framework. Our method achieves the state-of-the-art synthesis performance for the irregular and intermittent time-series synthesis task.

One-sentence Summary: Synthesize irregular and intermittent time-series