A data imputation method for multivariate time series based on generative adversarial network

Abstract

Multivariate time series (MTS) processing plays an important role in many fields such as industry, finance, medical, etc. However, the presence of missing data in MTS makes data analysis process more complicated. To address this issue, MTS imputation reconstructs missing data with a high accuracy by exploiting a precise model of MTS distribution. Despite being an effective tool for modeling distribution on images, generative adversarial networks (GANs) have limitations in modeling MTS distribution. In this paper, to improve MTS imputation performance, a multivariate time series generative adversarial network (MTS-GAN) is proposed for MTS distribution modeling by introducing the multi-channel convolution into GANs. It is then applied to MTS imputation by formulating a constrained MTS generation task. Experimental results show that MTS-GAN performs well in modeling MTS distribution. Compared with several approaches, the proposed MTS-GAN based imputation method not only achieves a higher imputation accuracy under different missing rates, but also performs more robustly as the missing rate increases.