Univariate Time Series missing data Imputation using Pix2Pix GAN

The use of data is essential for the supply of business, scientific and other processes. Often the consumption of these data is hampered when there are sample losses. Aiming to recover values representative of these losses, there are several approaches for filling them. In this paper, we propose a new method for imputation of missing data that transforms time series into an image and thus performs imputation using the conditional generative adversarial network (cGAN) pix2pix GAN. The results of ASMAPE and MAE show that the network outperforms all methods in 50% of the datasets. It was also revealed that the proposed network can learn time series features and retain some advantages over traditional methods, such as imputing the data in its entirety and exploiting spatial and temporal features for imputation.

Automated summary

The use of data is crucial for various processes such as business and scientific endeavors. However, data consumption can be hindered by sample losses. To address this, we propose a new method that transforms time series into images and utilizes a conditional generative adversarial network (cGAN) pix2pix GAN for imputation. Results indicate that the network outperforms other methods in 50% of the datasets based on ASMAPE and MAE evaluations. Furthermore, the proposed network demonstrates its ability to learn time series features and capitalize on spatial and temporal features for imputation.