Continuous imputation of missing values in time series via Wasserstein generative adversarial imputation networks and variational auto-encoders model

Abstract

The occurrence of missing values in time series is a common phenomenon attributed to equipment malfunction during data acquisition and transmission errors. However, most imputation methods overlook the continuous imputation of missing values, significantly losing local information resources. In this sense, we propose a new Wasserstein Generative Adversarial Imputation Network coupled with a Variational Autoencoder model (WGAIN-VAE). WGAIN-VAE utilizes a VAE as the model’s generator, leveraging the VAE’s data reconstruction capability to explore and capture latent data evolution features beyond missing positions. Simultaneously, it introduces a signaling information matrix to exit the initial model training and optimizes the neural network structure of the discriminator along with the loss function calculation strategy. Our empirical evidence indicates that the proposed model supplanted classical methods based on evaluation metrics frequently used to measure the accuracy of forecasting or adjustment models in time series.