Research On Multivariate Time Series Imputation With Generative Adversarial Network

Recent advances in sensors and hardware promote the rapid emergence of electronic data especially the multivariate time series data.Applications of multivariate time series include predicting the health status of patients,weather forecasting and stock market analysis.The missing values,appeared in most of multivariate time series,prevent advanced analysis of multivariate time series data.Existing approaches try to deal with missing values by deletion,statistical imputation,machine learning based imputation and generative imputation.However,these methods are incapable of dealing with temporal information and modeling the nature of complex distribution in multivariate time series.In this paper,Gated Recurrent Units for Imputation(GRUI)is proposed to model the temporal irregularity of the incomplete time series.Then we propose a Generative Adversarial Networks(GAN)based multivariate time series imputation method.This method learns the distribution of multivariate time series with the help of GRUI.What's more,it also tries to find the best input vector for each time series so that the generated time series is most closely to the original one.The generated time series will be used to impute the incomplete time series.However,this method is a multistage method and costs too many time to train the imputation model.This paper also proposes an end-to-end generative model E~2GAN to impute missing values in multivariate time series.With the help of the Denoising Auto-Encoder,E~2GAN can impute the incomplete time series by the nearest and best-connected generated complete time series at one stage.In this way,E~2GAN gains better time efficiency than multistage method on the training of neural networks.Experiments on multiple real-world datasets show that our model outperforms the baselines on the imputation accuracy and achieves state-of-the-art classification/regression results on the downstream applications.