Multivariate time series data imputation using attention-based mechanism

Abstract

As the widely deployment of different sensors and Internet of Things, a large volume of multivariate time series has been collected. However, there are many missing values in the multivariate time series due to different reasons, such as sensor damage, environmental intrusion and machine failure. These missing values bring challenges to further analysis of multivariate time series. The existing methods for multivariate time series imputation either destroy the properties of the original data or fail to capture the correlations of the data effectively. In order to solve the problems in existing methods, we propose a novel imputation method for multivariate time series based on Generative Adversarial Networks (GAN). Specifically, we use Auto-Encoder (AE) as the generator (G), Recurrent Neural Networks (RNN) as the discriminator (D). Taking advantages of generative adversarial networks to impute the incomplete multivariate time series data. In addition, Attention Mechanism is integrated to the structure of the Encoder and Decoder to keep the correlations among data. By integrating the advantages of Attention Mechanism, it enables the generator to take advantage of the internal features of the original data when imputing the data. Therefore, the imputed multivariate time series are more reasonable and reliable. Furthermore, we also improve our model by integrating Self-Attention Mechanism to improve the accuracy for multivariate time series imputation. Experimental results on real datasets demonstrate that the performance of the proposed model exceeds existing models and achieves state-of-the-art performance.