Multivariate Time Series Imputation Based on Residual GRU and AANN

Abstract

The existence of missing values brings inconvenience to data mining. In this paper, we propose a residual gated recurrent unit (GRU) and auto-associative neural network (AANN) based imputation method to impute missing values of multivariate time series. Instead of only utilizing GRU to learn the dynamic law and estimate the missing values of time series, AANN is employed to improve the estimation accuracy. By using AANN, observed values of current time step can be applied to the estimation of missing values, which makes available information can be fully utilized. Moreover, outputs of adjacent recurrent units are connected to form temporal residual network to promote the learning ability of the entire model. The experiments on several datasets validate the effectiveness of proposed method for incomplete multivariate time series imputation.