Multivariate Time Series Imputation with Bidirectional Temporal Attention-Based Convolutional Network

Abstract

The problem of missing data in time series will make the process of analysis much more tough and challenging. Imputation of missing values in multivariate time series can effectively solve this problem. Recurrent neural networks (RNNs) are widely used in sequential data due to their properties of sequential modeling. However, RNN has some problems such as gradient and long calculation time. In recent years, time series modeling has been fully developed utilizing a feedforward model based on convolutional networks and an attention mechanism, which has the advantage of parallelism over RNNs. This paper proposes a multivariate time series imputation model (BTACN) based on Temporal Convolutional Networks (TCN) and attention mechanism. Multivariate time series features were extracted by bidirectional TCN, and then attention was weighted to capture the long-term and short-term dependence of time series. Minimizing both reconstruction and imputation loss is used to train the model. Experiments on real datasets and simulated datasets reveal the superiority of the proposed method in terms of imputation performance.