Multistate time series imputation using generative adversarial network with applications to traffic data

Time series missing data is a pervasive problem in many fields, especially in intelligent transportation system, which hinders the application of timing analysis methods and the fine adjustment of control strategies. The prevalent imputation approaches reconstruct missing data with a high accuracy by exploiting a precise distribution model. But the multistate characteristic of time series data and the uncertainty of imputation process increase the difficulty of modeling temporal data distribution and reduce the imputation performance. In this paper, a novel time series generative adversarial imputation network (TGAIN) model is proposed to deal with time series data missing problem. The model combines the advantages of GAN's data distribution modeling and multiple imputation's uncertainty handling. Specifically, the TGAIN network is designed and adversarial trained to learn the multistate distribution of missing time series data. Through the conditional vector constraint and adversarial imputation process, the latent distribution for each missing position under different states can be effectively estimated based on implicit relationships with partial observation information. Then the corresponding multiple imputation strategy is proposed to deal with the uncertainty of imputation process and it can determine the best fill value from the learned distribution. Furthermore, sufficient experiments have been conducted in two real traffic flow datasets. The comparative results show the proposed TGAIN not only has better ability on time series data distribution modeling and imputation uncertainty handling, but also performs more robustly and stability even with the missing rate increases.