An effective variational auto-encoder-based model for traffic flow imputation

Traffic flow imputation is crucial in modern intelligent transportation systems, due to frequent data missing caused by the failure of detectors or the influence of hostile external environment (e.g., signal strength). However, this work can be challenging for mainly three reasons: firstly, how to impute traffic flow that shows both universal complex spatio-temporal regularity and individual random authenticity is non-trivial. Secondly, though there are so many algorithms for traffic flow imputation, most of them either ignore different periodic dependencies or just deal with them independently. Lastly, the effectiveness of most deep learning models (etc., CNN, RNN) may be influenced by data missing in model training phase. To solve the problems above, an effective model traffic flow imputation variational auto-encoder (TFI-VAE) considering robust joint-periodic spatio-temporal features is proposed, which can impute missing value not only accurately but also realistically by introducing Gaussian mixture distribution enhanced VAE and normalization flows. Moreover, spatial missing oriented block and temporal missing oriented block are utilized in TFI-VAE to learn the spatial and temporal features of traffic flow data with ability to resist the negative effects of missing data. The experiments conducted on three real-world traffic flow datasets demonstrate that our TFI-VAE outperforms other classical imputation models from all aspects.