Content-Aware Traffic Data Completion in ITS Based on Generative Adversarial Nets

Abstract

Big data analytics has been rapidly integrated into Intelligent Transportation System (ITS), empowering diverse applications such as real-time traffic prediction and management. However, incomplete traffic time-series data during the data analysis are nearly inevitable due to the constraints of data collection or packet loss in the communication process. Existing tensor-based completion methods fail to perform well in consecutively missing (CM) cases where a sub-series of the traffic time series is completely missing. Moreover, their high computational complexity prevents the road network-level implementation. Therefore, to tackle these problems, a batch-oriented traffic data completion method for large-scale road networks is proposed in this paper. In order to preserve multi-way natures of traffic data, we first model the traffic data by tensors so that traffic data completion becomes a tensor completion problem. Several types of tensor structures are adopted in order to analyze their impacts on traffic data completion. Thereafter, a Content-Aware traffic data completion method is further developed based on the Generative Adversarial Net (CA-GAN). More specifically, the formulated traffic tensors are interpreted as samples from a high-dimensional traffic distribution and a GAN is then proposed to learn this distribution. By considering effects of traffic data from different days and links, a weighted loss function is employed to search for the raw traffic tensor in the obtained traffic distribution. Afterwards, the estimated raw traffic tensor is used to complete the missing data. Visualization results demonstrate that our GANs can produce verisimilar traffic patterns with different time spans. Finally, our experiments on real traffic datasets show that the proposed CA-GAN method can effectively batch incomplete traffic tensors in CM cases with fast recovery process.