3D Convolutional Generative Adversarial Networks for Missing Traffic Data Completion

Abstract

The problem of data missing is a common issue in practical traffic data collection for an Intelligent Transportation System. However, how to efficiently impute the missing entries of the traffic data is still a challenge. Previous works on missing traffic data imputation usually apply matrix or tensor completion based methods which are unable to fully exploit the spatio-temporal features of historical traffic data to achieve a satisfactory performance. In this paper, we propose a 3D convolutional generative adversarial networks to impute missing traffic data. We propose to use a fractionally strided 3D convolutional neural network to construct the generator network since it can upsample efficiently in a deep network and a 3D convolutional neural network to construct the discriminator network to fully capture spatio-temporal features of traffic data. We also present numerical results with real traffic flow dataset to show that the proposed model can significantly improve the performance in terms of recovery accuracy over the other existing tensor completion methods under various data missing patterns. We believe that the proposed approach provides a promising alternative for data imputation in ITS and other applications.