Network-wide traffic state reconstruction: An integrated generative adversarial network framework with structural deep network embedding

Abstract

Traffic data imputation plays a crucial role in Intelligent Transportation System (ITS) applications when handling missing data. Previous methods have primarily focused on restoring traffic states for road links equipped with sensors but sometimes suffer from the issue of large-scale missing data. However, these approaches are inadequate for addressing the special scenario where multiple road links lack sensors, resulting in a complete data absence. To tackle this issue, this study proposes an Integrated Deep Learning for Traffic State Reconstruction (IDL-TSR) framework, which aims to reconstruct the network-wide traffic state using sensor data from a limited number of links. Specifically, we employ the Structural Deep Network Embedding (SDNE) to first embed high-dimensional and sparse traffic data (i.e., spatiotemporal images) into a low-dimensional space. This embedding process assists in alleviating the negative impact of data sparsity on the generative capacity of the Generative Adversarial Network (GAN) and enhances the spatiotemporal mining capability of the IDL-TSR. Then, a typical GAN with Wasserstein divergence (WGAN-div) is used to reconstruct traffic states for road links without sensors by leveraging the observed traffic data from the other links. To evaluate the performance of the proposed IDL-TSR framework, we conducted numerical experiments using traffic speed data obtained from Didi Chuxing in Chengdu, China. The results demonstrate that the IDL-TSR framework effectively reconstructs traffic speed states at a network level, surpassing other counterparts even under extreme missing rates.