GCompletor: A Graph-Based Deep Learning Method for Traffic State Imputation on Urban Road Networks

Kaijie Li, Juanjuan Zhao, Li Yan, Xitong Gao, Ye Li, Kejiang Ye

Abstract: Complete traffic data is the premise for traffic strategy making. However, due to the constraints of data collection, communication failures and other reasons, we may collect incomplete traffic states inevitably. The common idea of existing completion methods is to learn the latent representation reflecting the spatiotemporal correlation in the traffic data. However, due to insufficient influencing factors considered and limited capability of spatiotemporal correlation modeling, existing methods need further improvement, especially for the scenes with high missing rates. In this paper, we propose a novel traffic state imputation method GCompletor using Graph-based Encoder-Decoder framework, which enriches the features of each road by considering the physical features (road grade, direction, etc.), and organize all traffic features into a graph-based sequence. Then the sequence is fed into a novelly designed Encoder-Decoder component, where the spatiotemporal dependencies of each road is learned through extended GAT and BiGRU-CNN hybrid method. Experimental results demonstrate that GCompletor achieves better imputation performance than the state-of-the-art approaches. The source code is available at https://github.com/zfrInSIAT/GCompletor.