Multi-attribute Graph Convolution Network for Regional Traffic Flow Prediction

Yue Wang, Aite Zhao, Jianbo Li, Zhiqiang Lv, Chuanhao Dong, Haoran Li

Abstract: In recent years, traffic flow prediction has been extensively explored in Intelligent Transportation Systems, which is beneficial for reducing traffic jams and accidents as well as optimizing traffic network resources. Most of the previous methods divide cities into equal-sized grids and predict flows within one grid. However, we believe that each area is not independent, and there are interactions between areas. And the interaction between areas belonging to different attributes is more regular. Therefore, we propose a Multi-Attribute Graph Convolutional Network (MAGCN) for regional traffic flow prediction. Based on the attributes to which the areas belong, we divide cities into unequal-sized grids, and then a matrix is constructed using the flow of Functional area-based Origin-Destination pairs. GCN and dilated causal convolution allow the model to capture the spatial correlation and temporal dependence between functional regions while overcoming the under-fitting of local peaks. Extensive experimental results and evaluation metrics on two real-world datasets show that the MAGCN outperforms the baselines and has a higher accuracy for traffic flow prediction.