Mutual influence among different transport activities is a crucial factor for heterogeneous traffic prediction in modern urban transport systems. The asymmetry of influence has been widely observed in reality but has not been properly introduced in the relevant studies. In this paper, we propose an asymmetric heterogeneous traffic flow prediction model, TEDGCN, which involves two major parts: constructing a directed heterogeneous traffic graph via transfer entropy (TE) as input and building a stacked network formed by directed graph convolutional network (DGCN) and gated recurrent unit (GRU) for extracting and representing asymmetric spatiotemporal characteristics. The results of the comparative experiment show that the proposed method outperforms the current mainstream prediction methods based on undirected graphs for heterogeneous traffic prediction. The ablation experiment further demonstrates that the proposed TE attention is more effective than general attention in constructing directed traffic graphs. Finally, the experiment on correlation analysis indicates that our method is more sensitive to changes in the asymmetry of heterogeneous traffic, thus performing better than the baseline method as the asymmetry of mutual influence in heterogeneous traffic environments grows.
TEDGCN: Asymmetric Spatiotemporal GNN for Heterogeneous Traffic Prediction
2023-09-24
1144580 byte
Conference paper
Electronic Resource
English
TRAFFIC FLOW PREDICTION SYSTEM USING SPATIOTEMPORAL STOCHASTIC MODEL
European Patent Office | 2015
|Spatiotemporal Patterns in Large-Scale Traffic Speed Prediction
Online Contents | 2014
|