Toward Directed Spatiotemporal Graph: A New Idea for Heterogeneous Traffic Prediction

Enhancing the accuracy of traffic prediction relies on building a graph that effectively captures the intricate spatiotemporal correlations in traffic data. It is a widely observed phenomenon that different urban traffic activities exhibit an asymmetric mutual influence. However, existing methods for graph construction largely overlook this characteristic. To bolster prediction performance, this article introduces an attention mechanism based on transfer entropy (TE) to quantify the complex and asymmetric spatiotemporal correlations among traffic data. Subsequently, a graph attention network based on simplified TE calculation is devised to construct directed spatiotemporal graphs from heterogeneous traffic data. Finally, a directed spatiotemporal graph neural network is employed for training and prediction purposes. Experimental results demonstrate that our approach surpasses existing mainstream methods for spatiotemporal graph construction, leading to significant improvements in predicting heterogeneous traffic data. Further analysis reveals that TE exhibits higher sensitivity to asymmetric spatiotemporal influences in heterogeneous traffic environments compared to commonly used data dependency inference algorithms. This finding further validates the feasibility and advancement of our method in predicting heterogeneous traffic data.