Learning Dynamic Spatial-Temporal Dependence in Traffic Forecasting

Accurate traffic forecasting is a key part of intelligent transport systems, facilitating a variety of urban application services such as trip alerting, route planning and traffic management. However, due to the spatial and temporal correlations of the dynamic, this issue is not well addressed. In this paper, we propose a Multi Scale Spatial-Temporal Recurrent Graph Network (MSSTRG), focusing on local temporal, multi-scale temporal and dynamic spatial correlation. Specifically, we designed a dynamic graph convolution module to model local and global spatial connections in terms of both road distance and adaptive correlation. To capture the non-linear temporal relationship, we model the forward and backward contextual links of the traffic flow sequence using the Bidirectional Gated Recurrent Layer. We also replaced the linear layer in the Gated Recurrent unit with a dynamic graph convolution operation to jointly model spatial-temporal correlation. Finally, we propose a temporal fusion layer with multi-scale features to model accurate temporal semantic information from contextual environment with different window sizes, to further obtain accurate prediction results. Experimental results on several traffic datasets show that MSSTRG outperforms the most advanced models and validates the effectiveness of the individual components proposed through ablation study.