Abstract
Predicting the road traffic speed is a challenging task due to differenttypes of roads, abrupt speed changes, and spatial dependencies between roads,which requires the modeling of dynamically changing spatial dependencies amongroads and temporal patterns over long input sequences. This paper proposes anovel Spatio-Temporal Graph Attention (STGRAT) that effectively captures thespatio-temporal dynamics in road networks. The features of our approach mainlyinclude spatial attention, temporal attention, and spatial sentinel vectors.The spatial attention takes the graph structure information (e.g., distancebetween roads) and dynamically adjusts spatial correlation based on roadstates. The temporal attention is responsible for capturing traffic speedchanges, while the sentinel vectors allow the model to retrieve new featuresfrom spatially correlated nodes or preserve existing features. The experimentalresults show that STGRAT outperforms existing models, especially in difficultconditions where traffic speeds rapidly change (e.g., rush hours). Weadditionally provide a qualitative study to analyze when and where STGRATmainly attended to make accurate predictions during a rush-hour time.