A graph attention fusion network for event-driven traffic speed prediction.

• A novel framework named Event-Aware Graph Attention Fusion Network is proposed. • A SEE block is proposed to learn the direct spatial correlations of event. • A TE-Attention block is proposed to learn the temporal correlations of event. • A gated fusion mechanism is proposed to fuse the features of road traffic and event. • A dynamic event graph constructor is proposed to construct dynamic graph for event. Accurate road traffic speed prediction has a critical role in intelligent transportation systems and smart cities. This task is very challenging because of the complexity of road network structures, as well as various other unpredictable and ad hoc factors. Most existing traffic speed approaches handle external factors such as weather, holidays, and traffic accidents to enhance prediction accuracy. However, they ignore the impacts of social events or only simply embed them instead of learning the spatio-temporal representations. To address this issue, we design a novel framework named an event-aware graph attention fusion network (EGAF-Net) to effectively capture the spatiotemporal dependencies, including event impacts, in road networks based on an encoder-decoder architecture for traffic speed prediction. First, we utilize an ST-Speed attention block to model the spatial correlations among road segments and capture traffic speed changes. Second, we develop a spatial event embedding block exploiting a novel algorithm based on the node2vec approach, a new dynamic event graph constructor which produces learnable graphs utilized in graph convolution layers, and a temporal event attention block to learn the spatial and temporal representations of events. Finally, we propose a gated fusion mechanism to fuse the spatio-temporal correlations in road networks and the representations of events. Extensive experiments conducted based on the Q-Traffic, Q-Eastern-Traffic and Q-Western-Traffic datasets demonstrate the effectiveness of EGAF-Net over robust baselines. [ABSTRACT FROM AUTHOR]