Adaptive Graph Co-Attention Networks for Traffic Forecasting

Traffic forecasting has remained a challenging topic in the field of transportation, due to the time-varying traffic patterns and complicated spatial dependencies on road networks. To address such challenges, we propose an adaptive graph co-attention network (AGCAN) to predict traffic conditions on a road network graph. In our model, an adaptive graph modelling method is adopted to learn a dynamic relational graph in which the links can capture the dynamic spatial correlations of traffic patterns among nodes, even though the adjacent nodes may not be physically connected. Besides, we propose a novel co-attention network targeting long- and short-term traffic patterns. The long-term graph attention module is used to derive periodic patterns from historical data, while the short-term graph attention module is employed to respond to sudden traffic changes, like car accidents and special events. To minimize the loss generated during the learning process, we adopt an encoder-decoder architecture, where both the encoder and decoder consist of novel hierarchical spatio-temporal attention blocks to model the impact of potential factors on traffic conditions. Overall, the experimental results on two real-world traffic prediction tasks demonstrate the superiority of AGCAN