Sparse Attention-based Graph Neural Networks for Traffic Forecasting

Traffic forecasting is a significant task for an intelligent traffic system, which requires models with high predictive power and needs to acquire the exact data dependence. But the complexity of traffic flow data makes this task quite challenging. Previous studies have shown the effectiveness of Transformer in improving the predictive ability of models. Sparse Attentional Graph Neural Network (SAGNN) based on Transformer is proposed for traffic forecasting, using dynamic graph convolution and a new attention mechanism to get the dependencies in the different dimensions, respectively. Specifically, the operations of sparsification and aggregation of contextual information to the self-attention mechanism are added to make it efficient and process the adjacency matrix before graph convolution to simulate the dynamics of the road network. In addition, SAGNN also considers the data's periodicity and optimizes the dynamic decoding process. Two open highway datasets are used to validate SAGNN. And the results show that SAGNN effectively extracts traffic flow spatio-temporal information. The forecasting results outperform the prevalent baselines.