Learning dynamic and multi-scale graph structure for traffic demand prediction

Traffic demand prediction plays a crucial role in developing modern transport systems, as it can alleviate the dilemma of demand-and-supply imbalances in urban traffic. However, most existing traffic demand prediction works lack the ability to (1) efficiently capture the dynamic and multi-scale spatial dependency and (2) effectively utilize multi-scale and inter-multi-scale temporal features. To address these challenges, this paper develops a Dynamic and Multi-scale Graph Learning method, referred to as DMGL, for traffic demand prediction. In DMGL, a dynamic graph generator module is initially devised to construct different-scale dynamic graphs through temporal feature decomposition and aggregation. Next, a novel multi-scale temporal representation method is introduced that simultaneously captures both multi-scale and inter-multi-scale temporal dependencies. Lastly, a graph convolution module is leveraged to model dynamic and multi-scale spatial dependencies. To showcase the effectiveness of DMGL, we conduct experiments on two datasets, and the results of DMGL surpass the baselines.