Adaptive Diffused Spatiotemporal Graph Convolution for Traffic Flow Forecasting

Recent research typically formulates traffic forecasting tasks as spatiotemporal graph modeling problems. For spatial correlations, researchers typically use predefined graphs to learn short-range spatial dependencies among traffic data, neglecting the learning of long-rang spatial features that are crucial for accurately capturing the dependencies among traffic data. For temporal correlations, studies tend to focus on local continuous correlations while overlooking global temporal correlations. To this end, this paper presents an in-depth study on long-range spatial dependencies and global temporal continuity. Specifically, a novel method called adaptive diffused spatiotemporal graph convolution network (ADSTGCN) is proposed for accurate traffic flow prediction. In this approach, an adaptive adjacency matrix–based graph convolution operation is introduced to learn local spatial features. Building on this, global spatial interdependencies between traffic series are represented through graph diffusion convolution, which also relies on the adaptive adjacency matrix. Additionally, a novel module for learning continuous temporal correlations is introduced to effectively capture the sequential patterns in traffic flow data. Additionally, multihead attention is employed to model global temporal dependencies among the traffic series. The experimental results show that the proposed approach achieves superior performance compared to other state-of-the-art methods.