Multi-Scale Convolution Multi-Graph Attention Neural Networks for Traffic Flow Forecasting

Spatial-temporal data has wide-ranging applications in our daily lives, and spatial-temporal forecasting has now evolved into a current hotspot of research. Traffic flow forecasting serves as a typical example in this domain. Traffic flow exhibits complexities, nonlinearity, and highly dynamic spatial-temporal correlations, which remain a primary challenge for current researchers. For this problem, we propose a Multi-Scale Convolution Multi-Graph Attention Network (MCMGAT). Specifically, in terms of temporal correlations, we first employ channel-wise attention to allocate appropriate weights to different time steps to strengthen informative time steps and suppress irrelevant ones. Then, we use temporal convolution modules that consist of multiple convolution kernels of varying sizes to capture temporal correlations across different time scales. Concerning spatial correlations, we first introduce two adjacency matrices to simultaneously model local spatial correlations and long-distance spatial similarities. Then we employ dualbranch graph attentions to capture dynamic spatial dependencies. Experiments on four real-world datasets indicate that MCMGAT outperforms all baselines. Additionally, we conduct visual analyses to enhance the model's interpretability.