Multihead Attention-Based Multiscale Graph Convolution Network for ITS Traffic Forecasting

Traffic forecasting is a challenging issue in the transportation field due to its high nonlinearity and complexity. The key to extract valuable information from traffic data is to characterize the spatial and temporal correlations in a proper way, but it is difficult to achieve accurate quantification on these correlations, especially the spatial ones. Since road network in reality follows non-Euclidean geometry, graph convolution network (GCN), a semisupervised neural network for non-Euclidean graph modeling, has widely been applied in traffic forecasting to capture the spatial correlation of traffic flow. However, most of these GCN-based methods use a single definition on spatial correlation, which cannot precisely reflect the complicated association of road network. Meanwhile, the traditional form of graph convolution is the aggregation of neighboring nodes information, which is equal to a smoothing operation. When this operation repeats, the original data gets smoother, and that may lead to oversmoothing problem and the loss of some important characteristics of data. In response to these issues, a novel multiscale graph convolution method is proposed, in which three representations of the spatial structure of road network are defined and integrated through the multihead attention mechanism. Meanwhile, to avoid oversmoothing, the calculation of graph convolution is redefined to fuse the results of graphs with different scales of convolution by trainable adjustment factors. The proposed method is verified by experiments from different aspects.