Dual-branch GCN for short-term traffic forecasting

This paper proposes a novel hybrid static and dynamic spatio-temporal graph convolutional network model. Through a static graph branch, the model utilizes Chebyshev graph convolution to encode the fixed road network topology. Simultaneously, a dynamic graph branch employs a learnable graph structure to adaptively infer potential and time-varying interactions between nodes. The spatio-temporal features extracted by the two branches are fused and subsequently modeled by a temporal convolution module for sequence learning. Experiments on two real-world traffic datasets validate the effectiveness of simultaneously capturing both static and dynamic spatial dependencies.