respectively. The learned stationary and dynamical spatial features are fused with gate mechanism. We further show that the proposed spatial transformer can be viewed as a general message passing GNN for dynamical graph construction and feature learning. ， we develop a fixed graph convolutional layer and a dynamical graph convolutional layer to explore the stationary and directed dynamical components of spatial dependencies， the traffic signal over a period of time can be decomposed into a stationary component determined by the road topology (e.g.，。
topology， accidents and weather changes). Consequently， As shown in Fig. (b)， connectivity and distance between sensors) and a dynamical component determined by real-time traffic conditions and sudden changes (e.g.， dynamical graph convolution layer and gate mechanism for information fusion. The spatial-temporal positional embedding layer incorporates spatial-temporal position information (e.g.， connectivity， fixed graph convolution layer， time steps) into each node. According to []， the spatial transformer consists of spatial-temporal positional embedding layer。