STCNet: Spatial-Temporal Convolution Network for Traffic Speed Prediction

With cars being a necessity of life, traffic congestion has been more and more serious in recent years. To solve such problems, it is promising to predict the speed of sensors in traffic networks, namely traffic speed prediction. While there are abundant spatial and temporal dependencies within in traffic networks, existing researches are unable to capture the structural information and dynamic evolution of sensors at the same time. In this paper, we propose a Spatial-Temporal Convolution Network (STCNet), which mainly consists of a temporal block and a spatial block. We first design the temporal traffic network to model the temporal information in the topological graph. And then, we design the temporal block to model the short-term and long-term dependencies via different receptive fields. We further present the spatial block employs the convolution operation on graph to capture the spatial dependencies among nodes. Finally, we integrate both temporal and spatial representations of sensors into a unified framework for optimization. Extensive experiments on traffic speed dataset demonstrate that our proposed STCNet model outperforms the state-of-the-art baselines.