TAST-GCN: Topology Adaptive Spatial-Temporal GCN for traffic accidents prediction

Traffic accidents usually lead to bodily injury and economic loss. Accurate prediction help to prevent accidents happening and lower possible loss. However, urban traffic condition is effected by complex factors such as road infrastructure, traffic flow speed and adverse weather conditions. Thus, accidents prediction is a challenging task. To solve above problems, we proposed TAST-GCN. TAST-GCN contains three module: spatial module, spatial-temporal module, natural FC module. Spatial module performs graph convolution operations to learn spatial correlations. Spatial-temporal module utilizes graph convolution and temporal pooling to learn spatial-temporal correlations. Natural FC module aims to obtain meaningful representations of natural weather data. To evaluate the proposed model, wide range of real-world data, including weather, traffic flow, POI, road networks and accidents historical records are collected. Experiments on real-world dataset prove that TAST-GCN outperforms traditional and state-of-the-art methods.