Traffic Graph Convolutional Network with Residual Connection for Accident Severity Prediction

Urban traffic accidents have seriously affected people's life and property. The characteristics of road structure have direct impacts on the severity of traffic accidents, while traditional machine learning algorithms are difficult to take into account the complex road network structure. Therefore, research on an effective method to extract deep information within the road network is necessary. This paper proposes a traffic graph convolutional network with residual connection (ResTGC), which can deepen the structure of the neural network and effectively extract vital information through residual connection. By evaluating a data set collected from Oklahoma City, the proposed ResTGC outperforms a series of machine learning algorithms, such as SVM, GBDT, Adaboost, GCN, and GraphSAGE, with an improvement of over 6.1%. The remarkable classification ability can effectively identify the risk degree of various road network structures, which can provide a vital reference for the vehicle routing problem, municipal planning, and so on.