Local and Global Spatial-Temporal Networks for Traffic Accident Risk Forecasting

Abstract

Traffic accident forecasting is very important for urban public security, emergency treatment and construc-tion planning. However, the following problems still exist when forecasting traffic accident risk. Firstly, traffic accidents are affected by multiple factors, such as weather and road conditions. Besides, there are multi-scale correlations in the spatial dimension, i.e. local region spatial-temporal correlation and global region spatial-temporal similarity. Meanwhile, there is zero-inflated issue in the forecasting because of few traffic accidents in reality. Therefore, it is very challenging to forecast traffic accidents accurately, and existing traffic accident forecasting methods cannot take all the above problems into account. A novel model, named local and global spatial-temporal networks (ST-RiskNet), for traffic accident risk forecasting is proposed. The ST-RiskNet takes multi-source factors that affect traffic accidents into consideration, such as time, weather, traffic flow. It uses a local region spatial-temporal correlation module and a global region spatial-temporal similarity module to model the multi-scale spatial-temporal correlation and similarity simultaneously. Meanwhile, a sample weighted loss function is designed to solve the zero-inflated problem, which pays more weights to the higher risk samples. Extensive experiments on two real-world traffic accident datasets demonstrate the effectiveness of the ST-RiskNet against the state-of-the-art baseline methods.

Keywords

• traffic accident forecasting
• multi-source spatial-temporal data
• zero-inflated
• graph convolutional network (gcn)