Score-Based Spatial-Temporal Point Process for Traffic Accident Prediction

Traffic prediction is a crucial aspect of modern traffic management and has been a research focus for decades. Unlike the forecasting of traffic flow, speed, and demand, traffic accidents occur irregularly and are highly unpredictable. As a result, developing theory-based methods for traffic accident prediction is particularly challenging. In this study, we model the occurrence of traffic accidents as a Spatial-Temporal Point Process (STPP). First, we decompose the intensity function of the STPP into a Neural Temporal Point Process (NTPP) and a conditional spatial distribution. To manage both discrete and continuous historical information, we propose a contextual embedding module utilizing multi-head self-attention. The TPP is then modeled as a Hawkes Process, with the intensity function generated by neural networks. Afterwards, we employ a score-based diffusion model to learn the conditional spatial distribution. In addition, we introduce a co-prediction module to forecast the severity and duration of future accidents. We verify the effectiveness of our model based on real-world traffic accident datasets from three cities. The results demonstrate that our model can capture the complicated spatial-temporal patterns of traffic accidents well and outperform current approaches.