It Is About Weather: Explainable Machine Learning for Traffic Accident Understanding

Road traffic accidents cause injuries, claim lives, and disrupt economic activities. It is among the key problems for intelligent transportation and smart cities. Cities, especially the mega ones, must strive for reducing accidents for public safety and sustainable growth, and the first task is to understand accidents. In this paper, we build up such understanding with the emerging explainable machine learning (ML) technique. We prepare a huge dataset with over two million accident records and use it to deliver ML models for accident modelling. Given ML models of high fidelity for mapping accident features and conditions to the accident severity, we apply several explainable ML techniques to explain the models and understand accidents. We first consider coarse granularity to capture the overall feature importance. Then we consider fine granularity methods including partial dependence plot and Shapley additive explanations. The former shows that feature impact varies at different feature values and uses a plot to reflect the impact changes. The latter puts attention to individual accident and quantifies the feature impact for the specific accident. Our core observation is that traffic accidents are often about weather, followed by location and road type. Extensive experimental study is performed to support our discussion and justify our conclusion. The deliverable of this paper offers an advanced way of understanding traffic accidents accurately in a quantitative manner and has great potential to be used for intelligent transportation and smart city applications.