Using Multidimensional Data to Analyze Freeway Real-Time Traffic Crash Precursors Based on XGBoost-SHAP Algorithm

The traditional freeway safety studies with “poststatic” thinking basically use cross-sectional data or panel data, which find it difficult to figure out real-time traffic crash risk factors. With the development of information collection technology, it is possible to obtain high-resolution traffic flow data currently, which provide a data basis for the dynamic traffic safety research towards freeways. This research aims at accurately identifying the real-time traffic crash precursors on freeways and addressing the shortcomings of conventional dynamic traffic safety research with the thinking of limited factor dimensions. In this research, dimensional data were applied as input model variables, the input dataset includes traffic crash data and the matched dynamic traffic flow data, and weather information and road characteristics were also considered to figure out the interaction effects between these dimensional factors. The XGBoost (eXtreme Gradient Boosting) was carried out to identify the dynamic crash-prone variables and the SHAP (SHapley Additive exPlanations) interpreter was introduced to interpret the XGBoost model, as well as the visualization of the influence of each eigenvalue on the traffic crash was realized. The results indicate that, in addition to traffic flow variables, road, weather, and temporal characteristics also have an impact on the traffic crash risk, and there is an interaction between each feature. The results of this research can provide the theoretical basis for freeway real-time traffic crash prediction and safety control.