SHAP-based interpretation of holistic traffic risk causation patterns using high-resolution trajectory data and risk...

Establishing the relationship between traffic states and vehicle conflicts using high-resolution trajectory data is an effective approach for assessing road traffic safety. However, predicting the likelihood of a single type of conflict cannot fully capture the holistic risk level of a road segment. This study aims to develop a method for evaluating road segment-level traffic risk irrespective of conflict types and to analyze the key factors influencing such risks, thereby exploring their underlying mechanisms.

We propose an integrated surrogate safety measure based on risk field theory, enabling comprehensive identification of vehicle conflicts regardless of relative motion states. Road segment-level holistic risk assessment models were then developed using binary logistic regression and five machine learning methods, with traffic state variables as inputs and conflict occurrence as the output. Resampling techniques were applied to mitigate the effect of imbalanced sample categories.

The random forest model trained on the 60-s oversampled dataset achieved the best performance, with an overall prediction accuracy of 80.6%, a precision of 89.0%, and a recall of 86.8% for identifying conflict cases.

Using SHAP analysis, we interpreted the contribution of individual traffic state variables and their interactions to traffic risk. The vehicle type distribution in traffic flow was identified as a key factor influencing conflict occurrence. This study not only provides accurate risk prediction but also enhances the interpretability of machine learning-based traffic risk models, bridging prediction and causal insight.