Task-Dependent Degradation of Data-Driven Safety Models at Unsignalized Intersections Under Multi-Granularity Data: An Interpretable Perspective

Unsignalized intersections involve complex interactions among heterogeneous road users and are associated with elevated safety risks. Although surrogate safety measures derived from high-resolution trajectories enable proactive safety assessment, such data are not widely available in routine monitoring systems, which often provide only coarse-grained traffic observations. This study examines how the inferability of surrogate safety information changes as the available traffic data become progressively coarser. Using the high-resolution inD dataset, we implement a controlled feature degradation framework across three nested levels of data granularity and develop intersection-specific models for three tasks: critical conflict detection, dominant direction classification, and vulnerable road user (VRU) involvement identification. Model performance and changes in variable importance are evaluated using PR-AUC and SHAP analysis. The results show clear task-dependent degradation. Models based on high-granularity data achieve strong overall performance, with an average PR-AUC above 0.88. Dominant direction classification remains relatively robust as data granularity decreases, with PR-AUC declining from 0.970 to 0.893, whereas VRU involvement identification deteriorates substantially, from 0.991 to 0.697. The results further indicate that vehicle-based traffic variables retain meaningful predictive value for conflict detection and direction classification but are insufficient for reliable inference of VRU-related risk. Interpretability analysis shows a progressive shift in model reliance from kinematic interaction variables to coarser exposure-related and structural descriptors as observability decreases. These findings clarify the relationship between data granularity and task-dependent surrogate safety inference at unsignalized intersections.