Paradiso S., Ziakopoulos A., Sideris H., Yannis G., “A Graph Transformer Approach for Modeling Crash Occurrence at Intersections using Telematics-Informed Road Networks”, Proceedings of the 10th Road Safety and Simulation International Conference, Naples, 23-26 June 2026.

Spatial analysis has long been used in road safety to identify high-risk locations and uncover spatial patterns associated with crashes or hazardous behavior. The field has recently expanded thanks to advances in artificial intelligence and the growing availability of telematics data, enabling the development of more spatially informed models. This study leverages smartphone-collected telematics data and historical crash data, mapped onto the road network, along with road network features from OpenStreetMap, to predict crash risk at intersections in the city centre of Athens. The problem is formulated as a binary classification task, where intersections are labelled according to whether at least one crash has occurred. Two modelling approaches were compared: Extreme Gradient Boosting (XGBoost), one of the most commonly used machine learning methods in road safety, and a Graph Neural Network (GNN) with TransformerConv layers, which is well-suited for road network analysis, since it handles data in graph form, thus preserving its structure. While XGBoost captures relationships at the feature level, the GNN leverages the road network topology, improves node representations by incorporating information from neighbouring nodes, and integrates edge features.  Both models were trained using optimized hyperparameters to maximize accuracy, and undersampling was applied to address severe class imbalance. The GNN outperformed XGBoost and was selected for further exploratory analysis. Based on model interpretations via Integrated Gradients, speed variability, harsh event intensity, and the number of streets at each intersection were identified as the most influential features, providing actionable insights for targeted interventions. This study demonstrates an approach to combine telematics and crash data within a graph-based AI framework, enabling crash risk estimation and interpretable insights at intersections. The proposed methodology highlights the potential of analyzing a telematics-informed road network with GNNs to support data-driven traffic safety analysis and prioritization of locations for further investigation.