Exploring the Potential and Utility of Graph Neural Networks in Crash Analysis and Safety Assessment

Increasing traffic injury and fatality numbers in Europe and all over the World compel researchers to explore novel techniques to be employed in traffic safety research, due to the limitations of traditional methods hindering an in-depth and holistic understanding of crashes. This research explores the potential and utility of graph neural network in crash prediction and safety assessment by identifying the most important features and the most important road segments affecting crashes. For this purpose, we analyzed the frequency of injury and fatal cyclist crashes in Amsterdam, Utrecht, Rotterdam, and The Hague in the Netherlands, with data from 2182, 1928, 4269, and 9858 road segments, respectively. Moreover, we used explainable artificial intelligence methods to identify important crash risk factors and network elements in crash occurrence. The findings show that the bicycle exposure, motorized vehicle exposure and the betweenness centrality of a road segment are the most important features affecting the cyclist crash likelihood. Furthermore, the GNN analysis reveals that the importance of features are not uniform across all roads. Such capability has the potential to be used in identifying the most critical links in the network that can have the highest impact on safety improvements. This study shows that graph neural networks can be promising tools for explaining safety and pinpointing the road segments that might be more problematic than others.