The use of Graph Neural Networks for Clustering in Road Safety Analysis

Abstract

Road safety analysis aims to reduce traffic crashes and enhance transportation systems' efficiency. Graph theory, and its related mathematical framework, offers tools to model complex road networks and analyze various dynamics such as traffic flow, routing, crashes correlation, etc. By representing road networks as graphs and integrating telematics data onto the graph, it becomes possible to capture behaviors on the network. Graph Neural Networks (GNNs) provide a robust framework for analyzing such data and enabling network-level inference. The study presents an approach to identify node clusters in a network using edge features. Specifically, selecting an area within the Athens metropolitan area, it compares a clustering on node features with a clustering based on node embeddings generated by a GNN that incorporates the attention mechanism. By leveraging this mechanism, edges features are not left out from the analysis and the clustering algorithm shows better overall performance in terms of cluster quality.