Integrating a non-gridded space representation into a graph neural networks model for citywide short-term crash risk prediction

Current applications of Graph Neural Networks in citywide short-term crash risk prediction have been limited by a gridded representation of space, which restricts the network’s capability to effectively capture the spatial and temporal dependency of crash occurrences. In addition, a grided representation does not match most geographic units used for administrative purposes, limiting the use of crash risk predictions by practitioners. This paper applies a gated localised diffusion graph neural network (GLDNet) model to compare the use of two alternative geographic units, Mesh Block (MB) and grid, to forecast locations where crashes are likely to occur in a future time window. The GLDNet relies on a graph-based representation of geographic units and a weighted loss function to address the sparsity of crash occurrences. The tests are performed using crash data from the City of Melbourne, Australia, over a period of one year. The predictions are made at six-hour intervals, and the results show that the GLDNet consistently outperforms baseline methods, with differences in prediction accuracy from 10% to 21% in relation to historical average and benchmark deep learning models. In terms of geographic units, the MB-based GLDNet performed better than its grid counterpart, with differences in prediction accuracy of up to 12.3%. The better performance stems from the underlying information attached to the MB units (i.e., land use) and the network properties (i.e., degree of centrality), which enhance the GLDNet capability to identify crash risk in both central and peripherical areas. Regarding its applicability, the MB-based GLDNet directly integrates with other data sources, which provides contextual information about crash hotspots that helps decision-makers develop police patrolling and rescuing strategies.