Determining Statistically Significant Road Accident Spatial Hotspots using Machine Learning Approaches

Spatial Clustering is a subclass of Unsupervised Machine Learning algorithms that utilizes a specified criteria to group a set of geographically dispersed data points together. Effective clustering algorithms are required in societal applications such as road safety to uncover useful patterns linked with the data. Many applications employ Spatial Scan Statistics to identify spatial clusters, however it needs users to first determine the shape of the cluster, which is ambiguous in the context of road safety and could result in unfavourable outcomes. The paper outlines a method for discovering statistically significant shape-invariant spatial clusters in order to identify high-risk road accident zones. The proposed approach incorporates the OPTICS and HDBSCAN clustering algorithms, as well as Cluster density and the Log Likelihood Ratio methods for evaluating the significance of spatial clusters. OPTICS and HDBSCAN are optimal for finding clusters of arbitrary shape and only require a single input parameter, making hyper-parameter tuning relatively easier. The paper incorporates the statistical significance of clusters to eliminate spurious patterns. To illustrate the obtained results, the proposed approach is applied to the UK Road-Accidents data. The paper also presented a comprehensive analysis of the seasonal and temporal variations of the significant spatial hotspots discovered.