Data Mining and Complex Networks Algorithms for Traffic Accident Analysis

摘要

The field of traffic accident analysis has long been dominated by traditional statistical analysis. With the recent advances in data collection, storage and archival methods, the size of accident datasets has grown significantly. This in turn has motivated research on applying data mining and complex network analysis algorithms, which are specifically designed to handle datasets with large dimensions, to traffic accident analysis. This paper explores the potential for using two such methods, namely a modularity-optimizing community detection algorithm and association rules learning algorithm, to identify important accident characteristics. As a case study, the algorithms are applied to an accident dataset compiled for Interstate 190 in the Buffalo-Niagara metropolitan area. Specifically, the community detection algorithm is used first to cluster the data in order to reduce the inherent heterogeneity, and then the association rule learning algorithm is applied to each cluster to discern meaningful patterns within each,particularly related to high accident frequency locations (hotspots) and incident clearance time. To demonstrate the benefits of clustering, the association rule algorithm is also applied to the whole dataset (before clustering) and the results are compared to those discovered from the clusters. The study results indicate that: (1) the community detection algorithm was quite effective in identifying clusters with discernible characteristics; (2) clustering helped in unveiling relationships and accident causative factors that remained hidden when the analysis was performed on the whole dataset; and (3) the association rule learning algorithm yielded useful insight into accident hotspots and incident clearance time along I-190.