Unsupervised Approach to Investigate Urban Traffic Crashes Based on Crash Unit, Crash Severity, and Manner of Collision

Abstract

Both crash frequency analysis (CFA) and real-time crash prediction models (RTCPMs) divide a highway into small segments with a constant length [typically 0.161 km (0.10 mi)] for data aggregation. Many previous studies refer to this constant length as the segment length for data aggregation, but this paper adopts fragment size to avoid confusion with aggregation based on highway geometric features. Several studies have shown that segmentation length impacts the studies’ results and recommend not using a length smaller than 0.161 km (0.10 mi) or greater than 0.402 km (0.25 mi) to segment and aggregate traffic data for urban/suburban highways and freeways. Despite the significant impact of the segmentation length on traffic crash aggregation, no specific recommendation for selecting or determining the segmentation length for crash data aggregation exists. This research investigates the impact of segmentation length on traffic crash data aggregation. It establishes a methodology for determining a recommended fragment size (RFS) using hidden heterogeneity in traffic crash data. The study defines featured traffic crash rates using three major traffic crash characteristics: number of vehicles in crash, manner of collision, and crash severity. The analysis uses the Laplacian score with distance-based entropy measure and K-means to cluster highway segments based on the featured crash rates (FCRs) and total crash rates (TCRs) for fragment sizes ranging from 0.161 to 0.402 km (0.10 to 0.25 mi) with an increment of 0.016 km (0.01 mi). The clustering results are compared using their silhouette coefficients. The sample results shows that FCR-based clustering outperforms TCR-based clustering by providing important traffic crash groups within a highway and the RFS to segment and aggregate traffic crash data. The proposed method provides a data-driven comparison of different fragment sizes, revealing the pattern of traffic crashes and a standardized approach for RFS, which reduces the likelihood of fragment misclassification and benefits traffic studies depending on segmentation length.