Behavioral insights and hotspot identification: Integrating natural language processing, machine learning and geospatial analyses of cyclist crashes

This is the authors' abstract. We don't add key points for this paper.

• Applied NLP and ML techniques to crash descriptions to identify risky behavioural and road design factors. • Demonstrated use of crowdsourced descriptions to highlight underreported crash risks and guide safety measures. • Identified collision hotspots linked to risky behaviors using advanced geospatial techniques. In response to the rising trend in the promotion and adoption of cycling, ensuring cyclist safety is paramount. Understanding behavioural causes of crashes and identifying collision hotspots is important; however, the efforts are hindered by underreporting and limited data on all types of incidents, including near misses. Addressing these challenges, this study analyses text data reported on dedicated active travel collision platforms to categorize incidents and uncover behavioural patterns contributing to collisions. The reported text data is grouped into distinct themes applying Term Frequency-Inverse Document Frequency (TF-IDF) vectorization, and clustering. Additionally, the advanced geospatial technique Getis-Ord Gi* statistic is computed to identify spatial clustering of collisions and categorize geographical regions as hotspots and cold spots. Key themes contributing to collisions are grouped as follows: ‘close pass incidents,’ ‘blocked bicycle lanes,’ ‘cyclist incidents on tram tracks,’ ‘roundabout incidents,’ ‘left turn incidents,’ ‘incidents between buses and cyclists,’ ‘incidents involving cyclists and trucks,’ ‘incidents related to traffic lights and pedestrian crossings,’ and ‘turning incidents at intersections.’ Moreover, the hotspots from these incidents are located at or near the intersections of regional roads in the Central Business District (CBD) and on the peripheral regional roads encapsulating the CBD in Dublin, Ireland. This study advances the state of the art by utilizing an alternative data source, ‘crash descriptions’ from cyclist crashes, through the application of innovative machine learning techniques and advanced geospatial analyses. The insights from the unique themes and identified hotspots enhance understanding of risky behaviours and their spatial distribution, contributing to ongoing efforts to foster a safer cycling environment.