A learning-to-rank method to identify crash hotspots based on large-scale ride-hailing crash data

Abstract

Machine learning have been widely used in crash hotspot identification due to its superior prediction accuracy. Existing studies mainly treat hotspot identification as a classification or regression problem. This paper proposed a learning-to-rank(LTR) method to identify hotspots on a single trip and deviced a risk warning system based on the method to verify its effectiveness in crash mitigation. Ride-hailing crashes for a year in China were used as training and testing data. Three kinds of features were extracted to describe the safety level of each road segments, namely, road design features, time-related features, and traffic features. LambdaMART, a pairwise LTR algorism was applied to rank the road segments based on the extracted features. The experiment results suggested that the proposed LTR model outperforms three traditional machine learning models in terms of NDCG@10. The proposed LTR risk warning system integrated with Didi's ride-hailing service outperforms traditional zone-based warning system and bring a significant drop in Average Death Rate per Billion Kilometers.