Transfer learning-based highway crash risk evaluation considering manifold characteristics of traffic flow. (April 2022)

• Record Type: Journal Article
• Authors: Liu, Qingchao; Li, Chun; Jiang, Haobin; Nie, Shiqi; Chen, Long
• Is Part Of: Accident analysis and prevention. Volume 168(2022)
• Publication Date: 2022-04
• Subjects: Transfer learning -- Temporal and spatial characteristics -- Crash risk evaluation -- Similarity measure -- Manifold distance; Accidents -- Prevention -- Periodicals; Accident Prevention -- Periodicals; Accidents -- Prévention -- Périodiques; 363.106
• DOI: 10.1016/j.aap.2022.106598
• Languages: English
• ISSNs: 0001-4575

Abstract

Highlights: LightGBM was used to analyze the importance of traffic flow at different times and locations. A novel crash risk probability calculation method based on manifold distance was developed. The manifold distance considered more features of traffic flow data when selecting neighbors. TransferBoost was employed to build a crash risk model. Abstract: The main objective of this study is to evaluate highway crash risk and improve the spatial and temporal transferability of crash risk models. The predictive performance is affected by the difficulty of existing models to quantify crash risk from historical traffic flow data at different locations and times and may not fully capture the complex nonlinear relationships between high-dimensional factors in traffic flow states. Oregon US26W freeway data from 2016 and 2017 and I5N freeway data from 2017 were used. Raw detector data collected from two consecutive detector stations upstream-downstream detector stations were converted into 30 traffic variables. The averages, standard deviations, and coefficients of variation were obtained by aggregating traffic values using each lane. Candidate variables for traffic flow were extracted, and the importance of each variable was calculated using LightGBM, which reveals that variable differences between lanes contributed more. The manifold distance was then applied to quantify the crash risk and classify traffic crashes or not. When the manifold distance is 0.4, it could effectively distinguish traffic crashes. TransferBoost was further employed to build a crash risk model. Modeling using 2016 and 2017 data from the US26W freeway revealed a significant decrease in AUC and a gradual decrease in the model's sensitivity. However, the crash risk prediction performance of TransferBoost improved by 5.2% when modeling using 2017 data from US26W and I5N freeways. The results show that the model developed for one time period cannot be directly used to predict crash risk for another period on the same freeway. However, the model developed for one highway cannot be directly used to predict the crash risk of another highway either, maintaining some transferability at low false alarm rates. TransferBoost provides a fresh perspective on the transferability of the model. The findings of this study could facilitate more accurate proactive safety management and improvement countermeasure development.