联系我们： 手动添加方式: 微信>添加朋友>企业微信联系人>13262280223 或者 QQ: 1483266981

An interpretable machine learning-based method for predicting the severity of urban road traffic accidents
Research background and topic
Globally, the number of people killed and injured in traffic accidents is rising every year. The incalculable loss of life and property to society, and even the decline in productivity due to the billions of dollars spent each year on medical care, are serious obstacles to the development of human society. It is important to analyse traffic accidents and to provide the relevant authorities with scientific and reasonable recommendations to reduce the severity of traffic accidents.
The study of accident severity prediction has been a hot topic in the field of traffic safety, and scholars have used different research methods to predict the severity of accidents in different regions and accident types, to reveal the mechanisms of accident occurrence, and to continuously improve and develop accident causation theories.
Research purpose
As the dimensionality and volume of data increases, the limitations of traditional models become apparent, and machine learning models based on the intrinsic relationship between the influencing factors of accidents become popular, providing new research ideas to reduce the severity of accidents.
Research objectives
(1) Pre-processing and imbalance processing of data to predict the severity of urban road accidents based on machine learning, and to analyse the advantages of machine learning models compared to traditional logistic regression models.
(2) Visualisation and interpretation of the machine learning prediction results to suggest traffic improvement measures for the relevant authorities to Reduce the severity of accidents
Problem solving
(1) Because machine learning does not require pre-assumptions about the correlation between accident characteristics and accident severity and has greater data processing power, it can solve the problem that traditional statistical models are less effective due to the large sample size and high data dimensionality of traffic accident data.
(2) Most studies on accident severity prediction do not take into account accident data imbalance. This paper will imbalance the accident data, which can improve the performance of the model and also explore the potential value of accident.
Literature review on previous research
Statistical analysis of accident severity
Scholars were the first to analyse the various factors influencing accident severity based on statistical models and to propose measures for improvement based on their findings.
Dawnl, et al. (1995) used statistical analysis to examine the gender distribution of crash severity and the distribution of drivers by age group, showing that older people had higher crash severity than younger people and men had higher crash severity than women. Carson and Mannering (2001) used a statistical model to analyse snow and ice crashes in Washington State and concluded that the severity of crashes could be reduced by putting up appropriate warning signs in snow and ice conditions.
Analysis of accident severity prediction based on traditional models
Scholars have analysed the relationship between accident severity and each influencing factor based on traditional logistic regression models, and have ranked the importance of accident characteristics to reasonably evaluate the influence of accident influencing factors on accident severity.
Kim, et al. (1994) analysed accident severity in Hawaii based on a traditional linear regression model. The predictive analysis showed that accident severity was related to the type of crash, with rollover and frontal crashes causing the greatest severity. Fountas et al. (2018) analysed the influence of different influences on accident severity based on an ordered probability model and ranked the importance of the influences.
Predictive analysis of accident severity based on machine learning and deep learning models
Milton et al. (2008) used the Washington State crash data to compare traditional statistical models with a hybrid Logit model, which concluded that the hybrid logit model predicted highway accident severity better than the traditional statistical model and has great potential for application. Castro et al. (2013) used the RPNB model to analyse the relationship between illumination and crash severity using crash data from Washington State, USA, and concluded that right-handed illumination increases crash severity.
In recent years, more and more scholars have combined simulation with vehicle crashes. Wang et al. (2018) combined VISSIM simulation with vehicle crashes. The results show that the annual crash frequency (EACF) based on the full calibration strategy can be effective in assessing the safety of simulated accidents.
Research method design
The method has been designed in four main areas.
(1) To analyse the method of collecting urban road accident data and the sources of accident data. Dataset was selected from the website https://open.canada.ca/data/en/dataset/1eb9eba7-71d1-
4b30-9fb1-30cbdab7e63a. This research selects Canadian traffic accident data from 2015 to 2017 as the basis for statistical analysis.
(2) The accident data required for the experiment will be pre-processed in terms of completeness and relationships between variables, and two cars were selected for the study. The correlation between the independent and dependent variables of the accident data will be analysed.
(3) The LightGBM model and the Random Forest model will be developed, and Accuracy, AUC and F1 were selected as the evaluation indicators of the model. The prediction results of the LightGBM and Random Forest models will be then compared with the logistic regression model.
(4) The LightGBM model with the best predictive indexes will be visualised and interpreted using the SHAP model to analyse the relationship between each factor and the severity of the accident from the whole to the individual.
Challenges and interesting part
Imbalance data handling, feature selection, model visualisation and putting the model into practical use will be the main challenges in this research process. Among them, model visualization will be the interesting part because it will combine visualisation techniques and machine learning to visually analyse the relationship between accident severity and the factors influencing the accident, providing a more user-friendly view and different dimensions to understand the data itself and the relationships between the data.
Milestone
1. 2022-10-14 Proposal
2. 2022-10-28 Literature review
3. 2022-11-07 Data collection, pre-process and analysis
4. 2022-12-07 Model development and training
5. 2022-12-31 Model analysis and visualization
6. 2023-03-01 Dissertation
References
Carson, J. & Mannering, F. (2001). The effect of ice warning signs on ice-accident frequencies and severities. Accident Analysis and Prevention, 33, 99 – 109.
Castro, M., Paleti, R. and Bhat, C.R. (2013) A spatial generalized ordered response model to examine highway crash injury severity. Accident Analysis and Prevention, 52, 188– 203.
Dawin, L. M., Campbell, K. L., and Williams, A. F. (1995). Traffic accident involvement rates by driver age and gender. Accident Analysis and Prevention, 27(1), 73-87.
Fountasa, G., Anastasopoulos, P. C., Mannering, F. L. (2018). Analysis of vehicle accident-injury severities: A comparison of segment- versus accident-based latent class ordered probit models with class-probability functions. Analytic Methods in Accident Research, 18, 15–32.
Miltona, J. C., Shankarb, V. N. and Mannering, F. L. (2008). Highway accident severities and the mixed Logit model: An exploratory empirical analysis. Accident Analysis and Prevention, 40, 260–266.
Kim, K., Nitz, L., Richardson, J., and Li, L. (1994). Analyzing the Relationship Between Crash Types and Injuries in Motor Vehicle Collisions in Hawaii. Transportation Research Record.
Wang. C., Xu, C. C., Xia, J. X. (2018). A combined use of microscopic traffic simulation and extremevalue methods for traffic safety evaluation. Transportation Research Part C, 90, 281–291.