zhanginnovativepredictioncausalanalysis2025.md 84 KB
category: literaturenote citekey: zhanginnovativepredictioncausalanalysis2025 title: Innovative prediction and causal analysis of accident vehicle towing probability using advanced gradient boosting techniques on extensive road traffic scene data authors: "Zhang, Ronghui; Liu, Yang; Wang, Zihan; Chen, Junzhou; Zeng, Qiang; Zheng, Lai; Zhang, Hui; Pei, Yulong" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.aap.2024.107909 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2024.107909" zotero_key: HEBABH9I zotero_storage: BZ9DJBVK collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Zhang, Ronghui"
status: converted
Contents lists available at ScienceDirect
Accident Analysis and Prevention
journal homepage: www.elsevier.com/locate/aap
Innovative prediction and causal analysis of accident vehicle towing probability using advanced gradient boosting techniques on extensive road traffic scene data
Ronghui Zhang a , Yang Liu a , Zihan Wang a , Junzhou Chen a ,∗ , Qiang Zeng b , Lai Zheng c , Hui Zhang d,e , Yulong Pei f
- a Guangdong Key Laboratory of Intelligent Transportation System, School of Intelligent Systems Engineering, Sun Yat-sen University, Guangzhou, 510275, Guangdong, China
- b School of Civil Engineering and Transportation, South China University of technology, Guangzhou, 510640, Guangdong, China
- c School of Transportation Science and Engineering, Harbin Institute of Technology, Harbin, 150040, Heilongjiang, China
- Intelligent Transportation Systems Research Center, Wuhan University of Technology, Wuhan, 430063, Hubei, China
- e Engineering Research Center of Transportation Information and Safety, Ministry of Education, Wuhan, 430063, Hubei, China
- f School of Civil Engineering and Transportation, Northeast Forestry University, Harbin, 150040, Heilongjiang, China
A R T I C L E I N F O
Keywords: Traffic safety Road crash vehicle towing Prior scene feature Causal factor
A B S T R A C T
Accurate prediction and causal analysis of road crashes are crucial for improving road safety. One critical indicator of road crash severity is whether the involved vehicles require towing. Despite its importance, limited research has utilized this factor for predicting vehicle towing probability and analyzing its causal factors. This study addresses this gap by predicting the probability of vehicle towing in road crashes based on road scene features and identifying key causal factors. Utilizing the Transportation Injury Mapping System (TIMS) dataset from California, USA, encompassing 12 years, 14 relevant features, and over 2 million road crash records, research team developed a prediction model using advanced gradient boosting techniques. Our model outperforms Random Forest, GBDT, and XGBoost in predictive accuracy. Employing the Shapley Additive Explanation (SHAP) method, researchers elucidate seven key factors influencing towing necessity. These findings introduce a novel predictive approach and offer valuable insights for road crash risk assessment and road safety planning.
1. Introduction
Road crashes are the main research object of road traffic safety. According to the Global Status Report on Road Safety released by the World Health Organization in 2023, road crashes result in approximately 1.19 million deaths in 2021. Road traffic injury remains the leading cause of death for people aged 5 – 29 (World Health Organization, 2023). Many victims lose their ability to work, causing huge losses to themselves, their families, and society. Millions of people fall into poverty as a result.
While road crashes have caused such great damage, our methods of identifying the causes of road crashes are still imperfect. For specific road crashes that have already occurred, on-site investigations and the determination of responsibility can identify the specific causes of the road crashes. However, it is difficult to obtain detailed information on the key causes of specific road crash types by relying on traditional statistical analyses.
As shown in Fig. 1, compared to traditional statistical methods, machine learning techniques can identify key causal factors from largescale road crash data and develop predictive models. With the development of intelligent transportation systems and the increasing availability of traffic data, the application of machine learning techniques in this field has become more suitable. Consequently, researchers can construct models based on road crash records collected from various departmental information databases, combined with existing theoretical tools, to predict road crashes and their severity, and to explore important factors affecting road crashes (Chiou and Fu, 2015; Pang et al., 2022). This approach involves collecting features of people, vehicles, roads, and the environment at the time of the road crashes (Fountas et al., 2020; Zeng et al., 2022). It builds a prediction model for road crash risk or frequency to analyze the correlation between these features and road crashes (Ihueze and Onwurah, 2018; Ju and Sohn,
E-mail address: chenjunzhou@mail.sysu.edu.cn (J. Chen).
https://doi.org/10.1016/j.aap.2024.107909
∗ Corresponding author.
Fig. 1. Comparison between traditional path and machine learning path for analyzing the causes of road crashes.
2011).
However, existing machine learning-based methods for analyzing road crash causes still face some issues. Traditional machine learning methods typically predict the severity of road crashes in non-specific scenarios over large areas (Ma et al., 2021). While these studies can derive general conclusions about factors affecting road crash severity, they often lack practical applicability due to the absence of targeted scenarios. Given the varying severity and main types of road crashes under different traffic flow conditions (Zhao et al., 2021), some current research requires traffic flow state data to assess traffic risks (Berghaus et al., 2024). These results may not be applicable to roads with different traffic conditions (Hussain et al., 2024). Other similar studies generally focus on road crashes in specific small-scale scenarios, such as vehiclepedestrian collisions at intersections in a particular city over a few years (Zeng et al., 2023b; Olowosegun et al., 2022; Gurumurthy et al., 2022).
To obtain generalizable results after analyzing road crash causes, this paper collected a substantial amount of road crash data from open-source data websites and selected priority features to construct a large dataset (Mannering et al., 2020). In addition, research team chose ''whether the road crash vehicle needs to be towed'' as a specific application scenario and used it as the label for modeling the data.
Choosing this application scenario as a label has the following advantages: Firstly, the label can clearly define whether a driver in a traffic network is incapacitated due to injuries sustained in a road crash and whether a vehicle is incapacitated due to a collision. Secondly, risk prediction and factor analysis of road crashes with this label can help predict vehicle incapacitation risk points in a traffic network, thus helping traffic management to optimize trailer deployment strategies.
This study makes the following significant contributions to the field of traffic safety research:
- (1) Comprehensive Dataset: Our team compiled a large-scale dataset of over 2 million road crash records from the Transportation Injury Mapping System (TIMS), covering California from 2011 to 2022. This extensive dataset includes diverse road types and critical network nodes, facilitating the discovery of intricate patterns in road crashes.
(2) Advanced Predictive Modeling: Research team developed a robust risk prediction model using advanced gradient boosting techniques, specifically designed to predict the likelihood of vehicle towing in road crashes. Our model outperforms Random Forest, GBDT, and XGBoost models in accuracy and reliability.
(3) Causal Factor Analysis: Utilizing the Shapley Additive Explanation (SHAP) method, researchers identified and analyzed seven key causal factors influencing vehicle towing necessity in road crashes. This analysis provides deeper insights into the underlying causes of severe road crashes, guiding targeted interventions.
(4) Practical Applications: The study's findings have significant practical implications for traffic management and safety planning. Research team created a high-risk map for vehicle towing in road crashes, which can inform the strategic deployment of towing services. This can lead to quicker response times, reduced traffic congestion, and improved road safety.
In order to demonstrate the scope and specific scenarios of our research, as shown in Fig. 2, researchers mark road crashes with towing labels on the map, zoom in and display specific road crash locations.
2. Literature review
The article provides a summary table of the latest similar research, as shown in Table 1.
In terms of dataset construction, above studies have explored the influencing factors of road crashes in different application scenarios using different methods based on the collected dataset. Most of current research datasets cover a small range and have a small amount of data. The trained models and analyzed factors can only be used within a certain range of blocks. The results obtained do not have universality over a large spatiotemporal range. To make the research results more convincing and applicable to more regions, this article collects 12 years of road crash injury data across California.
In terms of labeling method, current road crash causation studies typically use both objective records and subjective classifications to evaluate road crash severity. However, there is a serious imbalance in the road crash severity labeling data derived from both evaluation methods due to the excessive number of similar values in the evaluation results. This serious imbalance of data can affect the fitting performance of the model and even make the fitting results overfitted. In this paper, since the towing label is balanced in characterizing the severity of road crashes, this newly proposed label helps to improve the model fitting accuracy.
In terms of model selection, The commonly used machine learning model methods in current similar research include: methods based on Logistic Regression (LR), methods based on Neural Networks (NN) and methods based on Decision Tree (DT). The data collected in this article is presented in tabular form and contains a large number of discrete semantic features. Due to the large amount of tabular discrete nominal data in the dataset of this article, the research team needs methods that can adapt to this type of data format. In order to analyze the relationship between the results and corresponding features, this method also needs to have a relatively complete interpretability method. LR based methods are difficult to fit numerous discrete non numeric features. Methods based on NN are difficult to use for analyzing causes corresponding to the results. Compared with LR based methods and NN based methods, DT based methods have shown excellent performance in handling discrete nominal data in tabular form and have mature interpretability methods. Based on the above considerations and literature review, team chose two classic DT based methods, Random Forest and GBDT, as well as two novel DT based methods, XGBoost and HGBoost, for the experiment.
3. Methodological approach
This article have organized the experiment process into Fig. 3. Firstly, research team choose California as our research area and collect data from TIMS. Secondly, team select features by expert analysis and clean the data. Then, researchers use 4 machine learning models to fit the dataset and measure the fitting effect by metrics. Finally, article use SHAP method to explain the best matched model and analyze the causal factors.
Fig. 2. (a) Road crashes in California from 2011 to 2022 in which vehicles need to be towed (Lehto, 2017). (b) Road crashes in California from 2011 to 2022 in which vehicles do not need to be towed (BİLİM-TEKNOLOJİ, 2014).
3.1. Step one: Determine range and data collection
In order to obtain more applicable research results, a dataset with a larger scope and a longer time span is necessity. Ultimately, research team chose California, USA, a state with rapid ITS development and complex geographic road network conditions, as the study area. Different from some developing regions, California's transport network and its operation have been relatively stable over the past decade. Even with the long time span of the dataset, the impact brought by changes in the structure of its road network on the consistency of the data remains relatively small. As a result, we collected 12 years of road crash injury records from 2011 to 2022, totaling more than 2 million pieces of data from the California TIMS (Transportation Injury Mapping System, Safe Transportation Research and Education Center, University of California, Berkeley, 2023). All road crashes with coordinates are plotted on a map and shown in Fig. 4.
3.2. Step two: Feature selection and data cleaning
The raw data collected by TIMS comes from Statewide Integrated Traffic Records System (SWITRS), which stores data collected by the California Highway Patrol (CHP) and includes all road crashes reported by local and government agencies to the California Highway Patrol. The data volume is large and the recorded sample features are diverse,
Table 1
| Author | Research object | Method | Results of factors analysis |
|---|---|---|---|
| El-Basyouny et al. (2014) | Traffic collision under extreme rain and snow weather changes |
Multivariate Poisson log normal (MVPLN) model |
Discovering a high correlation between ''sudden changes from dry to rain or snow'' and Follow-Too-Close, Stop-Sign-Violation, Ran-Off-Road crashes. |
| Ihueze and Onwurah (2018) | Analyze road crashes based on time series by combining human, vehicle, road, and environment |
Autoregressive integrated moving average with explanatory variables (ARIMAX) |
Discovering factors such as weather condition, brake failure, loss of control is more likely to lead to an increase in the number of car road crashes. |
| Lee et al. (2018) | The impact of rainfall intensity and water depth level on road crashes |
Structural equation modeling (SEM) | A downward slope, right-hand curve, short curve length, compact cars, nighttime, younger drivers and female drivers have a significant effect on the level of road crash severity |
| Wen et al. (2019) | The correlation between collision frequency and possible risk factors |
A Bayesian spatial–temporal model | The interaction between weather and road environmental factors, such as wind speed and slope, precipitation and curve, visibility and slope, is significantly related to the increased risk of highway collisions. |
| Schlögl (2020) | Environmental effects on road crash occurrence |
A method for obtaining stable and unbiased results when the data is imbalanced and has high resolution |
Road geometry, surface conditions, traffic volume, and many weather variables are of great significance for |
| (Random Forest, XGBoost) | predicting road crashes. | ||
| Jha et al. (2021) | Predict hit and run road crashes | Six supervised learning models | The support vector machine model performs best in predicting hit and run road crashes in their dataset. |
| Li et al. (2021) | Prediction of intersection traffic violations and factors analysis. |
Logistic, ProWSyn, Random Forest | The midday period, weekends, residential districts, collector roads, congested traffic conditions, high traffic flow, lower wind speed and low temperature would increase traffic violations probability. |
| Ma et al. (2021) | Prediction of road crash injury severity based on contributing factors |
CatBoost, SHAP, SSAE | When the speed limit is low, the more vehicles involved in the road crash, the lower the degree of damage, and vice |
| Wen et al. (2021) | The effects of key risk factors on various types of roadway segment crashes |
LightGBM, SHAP | versa when the speed limit is high. Etc. Different types of roadway segment crashes have different sets of main contributing factors except for AADT, segment length and speed limit. |
| Yan et al. (2021) | Single-vehicle crash severity prediction and determinant extraction |
Tree-based model (RF, AdaBoost, etc.), non-tree-based model (SVM, KNN, etc.) |
Road crashes that occur on highways or beyond shoulders will be more severe. |
| Zhao et al. (2021) | The impact of traffic states on crash risk in the vicinities of ''Type A weaving segment'' |
Deep convolutional embedded clustering (DCEC), logistic regression, convolutional neural network with gated convolutional layers (G-CNN) |
The severity and main types of road crashes vary under different traffic flow states on ''Type A weaving segment''. |
| Zhu and Wan (2021) | Hit-and-run crashes | Cost-sensitive semi-supervised logistic regression (CS3 LR) |
The more pedestrians, cyclists, and passengers involved in a road crash, the higher the probability of driver will hit and run. |
| Liu et al. (2022) | Highway crash risk and improve universality of crash risk models |
TransferBoost, LightGBM | Differences in traffic flow parameters between different lanes contribute more to road crashes. |
| Olowosegun et al. (2022) | Pedestrian accident injury severity at different types of road junctions. |
Ordered probit model | The severity of road crash injury is influenced by road, location, weather, vehicle and driver characteristics, as well as time attributes (including the time and date of the road crash). |
| Yang et al. (2022) | Multiple types of road crash severity | Multi-task DNN model, layer-wise relevance propagation (LRP) |
Road protection facility, road alignment, intersection and central isolation facility have significant impact on all type of |
| Huang et al. (2023) | Freeway primary and secondary crash severity |
Copula-based joint model | road crash severity. Weather, whether towed away, unsafe speed,have significant impact on primary crash severity propensity and collision type, road width, road condition,have significant impact on |
| Kashifi (2023) | Two-wheelers risk factors for severe crashes |
XGBoost, SHAP | secondary crash severity propensity. Severe injuries are more likely to occur in two-wheeler crashes associated with rural areas, older riders, riders not wearing helmets, run-off-road crashes, and crossing roads. |
(continued on next page)
Table 1 (continued).
| Author | Research object | Method | Results of factors analysis |
|---|---|---|---|
| Nikolaou et al. (2023) | Comparing machine learning techniques for predictions of motorway segment crash risk level |
Logistic regression, Decision Tree, Random Forest, SVM and K-NN |
The developed RF model could serve as a quite auspicious proactive road safety tool that could be used for the identification and prioritization of potentially hazardous motorway segments. |
| Zeng et al. (2023a) | The injury severity of motor vehicle–pedestrian crashes at urban intersections |
Spatiotemporal logistic regression | Pedestrian age, head injury, pedestrian location, pedestrian actions, driver maneuvers, vehicle type, first point of collision, and traffic congestion status significantly affected the severity of pedestrian injuries. |
| Xue et al. (2024) | The leading visual risk factors of cycling crash rates at intersections in Manhattan |
Panoptic segmentation model, XGBoost, LightGBM, SHAP |
Road barrier proportion, open sky proportion, and visible truck number are leading vision factors contributing to high-risk cycling intersections. |
| Ye et al. (2024) | The association between the cycling safety level and the road environment |
Cycling safety level for road environment (CSL-RE), Proportion of cycle accidents (PCA), Severity of cycle accidents (SCA). SMOTE, XGBoost, SHAP |
The second road class, light condition, and speed limit and greenery coverage among the imagery feature of street view had the significant influence on the safety level of riding. |
Fig. 3. Experiment process flowchart of road crash factors analysis by data driven.
Table 2 Partial features instruction.
| Feature name | Value | Description | | | |----------------|-----------|-----------------------------------------|--|--| | CASE_ID | 5 061 493 | Road Crash ID: 5061493 | | | | COLLISION_TIME | 2215 | Occurrence time: 22 h15 min | | | | COLLISION_DATE | 2011/2/5 | Occurrence date: 2011/2/5 | | | | WEATHER_1 | C | Main weather: C: Raining | | | | ROAD_COND_1 | H | Road condition: H: No Unusual Condition | | | | | | | | |
making it suitable for research purposes. Partial features instruction of road crash record data on TIMS are shown in Table 2.
Since the data collected contained missing values and invalid features, it was necessary to clean the data and select the necessary features. First, based on the feature description document on TIMS website, research team selected ''pre-scene features that can be obtained before the road crash'' by expert voting. In addition, we constructed temporal features based on the temporal description of the road crash and used ''whether the road crash vehicle was towed or not'' as the fitting label, and constructed a dataset containing 14 features as shown in Table 4. Finally, in the data cleaning step, researchers removed unlabeled samples and replaced null values with zero.
Table 3
| Feature name | Data type | Number of categories | Description |
|---|---|---|---|
| SEASON | Nominal | 4 | Mar-May: Spring |
| MONTH | Numerical | 12 | Month in a year |
| 24HOURS | Numerical | 24 | Hours on one day |
| DAY_OF_WEEK | Numerical | 7 | Week number |
After understanding the meaning of the data feature representation based on the explanatory documents in the TIMS website, team first constructed the temporal features of the data based on the temporal description of the road crash. In this process, team obtain the month and seasonal features of the road crash based on the specific date of the road crash. In addition, team construct the hourly feature of the road crash based on the specific time of the road crash. The constructed temporal features are shown in Table 3.
The prior scene features are inherent scene features that can be obtained before an accident occurs. The input of the prediction model requires prior scene features. The raw data collected by the team from TIMS contains a large number of posterior conclusion features, which greatly affect the model's fitting of inherent scene features and
Fig. 4. Road crashes with injuries that occurred in California from 2011 to 2022.
corresponding results. The team needs to vote for prior scene features from the raw data to ensure that the model input is valid. The specific voting process is as follows:
Firstly, team members confirm the specific meaning of each feature based on the feature introduction. Secondly, all members independently determine whether each feature belongs to the prior scene features that comply with this study. Thirdly, each member casts one vote for all features they deem appropriate. Finally, the team selects features that receive votes more than half team members as input features for the model. For example, the team has 8 people participating in the voting. If the lighting feature receives 7 votes, it is selected as the input feature of the model. If the lighting feature only receives 2 votes, it will not be selected.
The word cloud map of expert voting is shown in Fig. 5. A total of 14 features are selected to form dataset with ''TOW_AWAY'' feature as the classification label. The feature types and data types are shown in Table 4.
3.3. Step three: Data modeling
In this paper, Random Forest model, GBDT model, XGBoost model and HGBoost model are introduced to fit the data. Random forest is an ensemble learning method based on DT. This model improves the accuracy of predictions by voting or weighted averaging the predictions of all trees. The Random Forest model can predict the probability of road crashes more accurately by analyzing a variety of traffic related features and integrating the results of multiple decision trees. As a commonly used model based on the DT method, Random Forest can serve as the baseline for the model in this study. GBDT is an ensemble method based on Boosting and CART, which uses the Forward
Table 4 Dataset features description.
| Feature type | Feature name | Data type | Categories | First and second majority values | Number | Proportion |
|---|---|---|---|---|---|---|
| Time | MONTH | Numerical | 12 | 10: October 9: September |
194 223 182 745 |
9.33% 8.78% |
| Time | SEASON | Numerical | 4 | 3: Autumn 2: Summer |
554 846 522 855 |
26.64% 25.11% |
| Time | 24HOURS | Numerical | 24 | 17: 5p.m. 15: 3p.m. |
169 425 156 134 |
8.14% 7.50% |
| Time | DAY_OF_WEEK | Numerical | 7 | 5: Friday 4: Thursday |
338 773 308 905 |
16.27% 14.83% |
| Road | INTERSECTION | Nominal | 2 | N: Not at intersection Y: At intersection |
1 438 557 633 871 |
69.08% 30.44% |
| Road | STATE_HWY_IND | Nominal | 2 | N: Not at state highway Y: At state highway |
1 376 802 704 251 |
66.12% 33.82% |
| Road | ROAD_SURFACE | Nominal | 5 | A: Dry B: Wet |
1 924 253 136 509 |
92.40% 6.56% |
| Road | ROAD_COND_1 | Nominal | 9 | H: No unusual condition D: Construction or repair zone |
2 008 970 33 013 |
96.47% 1.59% |
| Road | DIRECTION | Nominal | 5 | S: South N: North |
385 524 378 933 |
18.51% 18.20% |
| Environment | WEATHER_1 | Nominal | 8 | A: Clear B: Cloudy |
1 787 885 219 042 |
85.86% 10.52% |
| Environment | LIGHTING | Nominal | 6 | A: Daylight C: Dark-Street Lights |
1 392 698 219 042 |
66.88% 21.26% |
| Vehicle | STWD_VEHTYPE_AT_FAULT | Nominal | 16 | A: Passenger Car/Station Wagon - or blank: Not Stated |
1 341 999 274 637 |
64.44% 13.19% |
| Vehicle | CHP_VEHTYPE_AT_FAULT | Nominal | 85 | 01: Passenger Car, Station Wagon, or Jeep 07: Sport Utility Vehicle |
1 011 191 246 467 |
48.56% 11.84% |
| Label | TOW_AWAY | Nominal | 2 | Y: Yes N: No |
1 287 752 775 455 |
61.84% 37.24% |
Fig. 5. Expert voting word cloud map for feature selection.
Distribution Algorithm for Greedy Learning during the training process (Friedman, 2001). GBDT is an excellent model for classification of tabular discrete data, which is very suitable for the processing of road crash data in this paper. GBDT is the foundation of two novel methods and also serves as a benchmark for comparison. XGBoost is an improved engineering implementation of GBDT (Chen and Guestrin, 2016). Compared with GBDT, the XGBoost method greatly improves the accuracy and efficiency by introducing second-order derivatives, regularization terms and parallel operations. On the basis of having the same excellent adaptability to tabular discrete data as GBDT, XGBoost can process the large-scale data studied in this paper faster and more accurately.
Inspired by LightGBM (Ke et al., 2017), the HGBoost model is an improved model based on Gradient Boosting (GB). In addition to having the same advantages as the above methods that are applicable to the dataset in this article, HGBoost also has the following advantages. As
Fig. 6. The brief application diagram of histogram-based algorithm in HGBoost method.
shown in Fig. 6, instead of finding the split points on the sorted feature values, histogram-based algorithm buckets continuous feature values into discrete bins and uses these bins to construct feature histograms during training. The histogram structure also accelerates the calculation of node splitting. When node A is splitting, it only needs to calculate child node B. Child node C can be obtained by subtracting the histograms of nodes A and B.
Due to the use of histogram-based algorithm in the HGBoost model, the number of splitting points and calculate consumption is greatly reduced. The feature values are no need to sort at each splitting node. When processing large-scale dataset, it can be several orders of magnitude faster than the GB model. In addition, using the histogrambased algorithm can have a regularization effect and effectively prevent overfitting of the model.
3.4. Step four: Causal factors analysis
In road crash prediction, numerous features (such as weather conditions, road type, driver behavior, etc.) collectively affect the occurrence and severity of road crashes. To understand the specific contribution of each feature to the prediction results, SHAP method was used to explain HGBoost model and analyze the key contributing factors of road crashes that affect ''whether the road crash vehicle needs to be towed'' (Ma et al., 2021). SHAP can quantify the impact of each sample on model's prediction results. The summary graph of SHAP can reflect the size of feature values and effectively demonstrate the key contributing factors that impact road crashes.
SHAP mainly comes from the Shapley value, which was a method pioneered by Lloyd S. Shapley in 1953 in cooperative game theory to allocate expenses to game participants based on their specific contributions to total expenses. In this article, each feature (such as weather, road condition or vehicle type) can be considered a ''game participants'' that contributes to the ''total expense'' of road crash prediction. By calculating the Shapley values of each feature, researchers can quantify its impact on predicting the occurrence or severity of road crashes. SHAP approximates the interpretation of complex models through simple explanatory models. It characterizes Shapley values as an additive feature attribution method (Lundberg and Lee, 2017):
$$g(z') = \phi0 + \sum{j=1}^{M} \phi_j z'_j$$ (1)
Eq. (1) shows how SHAP approximates the predictions of complex models by the contribution of each feature. is the explanatory model; ′ ∈ {0, 1} represents whether the features can be observed (represented by 0 or 1); The observations here mainly focus on hidden and unstructured data. Every feature in structured data can be observed, and missing values are considered observable data because you observe that it is missing; is the total number of input features; ∈ R is the Shapley value of each feature; 0 is a constant, which is actually the predicted mean of all training samples.
The Shapley value is calculated according to equation (2) (Lundberg et al., 2018):
$$\phij = \sum{S \subseteq {x_1, \dots, x_p \mid \chi(x_j)}} \frac{|S|!(p - |S| - 1)!}{p!} \left( f_x(S \cup {x_j}) - f_x(S) \right) \tag{2}$$
Eq. (2) provides us with a detailed method for calculating the Shapley value of each feature. This can help us understand which factors are most likely to cause road crashes in a given traffic scenario. {1 , … , } is the set of all input features; is the total number of features; {1 *,*… , } ⧵ { } represents the set of all input features that do not include the feature subset ; ( ∪ { }) is a prediction of the union of feature subset and feature subset ; () is the prediction of feature subset ; Symbol '' ! '' representing factorial; ||!(−||−1)! ! is a weight value that represents the proportion of the combination of feature subset to all feature combinations.
4. Modeling and factors analysis
This experiment was run on Python 3.8.8. The main packages used are scikit-learn1.3.2 and shap 0.41.0.
Table 5 Model validation results.
| Model | Label | Accuracy | Precision | Recall | F1 score |
|---|---|---|---|---|---|
| Y | 0.69 | 0.84 | 0.76 | ||
| RandomForest | N | 0.664 | 0.57 | 0.37 | 0.45 |
| Y | 0.69 | 0.96 | 0.80 | ||
| GBDT | N | 0.697 | 0.78 | 0.26 | 0.39 |
| Y | 0.69 | 0.95 | 0.80 | ||
| XGBoost | N | 0.698 | 0.77 | 0.26 | 0.39 |
| Y | 0.69 | 0.96 | 0.80 | ||
| HGBoost | N | 0.698 | 0.78 | 0.26 | 0.39 |
Table 6 Model testing results.
| Model | Label | Accuracy | Precision | Recall | F1 score |
|---|---|---|---|---|---|
| RandomForest | Y | 0.611 | 0.64 | 0.82 | 0.72 |
| N Y |
0.52 0.63 |
0.29 0.94 |
0.37 0.75 |
||
| GBDT | N | 0.627 | 0.63 | 0.16 | 0.26 |
| XGBoost | Y | 0.632 | 0.64 | 0.88 | 0.74 |
| N Y |
0.59 0.64 |
0.26 0.90 |
0.36 0.75 |
||
| HGBoost | N | 0.637 | 0.62 | 0.24 | 0.35 |
4.1. Metrics
The commonly used metrics for machine learning models include accuracy (), precision ( ), recall (), and 1 score. These metrics are calculated by the following formula:
$$A = \frac{1}{ns} \sum{i=1}^{n_s} \mathbb{1}(\hat{y}_i = y)$$ (3)
is the total number of samples; 1(̂ = ))indicates that when the predicted value is the same as the label, output 1; otherwise, output 0.
$$P = \frac{TP}{TP + FP} \tag{4}$$
$$P = \frac{TP}{TP + FN} \tag{5}$$
$$F_1 = \frac{2 \times P \times R}{P + R} \tag{6}$$
The specific meanings of the four metrics are as follows: Accuracy represents the proportion of samples that the model predicts correctly for the entire dataset; Precision represents the accuracy of the model's prediction of a certain classification; Recall can indicate whether the model's predictions for all real samples in a certain category are complete; 1 score is the harmonic average of the first two metrics, representing the comprehensive performance of the two indicators.
4.2. Modeling
To ensure validity of testing set. The training set uses the data from 2011 to 2020. The testing set uses the data from 2021 to 2022. The training and validation sets are randomly divided using 7:3 ratio. Team normalize digital features and perform one-hot encoding on category features. After preprocessing, 1 221 641 training samples and 523 561 validation samples obtained. Research team construct, train and validate the Random Forest model, GBDT model, XGBoost model and HGBoost model. Then, we use testing set test the predictive effect of vehicle towing risk in road crashes. The validation results of each model are shown in Table 5 and the testing results are shown in Table 6.
In road crash prediction research, the goal of prediction models is to predict high-risk road crashes as thoroughly as possible. This
Table 7 Testing results of data from different periods on our method.
| Period | Label | Accuracy | Precision | Recall | F1 score |
|---|---|---|---|---|---|
| Normal period | Y | 0.68 | 0.96 | 0.80 | |
| N | 0.696 | 0.78 | 0.26 | 0.39 | |
| Y | 0.67 | 0.95 | 0.79 | ||
| COVID-19 period | N | 0.681 | 0.76 | 0.24 | 0.37 |
| Y | 0.68 | 0.95 | 0.79 | ||
| Entire dataset | N | 0.693 | 0.77 | 0.26 | 0.39 |
study aims to explore the key contributing factors that cause ''road crash vehicles need to be towed'', so it is necessary for the model to fit more data related to ''road crash vehicles need to be towed''. The recall rate of label Y can well reflect this situation. Based on the above two considerations, the recall rate of label Y is the primary evaluation indicator. According to Tables 5 and 6, the recall rate of GBDT model is slightly higher than that of HGBoost model. However, the GBDT model predicts too many instances of label as label Y, resulting in a much lower recall rate for label than the HGBoost model. Compared with XGBoost model, HGBoost model has higher Y label recall, accuracy and label precision. There is no significant difference in the label Y precision of each model. Through comprehensive comparison, HGBoost has the best fitting effect on the dataset of this study.
The dataset of this study includes the period of COVID-19 epidemic. To investigate the impact of this period on experiment results, the research team divided the dataset into non epidemic period (2011– 2019) and epidemic period (2020–2022). Train and test HGBoost model separately for these two periods, and compare them with experiments on the entire dataset. The specific experiment process is the same as above. The experiment results are shown in Table 7.
According to the experiment results, the difference in test results among the three periods is very small, indicating that our method has strong robustness whether in epidemic period or not. Compared to non epidemic periods, the accuracy of test results during epidemic periods has decreased. The team infers that the problem may stem from the number of training samples, and the data volume during non epidemic periods is more sufficient, resulting in better training and testing results. In response to this situation, researchers compared the annual total number of road crashes between epidemic and non epidemic periods in the dataset and found that road crashes data during the COVID-19 period only showed little change. After policy research, the team found that COVID-19 has little impact on the traffic flow in the study area. During the pandemic, the traffic flow in California did not decrease significantly. Based on the above results, team is confident that the impact of COVID-19 on this study is limited.
In order to further explore the impact of different road network structures on such road crashes, team classified the dataset based on highway labels and intersection labels. We use the same method to segment the training set, validation set and testing set as before. Due to the best fitting performance among the four models mentioned above, team continue to use our method to fit datasets under different road structures. The testing results for different road structure classifications are shown in Table 8.
Based on our dataset, team find that the model has better test performance after fitting each kind of dataset classified by road structure, especially for highway intersection scenarios. The results of highway intersection data not only have a very high positive label recall rate, but also have a significantly higher accuracy than other road structure classifications. This prominent experimental result shows that there is a certain correlation between the road crashes that vehicle needs to be towed and road network structure.
In order to further explore the impact of different weather conditions on this study, the team divided the dataset according to different weather conditions. We used the same method as before to train and test on datasets with different weather conditions. The testing results
Table 8 Testing results based on different road structure data.
| Label | Accuracy | Precision | Recall | F1 score |
|---|---|---|---|---|
| Y | 0.67 | 0.99 | 0.80 | |
| N | 0.74 | 0.07 | 0.13 | |
| Y | 0.63 | 0.95 | 0.76 | |
| N | 0.80 | 0.27 | 0.40 | |
| Y | 0.65 | 0.97 | 0.78 | |
| N | 0.84 | 0.20 | 0.32 | |
| Y | 0.65 | 0.96 | 0.77 | |
| N | 0.77 | 0.21 | 0.33 | |
| Y | 0.73 | 0.98 | 0.84 | |
| N | 0.78 | 0.16 | 0.26 | |
| Y | 0.66 | 0.99 | 0.79 | |
| N | 0.73 | 0.06 | 0.12 | |
| Y | 0.64 | 0.97 | 0.77 | |
| N | 0.84 | 0.20 | 0.33 | |
| Y | 0.63 | 0.93 | 0.75 | |
| N | 0.78 | 0.31 | 0.45 | |
| 0.671 0.657 0.670 0.660 0.736 0.664 0.662 0.654 |
Table 9 Testing results of data from different weather on our method.
| Main weather | Label | Accuracy | Precision | Recall | F1 score |
|---|---|---|---|---|---|
| Y | 0.64 | 0.96 | 0.77 | ||
| Clear | N | 0.657 | 0.79 | 0.21 | 0.33 |
| Y | 0.69 | 0.97 | 0.81 | ||
| Cloudy | N | 0.702 | 0.79 | 0.18 | 0.30 |
| Y | 0.69 | 0.97 | 0.81 | ||
| Raining | N | 0.692 | 0.74 | 0.16 | 0.26 |
| Y | 0.79 | 0.96 | 0.87 | ||
| Snowing | N | 0.773 | 0.46 | 0.11 | 0.17 |
| Y | 0.78 | 0.96 | 0.86 | ||
| Fog | N | 0.760 | 0.54 | 0.15 | 0.23 |
| Y | 0.70 | 0.83 | 0.76 | ||
| Wind | N | 0.654 | 0.48 | 0.30 | 0.37 |
for different road structure classifications are shown in Table 9.
According to Table 9, our method has better testing performance on subsets of data divided according to different weather conditions in our dataset. Especially in snowy and foggy weather, the accuracy and positive sample precision of the test results are much higher than in other weather conditions. This significant experiment result indicates that there is a certain correlation between the need for road crash vehicles to be towed and snowy or foggy weather in the dataset of this study.
4.3. Significant causal factors analysis
SHAP method is used to analyze the key causal factors of road crashes. 50000 training set samples are taken as feedback samples for the SHAP method to draw a SHAP summary diagram. The SHAP summary diagram shown in Fig. 7. The rows in the SHAP summary graph represent a certain feature, each point in the graph represents a feature of a sample. The color of the points represents the size of the sample feature values. And the abscissa of the points indicates the influence of the sample features on the model output results. The impact is divided into positive and negative effects, with positive effects leading to higher model output and negative effects leading to lower model output. The sum of the absolute values of positive and negative influences is divided by the number of points to obtain the comprehensive influence of a feature. The features are sorted from top to bottom according to the comprehensive influence.
For the convenience of reading the image, team provide a detailed explanation of each feature in Table 10. The features in the table are divided into four categories: scene, vehicle, people and interference
Fig. 7. SHAP summary diagram.
Fig. 8. A causal diagram hypothesis sample: ''STWD_VEHTYPE_ATFAULT_L'' as the analysis object.
Table 10 Features interpretation.
| Type | Feature name | Interpretation |
|---|---|---|
| Scene | STATE_HWY_IND_N | Road crash did not occur on highway |
| Vehicle | STWD_VEHTYPE_AT_FAULT_L | Bicycle at fault |
| Scene | 24HOURS | Hour of road crash occurrence |
| People | STWD_VEHTYPE_AT_FAULT_N | Pedestrian at fault |
| Scene | INTERSECTION_Y | Road crash occurred at intersection |
| – | CHP_VEHTYPE_AT_FAULT_99 | Unknown subject at fault in road crash |
| Vehicle | STWD_VEHTYPE_AT_FAULT_A | Passenger car/Station wagon at fault |
| Scene | LIGHTING_D | Dark - No street lights |
| – | STWD_VEHTYPE_AT_FAULT | Not stated in road crash record |
items.
According to the SHAP summary chart, the situations in Table 11 are key contributing factors affecting whether the road crash vehicle needs to be towed.
In order to verify the causal relationship between the key features obtained by the SHAP analysis method and the towing label, the team introduced a method based on causal graph model. The causal graph model method requires establishing causal hypotheses by treating the analysis object as a confounding factor. In this study, the team sequentially selected each key feature analyzed by SHAP method as a confounding factor. Each causal graph establishes causal hypotheses with towing labels as endpoints. The example is shown in Fig. 8.
After establishing causal graph hypothesis for all seven features mentioned above, method estimate the causal effects between the features and label. In the hypothesis verification stage, team uses two refute to verify the validation of the hypothesis: ''data subset validation'' and ''adding random common cause''. If the -value of the verification result is greater than 0.05, it indicates that the causal hypothesis is valid. The causal inference results are shown in Table 12.
According to the results shown in Table 12, all hypotheses satisfy the ''adding random common cause'' validation. It can be inferred that these seven features have a stable and certain impact on the towing label in the dataset. Only two features with the highest absolute values of causal effects did not pass the validation of the data subset. It can be considered that these two features are affected by data subset sampling. These two features' influence on towing label vary in different data subsets.
Compare the causal effects estimated by causal inference with the SHAP values on the -axis corresponding to each feature in the SHAP summary diagram (Fig. 7). The influence estimation of the corresponding features in the SHAP diagram is very similar to the causal effect estimation provided by causal inference. This result validates the effectiveness of the SHAP method's analysis results from another perspective. The causal inference results support the conclusion of the SHAP method.
Table 11 Key causal factors and reasonable explanation.
| Key Causal Factors | Reasonable Explanation |
|---|---|
| When a road crash occurs on a highway, the road crash vehicle is more likely to need to be towed, which may be related to the higher driving speed on the highway. |
High speeds on highways increase the likelihood of vehicle breakdowns and driver incapacitation during road crashes. |
| When the main responsible party for a road crash is a bicyclist or pedestrian, the probability of the road crash vehicle needing to be towed is low because the vehicle damage is small when such a road crash occurs. |
Pedestrian-vehicle collisions and cyclist-vehicle collisions cause little harm to the motor vehicle and its driver, so this type of road crash is highly correlated with the fact that the road crash vehicle does not need to be towed. |
| When a road crash occurs in the early morning hours, the probability of the road crash vehicle needing to be towed is relatively high. |
The feature of early morning hours has a significant impact on the results. During the early morning hours, drivers are prone to fatigue and have low environmental visibility, making it more prone to serious road crashes and causing road crash vehicles need to be towed. |
| When a vehicle experiences a road crash at an intersection, it is also highly likely to require towing. |
Intersections are more prone to broadside collision. The vehicle's broadside is more fragile than the front and rear. When a broadside collision happens, the vehicle is more like to be broken down. |
| When the main responsible party for the road crash is a car, the road crash vehicle is more likely to need to be towed. |
Road crashes in which small cars are primarily responsible have a tendency to cause serious harm to motor vehicles and their drivers, leading to an increase in the probability of the road crash vehicle needing to be towed. |
| In the absence of street lighting at night, road crash vehicles are more likely to need to be towed. |
Driving a vehicle in a dark environment without street light illumination is a high-risk behavior. Drivers can only perceive the road environment within the range of lamp illumination, which is prone to serious road crashes (Tamakloe et al., 2021). |
Table 12 Causal inference results between 7 features obtained from SHAP method and towing label.
| Causal inference | Estimate effect | | Refute: Use a subset of data | | Refute: Add random common cause | | |-------------------------|-----------------|------------|------------------------------|------------|---------------------------------|--| | Confounding Factor | Mean value | New effect | p value | New effect | p value | | | STWD_VEHTYPE_AT_FAULT_L | −0.50697 | −0.49268 | 0 | −0.50697 | 1 | | | STWD_VEHTYPE_AT_FAULT_N | −0.42661 | −0.46039 | 0 | −0.42661 | 1 | | | STATE_HWY_IND_N | −0.11878 | −0.11876 | 0.96 | −0.11878 | 1 | | | LIGHTING_D | 0.08963 | 0.08991 | 0.96 | 0.08963 | 1 | | | STWD_VEHTYPE_AT_FAULT_A | 0.08948 | 0.08928 | 0.56 | 0.08948 | 1 | | | INTERSECTION_Y | 0.03094 | 0.02978 | 0.08 | 0.03094 | 1 | | | HOURS | 0.02471 | 0.02460 | 0.96 | 0.02471 | 0.89 | |
4.4. Map visualization of analysis results
Based on the important scene features summarized above, team screen out the road crashes that occurred at highway intersection without street lights in the late night from the entire dataset. There are 596 road crashes in the entire dataset that meet the above filtering criteria. Among them, there are 522 road crashes requiring towing vehicles, accounting for over 85%. We mark this kind of road crashes on the map and draw a high-risk map of road crashes that may require towing vehicles. The map is shown in Fig. 9. Towing vehicles deployed near these high-risk points can reach the road crash site faster and help alleviate traffic congestion. The detail of two zone on the map is displayed and shown in Figs. 10, 11.
5. Conclusion
This article proposes a framework for road crash prediction and causal analysis based on the road traffic scene feature of ''whether vehicles need to be towed'' in a large spatiotemporal range.
The main contributions of this article are summarized as follows: Research team have constructed a dataset based on TIMS that can be used for modeling the prior features of road crashes. This dataset has at least 14 valid features covering the entire California region of the United States for 12 years, with a sample size of over 2 million road crash records. This dataset is very suitable for road crash analysis research based on historical road crash records.
This article choosed ''whether the vehicles involved in the road crash were towed away'' as a particularly important feature of road crashes. As a new approach and application scenario, predicting the towing probability of road crash vehicles through prior scene features of road traffic can assist in towing planning. Our prediction result can help towing vehicle arrive at the scene earlier and tow the road crash vehicle to alleviate traffic congestion. It can also provide a reference for evaluating the risk of road crashes.
The article proposes a method for predicting the towing of road crash vehicles based on the HGBoost model and uses the SHAP method to analyze and discover seven key contributing factors that affect whether the road crash vehicles need to be towed: traffic environmental factors account for the largest proportion, including four features: ''on state highway'', ''early hours at night'', ''at intersections'', and ''no street lighting at night''; There are vehicle factors, including two features: ''bicyclists bear the main responsibility'' and ''car bear the main responsibility''; Finally, there is a human factors ''pedestrians taking primary responsibility''. In the absence of driver factors, several factors of the traffic environment are still the dominant causes in such kind of road crashes, providing a useful reference for road traffic safety planning and protection. For example, adding street lights to highway intersections and their adjacent areas that lack nighttime lighting. We would like to release our dataset to support further research and development in this field.
5.1. limitations and future directions
This study proposes a method framework based on machine learning and its interpretability to explore the key causes of specific road crashes. In the future, researchers can attempt to fully integrate causal inference methods into this framework to enhance its effectiveness and robustness in exploring and analyzing key causes.
Fig. 9. The high-risk map of road crashes that require towing vehicles and map partial zoom display.
Fig. 10. Zoom display: the intersection of Cabrillo Highway and Oso Flaco Lake Road.
Fig. 11. Zoom display: the intersection of US-101 and Spence Road.
This study is a research on specific accidents across a large area of California. In future research, researchers can attempt to build a more universal cross region method based on big datasets that can adapt to small sample environments, building upon the foundation of this study. On the other hand, researchers can try to select specific blocks for cross regional comparison within California and include background variables. These approaches can improve the robustness of the model and promote the development of targeted and effective safety interventions in different geographical and regulatory landscapes.
CRediT authorship contribution statement
Ronghui Zhang: Writing – review & editing, Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Conceptualization. Yang Liu: Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Zihan Wang: Writing – review & editing, Writing – original draft, Validation, Methodology, Investigation, Formal analysis, Conceptualization. Junzhou Chen: Writing – review & editing, Validation, Supervision, Software, Methodology, Conceptualization. Qiang Zeng: Writing – review & editing, Supervision, Methodology, Investigation, Formal analysis, Conceptualization. Lai Zheng: Writing – review & editing, Methodology, Investigation, Formal analysis, Conceptualization. Hui Zhang: Writing – review & editing, Validation, Supervision, Investigation, Formal analysis, Conceptualization. Yulong Pei: Writing – review & editing, Supervision, Methodology, Investigation, Formal analysis, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This project is jointly supported by National Natural Science Foundation of China (Nos. 52172350, 51775565), Guangdong Basic and Applied Research Foundation (Nos. 2021B1515120032, 2022B1515120072), Guangzhou Science and Technology Plan Project (Nos.2024B01W0079, 202206030005), Nansha Key RD Program (No. 2022ZD014), Science and Technology Planning Project of Guangdong Province (No. 2023B1212060029).
Data availability
Data will be made available on request.
References
- Berghaus, Moritz, Lamberty, Serge, Ehlers, Jörg, Kalló, Eszter, Oeser, Markus, 2024. Vehicle trajectory dataset from drone videos including off-ramp and congested traffic – analysis of data quality, traffic flow, and accident risk. Commun. Transp. Res. 4, 100133. http://dx.doi.org/10.1016/j.commtr.2024.100133.
- BİLİM-TEKNOLOJİ, 2014. Otomobiller acil durumda çağrıyapacak. https://www. aljazeera.com.tr/haber/otomobiller-acil-durumda-cagri-yapacak.
- Chen, Tianqi, Guestrin, Carlos, 2016. XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD '16, Association for Computing Machinery, New York, NY, USA, pp. 785–794. http://dx.doi.org/10.1145/2939672.2939785.
- Chiou, Yu-Chiun, Fu, Chiang, 2015. Modeling crash frequency and severity with spatiotemporal dependence. Anal. Methods Accid. Res. 5–6, 43–58. http://dx.doi. org/10.1016/j.amar.2015.03.002.
El-Basyouny, Karim, Barua, Sudip, Islam, Md Tazul, 2014. Investigation of time and weather effects on crash types using full Bayesian multivariate Poisson lognormal models. Accid. Anal. Prev. 73, 91–99. http://dx.doi.org/10.1016/j.aap.2014.08. 014.
Fountas, Grigorios, Fonzone, Achille, Gharavi, Niaz, Rye, Tom, 2020. The joint effect of weather and lighting conditions on injury severities of single-vehicle accidents. Anal. Methods Accid. Res. 27, 100124. http://dx.doi.org/10.1016/j.amar.2020. 100124.
Friedman, Jerome H., 2001. Greedy function approximation: A gradient boosting machine. Ann. Statist. 29 (5), 1189–1232, URL http://www.jstor.org/stable/ 2699986.
Gurumurthy, Krishna Murthy, Bansal, Prateek, Kockelman, Kara M., Li, Zili, 2022. Modelling animal-vehicle collision counts across large networks using a Bayesian hierarchical model with time-varying parameters. Anal. Methods Accid. Res. 36, 100231. http://dx.doi.org/10.1016/j.amar.2022.100231.
Huang, Helai, Ding, Xizhi, Yuan, Chen, Liu, Xinyuan, Tang, Jinjun, 2023. Jointly analyzing freeway primary and secondary crash severity using a copula-based approach. Accid. Anal. Prev. 180, 106911. http://dx.doi.org/10.1016/j.aap.2022. 106911.
Hussain, Fizza, Li, Yuefeng, Haque, Md Mazharul, 2024. Integrating machine learning and extreme value theory for estimating crash frequency-by-severity via AI-based video analytics. Commun. Transp. Res. 4, 100147. http://dx.doi.org/10.1016/j. commtr.2024.100147.
Ihueze, Chukwutoo C., Onwurah, Uchendu O., 2018. Road traffic accidents prediction modelling: An analysis of anambra state, Nigeria. Accid. Anal. Prev. 112, 21–29. http://dx.doi.org/10.1016/j.aap.2017.12.016.
Jha, Alok Nikhil, Chatterjee, Niladri, Tiwari, Geetam, 2021. A performance analysis of prediction techniques for impacting vehicles in hit-and-run road accidents. Accid. Anal. Prev. 157, 106164. http://dx.doi.org/10.1016/j.aap.2021.106164.
Ju, Yong Han, Sohn, So Young, 2011. Quantification method analysis of the relationship between occupant injury and environmental factors in traffic accidents. Accid. Anal. Prev. 43 (1), 342–351. http://dx.doi.org/10.1016/j.aap.2010.09.002.
Kashifi, Mohammad Tamim, 2023. Investigating two-wheelers risk factors for severe crashes using an interpretable machine learning approach and SHAP analysis. IATSS Res. 47 (3), 357–371. http://dx.doi.org/10.1016/j.iatssr.2023.07.005.
Ke, Guolin, Meng, Qi, Finley, Thomas, Wang, Taifeng, Chen, Wei, Ma, Weidong, Ye, Qiwei, Liu, Tie-Yan, 2017. LightGBM: A highly efficient gradient boosting decision tree. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. NeurIPS '17, Curran Associates Inc., Red Hook, NY, USA, pp. 3149–3157, URL https://proceedings.neurips.cc/paper/2017/file/ 6449f44a102fde848669bdd9eb6b76fa-Paper.pdf.
Lee, Jonghak, Chae, Junghyo, Yoon, Taekwan, Yang, Hojin, 2018. Traffic accident severity analysis with rain-related factors using structural equation modeling – a case study of seoul city. Accid. Anal. Prev. 112, 1–10. http://dx.doi.org/10.1016/ j.aap.2017.12.013.
Lehto, Steve, 2017. The Ultimate Car Repo Horror Story. https://www.roadandtrack. com/car-culture/a12474968/the-ultimate-car-repo-horror-story/.
Li, Yunxuan, Li, Meng, Yuan, Jinghui, Lu, Jian, Abdel-Aty, Mohamed, 2021. Analysis and prediction of intersection traffic violations using automated enforcement system data. Accid. Anal. Prev. 162, 106422. http://dx.doi.org/10.1016/j.aap.2021.
Liu, Qingchao, Li, Chun, Jiang, Haobin, Nie, Shiqi, Chen, Long, 2022. Transfer learning-based highway crash risk evaluation considering manifold characteristics of traffic flow. Accid. Anal. Prev. 168, 106598. http://dx.doi.org/10.1016/j.aap. 2022.106598.
Lundberg, Scott M., Erion, Gabriel G., Lee, Su-In, 2018. Consistent individualized feature attribution for tree ensembles. http://dx.doi.org/10.48550/arXiv.1802.03888, arXiv preprint arXiv:1802.03888.
Lundberg, Scott M., Lee, Su-In, 2017. A unified approach to interpreting model predictions. In: Guyon, I., Von Luxburg, U., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), In: Advances in Neural Information Processing Systems, 30, Curran Associates, Inc., URL https://proceedings.neurips.cc/paper\_ files/paper/2017/file/8a20a8621978632d76c43dfd28b67767-Paper.pdf.
Ma, Zhengjing, Mei, Gang, Cuomo, Salvatore, 2021. An analytic framework using deep learning for prediction of traffic accident injury severity based on contributing factors. Accid. Anal. Prev. 160, 106322. http://dx.doi.org/10.1016/j.aap.2021. 106322.
Mannering, Fred, Bhat, Chandra R., Shankar, Venky, Abdel-Aty, Mohamed, 2020. Big data, traditional data and the tradeoffs between prediction and causality in highway-safety analysis. Anal. Methods Accid. Res. 25, 100113. http://dx.doi.org/ 10.1016/j.amar.2020.100113.
Nikolaou, Dimitrios, Ziakopoulos, Apostolos, Dragomanovits, Anastasios, Roussou, Julia, Yannis, George, 2023. Comparing machine learning techniques for predictions of motorway segment crash risk level. Safety 9 (2), 32. http://dx.doi.org/10.3390/
Olowosegun, Adebola, Babajide, Nathaniel, Akintola, Adeyemi, Fountas, Grigorios, Fonzone, Achille, 2022. Analysis of pedestrian accident injury-severities at road junctions and crossings using an advanced random parameter modelling framework: The case of Scotland. Accid. Anal. Prev. 169, 106610. http://dx.doi.org/10.1016/ j.aap.2022.106610.
Pang, Jiajun, Krathaus, Adam, Benedyk, Irina, Ahmed, Sheikh Shahriar, Anastasopoulos, Panagiotis Ch., 2022. A temporal instability analysis of environmental factors affecting accident occurrences during snow events: The random parameters hazardbased duration model with means and variances heterogeneity. Anal. Methods Accid. Res. 34, 100215. http://dx.doi.org/10.1016/j.amar.2022.100215.
Schlögl, Matthias, 2020. A multivariate analysis of environmental effects on road accident occurrence using a balanced bagging approach. Accid. Anal. Prev. 136, 105398. http://dx.doi.org/10.1016/j.aap.2019.105398.
Tamakloe, Reuben, Lim, Sungho, Sam, Enoch F., Park, Shin Hyoung, Park, Dongjoo, 2021. Investigating factors affecting bus/minibus accident severity in a developing country for different subgroup datasets characterised by time, pavement, and light conditions. Accid. Anal. Prev. 159, 106268. http://dx.doi.org/10.1016/j.aap.2021. 106268.
Transportation Injury Mapping System, Safe Transportation Research and Education Center, University of California, Berkeley, 2023, https://tims.berkeley.edu/.
Wen, Xiao, Xie, Yuanchang, Wu, Lingtao, Jiang, Liming, 2021. Quantifying and comparing the effects of key risk factors on various types of roadway segment crashes with lightgbm and SHAP. Accid. Anal. Prev. 159, 106261. http://dx.doi. org/10.1016/j.aap.2021.106261.
Wen, Huiying, Zhang, Xuan, Zeng, Qiang, Sze, N.N., 2019. Bayesian spatial-temporal model for the main and interaction effects of roadway and weather characteristics on freeway crash incidence. Accid. Anal. Prev. 132, 105249. http://dx.doi.org/10. 1016/j.aap.2019.07.025.
World Health Organization, 2023. Global Status Report on Road Safety 2023. Technical Repor, World Health Organization, Geneva, Licensed: CC BY-NC-SA 3.0 IGO.
Xue, Huiyuan, Guo, Peizhuo, Li, Yiyan, Ma, Jun, 2024. Integrating visual factors in crash rate analysis at intersections: An automl and SHAP approach towards cycling safety. Accid. Anal. Prev. 200, 107544. http://dx.doi.org/10.1016/j.aap.
Yan, Xintong, He, Jie, Zhang, Changjian, Liu, Ziyang, Qiao, Boshuai, Zhang, Hao, 2021. Single-vehicle crash severity outcome prediction and determinant extraction using tree-based and other non-parametric models. Accid. Anal. Prev. 153, 106034. http://dx.doi.org/10.1016/j.aap.2021.106034.
Yang, Zekun, Zhang, Wenping, Feng, Juan, 2022. Predicting multiple types of traffic accident severity with explanations: A multi-task deep learning framework. Saf. Sci. 146, 105522. http://dx.doi.org/10.1016/j.ssci.2021.105522.
Ye, Ying, Zhong, Chen, Suel, Esra, 2024. Unpacking the perceived cycling safety of road environment using street view imagery and cycle accident data. Accid. Anal. Prev. 205, 107677. http://dx.doi.org/10.1016/j.aap.2024.107677.
Zeng, Qiang, Wang, Fangzhou, Chen, Tiantian, Sze, N.N., 2023a. Incorporating real-time weather conditions into analyzing clearance time of freeway accidents: A grouped random parameters hazard-based duration model with time-varying covariates. Anal. Methods Accid. Res. 38, 100267. http://dx.doi.org/10.1016/j.amar.2023. 100267.
Zeng, Qiang, Wang, Qianfang, Wang, Xiaofei, 2022. An empirical analysis of factors contributing to roadway infrastructure damage from expressway accidents: A Bayesian random parameters tobit approach. Accid. Anal. Prev. 173, 106717. http://dx.doi.org/10.1016/j.aap.2022.106717.
Zeng, Qiang, Wang, Qianfang, Zhang, Keke, Wong, S.C., Xu, Pengpeng, 2023b. Analysis of the injury severity of motor vehicle–pedestrian crashes at urban intersections using spatiotemporal logistic regression models. Accid. Anal. Prev. 189, 107119. http://dx.doi.org/10.1016/j.aap.2023.107119.
Zhao, Jingya, Liu, Pan, Xu, Chengcheng, Bao, Jie, 2021. Understand the impact of traffic states on crash risk in the vicinities of type a weaving segments: A deep learning approach. Accid. Anal. Prev. 159, 106293. http://dx.doi.org/10.1016/j. aap.2021.106293.
Zhu, Siying, Wan, Jianwu, 2021. Cost-sensitive learning for semi-supervised hit-and-run analysis. Accid. Anal. Prev. 158, 106199. http://dx.doi.org/10.1016/j.aap.2021. 106199.