A highway traffic accident risk assessment method based on multi-model fusion

Abstract

This invention proposes a highway traffic accident risk assessment method based on multi-model fusion. It integrates multi-source heterogeneous spatiotemporal data, including accident information, highway data, meteorological data, and traffic data. Based on the data characteristics, a multi-model fusion highway network traffic accident risk assessment model is designed. The model extracts key semantic information from accident descriptions and integrates highway images, meteorological data, and traffic data. Risk assessment is performed based on the construction and training of a random forest model using historical accident data. The model calculates the traffic accident risk index for each road section and uses the quartile method to classify traffic accident risk levels as an evaluation criterion. The model estimation results are interpreted using the Additive Model Interpreter (SHAP). The method of the present invention considers multidimensional accident risk factors, thereby achieving a comprehensive assessment of traffic accident risk under the multi-faceted influence of people, vehicles, roads, and the environment, and has broad applicability.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic safety, and in particular relates to a highway traffic accident risk assessment method based on multi-model fusion. Background Art

[0002] In recent years, cities have placed a high priority on urban safety development and preventing and mitigating safety risks. However, a clear methodology for categorizing accident risk across road sections has yet to be established. Therefore, categorizing the risk levels of accident-prone sections can help decision-makers focus on high-risk areas and develop targeted traffic safety measures to improve road safety.

[0003] The occurrence of traffic accidents is often influenced by multiple factors. In the prior art, Li Daoxun et al. disclosed a real-time highway traffic accident risk assessment method based on deep learning (CN112562337A). The model incorporates traffic flow information, road information, accident information, and time series features to complete a real-time assessment of highway accident risk levels. Xie Bo et al. disclosed a method for urban road traffic accident risk assessment based on a gradient boosting decision tree (CN116777300A). By collecting accident influencing factors, they deeply explored the influence relationship between traffic accident risk factors and the number of traffic accidents, achieving single-dimensional and multi-dimensional traffic accident risk assessment.

[0004] However, existing traffic accident risk assessment technologies are often limited to one or several specific factors. Models constructed based on a single type of acquired data features make it difficult to comprehensively assess the traffic accident risk level at key locations in the road network from multiple perspectives, including people, vehicles, roads, and the environment. Summary of the Invention

[0005] To integrate multi-source heterogeneous spatiotemporal data, such as accident information, road network data, and meteorological data, and more comprehensively extract the factors influencing highway traffic accident risk, this paper proposes a highway traffic accident risk assessment method based on multi-model fusion. Furthermore, based on the characteristics of multi-source heterogeneous spatiotemporal data, a highway traffic accident risk assessment model based on the Latent Dirichlet Allocation (LDA) document topic generation model and a random forest model is designed. This method extracts key semantic information from accident descriptions and integrates highway profiles, meteorological data, and traffic data. Risk assessment is then performed based on the construction and training of a random forest model using historical accident data. This method significantly improves the efficiency of multi-class data fusion and has broad applicability.

[0006] The technical solutions of the present invention are as follows:

[0007] A highway traffic accident risk assessment method based on multi-model fusion includes the following steps:

[0008] Step 1: Collect highway traffic accident investigation data, road network data, meteorological data, and traffic data;

[0009] Step 2: Extract keywords from the accident investigation records, focusing on human and vehicle factors, and construct LDA topic model feature vectors:

[0010] Step 21: Build a text library based on the acquired traffic accident investigation records, and use the LDA topic model to perform unsupervised classification learning on the text library to identify key semantics;

[0011] Step 22. During the iterative calculation process, each document is represented as a topic distribution vector, and each topic is represented as a word distribution vector. The topic with the highest probability in the document is taken, and the probability of the words in the topic is obtained. After weighted summation, the LDA topic vector is obtained.

[0012] Step 3: Calculate the road and environmental impact factors based on the data obtained in step 1:

[0013] Step 31: Calculate the highway section type, steep slope, lane number change point, speed limit change point, bridge and tunnel section, etc., establish a sample data set, and form a road section-based portrait.

[0014] Step 32: Obtain highway meteorological information, including maximum temperature, minimum temperature, rain and snow conditions, and wind speed level, and convert it into one-hot encoding.

[0015] Step 33: Match highway traffic characteristics for each road section.

[0016] Step 4: Based on the results obtained in steps 2 and 3, the LDA topic vector, road information, and meteorological information data are fused to construct a fused dataset. The dataset contains four categories: people, vehicles, roads, and environment, which are potential influencing factors of traffic accidents.

[0017] Step 5: Taking year A as the study period, calculate the traffic accident risk index of each road section and use it as the dependent variable. Use the potential influencing factors of traffic accidents obtained in step 4 as independent variables to establish a traffic accident risk assessment machine learning model, complete model parameter tuning, generate a traffic accident risk assessment model based on machine learning, and obtain the traffic accident risk index.

[0018] Step 6: Divide the traffic accident risk index of each road section into four levels according to the 25%, 50% and 75% quartiles as the traffic accident risk level evaluation standard.

[0019] Step 7: Use the additive model interpreter (SHAP) to interpret the estimation results of the traffic accident risk assessment model based on machine learning; calculate the contribution of individual influencing factors of each road section to the traffic accident risk (Shapley value).

[0020] Step 8: Based on the Shapley values ​​of the four factors (people, vehicles, roads, and environment) obtained in Step 7, calculate the risk index for each road section in these four dimensions. Define sections of the road network classified as "high risk" or "medium risk" as accident-prone sections. Based on the contribution of the risk factors (Shapley values) in these accident-prone sections, rank the potential risks in these four dimensions, providing accurate underlying data support for relevant management departments to implement governance and improvements.

[0021] Beneficial effects of the present invention:

[0022] The present invention proposes a highway traffic accident risk assessment method based on multi-model fusion. From the perspective of macro-road space, it considers the impact of human and vehicle characteristics, road factors, meteorological factors, and traffic data on traffic accidents in actual traffic accidents, integrates multi-source data for analysis, and interprets accident risks in a more comprehensive and objective manner. The present invention uses the LDA topic model to extract information related to human and vehicle factors from the accident description text, and constructs a fused data set with the help of highway portraits, meteorological data, and traffic data. The random forest model is used to mine the nonlinear correlation between accident risk factors and accident risk index, thereby achieving a comprehensive assessment of traffic accident risks under the multi-faceted influence of people, vehicles, roads, and the environment. Through this innovation, it is possible to more comprehensively assess the traffic accident risks of important sections of highways, analyze potential risk factors, and identify sections with high accident risks, providing strong support for traffic safety management and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of a traffic accident risk assessment method based on multi-model fusion provided by an embodiment of the present invention;

[0024] Figure 2 This is an operational flow chart of an LDA topic model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The specific embodiments of the present invention are further described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, the present invention proposes a traffic accident risk level classification model based on multi-model fusion, which includes the following steps:

[0027] Step 1: Collect highway traffic accident data, road network data, meteorological data, etc.

[0028] Among them, traffic accident data is obtained from the accident investigation of the traffic management department. Specific information should include the number of historical accidents, accident location, time, vehicle status, casualties, etc. The accident location should be specific to the highway name and pile number; road network data is obtained from highway portraits and should include highway mileage, road name, pile number, special sections, etc.; meteorological data includes maximum / minimum temperature, rain and snow conditions, and wind level.

[0029] Step 2: Extract keywords from the accident investigation records, focus on human and vehicle factors, and construct LDA topic model feature vectors (the basic process is as follows Figure 2 shown):

[0030] Step 21: Build a text library based on the acquired traffic accident investigation records, and use the LDA topic model to perform unsupervised classification learning on the text library to identify key semantics;

[0031] First, the collected accident description texts are preprocessed to remove prepositions and conjunctions without obvious meanings. They also contain a large number of numbers, English, punctuation marks, or meaningless Chinese words that cannot express the subject content. The original corpus is segmented using the Jieba word segmentation method, a stop word list is set to remove Chinese words that are not related to the subject, and a keyword list is set to retain special nouns related to the subject.

[0032] LDA topic modeling requires determining the optimal number of topics K. Generally, perplexity is used to determine the optimal number of topic words. First, the perplexity corresponding to different numbers of topics is calculated, and the changing trend of perplexity with the number of topics is observed. Normally, perplexity decreases as the number of topics increases, but when the number of topics is too large, perplexity may begin to increase. It is generally believed that the smaller the perplexity, the better the classification effect. Therefore, the number of topics corresponding to the fastest decrease in perplexity and stabilization is selected as the optimal number of topics. Perplexity The calculation formula is as follows:

[0033]

[0034] Among them, M is the number of texts in the collection, N m is the length of document m. is the probability value of word w in the mth document.

[0035] After determining the optimal number of topics, the segmented text data is used to train the LDA topic model, resulting in two probability distributions: "topic-word" and "document-topic." The "topic-word" distribution provides the top n most probable words for each topic, identifying the most frequent words within each topic. The "document-topic" distribution provides the probability that a document belongs to a particular topic.

[0036] Training the LDA model: Use the determined optimal number of topics K, set the model parameters α and β, and the number of text iterations Y times. Extract the Z words with the highest probability under each topic, sort the words from high to low by frequency, and use them as representatives to describe the meaning of the topic. Combine these words to identify the topic.

[0037] Step 22: During the iterative calculation process, each document is represented as a topic distribution vector, and each topic is represented as a word distribution vector. The topic with the highest probability in the document is selected, and the probabilities of the words within the topic are obtained. After weighted summation, the LDA topic vector is obtained.

[0038] Based on the trained LDA topic model, a feature vector is constructed. During the iterative calculation process, each document is represented as a topic distribution vector, and each topic is represented as a word distribution vector. Therefore, the LDA model can output the word distribution of each topic and the topic distribution of each document. The formula for calculating the LDA topic feature vector is as follows:

[0039] x={V(ω1),V(ω2)…V(ω t )}

[0040] Where: x represents the word vector representation of all words, V(ω t ) is represented as a word vector, ω t It is represented as the word vector representation of the vocabulary, and t represents the total number of vocabulary.

[0041] According to the results of model training, take each document d i The most likely topic z max , defining the document as d z Then from topic z max Select the first n words (t1, t2, ..., tn) and their corresponding probability values ​​(p1, p2, ..., pn), normalize the corresponding probability values, and use them as the weights of the n words. The calculation formula is as follows:

[0042]

[0043] Where (q1, q2, ..., qn) represent the normalized values ​​of (p1, p2, ..., pn), representing the weights of the first n words. The LDA topic vector is obtained by weighted summing the word vectors (C(t1), C(t2), ..., C(tn)):

[0044]

[0045] Furthermore, the obtained LDA topic vector is expressed as a numerical variable, and each element is used to represent the degree of association between the document or topic and the corresponding topic, which is displayed as a normalized probability value with a value range between 0 and 1.

[0046] Step 3: Based on the data obtained in step 1, calculate or obtain the following information:

[0047] Step 31: Calculate the highway section type, steep slope, lane number change point, speed limit change point, bridge and tunnel section, etc., establish a sample data set, and form a road section-based portrait.

[0048] Step 32: Obtain highway meteorological information, including maximum / minimum temperature, rain / snow conditions, and wind speed level, and convert it into a one-hot encoding with encoding lengths of L1, L2, and L3, respectively.

[0049] Step 33: Match highway traffic characteristics for each road section.

[0050] Step 4: Based on the results obtained in steps 2 and 3, the LDA topic vector, road structure characteristics, traffic information, and meteorological information data are fused to form a fused data set.

[0051] Step 4 specifically includes extracting the accident location and time from the accident description, matching the administrative district name with latitude and longitude information by administrative district code. Based on the extracted accident time and administrative district name, weather data is crawled to automatically match accident data with meteorological data, completing the spatiotemporal alignment of meteorological information.

[0052] Furthermore, a fusion dataset of LDA topic vectors, road information and weather information is constructed.

[0053] The present invention uses traffic accident data of consecutive A years as samples, and 5 years can be taken as a research period.

[0054] Step 5: Using Year A as the study period, calculate the traffic accident risk index for each road section and use it as the dependent variable. Using the potential influencing factors of traffic accidents calculated in Step 4 as independent variables, establish a machine learning model for traffic accident risk assessment, optimize the model parameters, and generate a traffic accident risk assessment model. This present invention uses a random forest model to establish the traffic accident risk assessment model.

[0055] The model in step five uses the fused data set in step four as input, and outputs a traffic accident risk index.

[0056] Based on step 5, the fused dataset is split, and a portion of samples are randomly selected as the training set, and the remaining samples are used as the test set. Then, the possible value sets of model hyperparameters such as the number of decision trees (n_estimators), the maximum tree depth (max_depth), the minimum number of leaf node samples (min_samples_leaf), and the minimum number of split samples (min_samples_split) are determined. To find the optimal parameter values, GridSearchCV under sklearn is called to perform a grid search within an appropriate range, and the best performing parameters are found through loop traversal. 2 _score (coefficient of determination) is used as a standard to measure the performance of the model. The cross-validation method is used to verify the parameter combination and obtain the evaluation index. 2 The calculation method is as follows:

[0057]

[0058] Where f(x i ) is the predicted value of the model, y i The coefficient of determination generally ranges from 0 to 1. The closer the value is to 1, the better the independent variable explains the dependent variable in the regression analysis.

[0059] The model calculation process is as follows:

[0060] (1) The input sample data set is D = {a1, a2, ..., a n}, a i (i=1~n) are independent variables including people, cars, roads, weather and other factors, and the number of iterations of the weak classifier is T. When t=1,2,…,T, the final strong classifier generated is g(a i );

[0061] (2) Randomly sample the training data set for the tth time, and collect j samples in total to obtain the sampling volume Dt including j samples;

[0062] (3) Use the sampling amount Dt to train the t-th decision tree model. In the process of training the model nodes, a part of the features is randomly selected. Among the selected features, the best feature is selected as the left and right subtrees of the decision tree for division.

[0063] (4) The value obtained by taking the arithmetic average of the regression results obtained by T weak learners is the final result obtained by the random forest model.

[0064] Furthermore, based on the trained random forest model, the ROC curve was drawn and the AUC value was calculated to evaluate the predictive ability of the model.

[0065] Furthermore, a traffic accident risk index is calculated based on the number and severity of historical accidents. Data on various types of traffic accidents over the past five years is collected, and the accidents are classified and weighted based on the number of casualties. The weighted sum of the assigned accident levels is then used to calculate the accident risk index. The classification and weighting are shown in the table below.

[0066]

[0067]

[0068] Based on step 4, the traffic accident risk factors are represented as a fusion data set of four types of data: LDA topic vectors (representing human and vehicle characteristics), road information, meteorological data, and traffic data.

[0069] Step 6. Based on the traffic accident risk index obtained in step 5, the risk index is defined as four levels of "high risk, medium risk, low risk, and very low risk" using the quartile method as a traffic accident risk level evaluation standard.

[0070] The step six specifically includes: based on the traffic accident risk index obtained from the output of step five, sorting the risk index from small to large, and dividing it into four levels according to quartiles, defining less than 25% of the quartile as extremely low risk, greater than 25% and less than 50% as low risk, greater than 50% and less than 75% as medium risk, and greater than 75% as high risk.

[0071] Step 7: Use the additive model interpreter (SHAP) to interpret the model and identify the contribution (shapely value) of the four risk factors of people, vehicles, roads, and environment to the accident risk index.

[0072] The step seven specifically includes: calculating the Shapley value to obtain the contribution of a single factor to the accident risk, and the calculation formula is as follows:

[0073]

[0074] Where φ represents the contribution of the i-th feature, N{i} represents the set of all features, l represents the number of features, S represents the subset of features for a given prediction, and f(S∪{i}) and f(S) represent the model estimation results with or without the i-th feature, respectively.

[0075] Based on the calculation of Shapley values, we can calculate the Shapley value of each factor for road sections with higher risk indices—that is, key sections of the road network classified as "high risk" or "medium risk." This value represents the contribution of each influencing factor to the accident risk level. By comparing the Shapley values ​​corresponding to each risk factor, we can derive a risk ranking for people, vehicles, roads, and the environment. Larger absolute Shapley values ​​indicate a greater impact on the dependent variable.

[0076] Step 8: Based on the Shapley values ​​of the four factors (people, vehicles, roads, and environment) obtained in Step 7, calculate the risk index for each road section in these four dimensions. Define sections of the road network classified as "high risk" or "medium risk" as accident-prone sections. Based on the contribution of the risk factors (Shapley values) in these accident-prone sections, rank the potential risks in these four dimensions, providing accurate underlying data support for relevant management departments to implement governance and improvements.

[0077] The traffic accident risk assessment method provided by the present invention uses the LDA topic model to extract relevant information about factors such as people and vehicles from accident description text. It then constructs a fused accident risk dataset using highway imagery, meteorological information, and traffic data. A machine learning algorithm is then used to explore the nonlinear correlation between potential influencing factors and accident risk, thereby enabling a comprehensive assessment of traffic accident risk under the multi-faceted influence of people, vehicles, roads, and the environment.

[0078] This paper comprehensively considers the impact of factors such as people, vehicles, roads, and the environment on traffic accident risks, integrates multi-source data, and innovatively designs a highway network traffic accident risk assessment model based on document topic generation and random forest fusion. It analyzes the nonlinear correlation between traffic accident risks and potential influencing factors, and identifies accident-prone sections of highways and potential risk factors.

Claims

1. A highway traffic accident risk assessment method based on multi-model fusion, characterized by: The steps include: Step 1: Collect highway traffic accident investigation data, road network data, meteorological data, and traffic data; Step 2: Extract keywords from the accident investigation records, focusing on human and vehicle factors, and construct LDA topic model feature vectors: Step 21: Build a text library based on the acquired traffic accident investigation records, and use the LDA topic model to perform unsupervised classification learning on the text library to identify key semantics; Step 22: During the iterative calculation process, each document is represented as a topic distribution vector, and each topic is represented as a word distribution vector. The topic with the highest probability in the document is taken, and the probability of the words in the topic is obtained. After weighted summation, the LDA topic vector is obtained. Step 3: Calculate the road and environmental impact factors based on the data obtained in step 1: Step 31: Calculate the highway section type, steep slope, lane number change point, speed limit change point, and bridge and tunnel sections, establish a sample data set, and construct a road section-based portrait. Step 32: Obtain highway meteorological information, including maximum temperature, minimum temperature, rain and snow conditions, and wind speed level, and convert it into a one-hot encoding; Step 33: Match highway traffic characteristics for each road section; Step 4: Based on the results obtained in steps 2 and 3, the LDA topic vectors, road information, and meteorological information data are fused to construct a fused dataset. The dataset contains four categories: people, vehicles, roads, and environment, which are potential influencing factors of traffic accidents. Step 5: Using Year A as the study period, calculate the traffic accident risk index for each road section and use it as the dependent variable. Use the potential influencing factors of traffic accidents obtained in Step 4 as independent variables to establish a machine learning model for traffic accident risk assessment. Optimize the model parameters to generate a machine learning-based traffic accident risk assessment model and obtain the traffic accident risk index. Step 6: Divide the traffic accident risk index of each road section into four levels of "high risk, medium risk, low risk, and very low risk" according to the 25th, 50th, and 75th percentiles, which serve as the traffic accident risk level evaluation standard; Step 7: Use the additive model interpreter SHAP to interpret the estimation results of the traffic accident risk assessment model based on machine learning; Calculate the contribution of individual influencing factors on the traffic accident risk of each road section, that is, the Shapley value; Step 8: Based on the Shapley values ​​of the four factors (people, vehicles, roads, and environment) obtained in Step 7, calculate the risk index values ​​for each road section in these four dimensions; define the sections of the road network that are classified as "high risk" or "medium risk" as accident-prone sections. According to the contribution of risk factors of accident-prone sections, the potential risk ranking of the four dimensions of the accident-prone areas is obtained.

2. The highway traffic accident risk assessment method based on multi-model fusion according to claim 1 is characterized in that: The step 21 is specifically as follows: First, we preprocessed the collected accident description texts to remove prepositions and conjunctions without obvious meanings, characters that contain a large number of numbers, English characters, punctuation marks, or meaningless Chinese words that cannot express the subject content, and then segmented the original corpus using the Jieba word segmentation method. We set a stop word list to remove Chinese words that are not related to the subject, and set a keyword list to retain specific nouns related to the subject. LDA topic model building requires determining the optimal number of topics K, and using perplexity to determine the optimal number of topic words; first, calculate the perplexity corresponding to different numbers of topics, and observe the changing trend of perplexity with the number of topics; the smaller the perplexity, the better the classification effect, and select the number of topics corresponding to the fastest decrease in perplexity and stable as the optimal number of topics; perplexity The calculation formula is as follows: Among them, M is the number of texts in the collection, N m is the length of document m; is the probability value of word w in the mth document; After determining the optimal number of topics, the segmented text data is used to train the LDA topic model, resulting in two probability distributions: "topic-word" and "document-topic." The "topic-word" distribution gives the top n words with the highest probability for each topic, identifying the high-frequency words in each topic. The "document-topic" distribution gives the probability that a document belongs to a particular topic. Training the LDA model: Using the optimal number of topics K, set the model parameters α and β, and the number of text iterations Y, extract the Z most likely words under each topic, sort the words by frequency from highest to lowest, and use them as representatives to describe the meaning of the topic. Combine these words to identify the topic; The step 22 is specifically as follows: Based on the trained LDA topic model, a feature vector is constructed. During the iterative calculation process, each document is represented as a topic distribution vector, and each topic is represented as a word distribution vector. Therefore, the LDA model outputs the word distribution of each topic and the topic distribution of each document. The formula for calculating the LDA topic feature vector is as follows: x={V(ω1),V(ω2)…V(ω t )} Where: x represents the word vector representation of all words, V(ω t ) is represented as a word vector, ω t It is represented as the word vector representation of the vocabulary, and t represents the total number of vocabulary; According to the results of LDA topic model training, take each document d i The most likely topic z max , defining the document as d z Then from topic z max Select the first n words and their corresponding probability values, normalize the corresponding probability values, and use them as the weights of the n words. The calculation formula is as follows: Where: q1, q2, ..., qn represent the normalized values ​​of p1, p2, ..., pn, respectively, representing the weights of the first n words; the word vectors C(t1), C(t2), ..., C(tn) are weighted summed to obtain the LDA topic vector:

3. The highway traffic accident risk assessment method based on multi-model fusion according to claim 1 is characterized in that: In the step 5, a traffic accident risk assessment model is established using a random forest model, the fused dataset from step 4 is used as input to the model, and the output is a traffic accident risk index; The fused dataset is split, and a portion of samples is randomly selected as the training set, and the remaining samples are used as the test set. Then, the possible value sets of the model hyperparameters, including the number of decision trees, the maximum tree depth, the minimum number of leaf node samples, and the minimum number of partition samples, are determined. To find the optimal parameter values, GridSearchCV under sklearn is called to perform a grid search within an appropriate range, and the best performing parameters are found through loop traversal. The coefficient of determination is used as the standard for measuring model performance, and the parameter combination is verified by cross-validation to obtain the evaluation index; the coefficient of determination R 2 The calculation method is as follows: Where f(x i ) is the predicted value of the model, y i is the true value; the value of the coefficient of determination ranges from 0 to 1. The closer the value is to 1, the better the independent variable explains the dependent variable in the regression analysis; The calculation process of the traffic accident risk assessment model is as follows: (1) The input sample data set is D = {a1, a2, ..., a n }, a i is the independent variable including human, vehicle, road and weather factors, and the number of iterations of the weak classifier is T. When t=1,2,…,T, the final strong classifier generated is g(a i ); (2) Randomly sample the training data set for the tth time, and collect j samples in total to obtain the sampling volume Dt including j samples; (3) Use the sampling amount Dt to train the t-th decision tree model. In the process of training the model nodes, a part of the features is randomly selected. Among the selected features, the best feature is selected as the left and right subtrees of the decision tree for division. (4) The value obtained by taking the arithmetic average of the regression results obtained by T weak learners is the final result obtained by the random forest model.

4. The highway traffic accident risk assessment method based on multi-model fusion according to claim 1 is characterized in that: In step six, based on the traffic accident risk index obtained from the output of step five, the risk index is sorted from small to large and divided into four levels according to the quartiles. A risk less than 25% of the quartile is defined as extremely low risk, a risk greater than 25% and less than 50% is defined as low risk, a risk greater than 50% and less than 75% is defined as medium risk, and a risk greater than 75% is defined as high risk.

5. The highway traffic accident risk assessment method based on multi-model fusion according to claim 1 is characterized in that: The step seven is as follows: Calculate the Shapley value to obtain the contribution of a single factor to the accident risk. The calculation formula is as follows: Where φ represents the contribution of the i-th feature, N{i} represents the set of all features, l represents the number of features, S represents the subset of given prediction features, f(S∪{i}) and f(S) represent the model estimation results with or without the i-th feature, respectively; Based on the calculation of Shapley values, the Shapley value of each factor is calculated for road sections with higher risk indices, that is, key road sections in the road network where the accident risk level is divided into "high risk" and "medium risk". This value represents the contribution of each influencing factor to the accident risk level; the Shapley values ​​corresponding to each risk factor are compared to obtain the risk ranking of people, vehicles, roads, and the environment.

Images