Regional traffic accident risk analysis method based on large-scale floating car trajectory data

Abstract

The application discloses a kind of regional traffic accident risk analysis methods based on large-scale floating car trajectory data, comprising the following steps: S1, using the improved Dijkstra space-time trajectory reconstruction algorithm of broken GPS positioning fusion matching, city large-scale floating car trajectory data is matched to city road network space-time line;S2, using the week-hour-land use type multi-dimensional labeling processing method of all-weather-multiseason fusion, finally generate floating car all-weather multi-dimensional complex activity label;S3, text clustering analysis is carried out to all-weather activity label in each traffic partition, and the potential space-time activity characteristics of each traffic partition floating car are excavated;S4, the city road network regional traffic accident risk analysis model combined with floating car is constructed, and the importance of traffic accident risk influencing factor is obtained.The application is efficient, and has important practical significance for realizing the traffic safety evaluation of city road network.

Description

Technical Field

[0001] This invention relates to the field of urban road network traffic safety analysis, and in particular to a method for analyzing traffic accident risks in road network areas based on large-scale floating car trajectory data. Background Technology

[0002] In the field of traffic engineering, regional traffic accident risk analysis of urban road networks is a crucial research area in traffic safety. Urban road networks are complex, open, and massive systems influenced by multiple factors, exhibiting characteristics such as diversity, randomness, dynamism, and nonlinearity. Regional traffic accident risk analysis involves a comprehensive investigation and analysis of the urban road network system to identify risk sources and define potential risk characteristics, ultimately determining the conditions under which risks transform into accidents. Its core involves single-factor risk analysis of unsafe human behavior, unsafe vehicle conditions, unsafe road conditions, and external environmental influences, as well as risk modeling of their interrelationships and couplings. The ultimate goal of regional traffic accident risk analysis of urban road networks is to explore how to control risks within an acceptable level within the overall road network environment, achieving improvements and enhancements to identified risk factors and major hazards through strategic planning, road network layout, and effective monitoring and implementation.

[0003] An analysis of existing research reveals that current urban road network research focuses primarily on data collection and management, with insufficient attention paid to the application of next-generation information technologies such as traffic big data in traffic safety analysis. Identifying regional traffic accident risks through data-driven methods remains a significant challenge. Exploring an effective way to utilize large-scale floating car trajectory data in cities to generate simple, fast-running data processing methods, while also ensuring real-time traffic risk identification, is a topic that requires ongoing research in this field. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a road network regional traffic accident risk analysis method that uses large-scale, refined floating car trajectory data of urban road networks as data support and identifies regional traffic accident risks through a data-driven approach.

[0005] Technical solution: The regional traffic accident risk analysis method of the present invention includes the following steps:

[0006] S1. An improved Dijkstra spatiotemporal trajectory reconstruction algorithm based on broken GPS positioning fusion matching is used to perform spatiotemporal route matching of urban road network for large-scale floating car trajectory data in the city.

[0007] S2 combines urban built environment data and denoised spatiotemporal trajectory data, and adopts a multi-dimensional labeling processing method that integrates all-weather and multi-time period data, including weekday, hour, and land use type, to obtain complex multi-dimensional activity identification labels per unit time, and further generates complex multi-dimensional activity labels for floating cars in all-weather conditions.

[0008] S3 employs an upgraded Max SP latent Dirichlet allocation model that uses a hierarchical weighted evaluation of similarity and confusion as dual indicators to perform text clustering analysis on all-weather activity labels within each traffic zone, thereby mining the potential spatiotemporal activity characteristics of floating cars in each traffic zone.

[0009] S4, based on the DeepLift network, adopts a two-layer coupled graph convolutional neural network interpretable model to construct an urban road network regional traffic accident risk analysis model that combines the potential spatiotemporal activity characteristics of floating cars, road network attributes, traffic facilities, and socio-demographic and economic characteristics. This leads to the importance ranking of traffic accident risk influencing factors and the weights of the intermediate layer of the DeepLift two-layer coupled graph convolutional neural network interpretable model.

[0010] Furthermore, in step S1, the implementation steps for matching the spatiotemporal routes of the urban road network using large-scale floating car trajectory data are as follows:

[0011] S11, Select a floating car trajectory within a specific area and with a non-zero speed for a continuous period of time;

[0012] S12, Delete trajectory data that lacks necessary information;

[0013] S13, match the selected floating car trajectory to the urban road network to obtain the floating car's running trajectory.

[0014] Furthermore, the selected floating car trajectory is matched to the urban road network to obtain the specific implementation steps of the floating car's running trajectory, as follows:

[0015] S131, assume that the floating car does not travel the same road segment repeatedly within a certain time period. If it does travel the same road segment repeatedly, the first road segment record is retained.

[0016] S132, each trajectory point maps to a unique road network segment;

[0017] S133. Based on the principles of road uniqueness, road segment connectivity, and shortest path, the spatiotemporal trajectory of the floating car is obtained.

[0018] Furthermore, in step S2, the specific implementation steps for generating the floating car's all-weather, multi-dimensional, and complex activity tags are as follows:

[0019] S21. Key features of floating car trajectories include weekdays, hours, and land use types. Combining these key features allows for the extraction of complex and multi-dimensional activity labels that describe the vehicle's all-weather, multi-time period activities.

[0020] S22, Statistically analyze the most frequent activity tags in a certain time interval, and define the activity tags of the floating car in the statistical time interval;

[0021] S23, following steps S21 and S22, obtains the complex, multi-dimensional, and 24 / 7 activity labels for each floating car within the city. Further, in step S3, the specific implementation steps for mining the potential spatiotemporal activity characteristics of floating cars in each traffic zone are as follows:

[0022] S31, treat each label as a word of a topic, each activity as a topic of a document, and the trajectory record of each traffic zone as a document of a corpus; according to the definition of activity labels, assign the activity labels in the entire dataset to each traffic zone according to their spatial location;

[0023] S32, the Gibbs sampling method is used to estimate the Max SP to upgrade the latent Dirichlet allocation model; the expected distribution of the latent Dirichlet allocation model is:

[0024]

[0025]

[0026] Where K represents the number of potential vehicle spatiotemporal activity features, and L = {l1, l2, ..., l...} N} represents the thematic set of potential travel features generated from the trajectory record document, where N represents the number of labels for each potential vehicle spatiotemporal activity feature, and M represents the number of traffic zones. Represents the mixed weight of activity feature label l in each potential vehicle activity feature k. β represents the number of labels for activity feature l in each potential vehicle activity feature k. l The model training generates the prior word distribution corresponding to topic l, θ k，m The mixed weights represent the activity feature labels k in traffic partition m. α represents the number of labels for activity feature k in traffic partition m. k This represents the prior topic distribution corresponding to the text generated by LDA through training;

[0027] The Max SP function is used to obtain the optimal number of spatiotemporal activity features for the model:

[0028]

[0029]

[0030] Max SP=max(Similarity(D)-Perplexity(D) / 10x)

[0031] 0≤|Perplexity(D) / 10x|≤1,x∈N

[0032] Where D = {d1, d2, ..., d} M} is a collection of trajectory recording documents, d M Let N represent the trajectory record document of traffic zone M, p(w) represent the probability of generating word w from trajectory record document d, and N represent the probability of generating word w from trajectory record document d. d C represents the number of labels for each potential vehicle spatiotemporal activity feature in document d; C represents the sparse matrix generated by each traffic zone and the number of potential vehicle spatiotemporal activity features; a and b represent any two traffic zones; C ak C represents a sparse matrix generated by traffic zone a and the number of potential vehicle spatiotemporal activity features k. bk The sparse matrix generated by traffic zone b and the number of potential vehicle spatiotemporal activity features k.

[0033] Furthermore, in step S4, a traffic accident risk analysis model for the urban road network area is constructed. The specific implementation steps for obtaining the importance ranking of traffic accident risk influencing factors and the weights of the intermediate layers of the DeepLift two-layer coupled graph convolutional neural network interpretable model are as follows:

[0034] S41, firstly, obtain the features of each traffic zone and represent them with X, including the spatiotemporal travel activity characteristics of floating cars, road network, traffic facilities, weather conditions, geographic information data, and socioeconomic factors, and simultaneously calculate the adjacency matrix.

[0035]

[0036] in, A degree matrix representing the various traffic zones of the city.

[0037] S42, input the adjacency matrix into the DeepLift two-layer coupled graph convolutional neural network interpretable model f(x, A) to obtain the prediction result for each label:

[0038]

[0039] Among them, W 0 W is the weight matrix for the first layer, used to map the feature representations of traffic zones to the corresponding hidden states; 1 The weight matrix for the second layer is used to map the hidden layer representation of traffic zones to the corresponding outputs, and softmax and ReLU are the activation functions.

[0040] S43, in the DeepLift two-layer coupled graph convolutional neural network interpretable model, to explore the impact of each input indicator on regional accident risk, this model provides each input x i Assign a reference value r i , αx i =x i -ri , αx i The value represents the original input x. i The effect of setting it as a reference value; through the mapping function x = h x (x i Mapping to the original input x i The feature gradient 1 indicates that the input accepts its original value x. i The gradient of a feature is 0, indicating that the input accepts the reference value r. i The effect of each input reference value change is αx i The contributions are as follows:

[0041]

[0042] Where z = f(x, A) is the prediction result of the model, αZ = f(x, A) - f(r, A), and n is the number of features of the traffic zone; the DeepLift two-layer coupled graph convolutional neural network interpretable model ranks the effect of the feature attributes of each traffic zone and adds the effects of all feature attributes to approximate the output Z of the original model, so as to obtain the prediction result of each traffic zone and the importance ranking of each factor on the frequency of accidents, as well as visualize the intermediate layer weights of the DeepLift two-layer coupled graph convolutional neural network interpretable model.

[0043] Compared with the prior art, the significant advantages of this invention are as follows:

[0044] 1. This invention combines urban road network area traffic accident risks with floating car trajectory characteristics, and examines the impact of spatiotemporal activity characteristics on traffic safety from a macro perspective, providing methodological support for real-time risk analysis of traffic safety management systems;

[0045] 2. By capturing and analyzing the trajectory characteristics of floating cars, this invention can provide technical support for the precursor characteristics and evolution patterns of traffic accidents in urban road networks, and provide methodological and data support for traffic safety management and control systems. Attached Figure Description

[0046] Figure 1 This is the overall flowchart of the present invention;

[0047] Figure 2 This is a schematic diagram illustrating the all-weather, multi-time period, multi-dimensional, and complex activity tag extraction method of the present invention. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0049] Based on existing traffic safety analysis and research, this invention establishes a data-driven deep learning model using multi-source heterogeneous data such as high-precision vehicle trajectory data, geographic information data, and road network structure to characterize the coupling relationship between variables and traffic accident risks. Combined with an interpretable model, it reveals the influence mechanism of variable factors on traffic accident risks, and further explores the complex spatiotemporal correlation of urban road network accident risks, providing basic theories and methods for fields such as traffic planning and traffic safety design.

[0050] like Figure 1 As shown, a method for regional traffic accident risk analysis based on large-scale floating car trajectory data includes the following steps:

[0051] Step 1 involves using an improved Dijkstra spatiotemporal trajectory reconstruction algorithm based on broken GPS positioning fusion matching (see Table 1) to perform spatiotemporal route matching of urban road network on large-scale floating car trajectory data. This includes the following sub-steps:

[0052] Step 11: Select the floating car trajectory within the study area.

[0053] Step 12: Delete unreasonable floating car trajectories caused by vehicle instrument malfunctions.

[0054] Step 121: Delete trajectory data that is missing essential information, such as location latitude and longitude, time, speed, etc.

[0055] Step 122: If the speed is zero for a continuous period of time, the floating car trajectory is considered unreasonable and is deleted.

[0056] Step 13: Match the selected floating car trajectory to the urban road network to obtain the floating car's running trajectory.

[0057] Step 131: Assume that the floating car does not travel the same road segment repeatedly within each statistical time period (e.g., 8 minutes, 10 minutes, 15 minutes). If it does travel the same road segment repeatedly, then retain the first road segment record.

[0058] Step 132: Each trajectory point maps to a unique road network segment;

[0059] Step 133: Based on the principles of road uniqueness, road segment connectivity, and shortest path, the spatiotemporal trajectory of the floating car is obtained.

[0060] Table 1. Improved Dijkstra spatiotemporal trajectory reconstruction algorithm based on fault GPS positioning fusion matching.

[0061]

[0062]

[0063]

[0064] Step 2: Combining urban built environment data and denoised spatiotemporal trajectory data, a multi-dimensional labeling method based on weekday, hour, and land use type under all-weather, multi-time period fusion calculation is adopted to obtain complex multi-dimensional activity identification labels per unit time. Further, all-weather, multi-dimensional, complex activity labels for floating cars are generated. The specific implementation steps are as follows:

[0065] Step 21: Key features of the floating car trajectory are used to extract complex and multi-dimensional activity labels describing the vehicle's activities throughout the day and across multiple time periods. Time features of the day of the week and hour are extracted using the time information of the floating car trajectory. The location of the floating car trajectory is matched with the geographic information dataset to obtain the land use characteristics of the area where the floating car is located. Finally, these key characteristics—day of the week, hour, and land use type—are combined together.

[0066] Step 22: Analyze the most frequent activity tags within a given time interval, and define the activity tags for the floating car within that statistical period. For example, ... Figure 2 As shown, where Figure 2 See Table 2 for Chinese annotations:

[0067] Table 2 Floating Car Track and Label Annotation Table

[0068]

[0069] Defined as 15 minutes, a floating car n recorded its trajectory from 9:00 to 9:15, which was combined according to the week-hour-land use key features. Since the floating car was most active in the commercial area at 9:00 on Wednesday, the floating car activity within 15 minutes was labeled as Wednesday-9:00-commercial.

[0070] Step 23: Apply the methods of Step 21 and Step 22 to all time periods throughout the day to obtain complex, multi-dimensional activity labels for each floating car within the city.

[0071] Step 3: An upgraded Max SP latent Dirichlet assignment model, which uses a hierarchical weighted evaluation of similarity and perplexity, is employed to perform text clustering analysis on all-weather activity labels within each traffic zone, thereby mining the potential spatiotemporal activity characteristics of floating cars in each traffic zone. The specific implementation steps are as follows:

[0072] Step 31: In this embodiment, activity labels are the basic analytical unit for exploring activity patterns using an upgraded Max SP Latent Dirichlet Allocation Model based on a hierarchical weighted evaluation of similarity and perplexity. Each label is considered a word in a topic, each activity is considered a topic in a document, and the trajectory record of each traffic zone is considered a document in a corpus. Furthermore, according to the definition of activity labels, activity labels in the entire dataset are assigned to each traffic zone according to their spatial location.

[0073] Step 32: Estimate the Max SP of the upgraded latent Dirichlet allocation model using Gibbs sampling. The expected distribution of the latent Dirichlet allocation model is:

[0074]

[0075]

[0076] Where D = {d1, d2, ..., d} M} is a collection of trajectory recording documents, d M Let N represent the trajectory record document of traffic zone M, p(w) represent the probability of generating word w from trajectory record document d, and N represent the probability of generating word w from trajectory record document d. d C represents the number of labels for each potential vehicle spatiotemporal activity feature in document d, C represents the sparse matrix generated by each traffic zone and the number of potential vehicle spatiotemporal activity features, and a and b represent any two traffic zones. ak C represents a sparse matrix generated by traffic zone a and the number of potential vehicle spatiotemporal activity features k. bk This represents a sparse matrix generated by traffic zone b and the number of potential vehicle spatiotemporal activity features k. The Max SP function is used to obtain the optimal number of spatiotemporal activity features for the model.

[0077]

[0078]

[0079] Max SP=max(Similarity(D)-Perplexity(D) / 10x) (5)

[0080] 0≤|Perplexity(D) / 10x|≤1,x∈N

[0081] Where D = {d1, d2, ..., d} M} is a collection of trajectory recording documents, d M Let N represent the trajectory record document of traffic zone M, p(w) represent the probability of generating word w from trajectory record document d, and N represent the probability of generating word w from trajectory record document d. dC represents the number of labels for each potential vehicle spatiotemporal activity feature in document d, C represents the sparse matrix generated by each traffic zone and the number of potential vehicle spatiotemporal activity features, and a and b represent any two traffic zones. ak C represents a sparse matrix generated by traffic zone a and the number of potential vehicle spatiotemporal activity features k. bk The sparse matrix generated by traffic zone b and the number of potential vehicle spatiotemporal activity features k.

[0082] Step 4: Based on the DeepLift network, a two-layer coupled graph convolutional neural network interpretable model is proposed. This model constructs a regional traffic accident risk analysis model for urban road networks, incorporating indicators such as the potential spatiotemporal activity characteristics of floating cars, road network attributes, traffic facilities, and socio-economic characteristics. This leads to the importance ranking of traffic accident risk influencing factors and the weights of the intermediate layers of the DeepLift two-layer coupled graph convolutional neural network interpretable model. The specific implementation steps are as follows:

[0083] Step 41: First, obtain the features of each traffic zone and represent them with X. These features include the spatiotemporal travel activity characteristics of floating cars, road network, traffic facilities, weather conditions, geographic information data, and socioeconomic factors. Simultaneously, calculate the adjacency matrix.

[0084]

[0085] in, This is the degree matrix of the traffic zone map of the city.

[0086] Step 42: Then, the adjacency matrix is ​​input into the DeepLift two-layer coupled graph convolutional neural network interpretable model f(x, A) to obtain the prediction result for each label:

[0087]

[0088] Among them, W 0 W is the weight matrix for the first layer, used to map the feature representations of traffic zones to the corresponding hidden state. 1 is the weight matrix of the second layer, used to map the hidden layer representation of traffic zones to the corresponding outputs, and softmax and ReLU are activation functions.

[0089] Step 43: In the DeepLift two-layer coupled graph convolutional neural network interpretable model, to explore the impact of each input indicator on regional accident risk, this model assigns each input x... i Assign a reference value r i , αx i =x i -r i , αx i The value represents the original input x.i The effect of setting it as a reference value; through the mapping function x = h x (x i Mapping to the original input x i The feature gradient 1 indicates that the input accepts its original value x. i The gradient of a feature is 0, indicating that the input accepts the reference value r. i The effect of each input reference value change is Δx i The contributions are as follows:

[0090]

[0091] Where Z = f(x, A) is the model's prediction result, αZ = f(x, A) - f(r, A), and n is the number of features in the traffic zone. The DeepLift dual-layer coupled graph convolutional neural network interpretable model ranks the effects of the feature attributes of each traffic zone and sums the effects of all feature attributes to approximate the output Z of the original model. This yields the prediction results for each traffic zone and the importance ranking of each factor's impact on accident frequency, as well as the visualization of the intermediate layer weights of the DeepLift dual-layer coupled graph convolutional neural network interpretable model.

Claims

1. A method for regional traffic accident risk analysis based on large-scale floating car trajectory data, characterized in that, Includes the following steps: S1. An improved Dijkstra spatiotemporal trajectory reconstruction algorithm based on broken GPS positioning fusion matching is used to perform spatiotemporal route matching of urban road network for large-scale floating car trajectory data in the city. S2 combines urban built environment data and denoised spatiotemporal trajectory data, and adopts a multi-dimensional labeling processing method that integrates all-weather and multi-time period data, including weekday, hour, and land use type, to obtain complex multi-dimensional activity identification labels per unit time, and further generates complex multi-dimensional activity labels for floating cars in all-weather conditions. S3 employs an upgraded Max SP latent Dirichlet assignment model, which uses a hierarchical weighted evaluation of similarity and perplexity as dual indicators, to perform text clustering analysis on all-weather activity labels within each traffic zone, thereby mining the potential spatiotemporal activity characteristics of floating cars in each traffic zone. The specific implementation steps are as follows: S31, treat each label as a word of a topic, each activity as a topic of a document, and the trajectory record of each traffic zone as a document of a corpus; according to the definition of activity labels, assign the activity labels in the entire dataset to each traffic zone according to their spatial location; S32, the Gibbs sampling method is used to estimate the Max SP to upgrade the latent Dirichlet allocation model; the expected distribution of the latent Dirichlet allocation model is: ， ， in, The number representing the potential spatiotemporal activity characteristics of vehicles This is a collection of themes representing potential travel features generated from trajectory recording documents. The number of labels representing the spatiotemporal activity characteristics of each potential vehicle. Represents the activity characteristics of each potential vehicle Activity Feature Tags Mixed weights, Represents the activity characteristics of each potential vehicle Chinese activity characteristics The number of tags, Representative model training generates topics The corresponding prior word distribution, Represents traffic zones Activity Feature Tags Mixed weights, Represents traffic zones Chinese activity characteristics The number of tags, This represents the prior topic distribution corresponding to the text generated by LDA through training; The Max SP function is used to obtain the optimal number of spatiotemporal activity features for the model: ， ， ， ， in, A collection of trajectory recording documents. Represents traffic zones Tracking document, Represents the number of traffic zones; Representative trajectory record document Generate words The probability, Representative document The number of labels for each potential vehicle's spatiotemporal activity feature. A sparse matrix representing the number of spatiotemporal activity features of each traffic zone and potential vehicles; Represents any two traffic zones; Represents traffic zones Number of potential vehicle spatiotemporal activity characteristics The generated sparse matrix, Represents traffic zones Number of potential vehicle spatiotemporal activity characteristics The generated sparse matrix; S4, based on the DeepLift network, adopts a two-layer coupled graph convolutional neural network interpretable model to construct an urban road network regional traffic accident risk analysis model that combines the potential spatiotemporal activity characteristics of floating cars, road network attributes, traffic facilities, and socio-demographic and economic characteristics. This leads to the importance ranking of traffic accident risk influencing factors and the weights of the intermediate layer of the DeepLift two-layer coupled graph convolutional neural network interpretable model.

2. The regional traffic accident risk analysis method based on large-scale floating car trajectory data according to claim 1, characterized in that, In step S1, the implementation steps for spatiotemporal route matching of urban road network using large-scale floating car trajectory data are as follows: S11, Select a floating car trajectory within a specific area and with a non-zero speed for a continuous period of time; S12, Delete trajectory data that lacks necessary information; S13, match the selected floating car trajectory to the urban road network to obtain the floating car's running trajectory.

3. The regional traffic accident risk analysis method based on large-scale floating car trajectory data according to claim 2, characterized in that, The specific steps for matching the selected floating car trajectory to the urban road network to obtain the floating car's running trajectory are as follows: S131, assume that the floating car does not travel the same road segment repeatedly within a certain time period. If it does travel the same road segment repeatedly, the first road segment record is retained. S132, each trajectory point maps to a unique road network segment; S133. Based on the principles of road uniqueness, road segment connectivity, and shortest path, the spatiotemporal trajectory of the floating car is obtained.

4. The regional traffic accident risk analysis method based on large-scale floating car trajectory data according to claim 1, characterized in that, In step S2, the specific implementation steps for generating the floating car's all-weather, multi-dimensional, and complex activity tags are as follows: S21. Key features of floating car trajectories include weekday, hour, and land use type. Combining these key features allows for the extraction of complex and multi-dimensional activity labels that describe the vehicle's all-day, multi-time period activities. S22, Statistically analyze the most frequent activity tags in a certain time interval, and define the activity tags of the floating car in the statistical time interval; S23. Following steps S21 and S22, obtain the complex, multi-dimensional activity labels for each floating car within the city area, covering all weather conditions and time periods.

5. The regional traffic accident risk analysis method based on large-scale floating car trajectory data according to claim 1, characterized in that, In step S4, the specific implementation steps for obtaining the importance ranking of traffic accident risk influencing factors and the weights of the intermediate layers of the DeepLift two-layer coupled graph convolutional neural network interpretable model are as follows: S41, first obtain the features of each traffic zone and use... This includes the potential spatiotemporal travel characteristics of floating cars, road network attributes, transportation facilities, and socio-demographic and economic characteristics, while also calculating the adjacency matrix. : ， in, A degree matrix representing the various traffic zones of the city; S42, input the adjacency matrix into the DeepLift two-layer coupled graph convolutional neural network interpretable model. In the process, the prediction results for each label are obtained: ， in, This is the weight matrix for the first layer, used to map the feature representation of traffic zones to the corresponding hidden state; This is the weight matrix for the second layer, used to map the hidden layer representation of traffic zones to the corresponding output. and For activation functions; S43, in the DeepLift two-layer coupled graph convolutional neural network interpretable model, to explore the impact of each input indicator on regional accident risk, this model provides each input... Assign a reference value , , The value represents the original input. The effect of setting it as a reference value; through the mapping function Mapping to raw input The feature gradient 1 indicates that the input accepts its original value. The feature gradient of 0 indicates that the input accepts a reference value. The effect of each input reference value change The contributions are as follows: ， in, It is the model's prediction result. , The number of features for each traffic zone; the DeepLift two-layer coupled graph convolutional neural network can interpret the model's ranking of the effects of each traffic zone's feature attributes and sum the effects of all feature attributes to approximate the output of the original model. The prediction results for each traffic zone and the importance ranking of each factor on the frequency of accidents are obtained, as well as the intermediate layer weights of the DeepLift two-layer coupled graph convolutional neural network interpretable model are visualized.

Images