A key road section identification method based on road network dynamic feature sampling and aggregation

Abstract

This invention provides a method for identifying key road segments based on dynamic feature sampling and aggregation of road networks, comprising the following steps: selecting a traffic database X and preprocessing X using the networkx and pandas libraries to obtain a traffic network line graph G containing traffic features; constructing a key road identification model M, then inputting G into M, and obtaining a final list of key roads through a data preprocessing module, a GraphDFSAGE module, and a key list output module. This invention's method has higher accuracy in identifying key roads in unknown areas and uses fewer parameters, enabling more accurate key road recommendations even for traffic networks with limited data, providing better reference for traffic resource planning and utilization.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for identifying key road segments based on dynamic feature sampling and aggregation of road networks. Background Technology

[0002] During traffic control, road conditions often change rapidly due to unforeseen events. Therefore, it is crucial to quickly identify key road segments in the road network. How to reduce the complexity of extracting key road segments and respond more quickly to the requirements for obtaining key road segments in actual traffic control in a certain region, and to locate key road segments of the road network more quickly and accurately, is an urgent problem to be solved.

[0003] The Path2vec model, a key road segment identification model based on dynamic traffic flow direction walking sampling, is fundamentally based on an improvement to graph embedding methods using walking sequences in graph representation learning. In real-world traffic scenarios, traffic flow direction is often non-random, and vehicle path selection exhibits a certain tendency, representing the dynamic transitivity of traffic flow. By improving the traditional walking sequence process, a dynamic traffic flow direction walking path sequence matching historical traffic flow transition probabilities is obtained to extract the dynamic transitivity sequence features of traffic flow. Another core idea of ​​this model is to use a graph sampling aggregation model, leveraging node sampling and message passing mechanisms to acquire regional dynamic and static features, which are then combined with the dynamic transitivity sequence features to identify key road segments. The model's results on both public and private datasets have demonstrated high accuracy in selecting road segments.

[0004] Although relatively good results have been achieved, the following issues still need to be addressed with the Path2vec model:

[0005] (1) In the traditional graph sampling aggregation model, node sampling is achieved by uniformly and randomly sampling a fixed number of neighboring nodes. This lacks specificity for complex traffic networks and is prone to including noisy nodes, which affects the performance of the aggregation model.

[0006] (2) High-dimensional vector node features may introduce redundancy or noise, resulting in low model training efficiency and making them unsuitable for the lightweight requirements in practical applications. Summary of the Invention

[0007] In view of the above-mentioned problems in the existing technology, the technical problem to be solved by the present invention is: how to improve the accuracy and efficiency of locating key road segments.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a key road segment identification method based on dynamic feature sampling and aggregation of road networks, comprising the following steps:

[0009] Select the traffic database X for a certain region:

[0010] S100: The networkx and pandas libraries in Python are used to process X to obtain the traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic feature information.

[0011] After converting the traffic data information table corresponding to X into a matrix, the traffic network line diagram G is obtained.

[0012] The expression for G is as follows:

[0013]

[0014] Where V=( ) is a set of nodes. Let G be the nth node. This represents the total number of nodes in G. Let G represent the set of edges. In G and One edge between them, ; () represents the set of average velocities of each node in G within a specific time T. Let G represent the average speed of a vehicle passing through the nth node; () represents the set of traffic flow at each node in G within a specific time T. Let G represent the traffic flow passing through the nth node; ) represents the set of actual road segment lengths corresponding to nodes. This represents the length of the actual road segment corresponding to the nth node in G;

[0015] S200: Construct a critical road identification model M, which includes a data preprocessing module, a GraphDFSAGE module, and a key list output module;

[0016] The data preprocessing module is used to preprocess G, based on the data in G. Obtain the initial set of multidimensional region feature vectors for all nodes in G. ;

[0017] The GraphDFSAGE module is an improved GraphSAGE layer model that transforms the random sampling method of GraphSAGE into a traffic flow-based, biased sampling method; it is used to... The GraphDFSAGE module takes G as input and outputs the set of embedded features for all nodes. ;

[0018] The key list output module is used to output the key list. As input, the output is the final list of critical roads;

[0019] S300: Select the traffic database of the region Y to be tested, and repeat S100-S200 to obtain the list of key roads in Y.

[0020] Preferably, in step S200, the initial multidimensional region feature vector set of all nodes in G is obtained. The calculation process is as follows:

[0021] S210: Calculation initial feature vector The calculation expression is as follows:

[0022]

[0023] in, Indicates a connection operation. , , , ;

[0024] S211: Calculate the initial feature vectors of all nodes in V, and stack the initial multidimensional region feature vectors of all nodes row by row to form the initial multidimensional region feature vector set. The calculation expression is as follows:

[0025]

[0026] in, Represents a node exist The initial feature vector of the sampling layer, Represents a node exist The initial feature vector of the sampling layer.

[0027] Preferably, the node embedding vector is output in S200. The steps are as follows:

[0028] Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to z;

[0029] S220: Calculate the i-th node neighborhood , All nodes within are The neighboring nodes, The calculation expression is as follows:

[0030]

[0031] in, express and The distance between them express The j-th neighbor node;

[0032] S221: Calculation Sampling dynamic neighborhood The expression is as follows:

[0033]

[0034] in, express and Traffic flow between express and When traffic flow is sorted in descending order, the traffic flow value of the top z is the highest.

[0035] S222: Aggregation in All layers The embedding features are calculated using the following expression:

[0036]

[0037] in, Represents a node neighboring nodes exist Layer embedding features, express exist Layer embedding features Indicates the first Sampling layer, ;

[0038] S223: Aggregation and get The final embedding features The calculation expression is as follows:

[0039]

[0040] in, Represents any neighboring node , express In the Layer embedding features It is a non-linear activation function. It is the first The learnable weight matrix of the layer;

[0041] S224: Repeat S220-S223 to obtain the final embedding features of all nodes, and stack the final embedding features of all nodes row by row to obtain the final embedding feature set of all nodes in V. The calculation expression is as follows:

[0042]

[0043] in, Represents a node The final embedding feature at layer K.

[0044] Preferably, in step S200, the clustering algorithm is the kmeans algorithm, and λ=3 is set.

[0045] Preferably, the step in S200 to obtain the final list of critical roads is as follows:

[0046] S230: Set the clustering algorithm parameter λ, and use the clustering algorithm to... Clustering is performed to obtain λ-class traffic segments;

[0047] S231: Calculate the average traffic demand index of all nodes in each type of traffic segment obtained from S230. The calculation formula is as follows:

[0048]

[0049]

[0050] in, h is Total number of nodes in traffic segments , representing the length of the actual road segment corresponding to the h-th node. This represents the traffic flow passing through the h-th node. This represents the average speed of the vehicle as it passes the h-th node.

[0051] S232: Select the category with the highest average traffic demand index as the key category, sort the traffic demand index of all nodes in the key category in descending order, select the actual road segments corresponding to the traffic demand index of the top W nodes as key road segments, output the W key road segments, and obtain the final list of key roads.

[0052] Compared with the prior art, the present invention has at least the following advantages:

[0053] 1. The technical method mentioned in this invention can extract traffic information from road networks more accurately. Traditional GraphSAGE uses random sampling, where all neighbors are sampled with equal probability. At each sampling layer, for each central node, its neighbors are selected as candidate nodes, and a certain number of nodes are randomly selected from these candidate nodes as sampling nodes. The sampled nodes are then aggregated to represent the central node. GraphDFSAGE, however, transforms GraphSAGE's random sampling method into a traffic flow-based and biased sampling method: at each sampling layer, for each central node, all neighbors are first collected, sorted according to the traffic flow between the central node and its neighbors, and the nodes with the highest traffic flow ranking are selected as candidate sampling nodes. Nodes obtained through this sampling method are then aggregated to represent the central node. By improving the node sampling method, the quality of information aggregation can be improved. Points with high traffic flow weights usually represent stronger interactions with the central node. Aggregating and filtering neighbors avoids noise interference and generates higher-quality node embedding vectors. By selectively filtering neighbors with high traffic flow weights, gradient updates of invalid data are reduced, while model convergence is accelerated. The model can learn key patterns faster, reducing the number of model iterations. Traditional random sampling methods may sample neighboring nodes with low relevance. These low-relevance nodes have low value and affect the generalization of the model. However, the method of this invention filters out low-relevance nodes, allowing the model to focus on strong semantic relevance. When semantic relevance is applied to traffic networks, it is manifested as high-traffic roads having higher relevance to surrounding roads.

[0054] 2. To address the problem of low training efficiency caused by the single feature extraction and lack of targeted neighbor sampling in traditional graph sampling aggregation models, this invention innovatively proposes the GraphDFSAGE dynamic feature sampling aggregation model for road networks. This model achieves the acquisition of multi-dimensional dynamic and static features of regions from the neighborhood of dynamically sampled nodes with bias, and obtains key road segments in the network by clustering the obtained embedding vectors. This invention uses a Python library in the data retrieval process, and the trained results are directly output when the Python library is called.

[0055] 3. From the perspective of dynamic and static characteristics of the traffic network, the accuracy of selected road segments is evaluated based on node network efficiency, network resilience, and dynamic contagion indicators. Compared with traditional indicator evaluation methods and graph representation learning methods, the method proposed in this invention has higher accuracy, significantly outperforming traditional methods in key road segment acquisition for network efficiency indicators. It also outperforms traditional methods in terms of subgraph quantity and node quantity ratio, demonstrating a significant advantage in network efficiency indicators compared to other methods. This indicates that the model can fully extract and aggregate road network node features, having a greater impact on the efficiency of information exchange and the integrity of the road network, thus enhancing its influence on dynamic and static network indicators.

[0056] 4. This invention improves upon the traditional GraphSAGE, which has 3 sampling layers and 5 samples per layer. However, the improved GraphDFSAGE in this invention has only 2 sampling layers and 3 samples per layer, reducing the variables of sampling layers and sample count. Furthermore, the initial node features only include flow rate, speed, and length, resulting in a reduction in dimensionality. This reduces sampling complexity and noise. A line graph mapping method based on traffic networks is used, where road connections represent nodes, and their connections are represented by edges. This integrates the traffic flow, speed, and length attributes of road segments into the node attributes of the traffic network, thereby achieving a lightweight representation of node features.

[0057] 5. While ensuring the improvement of dynamic and static indicators, the number of sampling neighborhood layers and the number of neighboring nodes in the model graph are reduced, which effectively reduces the complexity of sampling, improves the embedding quality of the model, and improves the running speed of the model. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the method architecture of the present invention;

[0059] Figure 2 This is a schematic diagram of the dynamic feature sampling layer structure;

[0060] Figure 3 This is a graph showing the network efficiency changes during the midday peak in PEMS08.

[0061] Figure 4 This is a graph showing the change in the proportion of noon peak hours in PEMS08.

[0062] Figure 5 This is a graph showing the change in the number of sub-plots for the midday peak in PEMS08.

[0063] Figure 6 This is a graph showing the changes in network efficiency during the evening rush hour in a certain region.

[0064] Figure 7 This represents the total network efficiency during the evening peak hours in a certain region.

[0065] Figure 8 This is a graph showing the change in the proportion of peak hours during the evening rush hour in a certain region.

[0066] Figure 9 A map showing the proportion of peak hours in a certain region.

[0067] Figure 10 This is a graph showing the change in the number of sub-maps during the evening rush hour in a certain region.

[0068] Figure 11 This is a sub-map showing the total number of passengers during the evening rush hour in a certain region.

[0069] Figure 12 The trend of evening rush hour congestion points in a certain region;

[0070] Figure 13 This is a road network location map of the top 15 key road segments in the region. Detailed Implementation

[0071] The present invention will now be described in further detail.

[0072] Select the traffic database X for a certain region:

[0073] S100: The networkx and pandas libraries in Python are used to process X to obtain the traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic feature information.

[0074] After converting the traffic data information table corresponding to X into a matrix, the traffic network line diagram G is obtained; the networkx library and pandas library are existing technologies;

[0075] The expression for G is as follows:

[0076]

[0077] Where V=( ) is a set of nodes. Let G be the nth node. This represents the total number of nodes in G. Let G represent the set of edges. In G and One edge between them, ; () represents the set of average velocities of each node in G within a specific time T. Let G represent the average speed of a vehicle passing through the nth node; () represents the set of traffic flow at each node in G within a specific time T. Let G represent the traffic flow passing through the nth node; ) represents the set of actual road segment lengths corresponding to nodes. This represents the length of the actual road segment corresponding to the nth node in G;

[0078] S200: Construct a critical road identification model M, which includes a data preprocessing module, a GraphDFSAGE module, and a key list output module;

[0079] The data preprocessing module is used to preprocess G, based on the data in G. Obtain the initial set of multidimensional region feature vectors for all nodes in G. ;

[0080] The GraphDFSAGE module is an improved GraphSAGE layer model that transforms the random sampling method of GraphSAGE into a traffic flow-based, biased sampling method; it is used to... The GraphDFSAGE module takes G as input and outputs the set of embedded features for all nodes. The GraphDFSAGE module, also known as the dynamic feature sampling layer, primarily acquires the dynamic neighborhood and the domain feature aggregation. In acquiring the dynamic neighborhood, based on the number of sampling layers K and the number of samples n, the top n neighbors by traffic in each layer of the central node are selected as the dynamic neighborhood. In each layer of domain feature aggregation, for the central node, according to the selected aggregation function, the features of the nodes within that layer's dynamic neighborhood are aggregated and connected to the node's embedding features in the previous layer, and then linearly represented by an activation function to obtain the final embedding vector.

[0081] The key list output module is used to output the key list. As input, the output is the final list of critical roads;

[0082] In step S200, the initial multidimensional region feature vector set of all nodes in G is obtained. The calculation process is as follows:

[0083] S210: Calculation initial feature vector The calculation expression is as follows:

[0084]

[0085] in, Indicates a connection operation. , , , ;

[0086] S211: Calculate the initial feature vectors of all nodes in V, and stack the initial multidimensional region feature vectors of all nodes row by row to form the initial multidimensional region feature vector set. The calculation expression is as follows:

[0087]

[0088] in, Represents a node exist The initial feature vector of the sampling layer, Represents a node exist The initial feature vector of the sampling layer.

[0089] The node embedding vector is output in S200. The steps are as follows:

[0090] Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to z;

[0091] S220: Calculate the i-th node neighborhood , All nodes within are The neighboring nodes, The calculation expression is as follows:

[0092]

[0093] in, express and The distance between them and There are multiple paths between them, and the shortest path is... and The distance between them; express The j-th neighbor node;

[0094] S221: Calculation Sampling dynamic neighborhood The expression is as follows:

[0095]

[0096] in, express and Traffic flow between express and When traffic flow is sorted in descending order, the traffic flow value of the top z is the highest.

[0097] S222: Aggregation in All layers The embedding features are calculated using the following expression:

[0098]

[0099] in, Represents a node neighboring nodes exist Layer embedding features express exist Layer embedding features Indicates the first Sampling layer, ; yes The initial value of a neighboring node is equal to its initial value when it is a node itself.

[0100] S223: Aggregation and get The final embedding features The calculation expression is as follows:

[0101]

[0102] in, Represents any neighboring node , express In the Layer embedding features It is a non-linear activation function. It is the first The learnable weight matrix of the layer; when hour, Take the calculated value from step S310;

[0103] S224: Repeat S220-S223 to obtain the final embedding features of all nodes, and stack the final embedding features of all nodes row by row to obtain the final embedding feature set of all nodes in V. The calculation expression is as follows:

[0104]

[0105] in, Represents a node The final embedding feature at layer K.

[0106] In S200, the clustering algorithm is the kmeans algorithm, and λ=3 is set.

[0107] The steps for obtaining the final list of critical roads in S200 are as follows:

[0108] S230: Set the clustering algorithm parameter λ, and use the clustering algorithm to... Clustering is performed to obtain λ-class traffic segments;

[0109] S231: Calculate the average traffic demand index of all nodes in each type of traffic segment obtained from S230. The calculation formula is as follows:

[0110]

[0111]

[0112] in, h is Total number of nodes in traffic segments , representing the length of the actual road segment corresponding to the h-th node. This represents the traffic flow passing through the h-th node. This represents the average speed of the vehicle as it passes the h-th node.

[0113] S232: Select the category with the highest average traffic demand index as the key category, sort the traffic demand index of all nodes in the key category in descending order, select the actual road segments corresponding to the traffic demand index of the top W nodes as key road segments, output the W key road segments, and obtain the final list of key roads.

[0114] S300: Select the traffic database of the region Y to be tested, and repeat S100-S200 to obtain the list of key roads in Y.

[0115] Experiments and Analysis:

[0116] 1. Dataset Description

[0117] This invention uses the PEMS08 public dataset and a private dataset of a certain region to process data to obtain traffic flow data, average road speed data, road length data, and other traffic features required for traffic network diagrams for each external road detection device during the morning peak (8:00-9:00), noon peak (12:00-13:00), and evening peak (17:00-18:00).

[0118] PEMS08 dataset: PEMS08 is California Highway Network PeMS Traffic Flow Dataset PEMS08 is a dataset generated by 170 detectors collecting data every 5 minutes for 62 days. This dataset contains traffic flow, road connectivity, and road segment lengths for the California highway network. After constructing the traffic network map, K was set to 30 as an appropriate number of critical road segments.

[0119] A dataset from a specific region: The same experimental validation was conducted on a private dataset from a specific region. This dataset contains raw traffic data acquired by external road detection devices in the surrounding road network from July 1st to July 31st, 2024. After data cleaning and processing, traffic flow data, average road speed data, and road length data for each external road detection device during morning, noon, and evening peak hours were obtained. K was set to 15 as an appropriate number of critical road segments.

[0120] 2. Comparison Model

[0121] This invention compares the key road segment selection methods proposed in this paper: one based on a dynamic feature sampling and aggregation model of the road network, another based on a centrality index, and a third based on graph representation learning. The accuracy of the selected key road segments is evaluated using two indicators related to overall network efficiency and network resilience: the number of subgraphs remaining after node deletion and the proportion of nodes in the largest connected subgraph after node deletion. A lower overall network efficiency index indicates a greater impact of the node on overall network efficiency. A larger number of subgraphs remaining after node deletion indicates a more dispersed network segmentation by the node, suggesting the node is more important. A lower proportion of nodes in the largest connected subgraph after node deletion indicates a greater impact of the node on the network's global throughput, suggesting the node is more important. A higher network infectiousness index indicates more congested road segments in the network, suggesting a greater influence of the node on the spread of congestion across the entire network.

[0122] Nine models were selected as comparative models for the comparative experiment. These models are highly representative, including key road segment identification models based on flow or centrality selection, key road segment identification models based on random walks in graph representation learning, and key road segment identification models based on graph neural network methods in graph representation learning. A brief introduction to these baseline models follows:

[0123] (1) Onlyflow: This method uses only traffic flow as an indicator to determine the importance of traffic segments and selects key segments based on their traffic flow ranking. Traffic flow is an important data point in traffic flow, and traffic segments with high traffic flow often have a greater impact on the traffic network. This method selects the top K traffic segments in the network based on their traffic flow ranking.

[0124] (2) OnlyDegree: This method uses only the degree of a node as an indicator to judge the importance of traffic segments, and selects key segments based on the degree ranking. The degree of a node is one of the indicators of the centrality of a network node. The degree reflects the number of connections with nodes in the network. The higher the degree of a node, the more edges are connected to that node. When a node fails, the number of nodes that may be affected will be greater. This method selects the top K traffic segments in the network in terms of degree.

[0125] (3) OnlyBetweenness: This method uses only the betweenness number of nodes as an indicator to judge the importance of traffic segments, selecting key segments based on the ranking of betweenness numbers. The betweenness number of a node is one of the indicators of network node centrality. Nodes with high betweenness numbers are usually bridges in the network, connecting different communities or groups, such as information dissemination intermediaries in social networks. This method selects the top K traffic segments in terms of the number of betweenness numbers in the network.

[0126] (4) OnlyCloseness: This method uses node proximity as the sole indicator of traffic segment importance, selecting key segments based on their proximity ranking. Node proximity is one of the indicators of network node centrality, representing the reciprocal of the average shortest distance from a node to all other nodes, reflecting the efficiency of information propagation. Nodes with high proximity can quickly reach other nodes. This method selects the top K traffic segments in the network based on their proximity ranking.

[0127] (5) Deepwalk: It is a purely random walk model. The traffic network line map is input into Deepwalk, and the walk path sequence is obtained by randomly walking to each node. The embedding vector is learned through the word vector model, and important road segments are obtained through clustering. The top K traffic segments are selected from them.

[0128] (6) Node2vec: a biased random walk model. The traffic network line graph is input into Node2vec. A biased random walk is performed on each node to obtain the walk path sequence. The word vector model is used to learn the embedding vector. Important road segments are obtained through clustering. The top K traffic segments are selected from them.

[0129] (7) Path2vec: The dynamic traffic flow direction wandering model proposed in Chapter 3 obtains the dynamic traffic flow direction wandering path sequence that matches the historical traffic flow transfer probability. Then, the Skip-gram model is used to learn the low-dimensional vector representation of the nodes, capture the dynamic transmission sequence characteristics of traffic flow embedded vectors, cluster the vectors to obtain important road segments, and select the top K traffic road segments from them.

[0130] (8) OnlyGraphSAGE: A graph sampling aggregation model that integrates dynamic attributes such as flow and speed features of nodes with static attributes such as node degree into a feature matrix. By sampling a certain number of neighboring nodes and aggregating them, the embedding vectors of the nodes are obtained. After clustering, important road segments are obtained, and the top K traffic segments are selected from them.

[0131] (9) GraphDFSAGE: The road network dynamic feature sampling and aggregation model proposed in this invention integrates the dynamic transitivity sequence features of dynamic traffic flow with biased multidimensional regional features, and samples more important nodes through dynamic graph sampling to obtain embedding vectors. Key road segments are then obtained through clustering. The top K traffic segments are selected from these.

[0132] 3. Model Training Settings

[0133] The parameters of each model in the experiment are as follows: K is set to 30 in the PEMS08 dataset and 15 in a certain region dataset, where K represents the number of key road segments selected. Deepwalk, Node2vec, and Path2vec embedding vectors have a length of 128, a walk sequence length of 80, and 10 walks. GraphSAGE embedding vectors have a length of 128, a model sampling layer of 3, and a sampling neighbor count of 5. The infection rate is set to 0.8, the recovery rate to 0.2, the time step to 1, and the total time T to 15, simulating a continuous propagation scenario with a high infection rate and a low recovery rate. The proposed road network dynamic feature sampling aggregation model, GraphDFSAGE, has a sampling layer of 2 and a sampling node count of 3 in the experiment, which is lower than that of GraphSAGE, resulting in fewer layers and fewer nodes.

[0134] 4. Experimental Results

[0135] During the morning peak hours of 8:00-9:00, the afternoon peak hours of 12:00-13:00, and the evening peak hours of 17:00-18:00, key road segments were identified using traditional flow or centrality indicator models, traditional graph representation learning models, and the aggregation model method proposed in this invention. The network efficiency after key road segment failure, the number of subgraphs remaining in the network after node deletion, the ratio of the number of nodes in the largest connected component of the network after node deletion to the number of nodes in the original graph, and the number of infected nodes in the network in the simulated infectious disease model were calculated. The average value of these indicators in each result and the total value of these indicators after 100 experiments were obtained. The traffic network line graph constructed based on the PEMS08 dataset contains a total of 277 road segment nodes. In actual traffic control and planning, due to cost considerations, too many nodes are not selected as key nodes. Therefore, K is set to 30 in the PEMS08 dataset model as a suitable number of key road segments, and the same experiment was conducted on a private dataset in a certain region. The experimental results on the public dataset will be briefly introduced first, followed by a detailed introduction to the experimental results on the private dataset.

[0136] (1) Experimental results using the public dataset PEMS08

[0137] Figures 3-5This paper demonstrates the trends of network efficiency metrics, subgraph node ratio, and network subgraph quantity metrics of the dynamic feature sampling aggregation model on the public dataset PEMS08 during the morning peak hours, after 100 experiments. From the perspectives of network efficiency metrics and subgraph node ratio, the network efficiency of the dynamic feature sampling aggregation model is significantly lower than that of traditional metric methods and traditional graph representation learning methods. This indicates that the key road segments obtained by the dynamic feature sampling aggregation model have a greater impact on the overall network propagation efficiency and global network throughput. Overall, the dynamic feature sampling aggregation model proposed in this invention significantly outperforms other comparative methods in all evaluation metrics on the public dataset, demonstrating that the model can effectively solve the key road segment classification problem in standard application scenarios.

[0138] Table 1 Performance Indicators of PEMS08 Models

[0139]

[0140] Table 1 shows the performance of each model on the PEMS08 dataset in terms of validation metrics. To make the results more intuitive, the value of each metric is the sum of the results after 100 simulations. The GraphDFSAGE model proposed in this invention outperforms all other baseline methods in both dynamic and static metrics for key road segments. Compared with the dynamic traffic flow walk model in Chapter 3, it shows an improvement in the metric of the ratio of the number of nodes in the largest connected component of the network to the number of nodes in the original graph after node deletion, making it the best among the baseline methods.

[0141] (2) Experimental results of a certain region using a private dataset

[0142] The GraphDFSAGE model proposed in this invention has further improved the performance of indicators on a certain regional dataset. Figure 6 , Figure 7 The data illustrates network efficiency metrics for a certain region during the evening rush hour (5:00 PM - 6:00 PM). The proposed GraphDFSAGE model exhibits the lowest network efficiency after key road segments fail, followed by the GraphSAGE and Path2vec models. Methods based on flow and centrality selection also show poor network efficiency after road segment failures. Compared to flow-based methods, GraphDFSAGE significantly improves the impact on network efficiency metrics by 20%, outperforming the other eight methods. This demonstrates that the key road segments identified by the proposed method have a greater global impact on the efficiency of network information exchange.

[0143] Figures 8-11The data demonstrates the resilience indicators of a certain area during the evening rush hour (17:00-18:00). Compared to the model proposed in Chapter 3, the GraphDFSAGE model shows a significant improvement in the node ratio indicator, while further reducing the node ratio. In terms of the number of subgraphs, its performance is close to that of the method proposed in Chapter 3. The node ratio indicator of the GraphDFSAGE model proposed in this invention is superior to the other eight methods, indicating that the proposed method simultaneously considers local connectivity patterns and global path dependencies, providing a more comprehensive evaluation of the function of nodes in complex traffic networks. Furthermore, the proposed model obtains key road segments with higher network segmentation capabilities, but the network integrity cannot be maintained after the deletion of key road segments, resulting in greater disruption to network connectivity. It also demonstrates stronger ability to disrupt network integrity after node failure.

[0144] Figure 12 This paper illustrates the changing trends of congested nodes in a traffic congestion contagion model during the evening rush hour. The blue area represents the GraphDFSAGE model proposed in this invention. The trends show that the node with the highest degree experiences the first congestion outbreak in the contagion model, with earlier congestion times and a higher number of congested nodes. The node with the highest degree also has the most connected edges, thus conforming to the logic of rapid infection and a high number of infections in the traffic congestion contagion contagion model. The peak infection time of the GraphDFSAGE model is later than that of the node with the highest degree, and the number of infected nodes at the peak is slightly lower than that of the node with the highest degree, but still significantly higher than the nodes selected by other baseline methods. The trend of the number of infected nodes shows that although the nodes selected in this invention are not the nodes with the highest degree, they are closely related to the nodes with the highest degree. Once infected, the nodes in this invention quickly spread to nodes with higher degree, and their ability to continuously infect the entire road network is also superior to that of nodes with higher degree. This demonstrates that the method of this invention successfully selects traffic segments with strong dynamic contagion capabilities.

[0145] Figure 13The graph shows the locations of the top 15 critical road segments obtained by the GraphDFSAGE model proposed in this invention during the evening rush hour in a certain region, compared to the locations of the top 15 critical road segments obtained through traffic flow analysis. These segments are depicted in red on the graph. In the actual road network, critical road segments may be distributed along main roads connecting different areas, bridges, tunnels, or connecting roads around traffic hubs. The results show that traffic flow-based critical road segments only include high-volume main roads and road segments with less impact on the actual road network, such as Blue 11 and Blue 26. However, the critical road segments obtained by the GraphDFSAGE method proposed in this invention not only include main roads but also important weaving areas such as Red 9 and Blue 2, and drop-off platform areas such as Red 4 and Red 5. This better matches frequently congested traffic segments and aligns more closely with empirically observed road segment locations. Traffic flow-based methods, on the other hand, lack the ability to identify complex weaving areas and important road network connections, primarily only identifying high-volume main roads.

[0146] Table 2 Performance Indicators of Various Models in a Certain Region

[0147]

[0148] Table 2 shows the model performance indicators for a certain region during three peak periods. It can be seen that the GraphDFSAGE model proposed in this invention further improves network efficiency, network resilience, and network dynamic contagion compared to the model proposed in Chapter 3, especially in terms of network efficiency. Consistent with the sampling logic of the dynamic feature sampling model, the dynamic feature sampling process simulates the transmission trend of traffic flow within a region, indirectly reflecting the process of information exchange in the network, thus leading to a significant improvement in network efficiency indicators. Through data analysis, the model proposed in this invention fully captures the dynamic transmission sequence features of traffic flow, possessing the highest network contagion index. The model also emphasizes the aggregation of dynamic and static regional features of the road network. Compared with traditional graph sampling aggregation methods, it reduces the complexity of sampling aggregation while increasing the influence on dynamic and static network indicators, and the selected key road segments are more accurate.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying key road segments based on dynamic feature sampling and aggregation of road networks, characterized in that, Includes the following steps: Select the traffic database X for a certain region; S100: The networkx and pandas libraries in Python are used to process X to obtain the traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic feature information. After converting the traffic data information table corresponding to X into a matrix, the traffic network line diagram G is obtained. The expression for G is as follows: Where V=( ) is a set of nodes. Let G be the nth node. This represents the total number of nodes in G. Let G represent the set of edges. In G and One edge between them, ; () represents the set of average velocities of each node in G within time T. Let G represent the average speed of a vehicle passing through the nth node; () represents the set of traffic flows at each node in G within time T. Let G represent the traffic flow passing through the nth node; This represents the set of actual road segment lengths corresponding to each node. This represents the length of the actual road segment corresponding to the nth node in G; S200: Construct a critical road identification model M, which includes a data preprocessing module, a GraphDFSAGE module, and a key list output module; The data preprocessing module is used to preprocess G, based on the data in G. Obtain the initial set of multidimensional region feature vectors for all nodes in G. ; The GraphDFSAGE module is an improved GraphSAGE layer model that transforms the random sampling method of GraphSAGE into a traffic flow-based, biased sampling method; it is used to... The GraphDFSAGE module takes G as input and outputs the set of embedded features for all nodes. ; The key list output module is used to output the key list. As input, the output is the final list of critical roads; The steps for obtaining the final list of critical roads in S200 are as follows: S230: Set the clustering algorithm parameter λ, and use the clustering algorithm to... Clustering is performed to obtain λ-class traffic segments; S231: Calculate the average traffic demand index of all nodes in each type of traffic segment obtained in S230. The calculation formula is as follows: in, h is The total number of nodes in a traffic segment This indicates the length of the actual road segment corresponding to the h-th node. This represents the traffic flow passing through the h-th node. This represents the average speed of the vehicle as it passes the h-th node. S232: Select the category with the highest average traffic demand index as the key category, sort the traffic demand index of all nodes in the key category in descending order, select the actual road segments corresponding to the traffic demand index of the top W nodes as key road segments, output the W key road segments, and obtain the final list of key roads. S300: Select the traffic database of the region Y to be tested, and repeat S100-S200 to obtain the list of key roads in Y.

2. The method for identifying key road segments based on dynamic feature sampling and aggregation of road networks as described in claim 1, characterized in that: In step S200, the initial multidimensional region feature vector set of all nodes in G is obtained. The calculation process is as follows: S210: Calculation initial feature vector The calculation expression is as follows: in, Indicates a connection operation. , , , ; S211: Calculate the initial feature vectors of all nodes in V, and stack the initial multidimensional region feature vectors of all nodes row by row to form the initial multidimensional region feature vector set. The calculation expression is as follows: in, Represents a node exist The initial feature vector of the sampling layer, Represents a node exist The initial feature vector of the sampling layer.

3. The method for identifying key road segments based on dynamic feature sampling and aggregation of road networks as described in claim 2, characterized in that: The output of S200 is the set of embedded features of all nodes. The steps are as follows: Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to z; S220: Calculate the i-th node neighborhood , All nodes within are The neighboring nodes, The calculation expression is as follows: in, express and The distance between them express The j-th neighbor node; S221: Calculation Sampling dynamic neighborhood The expression is as follows: in, express and Traffic flow between express and When traffic flow is sorted in descending order, the traffic flow value of the top z is the highest. S222: Aggregation in All layers The embedding features are calculated using the following expression: in, Represents a node neighboring nodes exist Layer embedding features, express exist Layer embedding features Indicates the first Sampling layer, ; S223: Aggregation and get The final embedding features The calculation expression is as follows: in, Represents any neighbor node , express In the Layer embedding features, It is a non-linear activation function. It is the first The learnable weight matrix of the layer; S224: Repeat S220-S223 to obtain the final embedding features of all nodes, and stack the final embedding features of all nodes row by row to obtain the final embedding feature set of all nodes in V. The calculation expression is as follows: in, Represents a node The final embedding feature at layer K.

4. The method for identifying key road segments based on dynamic feature sampling and aggregation of road networks as described in claim 3, characterized in that: In S200, the clustering algorithm is the kmeans algorithm, and λ=3 is set.

Images