Vehicle overlapping community detection method based on self-supervised contrastive graph autoencoder

Abstract

This invention discloses a vehicle overlapping community detection method based on a self-supervised contrastive graph autoencoder, belonging to the fields of vehicle networking and graph neural network deep learning. The method includes: acquiring vehicle trajectory data; constructing a vehicle social network graph based on the vehicle trajectory data; inputting two enhanced views into a graph attention encoder with shared parameters to obtain corresponding node embedding representations; training and optimizing all parameters of the graph attention encoder and decoder using a combination of self-supervised contrastive loss, structural reconstruction loss, modularity enhancement loss, and semi-supervised cross-entropy loss; inputting test node features into the trained graph attention encoder, obtaining a continuous membership matrix through continuous membership mapping, traversing the continuous membership matrix, and outputting the overlapping membership identity and membership strength of each vehicle. This invention can still accurately and continuously characterize the overlapping community structure of vehicle social networks even in environments with extremely sparse supervision signals.

Description

Technical Field

[0001] This invention relates to the fields of vehicle networking and graph neural network deep learning, and particularly to a method for detecting overlapping vehicle communities based on a self-supervised contrastive graph autoencoder. Background Technology

[0002] With the rapid evolution of intelligent transportation systems, vehicles are interconnected through onboard communication units, forming vehicular social networks with social attributes. In vehicular social networks (VSNs), community detection technology is fundamental for achieving efficient content distribution, traffic flow optimization, and privacy protection. Due to the multiplicity of vehicle nodes in terms of spatial trajectory and social intent, overlapping communities are prevalent in VSNs.

[0003] In recent years, graph neural networks (GNNs) based on deep learning have shown significant advantages in processing high-dimensional network data. Existing techniques have improved the ability to extract social features of vehicle nodes by introducing attention mechanisms. However, in practical applications, the above techniques still face the following pressing problems:

[0004] (1) Model collapse due to lack of supervision signals: Existing semi-supervised methods rely heavily on manually labeled community tags, but in real VSNs environments, it is extremely difficult to obtain real labels for vehicle communities. When available tags are extremely sparse, the model's feature representation collapses due to a lack of sufficient supervision signals, making it unable to uncover the hidden topological structure.

[0005] (2) Ambiguity in overlapping boundary identification: Existing technologies mostly use hard thresholds to determine node affiliation. This discrete division method will lead to gradient loss during backpropagation, and cannot accurately characterize the change in the affiliation strength of vehicles at the community boundary, making it difficult to break through the overlap accuracy index such as Jaccard coefficient.

[0006] (3) Competition and oscillation in multi-task training: When additional auxiliary tasks are introduced, gradient conflicts are likely to occur between different loss functions, making it difficult for the model to converge stably when dealing with large-scale vehicle networks. Summary of the Invention

[0007] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a vehicle overlap community detection method based on a self-supervised comparison map autoencoder.

[0008] Technical solution: The vehicle overlap community detection method based on self-supervised comparison map autoencoder of the present invention includes the following steps:

[0009] Step 1: Obtain vehicle trajectory data, construct a vehicle social network graph based on the vehicle trajectory data, generate an attribute matrix based on the vehicle multimodal spatiotemporal data, and construct two enhanced views that are complementary in topology and node attributes through the original adjacency matrix and attribute matrix of the vehicle social network graph.

[0010] Step 2: Input the two augmented views into the graph attention encoder with shared parameters to obtain the corresponding node embedding representations, which are denoted as the main view feature vector and the perturbation augmented view feature, respectively.

[0011] Step 3: Train and optimize all parameters of the graph attention encoder and decoder by combining self-supervised contrastive loss, structural reconstruction loss, modularity enhancement loss and semi-supervised cross-entropy loss;

[0012] Step 4: Input the test node features into the trained graph attention encoder, obtain the continuous membership matrix through continuous membership mapping, traverse the continuous membership matrix, and output the overlapping membership identity and membership strength of each vehicle.

[0013] Further, step 1 includes:

[0014] Construct a vehicle social network graph that includes node features and topological structure, using vehicles as nodes and the relationships between vehicles as edges. ,in For the set of vehicle nodes, For the vehicle edge set, the network graph The original adjacency matrix is ​​denoted as , Represents a node With nodes There must be a communication or social relationship between them, otherwise , This represents the number of vehicle nodes.

[0015] For the original adjacency matrix According to probability Randomly delete some edges to obtain the mask matrix. The elements in this matrix The perturbated adjacency matrix is ​​denoted as The corresponding first enhanced view is denoted as ;

[0016] For attribute matrix According to probability Randomly set the eigenvalues ​​to zero or add Gaussian noise to obtain the perturbed attribute matrix. The corresponding second enhanced view is denoted as .

[0017] Furthermore, in step 2, for any enhanced view, the node is calculated. Its neighboring nodes Attention coefficient between and normalized weights The calculation formulas are as follows:

[0018] ,

[0019] ,

[0020] In the formula, It is a linear rectified activation function with leakage, used to introduce nonlinearity and preserve gradient information in the negative region; For attention weight vectors, It is a linear transformation matrix. For nodes The input feature vector; Represents a node The set of neighboring nodes;

[0021] The features of neighboring nodes are weighted and aggregated according to normalized weights to generate the embedded representation of the current node.

[0022] Furthermore, in step 3, the total loss function for training optimization is:

[0023] ,

[0024] in, The expression for the self-supervised comparison loss is:

[0025] ,

[0026] In the formula, This is the cosine similarity calculation function. Indicates the total number of vehicles; Main view 1 eigenvector For perturbation enhancement view number 1 eigenvector;

[0027] in The modularity enhancement loss is calculated by interpolating the main view features with a predefined modularity matrix, and is expressed as follows:

[0028] ,

[0029] In the formula, Represents the trace operation of a matrix. For a predefined modularity matrix, Represents the feature vector of the main view;

[0030] in The structural reconstruction loss is expressed as:

[0031] ,

[0032] In the formula, This represents the Sigmoid activation function, used to map the input to the range (0,1); Denotes the Frobenius norm;

[0033] For semi-supervised cross-entropy loss, These are the weighting coefficients;

[0034] in and All depend on the number of training rounds The dynamically changing scheduling factor is expressed as:

[0035] ,

[0036] ,

[0037] In the formula, As the initial comparison weights, The attenuation rate, This is the modularity weighting factor. This represents the maximum number of training rounds.

[0038] Furthermore, in step 4, the continuous membership matrix By analyzing the feature vector of the main view The mapping yields the following expression:

[0039] ,

[0040] in, Indicates the total number of communities. Indicates vehicle Belonging to the community The continuous probability distribution value, with a range of values ​​as follows: , Indicates the first The weight vector of each community is used to embed nodes. The probability score mapped to that community; Indicates the first The bias term for each community is used to adjust the offset of the mapping.

[0041] Furthermore, in step 4, a membership determination threshold is set. Traverse the continuous membership matrix ,like Then determine the node Belongs to the community Finally, the overlapping membership identities and membership strength of each vehicle node are output.

[0042] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0043] 1. Significantly reduces reliance on manually labeled data, exhibiting extremely strong robustness against low-label data:

[0044] This invention introduces a self-supervised dual-view contrastive learning mechanism, enabling the model to spontaneously mine structural features from unlabeled vehicle social topology. Experimental data shows that in an extremely sparse scenario where the prior information ratio is only 2%, the overlap-normalized mutual information index of this invention reaches 0.4696, far exceeding the performance of similar semi-supervised methods under the same conditions, greatly solving the technical pain point of difficulty in obtaining real labels in vehicle social networks.

[0045] 2. Improved the accuracy of overlapping community boundary identification and solved the gradient distortion problem caused by hard partitioning:

[0046] This invention constructs a continuous membership mapping model based on Sigmoid activation, achieving a continuous mapping from high-dimensional embeddings to community distributions, effectively preserving the gradient information of boundary nodes. On the fb_1912 dataset with a 2% label rate, the Jaccard Index of this invention reaches 0.7467, a 3.9% improvement compared to the base model's 0.7184, enabling more accurate reconstruction of the complex overlapping social attributes of vehicle nodes.

[0047] 3. Enhanced convergence stability and model robustness in multi-task learning:

[0048] To address the potential gradient conflicts that may arise during joint training of contrastive loss, reconstruction loss, and modularity loss, this invention designs an adaptive dynamic loss weight scheduler. By dynamically decreasing self-supervised weights and increasing structured constraint weights with each training epoch, the scheduler guides the model smoothly from early global robust representation learning to later local fine-grained community clustering. This mechanism effectively avoids oscillations in the multi-objective optimization process, resulting in a smoother performance curve.

[0049] 4. It achieves efficient fusion representation of topology and node attributes.

[0050] This invention continues and optimizes the graph attention mechanism. By assigning differentiated attention weights to different neighbor nodes and coordinating self-supervised contrast enhancement, it can integrate vehicle driving intention, occupation, social preferences and other attribute information at a deeper level, and the generated node representation is more discriminative. Attached Figure Description

[0051] Figure 1 This is a flowchart of the present invention;

[0052] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0053] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0054] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0055] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0056] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0057] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0058] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0059] This invention addresses the performance collapse at extremely low labeling rates caused by over-reliance on manual annotation in existing vehicle overlapping community detection technologies, as well as the technical bottlenecks such as blurred overlapping boundary recognition and loss of gradient information due to hard thresholding. It provides a vehicle overlapping community detection method based on a self-supervised contrastive graph autoencoder (AC-CDGAAE). This invention aims to achieve high-precision and continuous characterization of the overlapping community structure of vehicle social networks even in environments with extremely sparse supervision signals.

[0060] Combination Figure 1 and Figure 2 As shown in this embodiment, the vehicle overlap community detection method based on a self-supervised comparison map autoencoder includes the following steps:

[0061] Step 1: Obtain vehicle trajectory data, construct a vehicle social network graph based on the vehicle trajectory data, generate an attribute matrix based on the vehicle multimodal spatiotemporal data, and construct two enhanced views that are complementary in topology and node attributes using the original adjacency matrix and attribute matrix of the vehicle social network graph.

[0062] Further, step 1 includes:

[0063] Using vehicles as nodes and the relationships between vehicles as edges, where relationships include, but are not limited to, one or more of the following: communication relationships, spatial co-occurrence relationships, trajectory similarity relationships, or social attribute relationships, a vehicle social network graph containing node features and topology is constructed. ,in For the set of vehicle nodes, For the vehicle edge set, the network graph The original adjacency matrix is ​​denoted as , Represents a node With nodes There must be a communication or social relationship between them, otherwise , This represents the number of vehicle nodes.

[0064] For the original adjacency matrix According to probability Randomly delete some edges to obtain the mask matrix. The elements in this matrix The perturbated adjacency matrix is ​​denoted as The corresponding first enhanced view is denoted as ;

[0065] For attribute matrix According to probability Randomly set the eigenvalues ​​to zero or add Gaussian noise to obtain the perturbed attribute matrix. The corresponding second enhanced view is denoted as .

[0066] Specifically, attribute matrix It includes known information such as the vehicle's speed, location, driving behavior characteristics, and the driver's social attributes. This represents the initial feature dimension of the node.

[0067] By providing two enhanced views and Different random seeds are set for each, so that the two complement each other in terms of topology and node attributes, for example... Use topological perturbation while preserving original properties. The topology remains original while using attribute perturbations.

[0068] Step 2: Input the two augmented views into the graph attention encoder with shared parameters to obtain the corresponding node embedding representations, which are denoted as the main view feature vector and the perturbation augmented view feature, respectively.

[0069] First enhanced view Second Enhanced View The graph attention encoder (GAT) with shared parameters is input separately. The GAT outputs the corresponding node embedding representation for each augmented view, denoted as the main view feature vector. and perturbation-enhanced view features .

[0070] Furthermore, in step 2, for any enhanced view, the node is calculated. Its neighboring nodes Attention coefficient between and normalized weights The calculation formulas are as follows:

[0071] ,

[0072] ,

[0073] In the formula, It is a linear rectified activation function with leakage, used to introduce nonlinearity and preserve gradient information in the negative region; For attention weight vectors, It is a linear transformation matrix. For nodes The input feature vector (if it is the first layer of the encoder, it is taken from the attribute matrix; otherwise, it is the node embedding output from the previous layer). Represents a node The set of neighboring nodes;

[0074] The features of neighboring nodes are weighted and aggregated according to normalized weights to generate the embedded representation of the current node.

[0075] Step 3: Train and optimize all parameters of the graph attention encoder and decoder by combining self-supervised contrast loss, structural reconstruction loss, modularity enhancement loss and semi-supervised cross-entropy loss.

[0076] Furthermore, in step 3, the total loss function for training and optimizing the graph attention encoder is:

[0077] ,

[0078] End-to-end joint optimization of all parameters of the graph attention encoder and decoder is performed by minimizing this loss function.

[0079] in For self-supervised contrastive loss, the model is driven to capture the social skeleton by maximizing the consistency between the two, expressed as:

[0080] ,

[0081] In the formula, This is the cosine similarity calculation function. Indicates the total number of vehicles; Main view 1 eigenvector For perturbation enhancement view number The mechanism improves the robustness of features by forcing the model to maintain representation invariance under attribute noise interference.

[0082] in The modularity enhancement loss is calculated by interpolating the main view features with a predefined modularity matrix, and is expressed as follows:

[0083] ,

[0084] In the formula, Represents the trace operation of a matrix. The predefined modularity matrix is ​​based on the original adjacency matrix. The pre-calculated topological attribute matrix, Represents the feature vector of the main view;

[0085] in The structural reconstruction loss is expressed as:

[0086] ,

[0087] In the formula, This represents the Sigmoid activation function, used to map the input to the range (0,1); Denotes the Frobenius norm;

[0088] For semi-supervised cross-entropy loss, These are the weighting coefficients;

[0089] in and All depend on the number of training rounds The dynamically changing scheduling factor is expressed as:

[0090] ,

[0091] ,

[0092] In the formula, As the initial comparison weights, The attenuation rate, This is the modularity weighting factor. This represents the maximum number of training epochs. In one example, possible values ​​are: , , .

[0093] Step 4: Input the test node features into the trained graph attention encoder, obtain the continuous membership matrix through continuous membership mapping, traverse the continuous membership matrix, and output the overlapping membership identity and membership strength of each vehicle.

[0094] Furthermore, in step 4, the continuous membership matrix By analyzing the feature vector of the main view The mapping yields the following expression:

[0095] ,

[0096] in, Indicates the total number of communities. Indicates vehicle Belonging to the community The continuous probability distribution value, with a range of values ​​as follows: , Indicates the first The weight vector of each community is used to embed nodes. The probability score mapped to that community; Indicates the first The bias term for each community is used to adjust the offset of the mapping.

[0097] Furthermore, in step 4, a membership determination threshold is set. Traverse the continuous membership matrix ,like Then determine the node Belongs to the community Finally, the overlapping membership identities and membership strengths of each vehicle node are output, where the membership strength is a continuous membership matrix. elements in The value of .

[0098] In one example, to further illustrate the superiority of the vehicle overlapping community detection method based on self-supervised contrastive graph autoencoder described in this invention, four representative real vehicle social network datasets, fb_1912, fb_1684, fb_414, and fb_348, were selected for comprehensive performance evaluation. The statistical information of each dataset is shown in Table 1.

[0099] Table 1. Statistical information of the dataset

[0100]

[0101] This example uses an NVIDIA RTX 3090 GPU and a PyTorch 1.12 and DGL 0.9 deep learning framework as the software environment. The parameter settings for the model in this invention include setting the hidden layer dimension. Characteristic perturbation probability The dynamic weight scheduler parameters are set as follows: initial comparison weights. attenuation rate Modularity weighting factor Overlap Normalized Mutual Information (ONMI), average F1 score, and Jaccard Index are used as the core evaluation criteria for the accuracy of community detection.

[0102] To verify the robustness of this invention (AC-CDGAAE) in scenarios with extremely sparse supervised signals, it is compared with the base model CDGAAE. CDGAAE is a semi-supervised overlapping community detection method based on graph attention autoencoders. This method integrates topological information and node attributes, uses a graph attention mechanism for encoding, and enhances the capture of overlapping community structures through modularity optimization. Compared with CDGAAE, this invention (AC-CDGAAE) further introduces a self-supervised dual-view comparison mechanism to enhance unsupervised feature learning capabilities, uses continuous membership mapping to replace hard threshold partitioning to preserve boundary gradient information, and designs an adaptive dynamic loss weight scheduler to alleviate gradient conflict problems in multi-task training. Experiments were conducted with prior information proportions set to 0% (purely unsupervised), 2%, 5%, and 10%. The performance of AC-CDGAAE and the base model CDGAAE was compared on four datasets, and the results are shown in Tables 2 to 5.

[0103] Table 2 Performance Comparison Results of fb_1912 Dataset

[0104]

[0105] Table 3 Performance Comparison Results of fb_414 Dataset

[0106]

[0107] Table 4 Performance Comparison Results of fb_1684 Dataset

[0108]

[0109] Table 5 Performance Comparison Results of fb_348 Dataset

[0110]

[0111] As shown in Tables 2 to 5, the method described in this invention exhibits low label robustness: on the fb_1912 dataset with 0% label rate (purely unsupervised), the ONMI of this invention reaches 0.4201, a 67.3% improvement compared to CDGAAE's 0.2511; the F1-score improves from 0.0688 to 0.5363, and the Jaccard Index improves from 0.1251 to 0.7294. This demonstrates that the dual-view self-supervised comparison module can lock social features solely based on the inherent correlation between topology and attributes, effectively solving the technical pain point of difficulty in obtaining real labels in vehicle social networks.

[0112] With a 2% labeling rate on the fb_1912 dataset, the Jaccard Index of this invention reaches 0.7467, which is close to the level of CDGAAE at a 10% labeling rate (0.7420). This proves that the continuous membership mapping model based on Sigmoid in this invention effectively preserves the boundary gradient information and significantly improves the recognition accuracy of overlapping boundaries.

[0113] On the fb_414, fb_1684, and fb_348 datasets, this invention significantly outperforms CDGAAE in terms of the Jaccard Index, especially in purely unsupervised scenarios, where the Jaccard Index of this invention remains consistently above 0.73, while CDGAAE's Jaccard Index is only 0.4306 and 0.4083 on fb_414 and fb_348, respectively. This fully verifies the generalization ability of this invention under different network structures.

[0114] To verify the contribution of the three core improvements of this invention to the overall performance, a stepwise ablation experiment was designed on the fb_1912 dataset (2% label rate). All configurations used continuous membership mapping as the basis, and a dual-view comparison module, a dynamic weight scheduler, and modularity enhancement loss were added sequentially. The experimental results are shown in Table 6.

[0115] Table 6 Comparison of overall performance in ablation experiments

[0116]

[0117] As shown in Table 6, when adding the contrastive loss alone under fixed weights, the model performance slightly decreases (ONMI drops from 0.3421 to 0.3338). This is because the contrastive loss and reconstruction loss compete for gradients in the early stages of optimization. This precisely illustrates the necessity of introducing a dynamic scheduler.

[0118] The complete AC-CDGAAE model, employing both a dynamic weight scheduler and modularity enhancement loss, achieves significant performance improvements, with an ONMI of 0.3892, a 13.8% improvement over the baseline. The F1-score and Jaccard Index also increase simultaneously. This demonstrates that the dynamic scheduler effectively coordinates multi-task learning, guiding the model smoothly from globally robust representations to locally refined clustering, while the modularity enhancement loss further strengthens the rationality of the community structure, making the detection results more consistent with topological compactness.

Claims

1. A method for detecting overlapping vehicle communities based on a self-supervised contrast map autoencoder, characterized in that, Includes the following steps: Step 1: Obtain vehicle trajectory data, construct a vehicle social network graph based on the vehicle trajectory data, generate an attribute matrix based on the vehicle multimodal spatiotemporal data, and construct two enhanced views that are complementary in topology and node attributes through the original adjacency matrix and attribute matrix of the vehicle social network graph. Step 2: Input the two augmented views into the graph attention encoder with shared parameters to obtain the corresponding node embedding representations, which are denoted as the main view feature vector and the perturbation augmented view feature, respectively. Step 3: Train and optimize all parameters of the graph attention encoder and decoder by combining self-supervised contrastive loss, structural reconstruction loss, modularity enhancement loss and semi-supervised cross-entropy loss; Step 4: Input the test node features into the trained graph attention encoder, obtain the continuous membership matrix through continuous membership mapping, traverse the continuous membership matrix, and output the overlapping membership identity and membership strength of each vehicle.

2. The vehicle overlap community detection method based on a self-supervised comparison map autoencoder according to claim 1, characterized in that, Step 1 includes: Construct a vehicle social network graph that includes node features and topological structure, using vehicles as nodes and the relationships between vehicles as edges. ,in For the set of vehicle nodes, For the vehicle edge set, the network graph The original adjacency matrix is ​​denoted as , Represents a node With nodes There must be a communication or social relationship between them, otherwise , This represents the number of vehicle nodes. For the original adjacency matrix According to probability Randomly delete some edges to obtain the mask matrix. The elements in this matrix The perturbated adjacency matrix is ​​denoted as The corresponding first enhanced view is denoted as ; For attribute matrix According to probability Randomly set the eigenvalues ​​to zero or add Gaussian noise to obtain the perturbed attribute matrix. The corresponding second enhanced view is denoted as .

3. The vehicle overlap community detection method based on a self-supervised comparison map autoencoder according to claim 2, characterized in that, In step 2, for any enhanced view, compute nodes. Its neighboring nodes Attention coefficient between and normalized weights The calculation formulas are as follows: ， ， In the formula, It is a linear rectified activation function with leakage, used to introduce nonlinearity and preserve gradient information in the negative region; For attention weight vectors, It is a linear transformation matrix. For nodes The input feature vector; Represents a node The set of neighboring nodes; The features of neighboring nodes are weighted and aggregated according to normalized weights to generate the embedded representation of the current node.

4. The vehicle overlap community detection method based on a self-supervised comparison map autoencoder according to claim 1, characterized in that, In step 3, the total loss function for training optimization is: ， in, The expression for the self-supervised comparison loss is: ， In the formula, This is the cosine similarity calculation function. Indicates the total number of vehicles; Main view 1 eigenvector For perturbation enhancement view number 1 eigenvector; in The modularity enhancement loss is calculated by interpolating the main view features with a predefined modularity matrix, and is expressed as follows: ， In the formula, Represents the trace operation of a matrix. For a predefined modularity matrix, Represents the feature vector of the main view; in The structural reconstruction loss is expressed as: ， In the formula, This represents the Sigmoid activation function, used to map the input to the range (0,1); Denotes the Frobenius norm; For semi-supervised cross-entropy loss, These are the weighting coefficients; in and All depend on the number of training rounds The dynamically changing scheduling factor is expressed as: ， ， In the formula, As the initial comparison weights, The attenuation rate, This is the modularity weighting factor. This represents the maximum number of training rounds.

5. The vehicle overlap community detection method based on a self-supervised comparison map autoencoder according to claim 4, characterized in that, In step 4, the continuous membership matrix By analyzing the feature vector of the main view The mapping yields the following expression: ， in, Indicates the total number of communities. Indicates vehicle Belonging to the community The continuous probability distribution value, with a range of values ​​as follows: , Indicates the first The weight vector of each community is used to embed nodes. The probability score mapped to that community; Indicates the first The bias term for each community is used to adjust the offset of the mapping.

6. The vehicle overlap community detection method based on a self-supervised comparison map autoencoder according to claim 5, characterized in that, In step 4, the membership determination threshold is set. Traverse the continuous membership matrix ,like Then determine the node Belongs to the community Finally, the overlapping membership identities and membership strength of each vehicle node are output.

Images