Vehicle overlapping group detection method based on graph attention automatic encoder

Abstract

The invention provides a vehicle overlapping group detection method based on a graph attention automatic encoder, and belongs to the technical field of Internet of Vehicles and the technical field of social networks. According to the technical scheme, the method comprises the following steps of: 1, inputting a network topology structure and node attribute information into an encoder part of a graph attention automatic encoder; step 2, encoding the input network topology structure and node attribute information by a graph attention mechanism of an encoder; step 3, the decoder combines the new node embedding representation with the original embedding representation, and restores the network topology structure and the node attribute information; step 4, node embedding is further adjusted through optimization constraints of a modularity optimization enhancement module; and 5, performing semi-supervised clustering by using prior information to obtain a node-based embedded representation to generate a community detection result. The method has the beneficial effect that the accuracy of overlapped community detection of the vehicle nodes can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of Internet of Vehicles and social network, and in particular to a method for detecting overlapping groups of vehicles based on a graph attention autoencoder. Background Art

[0002] With the development of intelligent transportation systems, vehicles can be connected to each other through wireless communication technology to form a self-organizing network that can support applications such as information sharing, navigation, and traffic management. This network uses communication between vehicles (V2V) and vehicles and infrastructure (V2I) to improve road safety, optimize traffic flow, and enhance environmental monitoring capabilities. Community detection, that is, identifying densely connected groups of nodes in the network, these communities are usually manifested as densely connected areas in the network, with close connections between nodes and relatively sparse connections with other communities. It is crucial for managing vehicles and maintaining safety within vehicle social networks. In vehicle social networks, communication efficiency can be improved, security and privacy protection can be enhanced, resource allocation can be optimized, and local management can be achieved.

[0003] Most existing community detection methods focus on identifying non-overlapping communities. For example, in the paper "Research on Hybrid Trust Evaluation Mechanism in In-vehicle Social Networks", a community discovery algorithm based on user interest similarity is proposed, but this algorithm can only detect non-overlapping vehicle communities. However, overlapping communities are very common in vehicle social networks. For vehicles, they may participate in different communities in different time periods or different situations. Summary of the invention

[0004] Community detection is particularly important in vehicle social networks, which helps to identify densely connected vehicle groups in the network; however, most existing community detection methods mainly focus on non-overlapping communities, and usually only pay attention to the topological structure of the network, while ignoring the attribute information of the nodes. In view of these limitations, the present invention provides a vehicle overlapping group detection method based on a graph attention autoencoder, which includes a graph attention autoencoder module, a modularity optimization enhancement module and a semi-supervised clustering module; first, the graph attention autoencoder module fuses topological information and node attribute information, and encodes the nodes using the graph attention mechanism; secondly, the modularity optimization enhancement module accurately captures the structure of overlapping communities; finally, the semi-supervised clustering module fuses prior information to improve the accuracy of community detection. Through the above three modules, the accuracy of overlapping community detection of vehicle nodes can be effectively improved.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a vehicle overlapping group detection method based on a graph attention autoencoder, comprising three modules, namely a graph attention autoencoder module, a modularity optimization enhancement module and a semi-supervised clustering module, specifically comprising the following steps:

[0006] Step 1: Input the network topology and node attribute information into the encoder part of the graph attention autoencoder. Through the autoencoder structure, not only can the potential representation of the graph be learned, but also the model can be optimized through reconstruction loss to better capture the key structural features in the graph.

[0007] Step 2: The encoder's graph attention mechanism encodes the input network topology and node attribute information. After being processed by the attention mechanism, it is represented as multiple different hidden layer embedding representations. These different embeddings capture the complex relationships and attribute differences between nodes. They are then nonlinearly transformed through an activation function to generate new node embedding representations.

[0008] Step 3: The decoder combines the new node embedding representation with the original graph embedding representation, restores the network topology and node attribute information through the decoding process, and compares them with the original topology and attribute information to evaluate the quality of reconstruction.

[0009] Step 4: Enhance the optimization constraints of the module through modularity optimization, further adjust the node embedding, ensure the rationality of the community structure, use the M / M / 1 queuing model, and set q ij ≥0 and Constraints are used to optimize, and finally a modularity optimization enhancement loss L is output. M , measure the rationality of community detection.

[0010] Step 5: Use prior information for semi-supervised clustering and reconstruct the loss L according to the graph R , modularity optimization enhancement loss L M , semi-supervised clustering loss L S After optimization, the node-based embedding representation is obtained to generate community detection results.

[0011] The present invention proposes a vehicle overlapping group detection method based on a graph attention autoencoder, and the specific steps of step 2 are as follows:

[0012] Step 2.1: In the multi-head attention mechanism, l independent attention heads are introduced, each with its own linear transformation matrix W (l) , used to transform the original node feature h i Convert to a new feature space. This transformation is performed on each node feature vector separately, and the formula is as follows:

[0013]

[0014] in is the new feature representation of node i in the lth head.

[0015] Step 2.2: For each node v i and v j Calculate an attention score, indicating that node v j For node v i The importance of this score. This score is calculated based on the feature representation after linear transformation. The specific formula is:

[0016]

[0017] where a (l) is the attention weight vector of the lth head, with a length of 2F', which is used to map the concatenated vector to a scalar, and || represents the concatenation operation, where F' is the new feature space dimension of each head. LeakyReLU is the activation function.

[0018] Step 2.3: Use the softmax function to normalize the original attention scores of all neighbor nodes to obtain the final attention coefficient. This process ensures that the sum of the attention coefficients of all neighbor nodes is 1:

[0019]

[0020] Step 2.4: For each attention head l, use the attention coefficient Perform weighted summation on neighbor node features to obtain the feature output of each head l

[0021]

[0022] Step 2.5: Average the output features of all heads to obtain the final node representation:

[0023]

[0024] The present invention proposes a vehicle overlapping group detection method based on a graph attention autoencoder, and the specific steps of step 3 are as follows:

[0025] Step 3.1. Reconstruct the adjacency matrix A of the graph through the latent representation Z of the node, where A is the predicted adjacency matrix. When using the multi-head attention mechanism, the latent representation z i Usually the output features z from multiple attention heads i (l) Aggregated. The output features of multiple heads are averaged, and the potential representation of each node is:

[0026]

[0027] Step 3.2: Potential representation z of the node i It is still an F'-dimensional vector. The formula of the decoder remains unchanged and is calculated as follows:

[0028]

[0029] in is the existence of an edge between node i and node j predicted by the decoder. i and z j are the potential representations of nodes i and j respectively, σ is the activation function, and the Sigmoid function is usually chosen:

[0030]

[0031] Step 3.3: The final result of the decoder is used as the reconstructed node attribute value, and the loss function of the reconstructed node attribute value is defined as Equation 9.

[0032]

[0033] Where X i is the original binary attribute of the ith node, is the probability that the model predicts that the attribute of the i-th node is 1. This loss function calculates the binary cross entropy for the binary attribute reconstruction of each node, and optimizes the model's reconstruction effect on node attributes by minimizing the loss.

[0034] Step 3.4, the loss function of adjacency matrix prediction is defined as Equation 10:

[0035]

[0036] Among them A ij represents the elements in the original adjacency matrix, specifically whether there is an edge between node i and node j. ij =1, indicating that there is an edge between node i and node j; if A ij =0, indicating no edge. It is an element in the adjacency matrix predicted by the model, indicating the probability that the model predicts that there is an edge between node i and node j.

[0037] Step 3.5: The final reconstruction loss function is formed by combining the reconstruction losses of the topological structure and node attributes.

[0038] L R =αL attr +βL top (11)

[0039] Where α and β are hyperparameters used to balance the impact of node attribute reconstruction loss and adjacency matrix prediction loss.

[0040] The present invention proposes a vehicle overlapping group detection method based on a graph attention autoencoder, and the specific steps of step 4 are as follows:

[0041] Step 4.1: The final node representation matrix H output by the encoder is further refined through the modularity optimization enhancement module. Modularity is an indicator of the strength of the network community structure and can be defined as:

[0042]

[0043] where d i and d j is the degree of node i and node j, H i and H j Represents the attribute vector of nodes i and j. Indicates the attribute similarity between nodes i and j.

[0044] Step 4.2: Add a Markov queueing system to the modularity maximization module to become a modularity optimization enhancement module. In the Markov queueing system, the state transition between nodes can be represented by the state transfer matrix P. The enhanced modularity optimization objective is defined as:

[0045]

[0046] Where P ij represents the probability of node i transferring to node j, A ij P ij It means that not only whether there are direct edges between nodes, but also the strength of dynamic interactions between nodes is considered.

[0047] Step 4.3: In order to optimize the enhanced modularity, the corresponding loss function is defined as:

[0048] L M =-Q M (14)

[0049] The present invention proposes a vehicle overlapping group detection method based on a graph attention autoencoder, and the specific steps of step 5 are as follows:

[0050] Step 5.1: Combine the known node label information Y through the semi-supervised clustering module to finally get the community detection result. The semi-supervised strategy can maintain the consistency of known labels and simultaneously explore the potential community affiliation of unlabeled nodes, making the entire community division more reasonable. Therefore, the cross entropy loss used as the semi-supervised loss L S :

[0051]

[0052] in is a set of labeled data, Y ij The true label indicating whether the marked node i belongs to community j.

[0053] Step 5.2: In order to obtain a more accurate community member representation matrix H, it is not enough to use the graph attention autoencoder alone. Integrating the graph attention autoencoder, modularity optimization enhancement, and semi-supervised clustering module into a whole can obtain more accurate community detection results. The final objective function is defined as in Equation 16.

[0054] L=L R +γL S -λL M (16)

[0055] Where γ and λ are hyperparameters representing different loss weights, which are used to adjust the contribution of the corresponding modules. R , L M and L S They represent the losses of the graph attention autoencoder, modularity optimization enhancement, and self-trained clustering module, respectively.

[0056] Step 5.3: Train the model according to the final loss function L to obtain the final community detection result.

[0057] The main symbols and their related explanations are shown in Table 1 below.

[0058] Table 1 Symbols

[0059]

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. Integration of topological and attribute information: The method proposed in the present invention utilizes the graph attention autoencoder (GAAE) to effectively integrate the topological structure and node attribute information, and can capture more complex network features. This integration can provide a more comprehensive understanding of the relationships and characteristics of vehicles in social networks, thereby achieving more accurate community detection. In vehicle social networks, vehicles not only interact with each other through topological connections, but may also be affected by the driver's social attributes (such as occupation, interests, etc.). Such integration makes community detection more accurate, can better reflect the real vehicle community structure, and contributes to more efficient traffic management and resource optimization.

[0062] 2. Modularity optimization and enhancement module: This module introduces the Markov queuing model based on modularity maximization. This innovation allows for more flexible representation of overlapping communities, allowing nodes to belong to multiple communities at the same time, which is particularly important in vehicle networks because the social relationships of vehicles may change over time. Traditional community detection methods often strictly divide nodes into a single community and cannot effectively identify overlapping communities. By introducing the Markov queuing model, this module can flexibly represent the possibility of nodes in multiple communities. This method is suitable for vehicle social networks because vehicles may participate in different communities at different times and occasions. The optimized modularity method enhances the ability to model dynamic interactions and relationships, thereby improving the accuracy and reliability of community detection.

[0063] 3. Semi-supervised clustering module: This module uses prior information (known community labels) to guide community detection. By introducing known community labels as prior information, this module can significantly improve the accuracy of community detection, ensure that the detected community is more consistent with the actual community structure, and solve the limitations of pure unsupervised methods. The semi-supervised method combines a small amount of labeled information and the potential community affiliation of unlabeled data, which can make community divisions more reasonable. This method is particularly important in vehicle social networks, because in actual applications, the social behavior and interaction patterns of vehicles may be affected by different factors. Known label information can effectively guide model learning and improve the final detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0065] Figure 1 It is a model framework diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0067] Example 1

[0068] See also Figure 1 The present invention provides a technical solution as follows: the present invention is a vehicle overlapping group detection method based on graph attention autoencoder (CDGAAE), and the overall process of the method is as shown in the attached Figure 1 The technical solution of the present invention is further described below in conjunction with specific embodiments:

[0069] Step 1: Input the network topology and node attribute information into the encoder part of the graph attention autoencoder. Through the autoencoder structure, not only can the potential representation of the graph be learned, but also the model can be optimized through reconstruction loss to better capture the key structural features in the graph.

[0070] Step 2: The encoder's graph attention mechanism encodes the input network topology and node attribute information. After being processed by the attention mechanism, it is represented as multiple different hidden layer embedding representations. These different embeddings capture the complex relationships and attribute differences between nodes. Then, a nonlinear transformation is performed through an activation function to generate a new node embedding representation.

[0071] Step 3: The decoder combines the new node embedding representation with the original graph embedding representation, and restores the network topology and node attribute information through the decoding process. It is compared with the original topology and attribute information to evaluate the quality of the reconstruction.

[0072] Step 4: Enhance the optimization constraints of the module through modularity optimization, further adjust the node embedding, and ensure the rationality of the community structure. Use the M / M / 1 queuing model and set q ij ≥0 and Constraints are used for optimization. Finally, a modularity optimization enhancement loss L is output. M , measure the rationality of community detection.

[0073] Step 5: Use prior information for semi-supervised clustering and reconstruct the loss L according to the graph R , modularity optimization enhancement loss L M , semi-supervised clustering loss L S After optimization, the node-based embedding representation is obtained to generate community detection results.

[0074] The specific steps of step 2 are as follows:

[0075] Step 2.1: In the multi-head attention mechanism, l independent attention heads are introduced, each with its own linear transformation matrix W (l) , used to transform the original node feature h i Convert to a new feature space. This transformation is performed on each node feature vector separately, and the formula is as follows:

[0076]

[0077] in is the new feature representation of node i in the lth head.

[0078] Step 2.2: For each node v i and v j Calculate an attention score, indicating that node vj For node v i The importance of this score. This score is calculated based on the feature representation after linear transformation. The specific formula is:

[0079]

[0080] where a (l) is the attention weight vector of the lth head, with a length of 2F', which is used to map the concatenated vector to a scalar, and || represents the concatenation operation, where F' is the new feature space dimension of each head. LeakyReLU is the activation function.

[0081] Step 2.3: Use the softmax function to normalize the original attention scores of all neighbor nodes to obtain the final attention coefficient. This process ensures that the sum of the attention coefficients of all neighbor nodes is 1:

[0082]

[0083] Step 2.4: For each attention head l, use the attention coefficient Perform weighted summation on neighbor node features to obtain the feature output of each head l

[0084]

[0085] Step 2.5: Average the output features of all heads to obtain the final node representation:

[0086]

[0087] The specific steps of step 3 are as follows:

[0088] Step 3.1. Reconstruct the adjacency matrix A of the graph through the latent representation Z of the node, where A is the predicted adjacency matrix. When using the multi-head attention mechanism, the latent representation z i Usually the output features from multiple attention heads Aggregated. The output features of multiple heads are averaged, and the potential representation of each node is:

[0089]

[0090] Step 3.2: Potential representation z of the node i It is still an F'-dimensional vector. The formula of the decoder remains unchanged and is calculated as follows:

[0091]

[0092] in is the existence of an edge between node i and node j predicted by the decoder. i and z j are the potential representations of nodes i and j respectively, σ is the activation function, and the Sigmoid function is usually chosen:

[0093]

[0094] Step 3.3: The final result of the decoder is used as the reconstructed node attribute value, and the loss function of the reconstructed node attribute value is defined as Equation 9.

[0095]

[0096] Where X i is the original binary attribute of the ith node, is the probability that the model predicts that the attribute of the i-th node is 1. This loss function calculates the binary cross entropy for the binary attribute reconstruction of each node, and optimizes the model's reconstruction effect on node attributes by minimizing the loss.

[0097] Step 3.4, the loss function of adjacency matrix prediction is defined as Equation 10:

[0098]

[0099] Among them A ij represents the elements in the original adjacency matrix, specifically whether there is an edge between node i and node j. ij =1, indicating that there is an edge between node i and node j; if A ij =0, indicating no edge. It is an element in the adjacency matrix predicted by the model, indicating the probability that the model predicts that there is an edge between node i and node j.

[0100] Step 3.5: The final reconstruction loss function is formed by combining the reconstruction losses of the topological structure and node attributes.

[0101] L R =αL attr +βL top (11)

[0102] Where α and β are hyperparameters used to balance the impact of node attribute reconstruction loss and adjacency matrix prediction loss.

[0103] The specific steps of step 4 are as follows:

[0104] Step 4.1: The final node representation matrix H output by the encoder is further refined through the modularity optimization enhancement module. Modularity is an indicator of the strength of the network community structure and can be defined as:

[0105]

[0106] where d i and d j is the degree of node i and node j, H i and H j Represents the attribute vector of nodes i and j. Indicates the attribute similarity between nodes i and j.

[0107] Step 4.2: Add a Markov queueing system to the modularity maximization module to become a modularity optimization enhancement module. In the Markov queueing system, the state transition between nodes can be represented by the state transfer matrix P. The enhanced modularity optimization objective is defined as:

[0108]

[0109] Where P ij represents the probability of node i transferring to node j, A ij P ij It means that not only whether there are direct edges between nodes, but also the strength of dynamic interactions between nodes is considered.

[0110] Step 4.3: In order to optimize the enhanced modularity, the corresponding loss function is defined as:

[0111] L M =-Q M (14)

[0112] The specific steps of step 5 are as follows:

[0113] Step 5.1: Combine the known node community label information Y through the semi-supervised clustering module to finally obtain the community detection result. The semi-supervised strategy can maintain the consistency of known labels and simultaneously explore the potential community affiliation of unlabeled nodes, making the entire community division more reasonable. Therefore, the cross entropy loss used as the semi-supervised loss L S :

[0114]

[0115] in is a set of labeled data, Y ij The true label indicating whether the marked node i belongs to community j.

[0116] Step 5.2: In order to obtain a more accurate community member representation matrix H, it is not enough to use the graph attention autoencoder alone. Integrating the graph attention autoencoder, modularity optimization enhancement, and semi-supervised clustering module into a whole can obtain more accurate community detection results. The final objective function is defined as in Equation 16.

[0117] L=L R +γL S -λL M (16)

[0118] Where γ and λ are hyperparameters representing different loss weights, which are used to adjust the contribution of the corresponding modules. R , L M and L S They represent the losses of the graph attention autoencoder, modularity optimization enhancement, and self-trained clustering module, respectively.

[0119] Step 5.3: Train the model according to the final loss function L to obtain the final community detection result.

[0120] Example 2

[0121] The social network dataset Facebook414 was selected for comparative experiments. Three widely used evaluation indicators were introduced for the detection accuracy of overlapping communities: Overlap Normalized Mutual Information (ONMI), Average F1 Score (F1 for short) and JaccardIndex. At the same time, five competing methods were compared. According to the principle of whether to integrate prior information, these five methods include unsupervised and semi-supervised methods. The proportion of prior information was changed from 0% to 10%, with a step size of 2%. The results of ONMI, F1 and JaccardIndex are shown in Tables 2, 3 and 4 respectively.

[0122] Table 2fb414_ONMI

[0123] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.2 0.2 0.2 0.2 0.2 0.2 SCI 0.28 0.28 0.28 0.28 0.28 0.28 Bespoke 0.15 0.15 0.15 0.15 0.15 0.15 NOCD 0.5 0.5 0.5 0.5 0.5 0.5 SSGCAE 0.26 0.31 0.35 0.36 0.37 0.4 CDGAE 0.4 0.5 0.51 0.52 0.53 0.55

[0124] Table 3fb414_F1

[0125] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.49 0.49 0.49 0.49 0.49 0.49 SCI 0.19 0.19 0.19 0.19 0.19 0.19 Bespoke 0.39 0.39 0.39 0.39 0.39 0.39 NOCD 0.41 0.41 0.41 0.41 0.41 0.41 SSGCAE 0.4 0.6 0.61 0.61 0.61 0.61 CDGAE 0.48 0.7 0.73 0.77 0.8 0.81

[0126] Table 4fb414_Jaccard Index

[0127] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.39 0.39 0.39 0.39 0.39 0.39 SCI 0.27 0.27 0.27 0.27 0.27 0.27 Bespoke 0.3 0.3 0.3 0.3 0.3 0.3 NOCD 0.22 0.22 0.22 0.22 0.22 0.22 SSGCAE 0.53 0.67 0.69 0.69 0.69 0.7 CDGAE 0.58 0.75 0.8 0.82 0.83 0.85

[0128] It can be seen that on the Facebook414 dataset, the performance of CDGAAE becomes increasingly superior as the proportion of prior information increases. Based on the results in Tables 2, 3, and 4, we can draw the following conclusions.

[0129] (1) The performance of CDGAAE is better than that of unsupervised methods such as CESNA and NOCD, which shows the effectiveness of incorporating prior information in our method. Compared with these unsupervised methods, the performance of CDGAAE becomes increasingly superior as the proportion of prior information increases.

[0130] (2) The performance of CDGAAE is better than that of the SSGCAE semi-supervised method, which shows the importance of extracting community structure and reconstructing relevant attribute features. CDGAAE introduces a graph attention mechanism that can dynamically assign different weights to each node and uses an optimized modularity metric to obtain node representation, which helps to obtain topologically tight community results.

[0131] Example 3

[0132] The social network dataset Facebook1684 was selected for the second comparative experiment. The evaluation indicators used were overlap normalized mutual information (ONMI), average F1 score (F1 for short) and Jaccard Index. It was also compared with five competing methods. According to the principle of whether to integrate prior information, these five methods include unsupervised and semi-supervised methods. The proportion of prior information was changed from 0% to 10%, with a step size of 2%. The results of ONMI, F1 and Jaccard Index are shown in Tables 5, 6 and 7 respectively.

[0133] Table 5fb1684_ONMI

[0134] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.19 0.19 0.19 0.19 0.19 0.19 SCI 0.24 0.24 0.24 0.24 0.24 0.24 Bespoke 0.15 0.15 0.15 0.15 0.15 0.15 NOCD 0.26 0.26 0.26 0.26 0.26 0.26 SSGCAE 0.27 0.42 0.43 0.45 0.45 0.46 CDGAE 0.27 0.45 0.46 0.47 0.48 0.5

[0135] Table 6fb1684_F1

[0136] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.37 0.37 0.37 0.37 0.37 0.37 SCI 0.15 0.15 0.15 0.15 0.15 0.15 Bespoke 0.31 0.31 0.31 0.31 0.31 0.31 NOCD 0.15 0.15 0.15 0.15 0.15 0.15 SSGCAE 0.1 0.4 0.41 0.5 0.51 0.61 CDGAE 0.33 0.6 0.62 0.64 0.68 0.7

[0137] Table 7fb1684_Jaccard Index

[0138] method 0％ 2％ 4％ 6％ 8％ 10％ CESNA 0.37 0.37 0.37 0.37 0.37 0.37 SCI 0.22 0.22 0.22 0.22 0.22 0.22 Bespoke 0.26 0.26 0.26 0.26 0.26 0.26 NOCD 0.15 0.15 0.15 0.15 0.15 0.15 SSGCAE 0.49 0.5 0.51 0.52 0.6 0.61 CDGAE 0.6 0.76 0.79 0.8 0.81 0.82

[0139] The results in Table 5, Table 6, and Table 7 are consistent with those obtained in Example 2. On the Facebook414 dataset, the performance of CDGAAE becomes increasingly superior as the proportion of prior information increases. It is superior to unsupervised methods such as CESNA and NOCD, proving the effectiveness of integrating prior information; it is superior to the semi-supervised method SSGAAE, proving that the graph attention mechanism can dynamically assign different weights to each node, and the importance of the optimized modularity metric to node representation.

[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A vehicle overlapping group detection method based on graph attention autoencoder, characterized in that: The following steps are involved: Step 1: Input the network topology and node attribute information into the encoder part of the graph attention autoencoder. Through the autoencoder structure, not only the potential representation of the vehicle node is learned, but also the reconstruction loss is used to optimize the model so that it captures the key structural features. Step 2: The encoder’s graph attention mechanism encodes the input network topology and node attribute information. After being processed by the attention mechanism, it is represented as multiple different hidden layer embedding representations. These different embeddings capture the complex relationships and attribute differences between nodes. Then, a nonlinear transformation is performed through an activation function to generate a new node embedding representation. Step 3: The decoder combines the new node embedding representation with the original embedding representation, restores the network topology and node attribute information through the decoding process, and compares them with the original topology and attribute information to evaluate the quality of reconstruction; Step 4: Enhance the optimization constraints of the module through modularity optimization, further adjust the node embedding, ensure the rationality of the community structure, use the M / M / 1 queuing model, and set q ij ≥0 and Constraints are used to optimize, and finally a modularity optimization enhancement loss L is output. M , measure the rationality of community detection; Step 5: Use prior information for semi-supervised clustering and reconstruct the loss L according to the graph R , modularity optimization enhancement loss L M , semi-supervised clustering loss L S After optimization, the node-based embedding representation is obtained to generate community detection results.

2. The vehicle overlapping group detection method based on graph attention autoencoder according to claim 1 is characterized in that: The specific steps of step 2 are as follows: Step 2.1: In the multi-head attention mechanism, l independent attention heads are introduced, each with its own linear transformation matrix W (l) , used to transform the original node feature h i Convert to the new feature space. This transformation is performed on each node feature vector separately. The formula is as follows: in is the new feature representation of node i in the lth head; Step 2.2: For each node v i and v j Calculate an attention score, indicating that node v j For node v i The importance of this score is based on the feature representation after linear transformation and is calculated as follows: where a (l) is the attention weight vector of the lth head, with a length of 2F', which is used to map the concatenated vector to a scalar, and || represents the concatenation operation, where F' is the new feature space dimension of each head, and LeakyReLU is the activation function; Step 2.3: Use the softmax function to normalize the original attention scores of all neighbor nodes to obtain the final attention coefficient. This process ensures that the sum of the attention coefficients of all neighbor nodes is 1: Step 2.4: For each attention head l, use the attention coefficient Perform weighted summation on neighbor node features to obtain the feature output of each head l Step 2.5: Average the output features of all heads to get the final node representation:

3. The vehicle overlapping group detection method based on graph attention autoencoder according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.

1. Reconstruct the adjacency matrix A through the potential representation Z of the node, where A is the predicted adjacency matrix. When using the multi-head attention mechanism, the potential representation z i is the output feature from multiple attention heads Aggregated, the output features of multiple heads are averaged, and the potential representation of each node is: Step 3.2: Potential representation z of the node i It is still an F'-dimensional vector, and the formula of the decoder remains unchanged. The calculation formula is as follows: in is the existence of an edge between node i and node j predicted by the decoder, z i and z j are the potential representations of nodes i and j respectively, σ is the activation function, and the Sigmoid function is usually chosen: Step 3.3, the final result of the decoder is used as the reconstructed node attribute value, and the loss function of the reconstructed node attribute value is defined as Equation 9; Where X i is the original binary attribute of the ith node, is the probability that the model predicts that the attribute of the i-th node is 1. This loss function calculates the binary cross entropy for the binary attribute reconstruction of each node, and optimizes the model's reconstruction effect on node attributes by minimizing the loss; Step 3.4, the loss function of adjacency matrix prediction is defined as Equation 10: Among them A ij Indicates the elements in the original adjacency matrix, whether there is an edge between node i and node j, if A ij =1, indicating that there is an edge between node i and node j; if A ij =0, indicating no edge, is an element in the adjacency matrix predicted by the model, indicating the probability that the model predicts that there is an edge between node i and node j; Step 3.5, the final reconstruction loss function is formed by combining the reconstruction losses of the topological structure and node attributes; L R =αL attr +βL top (11) Where α and β are hyperparameters used to balance the impact of node attribute reconstruction loss and adjacency matrix prediction loss.

4. The vehicle overlapping group detection method based on graph attention autoencoder according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1: The final node representation matrix H output by the encoder is further refined through the modularity optimization enhancement module. The modularity is an indicator of the strength of the network community structure and is defined as: where d i and d j is the degree of node i and node j, H i and H j represents the attribute vector of nodes i and j, Indicates the attribute similarity between nodes i and j; Step 4.2: Add a Markov queueing system to the modularity maximization module to become a modularity optimization enhancement module. In the Markov queueing system, the state transition between nodes is represented by the state transfer matrix P. The enhanced modularity optimization objective is defined as: Where P ij represents the probability of node i transferring to node j, A ij P ij It means that not only whether there are direct edges between nodes, but also the strength of dynamic interactions between nodes is considered; Step 4.3: In order to optimize the enhanced modularity, the corresponding loss function is defined as: L M =-Q M (14)。 5. The vehicle overlapping group detection method based on graph attention autoencoder according to claim 1, characterized in that: The step 5 comprises the following steps: Step 5.1: Combine the known node label information Y through the semi-supervised clustering module to finally obtain the community detection result. Through the semi-supervised strategy, the consistency of the known labels is maintained, and the potential community affiliation of the unlabeled nodes is explored at the same time, making the entire community division more reasonable. Therefore, the cross entropy loss used is used as the semi-supervised loss L S : in is a set of labeled data, Y ij The true label indicating whether the marked node i belongs to community j; Step 5.2: In order to obtain a more accurate community member representation matrix H, the graph attention autoencoder, modularity optimization enhancement and semi-supervised clustering module are integrated into a whole to obtain a more accurate community detection result. The final objective function is defined as in Equation 16: L=L R +γL S -λL M (16) Where γ and λ are hyperparameters representing different loss weights, which are used to adjust the contribution of the corresponding modules. R , L M and L S They represent the losses of the graph attention autoencoder, modularity optimization enhancement, and self-training clustering module respectively; Step 5.3: Train the model according to the final loss function L to obtain the final community detection result.

Images