Social network analysis method based on depth attribute graph clustering

By adopting dual-path parallel adaptive depth attribute graph clustering network in social network analysis, the problem of integration of graph data noise and sparsity and multi-scale information is solved, and a more accurate and comprehensive general social network analysis is achieved.

CN118982435APending Publication Date: 2024-11-19SHANXI UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411201690.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing social network data analysis methods are difficult to effectively process noise and sparsity in graph data, and ignore multi-level multi-scale information, which affects the comprehensiveness and accuracy of the clustering model.

Method used

Adaptive depth attribute map clustering network with dual path parallelism, including autoencoder module, adaptive map autoencoder module, attention fusion module and dual self-supervised clustering module, the network is minimized by minimizing the losses of each module. Training, dynamically update graph structure and feature representations, and fuse multi-scale information.

Benefits of technology

It improves the accuracy and robustness of social network analysis, can better capture potential connections and complete information between nodes, and improves the quality of data analysis and decision-making support capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982435A_ABST
    Figure CN118982435A_ABST
Patent Text Reader

Abstract

The invention discloses a social network analysis method based on depth attribute graph clustering, and belongs to the technical field of clustering analysis and social network analysis. Aiming at the problems that image data has noise and sparsity and neglects shallow layer features, the method comprises the following steps: preprocessing data to obtain an attribute feature matrix and an adjacent matrix; constructing a dual-path parallel adaptive depth attribute graph clustering network; in the first path, using an auto-encoder module to extract potential features of the data; in the second path, the self-adaptive graph self-encoder module is used for alternately updating the graph structure and the data representation layer by layer; performing weighted fusion on representations learned by the adaptive graph auto-encoder module and the auto-encoder module by using an attention fusion module; and clustering the fused representation by using a dual self-supervised clustering module to obtain a pseudo tag to finish analysis. According to the method, the graph structure information and the feature information of the data can be dynamically and adaptively updated, the complex relation in the social network can be reflected, and the analysis accuracy of the social network is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of cluster analysis and social network analysis, and specifically relates to a social network analysis method based on deep attribute graph clustering. Background Art

[0002] Social network data refers to the collection of various types of information generated by users on social platforms, including user profiles, published content (such as text, pictures, videos), interactive behaviors (likes, comments, shares, follows, etc.) and the social relationship networks formed. The purpose of these data is to reveal the characteristics of users' interests, behavior patterns, social structures, etc., and provide data support for personalized recommendations, targeted advertising, social trend analysis, public opinion monitoring, etc. Its significance lies in helping individuals better manage their social lives, and also providing decision-making references and market insights for enterprises, research institutions, etc. However, social network data in practical problems often contain complex node (such as users) and edge (such as relationships, interactions) structures, which makes it difficult for existing methods to model them. In addition, there are often a large number of sparsity and missing values ​​in social network data, which makes it impossible for existing methods to effectively capture the potential connections and complete information between nodes, thus affecting the accuracy and reliability of data analysis.

[0003] Deep attribute graph clustering refers to learning node representations of graph data through deep learning models, and using these representations to perform clustering analysis on the graph. It uses deep learning technology to automatically learn and extract complex structural features in social network data, and can fill and repair sparse data to a certain extent by learning the attribute features and neighbor information of nodes. Therefore, social network data can be modeled more accurately through deep attribute graph clustering methods. Although deep attribute graph clustering has shown good results in the field of social network data analysis, it still faces two major challenges: First, the noise and sparsity of graph data lead to missing connections or wrong edges between similar nodes, which affects the information transmission in the graph neural network and reduces the quality of node representation learning; second, existing methods often ignore the integration of multi-scale information in multiple layers and ignore shallow features, which limits the comprehensiveness and accuracy of clustering models, thereby affecting the final analysis effect and decision support capabilities. Summary of the invention

[0004] In order to solve the problem that graph data has noise and sparsity and ignores shallow features, this paper proposes a social network analysis method based on deep attribute graph clustering. This method not only integrates multi-scale information at different levels, but also adaptively updates the graph structure based on the integrated feature representation, so that it can learn better feature representation, which is helpful for social network analysis.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A social network analysis method based on deep attribute graph clustering includes the following steps:

[0007] Step 1: Obtain a social network data set, preprocess the data set, and obtain an attribute feature matrix and an adjacency matrix;

[0008] Step 2, construct a dual-path parallel adaptive deep attribute graph clustering network, which includes an autoencoder module, an adaptive graph autoencoder module, an attention fusion module, and a dual self-supervised clustering module;

[0009] Step 3: Input the attribute feature matrix and the adjacency matrix into the dual-path parallel adaptive deep attribute graph clustering network, train the network by minimizing the loss of each module, obtain pseudo labels, and complete social network analysis.

[0010] Furthermore, the specific method of step 1 is: using the social network data set provided by the social website or the user information and relationship network data obtained on the social media platform through the API interface using web crawler technology to form a social network data set, preprocessing the data set to obtain the attribute feature matrix and the adjacency matrix.

[0011] Further, the specific method of step 2 includes the following steps:

[0012] Step 2.1, reconstruct the attribute feature matrix through the autoencoder module to extract the potential attribute feature representation;

[0013] Step 2.2, reconstruct the attribute feature matrix X and the adjacency matrix A through the adaptive graph autoencoder module to capture the structural feature representation of the data;

[0014] Step 2.3, the feature representations learned by the autoencoder module and the encoders of the same layer in the adaptive graph autoencoder module are subjected to attention fusion through the attention fusion module to obtain a fused feature representation, and a new graph structure is obtained based on the fused feature representation learning; the newly constructed graph structure is fused with the original graph structure to update the original graph structure;

[0015] In step 2.4, the network is optimized by using a dual-path collaborative training method through a dual self-supervised clustering module.

[0016] Furthermore, the step 2.1, reconstructing the attribute feature matrix through the autoencoder module, and extracting the potential attribute feature representation specifically includes the following steps:

[0017] Step 2.1.1, pass the attribute feature matrix X into the autoencoder module, and learn useful attribute feature representations through the encoder;

[0018] Specifically, the attribute feature representation H learned by the (l)th layer in the encoder part is(l) :

[0019]

[0020] Among them, ψ(·) represents the activation function of the leaky rectified linear unit (Leaky Rectified LinearUnit), and denote the weight matrix and bias of the encoder layer (l) respectively;

[0021] H (0) Represented as attribute feature matrix X;

[0022] Step 2.1.2: pass the potential attribute feature representation into the decoder to reconstruct the attribute feature matrix X and obtain the reconstructed attribute feature matrix

[0023]

[0024] in, and denote the weight matrix and bias of the decoder layer (l), respectively;

[0025] H (2L) Represented as the reconstructed attribute feature matrix

[0026] The reconstruction loss of its autoencoder module is as follows:

[0027]

[0028] in, represents the reconstruction loss of the autoencoder module, N represents the number of samples, and x i represents the attribute characteristics of sample i, represents the reconstructed attribute features of sample i, represents the F-norm.

[0029] Furthermore, the step 2.2, reconstructing the attribute feature matrix X and the adjacency matrix A through the adaptive graph autoencoder module, and capturing the structural feature representation of the data specifically includes the following steps:

[0030] Step 2.2.1: Pass the attribute feature matrix X and adjacency matrix A in the graph data into the adaptive graph autoencoder module, and learn the structural feature representation Z through the graph encoder. (l) :

[0031]

[0032] Among them, φ(·) represents the activation function of the Rectified Linear Unit. I is the identity matrix, and represents the weight matrix of the graph encoder. In addition, Z (0) Represented as attribute feature matrix X;

[0033] Step 2.2.2, starting from the second layer, use the fusion representation F of the previous layer (l) And the updated graph structure A (l) , its expression becomes as follows:

[0034]

[0035] After passing through the fifth layer of the graph encoder, the final joint feature representation Z is obtained;

[0036] Step 2.2.3, pass the final joint feature representation Z into the graph decoder to obtain the reconstructed feature matrix

[0037]

[0038] in, represents the weight matrix of the graph decoder. In addition, Z (L) Represents the reconstructed feature matrix

[0039] Step 2.2.4, reconstruct the attribute feature matrix X and the adjacency matrix A:

[0040]

[0041] Among them, the reconstructed adjacency matrix is the inner product of the final joint feature representation Z and the reconstructed feature matrix The sum of the inner products of

[0042] The reconstruction loss of its adaptive graph autoencoder module is as follows:

[0043]

[0044]

[0045]

[0046] Furthermore, in step 2.3, the feature representations learned by the autoencoder module and the encoders of the same layer in the adaptive graph autoencoder module are subjected to attention fusion through the attention fusion module to obtain a fused feature representation, and a new graph structure is obtained through the fused feature representation, which is fused with the original graph structure, thereby updating the graph structure. Specifically, the following steps are included:

[0047] Step 2.3.1: Represent the potential attribute features learned by the autoencoder as H (l) Passed into the two-layer GNN to learn better attribute feature representation

[0048]

[0049] in, Represents the weight matrix in GNN;

[0050] Step 2.3.2: Represent the structural features learned by the adaptive graph autoencoder module as Z (l) and better attribute feature representation Connect together to get the cascade feature representation Y (l) :

[0051]

[0052] Among them, [,] represents the connection operation;

[0053] Step 2.3.3, using cascade feature representation, according to Z (l) and The importance of learning the corresponding weights and obtaining the fusion feature representation F (l) :

[0054] e=softmax(sigmoid(f(Y (l) ) / τ)

[0055] W = mean(e, dim = 0)

[0056]

[0057] Where f(·) represents a three-layer fully connected network, τ is the calibration factor; the learned weight W = [W1, W2]; in addition, the sigmoid(·) function and the calibration factor are to avoid assigning scores that are almost "1" to the most relevant feature representations;

[0058] Step 2.3.4, when fusing feature representations in the last layer, the feature representations of different scales learned in the previous layers are fused. Since the feature dimensions of different layers are different, the multi-scale features are aggregated in a cascade manner:

[0059] Z'=[Z(1) ||…||Z (l) ||…||Z (L) ||Z (L+1) ]

[0060] Among them, Z (L+1) express

[0061] Dynamically combine features of each scale through the attention mechanism:

[0062]

[0063] Where M = [m1||…||m l ||…||m L ||m L+1 ],m l It's Z (l) The attention coefficient, express Normalization, W s Represents the weight matrix, which is used to capture the relationship between features of different layers;

[0064] The learned attention coefficients are used to weight the multi-scale features and the fused features are represented as F (L) is formulated as:

[0065] F (L) =[(m111)⊙Z (1) ||…||(m l 1 l )⊙Z (l) ||…||(m L 1 L )⊙Z (L) ||(m L+1 1 L+1 )⊙Z (L+1) ]

[0066] Step 2.3.5, construct a similarity graph by fusion representation:

[0067]

[0068] Among them, the normalized adjacency matrix is ​​generated using the softmax(·) function The learned adaptive graph Add to original image , to update the graph structure as follows:

[0069]

[0070] Here, ∈ is the equilibrium parameter.

[0071] Furthermore, the step 2.4, guiding the optimization training of the networks of the two paths simultaneously by the dual self-supervised clustering module, specifically comprises the following steps:

[0072] Step 2.4.1, for the i-th sample and the j-th cluster, use the Student's-t distribution as the kernel to measure the feature representation h i and cluster center vector u j Similarities between:

[0073]

[0074] Among them, h i Yes H (L) The i-th row, u j is the K-means initialization of the representation learned by the pre-trained autoencoder, v is the degrees of freedom of the Student's-t distribution; q ij represents the probability of assigning sample i to cluster j, that is, Q = [q ij ] is considered as the distribution of all samples;

[0075] Step 2.4.2, after obtaining the clustering result distribution Q, our goal is to optimize the feature representation by learning from the high confidence distribution; specifically, we hope that the feature representation is closer to the cluster center, thereby improving the aggregation of clusters. Therefore, calculate the target distribution P:

[0076]

[0077] Among them, f j =Σ i q ij ; In the target distribution P, each assignment in Q is squared and normalized so that the assignment has a higher confidence, resulting in the following objective function:

[0078]

[0079] Step 2.4.3, by minimizing the KL divergence between the distributions of Q and P, the target distribution P can help the autoencoder module learn a better representation of the clustering task, that is, making the feature representation closer around the cluster center. The adaptive graph autoencoder module also provides the cluster distribution Z; therefore, the distribution P can be used to supervise the distribution Z, resulting in the following objective function:

[0080]

[0081] Further, the specific implementation method of step 3 is as follows:

[0082] The attribute feature matrix and the adjacency matrix are input into a dual-path parallel adaptive deep attribute graph clustering network, and the network is trained by minimizing the loss of each module to obtain pseudo labels.

[0083] Furthermore, the overall target loss function of the network in the training phase consists of three parts, namely the reconstruction loss of the autoencoder module, the reconstruction loss of the adaptive graph autoencoder module, and the clustering loss of the dual self-supervised clustering module. The overall target loss function is as follows:

[0084]

[0085] in, is the reconstruction loss of the autoencoder module, is the reconstruction loss of the adaptive graph autoencoder module, is the clustering loss of the dual self-supervised clustering module, λ, α, and β are all used to balance the importance of reconstruction loss and clustering loss of the network, N is the number of samples, X is the attribute feature matrix, is the reconstructed attribute feature matrix, A is the adjacency matrix, is the reconstructed feature matrix, is the reconstructed adjacency matrix, q ij According to h i represents the probability of assigning sample i to cluster j, f j =∑ i q ij , z ij According to z i represents the probability of assigning sample i to cluster j.

[0086] Compared with the prior art, the present invention has the following advantages:

[0087] The social network analysis method based on deep attribute graph clustering proposed in the present invention can dynamically and adaptively update the graph structure information and feature information of the data, integrates multi-scale information at different levels, and adaptively integrates feature representations learned from two networks, which helps to reflect the complex relationships in the social network, enables it to learn better feature representations, and improves the accuracy and robustness of social network analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a network architecture diagram of a social network analysis method based on deep attribute graph clustering;

[0089] Figure 2 It is a flowchart of the social network analysis method based on deep attribute graph clustering. DETAILED DESCRIPTION

[0090] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0091] like Figure 1 As shown, a social network analysis method based on deep attribute graph clustering includes the following steps:

[0092] Step 1: Obtain a social network dataset, preprocess the dataset, and obtain an attribute feature matrix and an adjacency matrix; use the social network dataset provided by the social website or obtain user information (basic user information, published content, likes, comments), and relationship network (follow-up relationship) data on the social media platform through the API interface using web crawler technology to form a social network dataset, preprocess the dataset, and obtain an attribute feature matrix and an adjacency matrix.

[0093] Step 2, construct a dual-path parallel adaptive deep attribute graph clustering network, which includes an autoencoder module, an adaptive graph autoencoder module, an attention fusion module, and a dual self-supervised clustering module;

[0094] Step 2.1, reconstruct the attribute feature matrix through the autoencoder module to extract the potential attribute feature representation;

[0095] Step 2.1.1, pass the attribute feature matrix X into the autoencoder module, and learn useful attribute feature representations through the encoder;

[0096] Specifically, the attribute feature representation H learned by the (l)th layer in the encoder part is (l) :

[0097]

[0098] Among them, ψ(·) represents the activation function of the leaky rectified linear unit (Leaky Rectified LinearUnit), and denote the weight matrix and bias of the encoder layer (l) respectively;

[0099] H (0) Represented as attribute feature matrix X;

[0100] Step 2.1.2: pass the potential attribute feature representation into the decoder to reconstruct the attribute feature matrix X and obtain the reconstructed attribute feature matrix

[0101]

[0102] in, and denote the weight matrix and bias of the decoder layer (l), respectively;

[0103] H (2L) Represented as the reconstructed attribute feature matrix

[0104] The reconstruction loss of its autoencoder module is as follows:

[0105]

[0106] in, represents the reconstruction loss of the autoencoder module, N represents the number of samples, and x i represents the attribute characteristics of sample i, represents the reconstructed attribute features of sample i, represents the F-norm.

[0107] Step 2.2, reconstruct the attribute feature matrix X and the adjacency matrix A through the adaptive graph autoencoder module to capture the structural feature representation of the data;

[0108] Step 2.2.1: Pass the attribute feature matrix X and adjacency matrix A in the graph data into the adaptive graph autoencoder module, and learn the structural feature representation Z through the graph encoder. (l) :

[0109]

[0110] Among them, φ(·) represents the activation function of the Rectified Linear Unit. I is the identity matrix, and represents the weight matrix of the graph encoder. In addition, Z (0) Represented as attribute feature matrix X;

[0111] Step 2.2.2, starting from the second layer, use the fusion representation F of the previous layer (l) And the updated graph structure A (l) , its expression becomes as follows:

[0112]

[0113] After passing through the fifth layer of the graph encoder, the final joint feature representation Z is obtained;

[0114] Step 2.2.3, pass the final joint feature representation Z into the graph decoder to obtain the reconstructed feature matrix

[0115]

[0116] in, represents the weight matrix of the graph decoder. In addition, Z (L) Represents the reconstructed feature matrix

[0117] Step 2.2.4, reconstruct the attribute feature matrix X and the adjacency matrix A:

[0118]

[0119] Among them, the reconstructed adjacency matrix is the inner product of the final joint feature representation Z and the reconstructed feature matrix The sum of the inner products of

[0120] The reconstruction loss of its adaptive graph autoencoder module is as follows:

[0121]

[0122]

[0123]

[0124] Step 2.3, the feature representations learned by the autoencoder module and the encoders of the same layer in the adaptive graph autoencoder module are subjected to attention fusion through the attention fusion module to obtain a fused feature representation, and a new graph structure is obtained based on the fused feature representation learning; the newly constructed graph structure is fused with the original graph structure to update the original graph structure;

[0125] Step 2.3.1: Represent the potential attribute features learned by the autoencoder as H (l) Passed into the two-layer GNN to learn better attribute feature representation

[0126]

[0127] in, Represents the weight matrix in GNN;

[0128] Step 2.3.2: Represent the structural features learned by the adaptive graph autoencoder module as Z (l) and better attribute feature representation Connect together to get the cascade feature representation Y (l) :

[0129]

[0130] Among them, [,] represents the connection operation;

[0131] Step 2.3.3, using cascade feature representation, according to Z(l) and The importance of learning the corresponding weights and obtaining the fusion feature representation F (l) :

[0132] e=softmax(sigmoie(f(Y (l) ) / τ)

[0133] W = mean(e, dim = 0)

[0134]

[0135] Where f(·) represents a three-layer fully connected network, τ is the calibration factor; the learned weight W = [W1, W2]; in addition, the sigmoid(·) function and the calibration factor are to avoid assigning scores that are almost "1" to the most relevant feature representations;

[0136] Step 2.3.4, when fusing feature representations in the last layer, the feature representations of different scales learned in the previous layers are fused. Since the feature dimensions of different layers are different, the multi-scale features are aggregated in a cascade manner:

[0137] Z'=[Z (1) ||…||Z (l) ||…||Z (L) ||Z (L+1) ]

[0138] Among them, Z (L+1) express

[0139] Dynamically combine features of each scale through the attention mechanism:

[0140]

[0141] Where M = [m1||…||m l ||…||m L ||m L+1 ],m l It's Z (l) The attention coefficient, express Normalization, W s Represents the weight matrix, which is used to capture the relationship between features of different layers;

[0142] The learned attention coefficients are used to weight the multi-scale features and the fused features are represented as F (L) is formulated as:

[0143] F (L) =[(m111)⊙Z (1) ||…||(ml 1 l )⊙Z (l) ||…||(m L 1 L )⊙Z (L) ||(m L+1 1 L+1 )⊙Z (L+1) ]

[0144] Step 2.3.5, construct a similarity graph by fusion representation:

[0145]

[0146] Among them, the sigmoid(·) function is used to generate the normalized adjacency matrix The learned adaptive graph Add to original image , to update the graph structure as follows:

[0147]

[0148] Here, ∈ is the equilibrium parameter.

[0149] In step 2.4, the network is optimized by using a dual-path collaborative training method through a dual self-supervised clustering module.

[0150] Step 2.4.1, for the i-th sample and the j-th cluster, use the Student's-t distribution as the kernel to measure the feature representation h i and cluster center vector u j Similarities between:

[0151]

[0152] Among them, h i Yes H (L) The i-th row, u j is the K-means initialization of the representation learned by the pre-trained autoencoder, v is the degrees of freedom of the Student's-t distribution; q ij represents the probability of assigning sample i to cluster j, that is, Q = [q ij ] is considered as the distribution of all samples;

[0153] Step 2.4.2, after obtaining the clustering result distribution Q, calculate the target distribution P:

[0154]

[0155] Among them, f j =Σ i q ij; In the target distribution P, each assignment in Q is squared and normalized so that the assignment has a higher confidence, resulting in the following objective function:

[0156]

[0157] Step 2.4.3, by minimizing the KL divergence between the distributions of Q and P, the target distribution P can help the autoencoder module learn a better representation of the clustering task, that is, making the feature representation closer around the cluster center. The adaptive graph autoencoder module also provides the cluster distribution Z; therefore, the distribution P can be used to supervise the distribution Z, resulting in the following objective function:

[0158]

[0159] Step 3: Input the attribute feature matrix and the adjacency matrix into the dual-path parallel adaptive deep attribute graph clustering network, train the network by minimizing the loss of each module, obtain pseudo labels, and complete social network analysis.

[0160] The overall target loss function of the network in the training phase consists of three parts, namely the reconstruction loss of the autoencoder module, the reconstruction loss of the adaptive graph autoencoder module, and the clustering loss of the dual self-supervised clustering module. The overall target loss function is as follows:

[0161]

[0162] in, is the reconstruction loss of the autoencoder module, is the reconstruction loss of the adaptive graph autoencoder module, is the clustering loss of the dual self-supervised clustering module, λ, α, and β are all used to balance the importance of reconstruction loss and clustering loss of the network, N is the number of samples, X is the attribute feature matrix, is the reconstructed attribute feature matrix, A is the adjacency matrix, is the reconstructed feature matrix, is the reconstructed adjacency matrix, q ij According to h i represents the probability of assigning sample i to cluster j, f j =Σ i q ij , z ij According to z i represents the probability of assigning sample i to cluster j.

[0163] Reconstruction loss of the autoencoder module Used to learn latent attribute feature representation. Reconstruction loss of adaptive graph autoencoder module Used to learn structural feature representation of data. Clustering loss of dual self-supervised clustering module It is used to jointly optimize the network of the two paths for training, so as to achieve better cluster assignment.

[0164] In order to better demonstrate the clustering effect of this deep attribute graph clustering method, the clustering results of this method and the existing methods are compared on the ACM and DBLP datasets. The following is a detailed description of the two datasets:

[0165] ACM dataset: This dataset consists of 3025 articles collected from the ACM database, where nodes represent papers and if two papers have the same author, there is an edge between them. According to the research field, these articles can be divided into three categories: database, wireless communication, and data mining. The features are bag-of-words representations of the paper keywords. DBLP dataset: This dataset is collected from the DBLP integrated database system and contains 4058 authors. If two authors are co-authors, there is an edge between them. According to the research direction, these authors can be divided into four categories: database, data mining, machine learning, and information retrieval, and the categories to which the authors belong are marked by the conferences they attended. The author's features are also bag-of-words representations of keywords.

[0166] For the evaluation of clustering results, the evaluation indicators used in the present invention include Normalized Mutual Information (NMI), Accuracy (ACC), Adjusted Rand Index (ARI) and Macro-F1 index.

[0167] The present invention selects seven state-of-the-art clustering methods as baselines, which are introduced in detail as follows: AE: It is a two-stage deep clustering algorithm that performs K-means on the representations learned by the autoencoder. GAE&VGAE: It is a deep network embedding method based on the autoencoder. GCN is used as an encoder to utilize attribute information and structural information. DAEGC: It uses a variant of the graph attention network to fuse attribute features and structures. Unlike the previous stage-by-stage processing method, a unified two-stage approach is used to handle the attribute graph clustering task. SDCN: It is an end-to-end graph clustering method that introduces an autoencoder to help GCN learn better node representations. AGCN: This method proposes two specific fusion modules driven by the attention mechanism to learn representations, and uses a self-supervised module to train the clustering process. AGCC: This method performs DAE and GCN feature fusion and graph construction at each layer to improve representation capabilities.

[0168] The clustering results of this method and the above seven methods on the ACM and DBLP datasets are compared as follows:

[0169] Table 1: Clustering comparison results on the ACM dataset

[0170]

[0171] Table 2: Clustering comparison results on the DBLP dataset

[0172]

[0173] From the above two tables, it can be found that the clustering results of the deep attribute graph clustering method proposed in the present invention on the two data sets are better than those of the existing methods.

[0174] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A social network analysis method based on deep attribute graph clustering, characterized in that: The following steps are involved: Step 1: Obtain a social network data set, preprocess the data set, and obtain an attribute feature matrix and an adjacency matrix; Step 2, construct a dual-path parallel adaptive deep attribute graph clustering network, which includes an autoencoder module, an adaptive graph autoencoder module, an attention fusion module, and a dual self-supervised clustering module; Step 3: Input the attribute feature matrix and the adjacency matrix into the dual-path parallel adaptive deep attribute graph clustering network, train the network by minimizing the loss of each module, obtain pseudo labels, and complete social network analysis.

2. A social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The specific method of step 1 is: using the social network data set provided by the social networking website or the user information and relationship network data obtained on the social media platform using web crawler technology to form a social network data set, preprocessing the data set to obtain an attribute feature matrix and an adjacency matrix.

3. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The specific method of step 2 comprises the following steps: Step 2.1, reconstruct the attribute feature matrix through the autoencoder module to extract the potential attribute feature representation; Step 2.2, reconstruct the attribute feature matrix X and the adjacency matrix A through the adaptive graph autoencoder module to capture the structural feature representation of the data; Step 2.3, the feature representations learned by the autoencoder module and the encoders of the same layer in the adaptive graph autoencoder module are subjected to attention fusion through the attention fusion module to obtain a fused feature representation, and a new graph structure is obtained based on the fused feature representation learning; the newly constructed graph structure is fused with the original graph structure to update the original graph structure; In step 2.4, the network is optimized by using a dual-path collaborative training method through a dual self-supervised clustering module.

4. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The step 2.1, reconstructing the attribute feature matrix through the autoencoder module, and extracting the potential attribute feature representation specifically includes the following steps: Step 2.1.1, pass the attribute feature matrix X into the autoencoder module, and learn the potential attribute feature representation through the encoder; Specifically, the attribute feature representation H learned by the (l)th layer in the encoder part is (l) : Among them, ψ(·) represents the activation function of the leaky rectified linear unit (Leaky Rectified LinearUnit), and denote the weight matrix and bias of the encoder layer (l) respectively; H (0) Represented as attribute feature matrix X; Step 2.1.2: pass the potential attribute feature representation into the decoder to reconstruct the attribute feature matrix X and obtain the reconstructed attribute feature matrix in, and denote the weight matrix and bias of the decoder layer (l), respectively; H (2L) Represented as the reconstructed attribute feature matrix The reconstruction loss of its autoencoder module is as follows: in, represents the reconstruction loss of the autoencoder module, N represents the number of samples, and x i represents the attribute characteristics of sample i, represents the reconstructed attribute features of sample i, represents the F-norm.

5. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The step 2.2, reconstructing the attribute feature matrix X and the adjacency matrix A through the adaptive graph autoencoder module, and capturing the structural feature representation of the data specifically includes the following steps: Step 2.2.1: Pass the attribute feature matrix X and adjacency matrix A in the graph data into the adaptive graph autoencoder module, and learn the structural feature representation Z through the graph encoder. (l) : Among them, φ(·) represents the activation function of the Rectified Linear Unit. I is the identity matrix, and represents the weight matrix of the graph encoder. In addition, Z (0) Represented as attribute feature matrix X; Step 2.2.2, starting from the second layer, use the fusion representation F of the previous layer (l) And the updated graph structure A (l) , its expression becomes as follows: After passing through the fifth layer of the graph encoder, the final joint feature representation Z is obtained; Step 2.2.3, pass the final joint feature representation Z into the graph decoder to obtain the reconstructed feature matrix in, represents the weight matrix of the graph decoder. In addition, Z (L) Represents the reconstructed feature matrix Step 2.2.4, reconstruct the attribute feature matrix X and the adjacency matrix A: Among them, the reconstructed adjacency matrix is the inner product of the final joint feature representation Z and the reconstructed feature matrix The sum of the inner products of The reconstruction loss of its adaptive graph autoencoder module is as follows:

6. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The step 2.3, through the attention fusion module, performs attention fusion on the feature representations learned by the encoders of the same layer in the autoencoder module and the adaptive graph autoencoder module to obtain a fused feature representation, and learns a new graph structure through the fused feature representation, which is fused with the original graph structure, thereby updating the graph structure, specifically including the following steps: Step 2.3.1: Represent the potential attribute features learned by the autoencoder as H (l) Passed into the two-layer GNN to learn better attribute feature representation in, Represents the weight matrix in GNN; Step 2.3.2: Represent the structural features learned by the adaptive graph autoencoder module as Z (l) and better attribute feature representation Connect together to get the cascade feature representation Y (l) : Among them, [,] represents the connection operation; Step 2.3.3, using cascade feature representation, according to Z (l) and The importance of learning the corresponding weights and obtaining the fusion feature representation F (l) : e=softmax(sigmoid(f(Y (l) ) / τ) W = mean(e, dim = 0) Where f(·) represents a three-layer fully connected network, τ is the calibration factor; the learned weight W = [W1, W2]; in addition, the sigmoid(·) function and the calibration factor are used to avoid assigning scores that are almost "1" to the most relevant feature representations; Step 2.3.4, when fusing feature representations in the last layer, the feature representations of different scales learned in the previous layers are fused. Since the feature dimensions of different layers are different, the multi-scale features are aggregated in a cascade manner: Z'=[Z (1) ||…||From (l) ||…||From (L) ||From (L+1) ] Among them, Z (L+1) express Dynamically combine features of each scale through the attention mechanism: Where M = [m1||…||m l ||…||m L ||m L+1 ],m l It's Z (l) The attention coefficient, express Normalization, W s Represents the weight matrix, which is used to capture the relationship between features of different layers; The learned attention coefficients are used to weight the multi-scale features and the fused features are represented as F (L) is formulated as: F (L) =[(m111)⊙Z (1) ||…||(m l 1 l )⊙Z (l) ||…||(m L 1 L )⊙Z (L) ||(m L+1 1 L+1 )⊙Z (L+1) ] Step 2.3.5, construct a similarity graph by fusion representation: Among them, the normalized adjacency matrix is ​​generated using the softmax(·) function The learned adaptive graph Add to original image , to update the graph structure as follows: Here, ∈ is the equilibrium parameter.

7. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The step 2.4, guiding the networks of the two paths to perform optimization training simultaneously through the dual self-supervised clustering module, specifically includes the following steps: Step 2.4.1, for the i-th sample and the j-th cluster, use the Student's-t distribution as the kernel to measure the feature representation h i and cluster center vector u j Similarities between: Among them, h i Yes H (L) The i-th row, u j is the K-means initialization of the representation learned by the pre-trained autoencoder, v is the degrees of freedom of the Student's-t distribution; q ij represents the probability of assigning sample i to cluster j, that is, q = [q ij ] is considered as the distribution of all samples; Step 2.4.2, after obtaining the clustering result distribution Q, calculate the target distribution P: Among them, f j =∑ i q ij ; In the target distribution P, each assignment in Q is squared and normalized so that the assignment has a higher confidence, resulting in the following objective function: Step 2.4.3, by minimizing the KL divergence between the distributions of Q and P, the target distribution P can help the autoencoder module learn a better representation of the clustering task, that is, making the feature representation closer around the cluster center. The adaptive graph autoencoder module also provides the cluster distribution Z; therefore, the distribution P can be used to supervise the distribution Z, resulting in the following objective function:

8. The social network analysis method based on deep attribute graph clustering according to claim 1, characterized in that: The specific operation of step 3 is: the overall objective loss function of the network in the training phase includes three parts, namely, the reconstruction loss of the autoencoder module, the reconstruction loss of the adaptive graph autoencoder module and the clustering loss of the dual self-supervised clustering module; The overall objective loss function is as follows: in, is the reconstruction loss of the autoencoder module, is the reconstruction loss of the adaptive graph autoencoder module, is the clustering loss of the dual self-supervised clustering module, λ, α, and β are all used to balance the importance of reconstruction loss and clustering loss of the network, N is the number of samples, X is the attribute feature matrix, is the reconstructed attribute feature matrix, A is the adjacency matrix, is the reconstructed feature matrix, is the reconstructed adjacency matrix, q ij According to h i represents the probability of assigning sample i to cluster j, f j =Σ i q ij , z ij According to z i represents the probability of assigning sample i to cluster j.

Citation Information

Cited By

  • Community discovery method based on attention fusion and neighborhood information enhancement

    CN120030375A

  • A Community Discovery Method Based on Attention Fusion and Enhanced Neighborhood Information

    CN120030375B