A social network propagation backbone structure discovery method based on correlation analysis

Abstract

The application realizes a social network propagation backbone structure discovery method based on correlation analysis. The method is composed of a user importance measurement module and a propagation backbone structure learning module. First, input the social media specific topic data social network library, and extract the user representation of the social network. The user importance measurement module inputs the user representation. Then, select the top-k key users in the social network and the key user groups composed of the surrounding neighbor users of the key users, and input the key user groups into the propagation backbone structure learning module. The propagation backbone structure learning module includes a user correlation relationship submodule and a user interaction submodule, and outputs the propagation backbone structure. Thus, the correlation information of the surrounding user groups can be obtained while the key users in the social network are mined, and the effect of obtaining the information propagation backbone network under the specific topic of the social network is achieved.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, and in particular to a method for discovering the propagation backbone structure of social networks based on association analysis. Background Technology

[0002] In recent years, with the development and popularization of internet technology, more and more netizens are using social networks for communication, gradually changing the online public opinion ecosystem and promoting the transfer of public discourse power. In social networks, information spreads through user-to-user communication behaviors along the social networks formed by the relationships between users. The roles played by each individual and different groups in the network structure and function vary greatly. Key users are special individuals who can significantly influence the network structure and function formed during information dissemination. For example, influential Weibo users can accelerate the spread of facts or rumors on social networks. Therefore, in order to better guide online public opinion, it is necessary to accurately identify key users and related user groups from the massive user base of social networks, and to use the communication backbone structure formed by these users and their interrelationships to capture and predict the development trend of netizens' public opinion.

[0003] One important research direction in mining the backbone structure of communication is identifying opinion leaders in networks, who play a crucial mediating or filtering role in the formation of mass communication effects. Many social network opinion leader identification algorithms have been proposed. Classic methods based on network topology include degree centrality, betweenness centrality, proximity centrality, and eigenvector centrality. The most intuitive and basic method is degree centrality, which is simple and effective in calculation, but it only reflects local information in complex networks and does not consider the individual characteristics of nodes or the mutual influence between nodes. Furthermore, each evaluation method is based on statistics and has its own limitations, making it unsuitable for all types of networks. Other methods, such as cluster analysis and the PageRank algorithm, identify opinion leader nodes in networks. These traditional methods have the advantage of simple models and can converge relatively quickly when there are many samples. However, the clustering scale cannot be controlled, and nodes within the same category may not all have similar characteristics. The PageRank algorithm assumes that nodes have the same jump probability, resulting in non-unique ranking results. Some methods based on neural network models often focus on mining key user sets, but these methods cannot fully explore the backbone structure formed by key users and their neighbors, as well as the information of key user groups at different levels.

[0004] To address the problem that existing research cannot fully utilize social network information, leading to inaccurate identification of key users and incomplete discovery of user relationships, thus affecting the discovery of the main structure of social network propagation, this invention proposes to construct a social network based on user interactions under specific social media topics. By enriching the features of nodes through the inherent and social attributes of users in the social network, and comprehensively measuring the importance of a user in the social network from both local and global importance, the top-k key users are obtained. The reconstructed interaction relationships between users are obtained through user characteristics and topological distance features, thereby realizing the task of discovering the main structure of social network propagation. Summary of the Invention

[0005] To this end, this invention first proposes a method for discovering the backbone structure of social network propagation based on association analysis. It consists of a user importance measurement module and a propagation backbone structure learning module. First, a social network database of specific social media topics is input. If the topic does not exist, the process ends directly. If it exists, the information propagation network corresponding to the topic is extracted, a social network is constructed, and then the user representation of the social network is extracted.

[0006] The user importance measurement module inputs the user representation into the user local importance submodule and the user global importance submodule respectively, and obtains the importance score of each user based on the user local importance and user global importance; then, the importance scores of each user are sorted.

[0007] Then, key user groups consisting of key users in the top-k social networks and their surrounding neighbor users are selected and input into the propagation backbone structure learning module.

[0008] The propagation backbone structure learning module includes a user association submodule and a user interaction submodule. Based on the key users and their surrounding neighbor users selected from the top-k social networks in the user importance measurement module, the probability of a connection between each pair of users is obtained through the user association submodule and the user interaction submodule. The association relationships between these users are relearned to obtain a new adjacency matrix, which constitutes the propagation backbone structure.

[0009] Specifically, the method for extracting the information dissemination network corresponding to the topic involves first obtaining topic data for a preset time period based on topic keywords, and then using web crawling technology to obtain relevant content based on the pre-set time period and keywords. If the content is related by forwarding or commenting, then there is a connection between the corresponding two user nodes. The user's social attributes and self-attributes are extracted as the initial user representation of the user node.

[0010] The user importance measurement module takes as input the feature matrix X of user nodes in the social network and the relationship matrix A between the user nodes. Based on the constructed social network structure and the user node features obtained from the social media platform, it obtains the low-dimensional representation output of each node based on the graph convolutional neural network, calculates the correlation between the node and its neighboring nodes, and the information correlation degree of the user group in the entire social network, and outputs the importance score of each node.

[0011] The implementation of user local importance is as follows: Given a graph G = (A, X), define A as the adjacency matrix of the graph, X as the feature matrix of the graph, and design a discriminator T. θ Analyzing the correlation between each node and its corresponding r-order neighbors can reflect the node's ability to express neighbor information:

[0012]

[0013] Where N represents the set of all nodes in the graph, N r,v Let f represent the set of all neighbors of node v of order r, u represent any node in N that is not node v, and f represent the set of all neighbors of node v of order r. v This represents the feature vector of node v. Let represent the sum of the characteristics of all neighboring nodes of order r of node v. It is a discriminator; MLP stands for Multilayer Perceptron; σ represents the activation function. NMI The first part reflects the similarity between a node and its r-order neighbors, while the second part reflects the differences between a node and any other node's r-order neighbors. More important nodes need to describe their neighbor information to the greatest extent possible while minimizing similarity to the neighbor information of any other node. In training, this translates to maximizing T. θ Further considering the importance of different nodes in the graph, the local importance score of node v is designed primarily by calculating each node's ability to describe its neighbors, i.e., the discriminability of a node to local information.

[0014]

[0015] The implementation of the user group importance score is as follows: The r-order neighbor region of a node is considered a subgraph within a larger graph, with the node itself at the center of the subgraph. The node and link information contained within the subgraph constitute the propagation structure during information transmission. A user group importance score is designed to measure the importance of different subgraphs within the overall graph. This is primarily achieved by designing a discriminator T. δ The correlation between the r-order neighbor information of a node and the entire graph reflects the ability of the subgraph structure centered on each node to describe global information. Specifically, it is defined as:

[0016]

[0017] Where N represents the set of all nodes in the graph, N r,v Let v represent the set of all neighboring nodes of order r of node v, and u represent any node in N that is not node v. f represents the sum of the characteristics of all r-order neighboring nodes of node v. G Feature vectors representing social networks It is a discriminator; MLP stands for Multilayer Perceptron; σ represents the activation function. GMI The first part reflects the difference between a node-centric subgraph and global information, while the second part reflects the difference between a subgraph centered on any node and global information. More important subgraph structures need to describe global information to the greatest extent possible; in training, this translates to maximizing T. δ Further considering the importance of different subgraph structures within the graph, primarily by calculating the ability of each subgraph to describe global information, i.e., the discriminability of the subgraph to global information, a global importance score for nodes centered on node v is designed:

[0018]

[0019] The top-k key users can be selected in the final network layer based on user importance scores. However, selecting user nodes solely based on this information lacks the integrity of the graph structure, hindering further message passing. Therefore, a backbone structure learning layer is designed to relearn a new structure based on the existing nodes. The input consists of user nodes and their low-dimensional feature representations output by the user importance measurement module, taking into account user interactions and the relationships between each pair of users.

[0020] The propagation backbone structure learning module includes a user relationship learning sub-module and a user interaction analysis sub-module.

[0021] The user association learning submodule is implemented as follows: It utilizes a single-layer neural network to learn the similarity, or potential connection score, between two nodes, defined as: E(i,j)=σ(w T (x i ,x j )), where w T It is a learnable weight vector, and σ represents the activation function.

[0022] The user interactivity analysis submodule is implemented as follows: Considering two scenarios, one is the shortest distance between two nodes in the real-world graph; the greater the shortest distance, the less direct information interaction occurs between the two nodes. The other scenario utilizes the embedding space of nodes; two nodes may share the same information. A closer distance in the embedding space symbolizes potential information interaction between the nodes during information propagation. The Euclidean distance between two nodes in the embedding space is designed as a measure of the likelihood of information interaction. The user interactivity matrix A(i,j) is defined as follows: D ij d(v) represents the shortest distance between two nodes in a social network. i ,v j d(v) represents the Euclidean distance between two nodes in the latent space, and is calculated as: d(v) i ,v j = Euclidean(x) vi ,x vj ), where x vi Represents node v i Low-dimensional vector representation of features, Both ρ and ρ are hyperparameters.

[0023] To make similarity scores comparable across different nodes, the softmax function is used for regularization:

[0024]

[0025] Where ReLU(·) = max(0,·), and μ is a hyperparameter. This is used to obtain the new subgraph structure. Then, a new adjacency matrix is ​​used when aggregating neighbor information in the next graph convolutional layer. The output of the final layer of the model is the backbone structure of social network propagation.

[0026] The overall model completes the task of discovering the network propagation backbone structure in two phases: the training phase and the prediction phase.

[0027] During the training phase: For each specific topic event, multiple subgraphs consistent with the original social network data distribution are obtained through sampling and used as the training set. The labels of these subgraphs are consistent with the topic. For multiple topics, multiple subgraph sets with different labels can be obtained. These are fed into the model for training, and the parameter model is continuously optimized to obtain the trained social network propagation backbone structure discovery model.

[0028] In the prediction phase: Input a social network graph of a specific topic event, and use a trained model to calculate the corresponding key participants and the main structure of social network propagation.

[0029] The technical effects to be achieved by this invention are as follows:

[0030] By uncovering the backbone structure of multi-level information dissemination, this invention can not only identify key users in the spread of topics on social networks, but also obtain different relationships among users at multiple levels.

[0031] This invention proposes an end-to-end propagation backbone structure learning method based on a graph convolutional neural network model.

[0032] The method proposed in this invention has the following characteristics:

[0033] 1. This invention is the first to propose the problem of learning the backbone structure in social network propagation. While mining key users in social networks, it can also obtain the correlation information of surrounding user groups and obtain the backbone network of information propagation under specific topics in social networks.

[0034] 2. In the process of learning the backbone structure, two discriminators were designed to consider the influence of two different factors on the importance of users under different topics: local user importance and user group importance. At the same time, the structure of different backbone networks can be re-acquired through the relationship between users and user interaction. Attached Figure Description

[0035] Figure 1 A methodology for discovering the backbone structure of social network propagation based on correlation analysis; Detailed Implementation

[0036] The following are preferred embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0037] This invention proposes a method for discovering the propagation backbone structure of social networks based on association analysis.

[0038] This invention provides a method for discovering the propagation backbone structure of a social network based on multi-level user group association learning. While mining key users in a social network, it can also obtain association information of surrounding user groups and the propagation backbone structure, including two modules: a user importance measurement module and a propagation backbone structure learning module.

[0039] Specifically, the system first retrieves Weibo posts within a preset time period based on topic keywords. Then, using web scraping technology, it obtains relevant Weibo content based on the pre-set time period and keywords. Since the Weibo posts are related by reposts or comments, there is an edge connecting the corresponding two nodes.

[0040] The system searches the social network database based on the specific topic in the input Weibo post. If the topic does not exist, the process ends directly. If the topic exists, the information dissemination network corresponding to the topic is retrieved. Then, the initial user representation vector is input into the user importance measurement module.

[0041] The system searches a social network database based on a specific topic in the input Weibo post. If the topic does not exist, the process ends directly. If it does exist, the information propagation network corresponding to the topic is retrieved. Then, the initial user features and adjacency matrix are input into a graph convolutional layer to obtain the user's representation vector.

[0042] Next, the importance of a user in a specific topic event is calculated, and the resulting user representation vector is input into the user importance measurement module. The user importance measurement module includes a user local importance submodule and a user group importance submodule, which calculates the importance of each user in different topics based on the user local importance and the user group global importance.

[0043] User local importance is determined by designing a discriminant T. θ Analyzing the correlation between each node and its corresponding r-order neighbors can reflect the node's ability to express neighbor information:

[0044]

[0045] Where N represents the set of all nodes in the graph, N r,v Let f represent the set of all neighbors of node v of order r, u represent any node in N that is not node v, and f represent the set of all neighbors of node v of order r. v This represents the feature vector of node v. Let represent the sum of the characteristics of all neighboring nodes of order r of node v. It is a discriminator; MLP stands for Multilayer Perceptron; σ represents the activation function. NMI The first part reflects the similarity between a node and its r-order neighbors, while the second part reflects the differences between a node and any other node's r-order neighbors. More important nodes need to describe their neighbor information to the greatest extent possible while minimizing similarity to the neighbor information of any other node. In training, this translates to maximizing T. θ Further considering the importance of different nodes in the graph, the local importance score of node v is designed primarily by calculating each node's ability to describe its neighbors, i.e., the discriminability of a node to local information.

[0046]

[0047] The user group importance design is as follows: A node's r-order neighbor region is considered a subgraph within a larger graph, with the node itself at the center of the subgraph. The node and link information contained within the subgraph constitute the propagation structure during information transmission. A user group importance score is designed to measure the importance of different subgraphs within the overall graph. This is primarily achieved by designing a discriminator T. δ The correlation between the r-order neighbor information of a node and the entire graph reflects the ability of the subgraph structure centered on each node to describe global information. Specifically, it is defined as:

[0048]

[0049] Where N represents the set of all nodes in the graph, N r,v Let represent the set of all neighboring nodes of order r of node v, and let u represent any node in N that is not node v. f represents the sum of the characteristics of all r-order neighboring nodes of node v. G Feature vectors representing social networks It is a discriminator; MLP stands for Multilayer Perceptron; σ represents the activation function. GMI The first part reflects the difference between a node-centric subgraph and global information, while the second part reflects the difference between a subgraph centered on any node and global information. More important subgraph structures need to describe global information to the greatest extent possible; in training, this translates to maximizing T. δ Further considering the importance of different subgraph structures within the graph, primarily by calculating the ability of each subgraph to describe global information, i.e., the discriminability of the subgraph to global information, a global importance score for nodes centered on node v is designed:

[0050]

[0051] Combining user local importance score and user group importance score s = s local +s global Rank each user by importance, and select the top-k key users in the social network and their surrounding users as the key user group: idx = top rank(s,k) , among which top rank The function returns the indices of the k most important nodes, outputting the characteristics of k important users and their corresponding indices. k is a predefined value representing the number of nodes output. These k selected nodes and their surrounding neighborhoods constitute the k key user groups in the social network.

[0052] After selecting the nodes to be retained, the rows corresponding to the unselected nodes in the node feature matrix need to be removed to obtain a new feature representation H. Simultaneously, the corresponding graph adjacency matrix also changes. Nodes selected solely based on the above information do not have direct edges in the original graph, but they possess similar information, indicating potential information propagation. Connecting nodes solely based on existing edges would lose the integrity of the graph structure information, hindering further message transmission. Therefore, a backbone structure learning module is designed to relearn a new structure based on the existing sampled nodes.

[0053] The propagation backbone structure learning module includes a user relationship submodule and a user interactivity submodule. The input consists of the feature information and index information of k important users. Based on the user groups composed of the top-k key users in the social network selected from the user importance measurement module and their surrounding neighbor users, the module relearns the relationships between users and outputs a newly learned adjacency matrix.

[0054] The user association learning submodule is implemented as follows: It utilizes a single-layer neural network to learn the similarity, or potential connection score, between two nodes, defined as: E(i,j)=σ(w T (x i ,x j )), where w T It is a learnable weight vector, and σ represents the activation function.

[0055] The user interaction analysis submodule is implemented as follows: Consider two scenarios. One is the shortest distance between two nodes in the real-world graph; the greater the shortest distance, the less direct information interaction occurs between the two nodes. The other scenario utilizes the embedding space of nodes; two nodes may share the same information. A closer distance in the embedding space symbolizes potential information interaction during information propagation. The Euclidean distance between two nodes in the embedding space is designed as a measure of the likelihood of information interaction. The user interaction matrix A(i,j) is defined as follows: D ij d(v) represents the shortest distance between two nodes in a social network. i ,v j d(v) represents the Euclidean distance between two nodes in the latent space, and is calculated as: d(v) i ,v j = Euclidean(x) vi ,x vj ), where x vi Represents node v i Low-dimensional vector representation of features, Both ρ and ρ are hyperparameters.

[0056] To make similarity scores comparable across different nodes, the softmax function is used for regularization:

[0057]

[0058] Where ReLU(·) = max(0,·), and μ is a hyperparameter. This is used to obtain the new subgraph structure. Then, a new adjacency matrix is ​​used when aggregating neighbor information in the next graph convolutional layer. The output of the final layer of the model is the backbone structure of social network propagation, characterized by graph representation vector H. G .

[0059] To train the model, a graph classification task is used to construct the objective function. The feature representation of the entire graph is input into a three-layer perceptron to predict the graph label. A cross-entropy loss function for multi-class classification is designed for model training. After parameter tuning, the trained model is implemented and an interface is encapsulated. The overall loss function is designed as follows:

[0060] y' = softmax(MLP(H G ))

[0061]

[0062] Where MLP represents a multilayer perceptron, H G is the graph representation vector of the propagation backbone structure, M represents the number of labeled topic events, c represents the number of social network categories, and α and β are hyperparameters.

[0063] To achieve the above objectives, the present invention also provides a social network propagation backbone structure discovery device based on multi-level user group association learning, comprising: a processor and a memory; the memory is used for the method of discovering the social network propagation backbone structure based on multi-level user group association learning; the processor is connected to the memory and is used to execute the method stored in the memory, so that the social network propagation backbone structure discovery device based on multi-level user group association learning executes the above-mentioned functional modules.

[0064] This invention provides a method and apparatus for discovering the backbone structure of social network propagation based on multi-level user group association analysis, using methods from the field of data mining. It can discover key users in social networks while obtaining association information of surrounding user groups and multi-level information propagation backbone networks under different categories of topics. It consists of two modules: a user importance measurement module and a propagation backbone structure learning module.

[0065] The user importance measurement module includes a user local importance submodule and a user global importance submodule. First, it searches the social network database based on the specific topic in the input microblog. If the topic does not exist, the process ends directly. If it exists, it retrieves the information propagation network corresponding to the topic. Then, it inputs the initial user representation vector into the user importance measurement module and obtains the importance score of each user based on the user local importance and user global importance. After that, it sorts the importance scores of each user and selects the key user group consisting of the top-k key users in the social network and their surrounding neighbor users.

[0066] The propagation backbone structure learning module includes a user relationship submodule and a user interaction submodule. Based on the key users and their surrounding neighbor users selected from the top-k social networks in the user importance measurement module, the probability of a connection between each pair of users is obtained through the user relationship submodule and the user interaction submodule. The relationships between these users are then relearned to obtain a new adjacency matrix, which constitutes the propagation backbone structure.

[0067] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for discovering the backbone structure of social network propagation based on association analysis, characterized in that: It consists of a user importance measurement module and a propagation backbone structure learning module. First, it inputs a social network database of specific social media topics. If the topic does not exist, it ends directly. If it exists, it extracts the information propagation network corresponding to the topic, constructs a social network, and then extracts the user representation of the social network. The user importance measurement module inputs the user representation into the user local importance submodule and the user global importance submodule respectively, and obtains the importance score of each user based on the user local importance and user global importance; then, the importance scores of each user are sorted. Then, key user groups consisting of key users in the top-k social networks and their surrounding neighbor users are selected and input into the propagation backbone structure learning module. The propagation backbone structure learning module includes a user association submodule and a user interaction submodule. Based on the key users and their surrounding neighbor users selected in the top-k social networks from the user importance measurement module, the probability of a connection between each pair of users is obtained through the user association submodule and the user interaction submodule. The association between these users is relearned to obtain a new adjacency matrix, which constitutes the propagation backbone structure. The method for extracting the information dissemination network corresponding to the topic is as follows: First, topic data for a preset time period is obtained based on topic keywords. Relevant content is obtained using web crawling technology through the preset time period and keywords. If the content is related to forwarding or commenting, there is a connection between the two corresponding user nodes. The user's social attributes and self-attributes are extracted as the initial user representation of the user node. The user importance measurement module takes as input the feature matrix X of user nodes in the social network and the relationship matrix A between the user nodes. Based on the constructed social network structure and the user node features obtained from the social media platform, it obtains the low-dimensional representation output of each node based on the graph convolutional neural network, calculates the correlation between the node and its neighboring nodes, and the information correlation of the user group in the entire social network, and outputs the importance score of each node. The implementation of the user local importance calculation submodule is as follows: given a graph ,definition Represents the adjacency matrix of a graph. To represent the feature matrix of a graph, design a discriminant. Analyze each node and its corresponding Correlation between order neighbors: in, Represents the set of all nodes in the graph. Represents a node of The set of all neighboring nodes of the order, express China-Africa node any node, Represents a node eigenvectors, Represents a node of The sum of the characteristics of all neighboring nodes of order 1. It is a discriminator; MLP stands for Multilayer Perceptron. This represents the activation function, which maximizes the activation value during training. ; Node design is achieved by calculating each node's ability to describe its neighbors, i.e., the discriminability of a node to local information. Local importance score: ; The implementation method of the user global importance calculation submodule is as follows: The node's... The neighbor region is considered as a subgraph in the larger graph, with the node itself as the center of the subgraph. A discriminator is designed to address this. compute nodes The degree of correlation between the neighbor information and the whole graph is specifically defined as: in, Represents the set of all nodes in the graph. Represents a node of The set of all neighboring nodes of the order, express China-Africa node any node, Represents a node of The sum of the characteristics of all neighboring nodes of order 1. Feature vectors representing social networks It is a discriminator; MLP stands for Multilayer Perceptron. This represents the activation function, which needs to be maximized during training. By calculating the ability of each subgraph to describe global information, i.e., the discriminability of the subgraph to global information, a node-based design is developed. Global importance score of the central node: .

2. The method for discovering the backbone structure of social network propagation based on association analysis as described in claim 1, characterized in that: The specific method for generating the key user group is as follows: select the top-k key users in the last layer of the network based on the user importance metric score, design a backbone structure learning layer, relearn a new structure based on the existing nodes, and input the user nodes and their low-dimensional feature representations output by the user importance metric module.

3. The method for discovering the backbone structure of social network propagation based on association analysis as described in claim 2, characterized in that: The propagation backbone structure learning module extracts key participants from key users, inputs user relationship learning sub-module and user interaction analysis sub-module; The user association learning submodule is implemented as follows: It utilizes a single-layer neural network to learn the similarity, or potential connection score, between two nodes, defined as: ,in It is a learnable weight vector. Indicates the activation function; The user interaction analysis submodule is implemented by defining a user interaction matrix. in , This represents the shortest distance between two nodes in a social network. The Euclidean distance between two nodes in the latent space is calculated as follows: ,in Represents a node Low-dimensional vector representation of features, and These are all hyperparameters; To make similarity scores comparable across different nodes, the softmax function is used for regularization: in, , As a hyperparameter, in obtaining the new subgraph structure Then, when aggregating neighbor information in the next graph convolutional layer, the adjacency matrix of the backbone structure is updated using the new adjacency matrix. The output of the last layer of the model is the backbone structure of social network propagation.

Images