An expert influence maximization method based on graph convolution and DBSCAN algorithm

Abstract

The application discloses an expert influence maximization method based on graph convolution and DBSCAN algorithm, and relates to the field of influence maximization research under a citation network. First, each expert is regarded as a node, and a citation network is constructed according to the citation relationship between the experts, a graph convolution neural network is used to extract features of the citation network, and a feature representation of each expert node is obtained; then, a matrix is clustered by using a DBSCAN algorithm, different communities are obtained, and a spectral clustering method is used to divide the expert nodes into different clustering clusters, and each clustering cluster represents an expert group with high influence; finally, the expert nodes are sorted according to influence scores, and thus a list of experts with the most influence is obtained. Compared with the prior art, the method provided in the application can better identify experts with high influence in the citation network, and is a relatively excellent influence maximization algorithm.

Description

Technical Field

[0001] This invention belongs to the field of maximizing citation network influence technology, and specifically relates to a method for maximizing expert influence based on graph convolution and DBSCAN algorithm. Background Technology

[0002] Influence Maximization (IM) is a task designed to determine how to maximize the influence of a specific node or user in a social network or other type of graph network. With the rise of social media and online communities, influence maximization has become a significant problem because the influence of individuals and organizations in these networks can have a substantial impact on information dissemination and advertising. Influence maximization involves identifying a small set of influential nodes in a complex network to maximize the number of activated nodes. Due to key issues such as the accuracy, stability, and time complexity of seed set selection, numerous studies and algorithms have been proposed over the past decade. Most influence maximization algorithms have encountered challenges, such as the lack of optimal seed node selection, inappropriate influence propagation, and high time complexity.

[0003] The question of maximizing the influence of a specific expert within an expert citation network. In professional fields, expert influence is crucial for enhancing an organization's research capabilities, accelerating technological innovation, and helping individuals succeed in their careers.

[0004] Many methods have been proposed to address the problem of maximizing influence. One common approach is the CitationRank algorithm, a network analysis-based method for assessing expert influence. It evaluates an expert's influence by analyzing their citation relationships. The basic idea of ​​CitationRank is that an expert's influence depends on the number of times their papers are cited and the influence of the experts citing them. Specifically, for each paper, CitationRank defines its influence as the number of citations it receives; for each expert, its influence is defined as the sum of the influence of all their papers, where the influence of each paper is determined by both its citation count and the influence of the experts citing it. However, CitationRank's calculations rely solely on the number of citations, which is very limited. It cannot accurately predict expert influence and may underestimate the influence of experts in emerging fields.

[0005] Therefore, this paper proposes a method for maximizing expert influence based on graph convolution and the DBSCAN algorithm. This method extracts feature attributes including the number of papers published, the number of citations, and the publication date. It helps identify and maximize the influence of the most important nodes or users in the network, and more accurately predicts expert influence. By combining graph convolution and spectral clustering, we can more accurately predict and identify the most influential individuals and organizations in social networks, and provide them with better opportunities to expand their influence and engagement. Summary of the Invention

[0006] Purpose of the invention: In view of the problems pointed out in the background art, the present invention discloses an influence maximization method based on graph convolution and DBSCAN algorithm. By combining graph convolutional networks and DBSCAN algorithm, a more accurate and reliable method is provided to predict and identify the most influential experts in expert citation networks.

[0007] Technical Solution: This invention discloses a method for maximizing expert influence based on graph convolution and the DBSCAN algorithm, comprising the following steps:

[0008] S1. Construct an undirected graph based on the citation data, where each node corresponds to an expert and an edge represents a citation relationship between two experts. Extract the feature attributes of the experts themselves and construct a citation network. The feature attributes include the number of papers published, the number of citations of the papers, and the publication time of the papers.

[0009] S2. Use a graph convolutional neural network to extract features from the citation network to obtain the feature matrix representation of each expert;

[0010] S3. Perform spectral decomposition on the feature matrix of the experts using the DBSCAN algorithm to obtain clustering results: Perform symmetric normalization on the feature matrix, and then perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. Divide the experts into different clusters based on the eigenvectors.

[0011] S4. Calculate the influence score of each expert based on their different clusters, rank the experts according to their influence scores, and select the most influential expert sequence.

[0012] Further, step S1 specifically includes:

[0013] S1.1 Extract the expert node ID with the largest connected component from the citation dataset, as well as the expert attributes such as the number of published articles, the number of cited articles, and the publication time.

[0014] S1.2 Extract the reference relationships between expert nodes;

[0015] S1.3. Each expert is treated as a node, and a citation network is constructed based on the citation relationships between them.

[0016] Furthermore, the specific steps of step S2 are as follows:

[0017] S2.1, Perform data preprocessing: Construct an adjacency matrix A and a feature matrix X, where A is a... A is a symmetric matrix, where N is the number of nodes, and A is the th node. Each element represents whether node i and node j are connected by an edge; X is a... The matrix is ​​a matrix where each row represents the feature vector of a node, namely the number of expert articles, the expert's citation count, and the article creation time. F is the dimension of the feature vector.

[0018] S2.2. Perform a linear transformation on the feature matrix. We need to perform a linear transformation on the input feature matrix X to obtain a new feature matrix H, as shown in the following formula: C is the weight matrix to be learned, which performs a linear transformation on the input feature matrix X, mapping the feature vector of each node in the input to a new vector; the size of the weight matrix C is... ,in This represents the dimension of the input feature matrix. This indicates the dimension of the output feature matrix, where each element is a parameter that needs to be learned.

[0019] S2.3 Normalize the adjacency matrix by dividing the adjacency matrix A by the node degree matrix D;

[0020] S2.4 Perform convolution operation: After normalizing the adjacency matrix, perform a convolution operation between it and the feature matrix; define a filter, also called kernel W, with a size of... Where k represents the size of the filter, and F represents the dimension of the eigenvector of each node in the feature matrix; the normalized matrix after normalizing the filter W and the adjacency matrix. Perform convolution to obtain a new feature matrix Z:

[0021] .

[0022] Furthermore, the specific steps of step S3 are as follows:

[0023] S3.1 Normalize the new feature matrix obtained by graph convolution to obtain matrix Y, that is, subtract the mean of the feature column from each feature value, and then divide by the standard deviation of the feature column.

[0024] S3.2 Perform DBSCAN-based spectral clustering on Y to obtain different communities. In DBSCAN-based spectral clustering, we treat each row of matrix Y as a data point, that is, the vector representation of each node in the new feature space. We use these vectors as input and apply the DBSCAN algorithm for clustering. The DBSCAN algorithm determines which cluster each point belongs to based on the density around it.

[0025] Furthermore, the specific steps of step S4 are as follows:

[0026] S4.1. Using the DBSCAN algorithm, different communities are obtained, and each community contains a group of closely connected nodes;

[0027] S4.2 For each community, calculate the sum of the degrees of all its nodes to obtain the community degree. ;

[0028] S4.3 For each node i, calculate the sum of the number of edges connecting it to other nodes to obtain the node degree. ;

[0029] S4.4, Calculation Node Influence Index ;

[0030] S4.5 Sort all nodes according to their influence index and select the node with the greatest influence as the important node.

[0031] Beneficial effects:

[0032] (1) More accurate prediction of expert influence: This invention combines graph convolutional networks (GCNs) and the DBSCAN algorithm to more accurately predict and identify the most influential experts in expert social networks. GCNs can learn the relationships between nodes and node characteristics, thereby predicting the potential of a node to generate the greatest influence in the network. The DBSCAN algorithm can cluster nodes based on the density between nodes, thereby better discovering key figures in social networks.

[0033] (2) Efficient Algorithm: This invention provides an efficient method to solve the problem of maximizing expert influence, enabling the identification and prediction of key experts in an expert network within a shorter timeframe. This will help organizations better understand the key experts in their expert networks and improve their competitiveness in their respective fields.

[0034] (3) Scalability: The invention has good scalability and can be applied to different types of expert social networks, such as academic research and enterprise internal social networks.

[0035] (4) Practicality: The invention is highly practical and can help organizations and individuals better discover and utilize key experts in expert networks and improve their competitiveness and influence in their professional fields.

[0036] In summary, the invention of an expert influence maximization method based on graph convolution and DBSCAN algorithm has many advantages and can help improve the influence and competitiveness of organizations and individuals in their professional fields. Attached Figure Description

[0037] Figure 1 The flowchart shows a method for maximizing expert influence based on graph convolution and the DBSCAN algorithm.

[0038] Figure 2 A relationship graph of attributes such as expert papers in the citation dataset;

[0039] Figure 3 The structural diagram of the community after clustering;

[0040] Figure 4 Detailed flowcharts for each step;

[0041] Figure 5 This is a comparison chart of the propagation capability of this algorithm with other methods. Detailed Implementation

[0042] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

[0043] like Figure 1 As shown, this invention discloses a method for maximizing expert influence based on graph convolution and the DBSCAN algorithm. The specific steps are as follows:

[0044] S1. The citation dataset in this embodiment comes from the Microsoft Academic Graph data website, such as... Figure 2 As shown in the figure, the graph represents the relationship between attributes such as experts and papers. An undirected graph is constructed based on citation data, where each node corresponds to an expert, and edges represent citation relationships between two experts. Feature attributes of the experts are extracted, including the number of papers published, the number of citations, and the publication date. Each expert is considered a node, and a citation network is constructed based on the citation relationships between them.

[0045] S2. Use graph convolutional neural networks to extract features from the citation network to obtain the feature representation of each expert.

[0046] After initializing the citation network, important node sequences were selected using graph convolution and DBSCAN spectral clustering. The detailed process is attached. Figure 4 As shown:

[0047] S2.1, Perform data preprocessing: Construct an adjacency matrix A and a feature matrix X, where A is a... A is a symmetric matrix, where N is the number of nodes, and A is the th node. Each element represents whether node i and node j are connected by an edge, as shown in Table 1. X is a... The matrix is ​​a set of features, where each row represents a feature vector of a node (number of expert articles, expert citation count, and article creation time), and F is the dimension of the feature vector, as shown in Table 2:

[0048] Table 1

[0049] A B C D E A 0 0 0 0 1 B 0 0 0 1 1 C 0 0 0 1 1 D 0 1 1 0 1 E 1 1 1 1 0

[0050] Table 2

[0051] A -1.1 3.2 4.2 B 0.4 5.1 -1.2 C 1.2 1.3 2.1 D 1.4 -1.2 2.5 E 1.4 2.5 4.5

[0052] S2.2. Perform a linear transformation on the feature matrix. We need to perform a linear transformation on the input feature matrix X to obtain a new feature matrix H, as shown in the following formula: W is the weight matrix to be learned, which performs a linear transformation on the input feature matrix X, mapping the feature vector of each node in the input to a new vector. The size of the weight matrix W is... ,in This represents the dimension of the input feature matrix. This indicates the dimension of the output feature matrix, where each element is a parameter that needs to be learned.

[0053] S2.3 Normalize the adjacency matrix by dividing the adjacency matrix A by the node degree matrix D;

[0054] S2.4. Perform convolution operation. After normalizing the adjacency matrix, we can perform a convolution operation between it and the feature matrix. Specifically, we define a filter (also called a kernel) W, whose size is... Where k represents the filter size and F represents the dimension of the eigenvectors of each node in the feature matrix. Then, we can combine the filter W with the normalized adjacency matrix... Perform convolution to obtain a new feature matrix Z:

[0055]

[0056] S3. Perform spectral decomposition on the expert's feature matrix using the DBSCAN algorithm to obtain clustering results, as shown in the figure. Figure 3 As shown. Specifically, we need to perform symmetric normalization on the feature matrix, and then perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. Then, we divide the experts into different clusters based on the eigenvectors.

[0057] S3.1. Normalize the new feature matrix obtained through graph convolution to obtain matrix Y. This is done by subtracting the mean of the feature column from each eigenvalue and then dividing by the standard deviation of the feature column. This step ensures that all features of the feature matrix are on the same scale, facilitating subsequent processing.

[0058] S3.2. Perform DBSCAN-based spectral clustering on Y to obtain different communities. In DBSCAN-based spectral clustering, we treat each row of matrix Y as a data point, that is, a vector representation of each node in the new feature space. Then, we use these vectors as input and apply the DBSCAN algorithm for clustering. The DBSCAN algorithm determines which cluster each point belongs to based on the density around it.

[0059] S4. Calculate the influence score for each expert, rank the experts according to their influence scores, and select the most influential expert sequence.

[0060] S4.1 Using the DBSCAN algorithm, we can obtain different communities, each containing a group of closely connected nodes.

[0061] S4.2 For each node i, calculate the sum of the number of edges connecting it to other nodes to obtain the node degree.

[0062] S4.3 For each community, calculate the degree of all nodes within it. The sum of these values ​​yields the community degree. , ;

[0063] S4.4, Calculation Node Influence Index

[0064] S4.5 Sort all nodes according to their influence index and select the node with the greatest influence as the important node sequence.

[0065] like Figure 5 As shown, the nodes in the comparative experiment were selected using different methods to evaluate node importance. Random uses a random function, PageRank uses the PageRank algorithm to calculate the PR value of each node and sort them from largest to smallest, and Closeness Centrality calculates the closeness centrality of each node and sorts them from strongest to weakest.

[0066] The experimental data calculated by this algorithm are compared with the propagation capabilities of randomly selected nodes, nodes selected based on PR value, and nodes selected based on tight centrality. The propagation model of nodes in complex networks uses an independent cascade (IC) model to visualize propagation capabilities, with the x-axis representing the number of iterations of the key node and the y-axis representing the number of nodes infected by the key node. It can be seen that the experimental data of this algorithm achieves excellent results on the network, and the propagation efficiency of nodes is significantly improved.

[0067] This invention can be combined with a computer system to identify seed nodes. The expert influence maximization method based on graph convolution and the DBSCAN algorithm disclosed in this invention can be used for the identification of important nodes in major social networks.

[0068] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for maximizing expert influence based on graph convolution and DBSCAN algorithm, characterized in that, Includes the following steps: S1. Construct an undirected graph based on the citation data, where each node corresponds to an expert and an edge represents a citation relationship between two experts. Extract the feature attributes of the experts themselves and construct a citation network. The feature attributes include the number of papers published, the number of citations of the papers, and the publication time of the papers. S2. Use a graph convolutional neural network to extract features from the citation network to obtain the feature matrix representation of each expert; The specific steps of step S2 are as follows: S2.1, Perform data preprocessing: Construct an adjacency matrix A and a feature matrix X, where A is a... A is a symmetric matrix, where N is the number of nodes, and A is the th node. Each element represents whether node i and node j are connected by an edge; X is a... The matrix is ​​a matrix where each row represents the feature vector of a node, namely the number of expert articles, the expert's citation count, and the article creation time. F is the dimension of the feature vector. S2.

2. Perform a linear transformation on the feature matrix. We need to perform a linear transformation on the input feature matrix X to obtain a new feature matrix H, as shown in the following formula: C is the weight matrix to be learned, which performs a linear transformation on the input feature matrix X, mapping the feature vector of each node in the input to a new vector; the size of the weight matrix C is... ,in This represents the dimension of the input feature matrix. This indicates the dimension of the output feature matrix, where each element is a parameter that needs to be learned. S2.3 Normalize the adjacency matrix by dividing the adjacency matrix A by the node degree matrix D; S2.4 Perform convolution operation: After normalizing the adjacency matrix, perform a convolution operation between it and the feature matrix; define a filter, called kernel W, with a size of... Where k represents the size of the filter, and F represents the dimension of the eigenvector of each node in the feature matrix; the normalized matrix after normalizing the filter W and the adjacency matrix. Perform convolution to obtain a new feature matrix Z: ； S3. Perform spectral decomposition on the feature matrix of the experts using the DBSCAN algorithm to obtain clustering results: Perform symmetric normalization on the feature matrix, and then perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. Divide the experts into different clusters based on the eigenvectors. S4. Calculate the influence score of each expert based on their different clusters, rank the experts according to their influence scores, and select the most influential expert sequence.

2. The expert influence maximization method based on graph convolution and DBSCAN algorithm according to claim 1, characterized in that, Step S1 specifically involves: S1.1 Extract the expert node ID with the largest connected component from the citation dataset, as well as the expert attributes such as the number of published articles, the number of cited articles, and the publication time. S1.2 Extract the reference relationships between expert nodes; S1.

3. Each expert is treated as a node, and a citation network is constructed based on the citation relationships between them.

3. The expert influence maximization method based on graph convolution and DBSCAN algorithm according to claim 1, characterized in that, The specific steps of step S3 are as follows: S3.1 Normalize the new feature matrix obtained by graph convolution to obtain matrix Y, that is, subtract the mean of the column of each feature value from each feature value, and then divide by the standard deviation of the column of each feature value. S3.2 Perform DBSCAN-based spectral clustering on Y to obtain different communities. In DBSCAN-based spectral clustering, we treat each row of matrix Y as a data point, that is, the vector representation of each node in the new feature space. We use these vectors as input and apply the DBSCAN algorithm for clustering. The DBSCAN algorithm determines which cluster each point belongs to based on the density around it.

4. The expert influence maximization method based on graph convolution and DBSCAN algorithm according to any one of claims 1 to 3, characterized in that, The specific steps of step S4 are as follows: S4.

1. Using the DBSCAN algorithm, different communities are obtained, and each community contains a group of closely connected nodes; S4.2 For each community, calculate the sum of the degrees of all its nodes to obtain the community degree. ; S4.3 For each node i, calculate the sum of the number of edges connecting it to other nodes to obtain the node degree. ; S4.4, Calculation Node Influence Index ; S4.5 Sort all nodes according to their influence index and select the node with the greatest influence as the important node.

Images