Author collaboration network mining method integrating multi-view information-guided representation learning

Abstract

The present invention discloses a method for mining an author collaboration network that integrates multi-perspective information-oriented representation learning, extracts the largest subgraph from the author collaboration network, constructs a graph convolutional neural network guided by proximity and structure similarity within the framework of a variational autoencoder to aggregate node embedding vectors, captures the network structure features of the two orientations, and constructs corresponding loss terms to update model parameters; the node embeddings of the two orientations are used as input for clustering tasks, cluster analysis is performed using the k-means algorithm, and node community labels and node role labels are obtained; the mataplotlib library is used for visualization, and after analysis, the topic association rules in the current environment are obtained and the distribution of talents under interdisciplinary subjects is inferred. The training method of the present invention is unsupervised learning, and no manual labeling is required; thanks to the advantages of the variational autoencoder, its encoding is consistent with the probability distribution during training to ensure that it has good characteristics in the latent space, thereby avoiding overfitting.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and network representation learning, and aims to map the nodes in the graph into a low-dimensional dense vector space while retaining information such as the structural characteristics of the network and node attributes, and then apply it to some downstream machine learning tasks such as node classification and link prediction. Background Art

[0002] How to mine the underlying dynamic patterns in network systems has always been a research hotspot in graph data science. Roles (or structural equivalence) exist in a variety of real-world systems, including social networks, biological protein networks, air traffic networks, citation networks, etc., corresponding to different semantics in different scenarios. Mining the role characteristics of nodes can better help decision makers understand the function and role of each node in the network.

[0003] The generation of node edges is driven by two properties. The capture of one property can serve as a guide for the capture of another type of information. When part of the role information is known, when the central node of the field is known, such as the central figure is closely connected with other authors in the field, it is helpful to mine the community; on the contrary, if the community information is known, some nodes with a high probability of belonging to multiple communities at the same time and a large degree tend to be bridge nodes. Some studies have tried to utilize and learn the interaction patterns between roles and communities. The present invention mainly focuses on the scientific problem of network representation learning that integrates proximity and structural similarity, and with the help of the powerful learning and representation capabilities of deep learning graph neural networks, jointly learns two types of guided node vectors.

[0004] References:

[0005] [1]Refex[Henderson, Keith, et al. "It's who you know: graph mining using recursive structural features." Proceedings of the 17th ACM SIGKDDinternational conference on Knowledge discovery and data mining. 2011.]. Summary of the Invention

[0006] In response to the above-mentioned existing technologies, the present invention proposes a method for mining author collaboration networks that integrates multi-perspective information-guided representation learning. Its main purpose is to construct a citation network based on collaborators and study how to map nodes into a role vector space, where nodes with similar distances in the space have similar roles or functions; construct a new graph convolutional neural network corresponding to this task to effectively aggregate node information within the neighborhood; perform graph visualization, label various role nodes in the graph, and clearly identify the underlying structure and correlation in the graph data.

[0007] In order to solve the above technical problems, the present invention proposes a method for mining author collaboration networks that integrates multi-perspective information-guided representation learning, which mainly includes: constructing an author collaboration network after obtaining paper data, and extracting the largest subgraph of the author collaboration network; sorting the nodes of the largest subgraph in ascending order according to degree to perform node set segmentation; using Refex to count node attributes and node neighborhood attribute features, thereby extracting node features; constructing a generative model within the framework of a variational autoencoder: constructing graph convolutional neural networks guided by proximity and structural similarity respectively to aggregate node embedding vectors, capture the network structure features of the two orientations, and then construct corresponding loss terms for parameter updating; using the generative model to obtain the above-mentioned two-oriented node embeddings, using this as the input of the clustering task, using the k-means algorithm to perform clustering analysis, and obtaining node community labels and node role labels; using the mataplotlib library for visualization, obtaining the subject association rules in the current environment by analyzing the node community labels, and inferring the talent distribution under interdisciplinary subjects by analyzing the node role labels.

[0008] Furthermore, the author collaboration network mining method of the present invention includes the following specific steps:

[0009] Step 1: Obtaining data and processing data:

[0010] 1-1) Search and organize public research papers to obtain research paper data, delete incomplete data, remove duplicate data, and then extract author information from the papers, delete author information containing English names, and screen and decrypt private data involving personal privacy, ultimately obtaining an author data table that includes author name, institution affiliation, discipline of the author, published papers, and paper collaboration relationships;

[0011] 1-2) Using the paper collaboration relationships in the author data table as edges, we construct an author collaboration network, where the author's name and their affiliated institution are used as the unique identifiers of the nodes in the collaboration network;

[0012] 1-3) Extract the largest subgraph of the author collaboration network and use the adjacency matrix A as the interaction of the nodes in the largest subgraph. When node i interacts with node j, A ij =1;

[0013] Step 2: Sort the nodes of the largest subgraph described in step 1 in ascending order according to degree, and construct a mapping function f:deg(v i )→m, its function is to map N nodes to C={C1,...,C M}, M is the number of groups, which is a hyperparameter. M is set to 4. The specific construction process of the function is as follows:

[0014] 2-1) Given a proximity matrix A, calculate the degrees d of all nodes and arrange them in ascending order;

[0015] 2-2) Calculate the average node degree avg;

[0016] 2-3) The pointer j is assigned a value of 1;

[0017] 2-3) Pointer i is assigned a value of 1;

[0018] 2-4) When Group C i When the number of nodes is less than avg, add node j to C i middle;

[0019] 2-5) Repeat steps 2-1) to 2-4) above until all nodes are added to all groups, obtaining node group C;

[0020] Step 3: Use Refex to count the node attributes and node neighborhood attribute features to obtain the node feature matrix F. The specific steps are as follows:

[0021] 3-1) Pointer i is assigned a value of 1;

[0022] 3-2) Calculate the degree of node i;

[0023] 3-3) Calculate all edges of the egocentric network formed by node i;

[0024] 3-4) Calculate the degree of all nodes in the egocentric network formed by node i;

[0025] 3-5) Calculate the ratio of incoming edges to outgoing edges within the egocentric network;

[0026] 3-6) Calculate the ratio of edges within the non-egocentric network to all edges;

[0027] 3-7) Calculate the number of triangular subgraphs in the egocentric network;

[0028] 3-8) Calculate the clustering coefficient of node i;

[0029] 3-9) Repeat the above steps 3-1) to 3-8) until all nodes are calculated, and construct the node feature matrix F with the node attributes and node neighborhood attribute features;

[0030] Step 4: Construct a generative model: The generative model consists of an encoder and a decoder. The steps for constructing the generative model are as follows:

[0031] 4-1) Constructing a two-layer graph convolutional neural network guided by node structure similarity: With the adjacency matrix A as input and the node feature matrix F as the first layer embedding, when updating the node embedding matrix, the following steps are performed: global mapping is performed on the node embedding of the previous layer, local feature mapping is performed on the node embedding of the previous layer through the convolution kernel pointed to by the group number C; matrix summation of the global mapping and the local feature mapping is performed, and nonlinear mapping is performed through the RELU function to obtain the node embedding matrix of the new layer. The two-layer graph convolutional neural network guided by node structure similarity is in the following form:

[0032]

[0033] In formula (1), v is the node with the initial label r, h is the embedding vector, N(v) is the neighbor node of v, k represents the number of neural network layers, and W g is the global training matrix, W r is the training matrix corresponding to label r;

[0034] 4-2) Use GCN to build a two-layer graph convolutional neural network based on node proximity: by aggregating the features of the central node itself and its neighboring nodes, a new layer of network embedding is formed. The first layer embedding is the identity matrix, and its propagation formula is as follows:

[0035]

[0036] In formula (2), l is the l-th layer neural network, L is the graph normalized Laplace matrix, D is the node degree matrix, that is I is the identity matrix, W c is the proximity-guided training parameter matrix;

[0037] 4-3) Constructing an encoder: The encoder includes a proximity capture module and a structural similarity capture module;

[0038] The proximity capture module is composed of the two-layer graph convolutional neural network based on node proximity guidance described in step 4-2), which takes the adjacency matrix A as input and finally outputs the mean μ of the proximity capture module. c and variance δ c ;

[0039] The structural similarity capture module is composed of a two-layer graph convolutional neural network guided by the node structure similarity described in step 4-1), which takes the adjacency matrix A and the node feature matrix F as input and finally outputs the mean μ of the structural similarity capture module. r and variance δ r ;

[0040] Use the reparameterization technique to sample the node embedding Z c and Z r ,

[0041]

[0042]

[0043] In formula (3) and formula (4), Z c i is the node vector of node i based on proximity, u c i and σ c i is the mean and variance of the softmax Gaussian distribution of node i about proximity; Z r i is the node vector of node i based on structural similarity, u r i and σ r i is the mean and variance of the softmax Gaussian distribution of node i about structural similarity; ε is sampled from the standard normal distribution;

[0044] 4-4) Construct the decoder: Use formula (3) and formula (4) to get Z c and Z r After that, reconstruct the adjacency matrix

[0045] p(A ij |z i z j )=(z c i ) T z c j +f(z r i ) T f(z r j ) (5)

[0046] In formula (5), f is a multi-layer perceptron (MLP); z c It captures the node proximity. The smaller the vector distance, the more likely it is to generate an edge.r , use MLP to first perform nonlinear mapping and then perform inner product operation;

[0047] 4-5) Construct model loss function:

[0048]

[0049] In formula (5), q φ (Z|A, F) is the encoder posterior distribution, φ is the encoder parameter to be trained, p is the prior distribution, p θ (A|Z decoder distribution, θ is the decoder parameter, Z=(Z r ,Z c ), where the first KL loss is divided into proximity-guided loss and structural similarity-guided loss, which are:

[0050]

[0051]

[0052] The second term is the reconstruction error, using binary cross entropy as the loss, which can be written as:

[0053]

[0054] in, is the adjacency matrix reconstructed by the decoder; σ c0 and σ r0 is the parameter of the prior distribution p, which is randomly generated; 4-6) Model parameter update: In the forward propagation process, the encoder is used to sample and obtain the node embedding Z, which is then input into the decoder to reconstruct the adjacency matrix and update the loss function term described in formula (6) for reverse gradient backpropagation; repeat steps 4-6) 200 times to complete the construction of the generative model;

[0055] Step 5. Run the generative model constructed in step 4 on the largest subgraph extracted in steps 1-3) to obtain node embeddings based on proximity orientation and structural similarity orientation, and then use the k-means algorithm to perform cluster analysis respectively to obtain node community labels and node role labels; use the mataplotlib library for visualization and draw nodes, divide nodes with the same type of community labels into the same subgraph, and give nodes with the same type of role labels the same shape; by analyzing the community labels and author data tables, explore the topic association rules in the current environment; by analyzing the node role labels, explore the talent distribution in the current academic environment.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The method described in the present invention can be applied to networks in different scenarios, such as email sending networks and airport networks, and can automatically identify the node roles therein; the training method is unsupervised learning, which does not require manual labeling, greatly reducing the workload; thanks to the advantages of variational autoencoders, its encoding during training is consistent with the probability distribution, ensuring that it has good characteristics in the latent space, thereby avoiding overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the main flow chart of the method of the present invention;

[0059] Figure 2 for Figure 1 A detailed diagram of the main process shown;

[0060] Figure 3 A model structure diagram for implementing the method of the present invention;

[0061] Figure 4 It is the node role and community label in the embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention proposes an author collaboration network mining method that integrates multi-perspective information-oriented representation learning. Its main purpose is to mine the network's potential knowledge by focusing on the local structure of the node and the node set within the neighborhood, and to find the node set with commonality. The design concept of the present invention is: after arranging the public paper documents to obtain the paper data, the author collaboration network is constructed, and the largest subgraph of the author collaboration network is extracted; the nodes of the largest subgraph are sorted in ascending order according to the degree to perform node set segmentation; Refex is used to count the node attributes and node neighborhood attribute features to extract node features; for the author collaboration network, the node similarity is analyzed by mining the network structure features, the node interaction mode within the short or medium distance is focused on from the proximity perspective, and the node neighborhood is analyzed by structural similarity. The domain structure is used to focus on the functions and roles of nodes in the overall network, that is, within the framework of the variational autoencoder, a graph convolutional neural network guided by proximity and structural similarity is constructed to aggregate node embedding vectors, capture the network structure characteristics of the two orientations, and then construct corresponding loss terms to update parameters to obtain a generative model; the generative model is used to obtain the above-mentioned two-oriented node embeddings, which are used as the input of the clustering task, and the k-means algorithm is used for clustering analysis to obtain node community labels and node role labels. The mataplotlib library is used for visualization, and the topic association rules and scientific research topic distribution in the current environment are obtained by analyzing the node community labels. The distribution of talents under interdisciplinary subjects is inferred by analyzing the node role labels, such as Figure 1 shown.

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention in any way.

[0064] This paper proposes a method for mining author collaboration networks that integrates multi-perspective information-guided representation learning. First, for given paper data, authors are extracted as nodes, and then paper citations and collaboration relationships are extracted as edges to construct an author collaboration network. Within the framework of a variational autoencoder, graph neural networks are used to aggregate node embedding vectors and construct corresponding loss terms for parameter updates. Figure 2 As shown, the method of the present invention comprises the following steps:

[0065] Step 1: Data acquisition and data processing, including organizing public papers and literature, constructing the author collaboration network after obtaining the paper data, and extracting the largest subgraph of the author collaboration network; the specific steps are as follows:

[0066] 1-1) In this embodiment, the paper documents from 2020 to 2021 on the China National Knowledge Infrastructure (CNKI) website (https: / / www.cnki.net / ) are searched and sorted, incomplete data is deleted, duplicate data is eliminated, and then the author information in the paper is extracted, duplicate data is merged, author information containing English names is deleted, and private data involving personal privacy is screened and declassified, and finally an author data table is obtained, which includes the author's name, the author's institution, the author's discipline, published papers, and paper cooperation relationships.

[0067] 1-2) Extract the combination of the author name and the affiliated institution of all papers as the unique identifier of each node to distinguish authors with the same name from different affiliated institutions. Then use the author name and the affiliated institution as the unique identifier of the node of the collaboration network, and use the paper collaboration relationship in the author data table as the edge to construct the author collaboration network.

[0068] 1-3) Since there are some extremely small subgraphs in the network, which interfere with the mining of bridge nodes, there is no need to explore and analyze the properties of extremely small subgraphs. Therefore, the method of the present invention extracts the largest subgraph of the network for algorithm operation and analysis, that is, extracts the largest subgraph of the author collaboration network, and uses the adjacency matrix A as the interaction of the nodes in the largest subgraph. When node i interacts with node j, A ij =1.

[0069] Step 2: sort the nodes of the largest subgraph in ascending order according to degree to split the node set:

[0070] Since the distribution range of node degrees is generally very wide, it is necessary to sort the nodes of the largest subgraph in step 1 in ascending order according to degree and further divide them into groups, that is, to construct a mapping function f:deg(v i )→m, its function is to map N nodes to C={C1,...,C M}, M is the number of groups, which is a hyperparameter. In this invention, M is set to 4. The specific construction process of this function is as follows:

[0071] 2-1) Given a proximity matrix A, calculate the degrees d of all nodes and arrange them in ascending order;

[0072] 2-2) Calculate the average node degree avg;

[0073] 2-3) The pointer j is assigned a value of 1;

[0074] 2-3) Pointer i is assigned a value of 1;

[0075] 2-4) When Group C i When the number of nodes is less than avg, add node j to C i middle;

[0076] 2-5) Repeat steps 2-1) to 2-4) above until all nodes are added to all groups, obtaining node group C;

[0077] Step 3: Feature initialization. Good initial node features can make the model converge quickly and achieve better results. Therefore, in order to make the initial features contain information to measure structural similarity, the present invention uses the method of reference [1] to perform feature extraction. Refex is a method based on statistical node and node neighborhood attributes, which has extremely high computational efficiency. In the present invention, Refex is used to statistically analyze node attributes and node neighborhood attribute features to obtain the node feature matrix F. The specific steps are as follows:

[0078] 3-1) Pointer i is assigned a value of 1;

[0079] 3-2) Calculate the degree of node i;

[0080] 3-3) Calculate all edges of the egocentric network formed by node i;

[0081] 3-4) Calculate the degree of all nodes in the egocentric network formed by node i;

[0082] 3-5) Calculate the ratio of incoming edges to outgoing edges within the egocentric network;

[0083] 3-6) Calculate the ratio of edges within the non-egocentric network to all edges;

[0084] 3-7) Calculate the number of triangular subgraphs in the egocentric network;

[0085] 3-8) Calculate the clustering coefficient of node i;

[0086] 3-9) Repeat the above steps 3-1) to 3-8) until all nodes are calculated, and use these statistical values ​​(node ​​attributes and node neighborhood attribute features) to form the node feature matrix F.

[0087] Step 4: Construct a generative model within the variational autoencoder framework. This mainly involves constructing two graph convolutional neural networks, one guided by proximity and the other by structural similarity, to aggregate node embedding vectors, capturing the structural characteristics of the two networks, and then constructing corresponding loss terms for parameter updates. The details are as follows:

[0088] 4-1) Constructing a two-layer graph convolutional neural network guided by node structure similarity:

[0089] Traditional convolutional neural networks update the representation of central nodes by aggregating node features in the neighborhood. However, in character mining, considering the different neighborhood structures between different characters, different convolution kernels should be assigned to different characters during the convolution process. Therefore, the present invention constructs a graph neural network for this type of role task, taking the adjacency matrix as input. First, the node features of the structural similarity neural network module are initialized, and then a global mapping is performed. After that, the nodes with specific labels are input into the corresponding multi-layer perceptron, that is, nodes with the same label share the same training matrix when aggregating neighborhood information.

[0090] Taking the adjacency matrix A as input and the node feature matrix F as the first layer embedding, updating the node embedding matrix includes: performing global mapping on the node embedding of the previous layer, and performing local feature mapping on the node embedding of the previous layer through the convolution kernel pointed to by the group number C; performing matrix addition on the global mapping and the local feature mapping, and performing nonlinear mapping through the RELU function to obtain the node embedding matrix of the new layer. The two-layer graph convolutional neural network guided by node structure similarity is as follows:

[0091]

[0092] In formula (1), v is the node with the initial label r, h is the embedding vector, N(v) is the neighbor node of v, k represents the number of neural network layers, and W g is the global training matrix, W r is the training matrix corresponding to label r.

[0093] 4-2) Use GCN to build a two-layer graph convolutional neural network based on node proximity:

[0094] By mapping the features of the aggregation center node itself and its neighboring nodes, a new layer of network embedding is formed. The first layer embedding is the unit matrix. During the information propagation process, neighboring nodes will be represented by node vectors with similar distances. The propagation formula is as follows:

[0095]

[0096] In formula (2), l is the l-th layer neural network, L is the graph normalized Laplace matrix, D is the node degree matrix, that is I is the identity matrix, W c is the proximity-guided training parameter matrix.

[0097] 4-3) Constructing an encoder: The encoder includes a proximity capture module and a structural similarity capture module;

[0098] The proximity capture module is composed of the two-layer graph convolutional neural network based on node proximity guidance described in step 4-2), which takes the adjacency matrix A as input and finally outputs the mean μ of the proximity capture module. c and variance δ c .

[0099] The structural similarity capture module is composed of a two-layer graph convolutional neural network guided by the node structure similarity described in step 4-1), which takes the adjacency matrix A and the node feature matrix F as input and finally outputs the mean μ of the structural similarity capture module. r and variance δ r .

[0100] After using the neural network to obtain the mean and variance of the probability distribution of nodes based on proximity and structural similarity, the node embedding Z is obtained by sampling using the reparameterization technique. c and Z r , so that the model can perform gradient retransmission.

[0101]

[0102]

[0103] In formula (3) and formula (4), Z c i is the node vector of node i based on proximity, u c i and σ c i is the mean and variance of the softmax Gaussian distribution of node i about proximity; Z r i is the node vector of node i based on structural similarity, u ri and σ r i is the mean and variance of the softmax Gaussian distribution of node i about structural similarity; ε is sampled from the standard normal distribution;

[0104] 4-4) Construct the decoder: Use formula (3) and formula (4) to get Z c and Z r After that, reconstruct the adjacency matrix The formula is as follows:

[0105] p(A ij |z i z j )=(z c i ) T z c j +f(z r i ) T f(z r j ) (5)

[0106] In formula (5), f is a multi-layer perceptron (MLP). c It captures the node proximity. The smaller the vector distance, the more likely it is to generate an edge. However, for the role-oriented vector z r , the inner product operation cannot be used directly, but MLP needs to be used to perform nonlinear mapping first, and then the inner product operation.

[0107] 4-5) Construct model loss function:

[0108] In the present invention, the generative model consists of an encoder and a decoder. It is trained to reduce the reconstruction error between the decoded data and the initial data, where the encoded data is sampled at a certain point in the probability distribution, the prior distribution of the latent variable is a normal distribution, and the network is trained using the pytorch deep learning library, and its ELBO loss is:

[0109]

[0110] In formula (6), q φ (Z|A, F) is the encoder posterior distribution, φ is the encoder parameter to be trained, p is the prior distribution, p θ (A|Z) decoder distribution, θ is the decoder parameter, Z = (Z r ,Z c ), where the first KL loss is divided into proximity-guided loss and structural similarity-guided loss, which are:

[0111]

[0112]

[0113] The second term is the reconstruction error, using binary cross entropy as the loss, which can be written as:

[0114]

[0115] in, is the adjacency matrix reconstructed by the decoder; σ c0 and σ r0 is the parameter of the prior distribution p, which is randomly generated; the final structure of the model is as follows Figure 3 As shown, first, the adjacency matrix is ​​used as input to aggregate the node proximity vector Z c i , taking the node feature matrix F as input and the node group C as the basis to input into the corresponding convolution kernel, aggregate the node structure similarity vector Z r i , and finally Z c i and Z r i Input to the decoder to reconstruct the adjacency matrix A.

[0116] 4-6) Model parameter update: During the forward propagation process, the encoder is used to sample and obtain the node embedding Z, which is then input into the decoder to reconstruct the adjacency matrix and update the loss function term described in formula (6) for reverse gradient backpropagation; repeat steps 4-6) 200 times.

[0117] Step 5: Run the generative model constructed in step 4 on the largest subgraph extracted in steps 1-3) to obtain node embeddings based on proximity and structural similarity. Then, perform cluster analysis using the k-means algorithm to obtain node community labels and node role labels. Use the mataplotlib library for visualization and draw nodes. Nodes with the same community label are grouped into the same subgraph, and nodes with the same role label are given the same shape. The final result is as follows: Figure 4As shown, the network consists of three subgraphs and three types of nodes. Due to the close collaborative relationships between authors within each subgraph, there are fewer edges between subgraphs. Each subgraph can be considered a type of academic circle. By analyzing community labels and author data tables, we can further explore the patterns of topic associations in today's academic environment. Nodes with similar local structures in the overall network are assigned the same shape. For example, nodes that serve as bridges between academic circles are labeled as circles, likely representing leading figures in interdisciplinary fields. Nodes with strong connections within academic circles are labeled as rectangular rectangles, likely representing central tasks in their fields. The remaining rounded rectangles are edge nodes. By analyzing node role labels, we can explore the distribution of talent in today's academic environment.

[0118] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many variations without departing from the purpose of the present invention, and these are all protected by the present invention.

Claims

1. A method for mining author collaboration networks integrating multi-view information-guided representation learning, characterized by: Mainly include: Constructing an author collaboration network and extracting the largest subgraph of the author collaboration network; sorting the nodes of the largest subgraph in ascending order according to their degrees to perform node set segmentation; Utilize Refex to count node attributes and node neighborhood attribute features to extract node features; A generative model is constructed within the framework of a variational autoencoder. Proximity-oriented and structural similarity-oriented graph convolutional neural networks are constructed to aggregate node embedding vectors, capturing the structural characteristics of the two networks and constructing corresponding loss terms for parameter updates. The generative model is used to obtain the above two types of node embeddings, which are used as the input for the clustering task. The k-means algorithm is used for cluster analysis to obtain node community labels and node role labels. Use the mataplotlib library for visualization, analyze node community labels to obtain the topic association patterns in the current environment, and analyze node role labels to infer the distribution of talents in interdisciplinary disciplines.

2. The author collaboration network mining method according to claim 1, characterized in that: The following steps are involved: Step 1: Data acquisition and data processing, including: Step 1-1) Search and organize public papers and obtain paper data, delete incomplete data, remove duplicate data, and then extract author information from the papers, delete author information containing English names, and screen and decrypt private data involving personal privacy, ultimately obtaining an author data table, which includes author name, author affiliation, author discipline, published papers, and paper collaboration relationships; Step 1-2) Construct an author collaboration network using the paper collaboration relationships in the author data table as edges, where the author's name and their affiliated institution are used as the unique identifiers of the nodes in the collaboration network; Step 1-3) Extract the largest subgraph of the author collaboration network and use the adjacency matrix A as the interaction of the nodes in the largest subgraph. When node i interacts with node j, A ij =1; Step 2: Sort the nodes of the largest subgraph described in step 1 in ascending order according to degree, and construct a mapping function f:deg(v i )→m, its function is to map N nodes to C={C1,...,C M }, M is the number of groups, which is a hyperparameter. M is set to 4. The specific construction process of the function is as follows: Step 2-1) Given the proximity matrix A, calculate the degrees d of all nodes and arrange them in ascending order; Step 2-2) Calculate the average node degree avg; Step 2-3) Assign the pointer j to 1; Step 2-3) Assign pointer i a value of 1; Step 2-4) When Group C i When the number of nodes is less than avg, add node j to C i middle; Step 2-5) Repeat steps 2-1) to 2-4) above until all nodes are added to all groups, obtaining node group C; Step 3: Use Refex to count the node attributes and node neighborhood attribute features to obtain the node feature matrix F. The specific steps are as follows: Step 3-1) Assign pointer i a value of 1; Step 3-2) Calculate the degree of node i; Step 3-3) Calculate all edges of the egocentric network formed by node i; Step 3-4) Calculate the degree of all nodes in the egocentric network formed by node i; Step 3-5) Calculate the ratio of incoming edges to outgoing edges within the egocentric network; Steps 3-6) Calculate the ratio of edges within the non-egocentric network to all edges; Step 3-7) Calculate the number of triangular subgraphs of the egocentric network; Step 3-8) Calculate the clustering coefficient of node i; Step 3-9) Repeat the above steps 3-1) to 3-8) until all nodes are calculated, and construct the node feature matrix F with the node attributes and node neighborhood attribute features; Step 4: Construct a generative model, which consists of an encoder and a decoder. The steps for constructing the generative model are as follows: Step 4-1) Construct a two-layer graph convolutional neural network guided by node structure similarity: Taking the adjacency matrix A as input and the node feature matrix F as the first layer embedding, updating the node embedding matrix includes: performing global mapping on the node embedding of the previous layer, and performing local feature mapping on the node embedding of the previous layer through the convolution kernel pointed to by the group number C; performing matrix addition on the global mapping and the local feature mapping, and performing nonlinear mapping through the RELU function to obtain the node embedding matrix of the new layer. The two-layer graph convolutional neural network guided by node structure similarity is as follows: In formula (1), v is the node with the initial label r, h is the embedding vector, N(v) is the neighbor node of v, k represents the number of neural network layers, and W g is the global training matrix, W r is the training matrix corresponding to label r; Step 4-2) Use GCN to build a two-layer graph convolutional neural network based on node proximity: By mapping the features of the aggregation center node itself and its neighboring nodes, a new layer of network embedding is formed. The first layer embedding is the unit matrix, and its propagation formula is as follows: In formula (2), l is the l-th layer neural network, L is the graph normalized Laplace matrix, D is the node degree matrix, that is I is the identity matrix, W c is the proximity-guided training parameter matrix; Step 4-3) Build the encoder: The encoder includes a proximity capture module and a structural similarity capture module; The proximity capture module is composed of the two-layer graph convolutional neural network based on node proximity guidance described in step 4-2), which takes the adjacency matrix A as input and finally outputs the mean μ of the proximity capture module. c and variance δ c ; The structural similarity capture module is composed of a two-layer graph convolutional neural network guided by the node structural similarity described in step 4-1), which takes the adjacency matrix A and the node feature matrix F as input and finally outputs the mean μ of the structural similarity capture module. r and variance δ r ; Use the reparameterization technique to sample the node embedding Z c and Z r , In formula (3) and formula (4), Z c i is the node vector of node i based on proximity, u c i and σ c i is the mean and variance of the softmax Gaussian distribution of node i about proximity; Z r i is the node vector of node i based on structural similarity, u r i and σ r i is the mean and variance of the softmax Gaussian distribution of node i about structural similarity; ε is sampled from the standard normal distribution; Step 4-4) Build the decoder: Using formula (3) and formula (4) to get Z c and Z r After that, reconstruct the adjacency matrix p(A ij |z i z j )=(z c i ) T z c j +f(z r i ) T f(z r j ) (5) In formula (5), f is a multi-layer perceptron (MLP); z c It captures the node proximity. The smaller the vector distance, the more likely it is to generate an edge. r , use MLP to first perform nonlinear mapping and then perform inner product operation; Step 4-5) Construct the model loss function: In formula (5), q φ (Z|A, F) is the encoder posterior distribution, φ is the encoder parameter to be trained, p is the prior distribution, p θ (A|Z) decoder distribution, θ is the decoder parameter, Z = (Z r ,Z c ), where the first KL loss is divided into proximity-guided loss and structural similarity-guided loss, which are: The second term is the reconstruction error, using binary cross entropy as the loss, which can be written as: in, is the adjacency matrix reconstructed by the decoder; σ c0 and σ r0 is the parameter of the prior distribution p, which is randomly generated; Steps 4-6) Model parameter update During the forward propagation process, the encoder is used to sample and obtain the node embedding Z, which is then input into the decoder to reconstruct the adjacency matrix and update the loss function term described in formula (6) for reverse gradient backpropagation; steps 4-6) are repeated 200 times to complete the construction of the generative model; Step 5. Run the generative model constructed in step 4 on the largest subgraph extracted in steps 1-3) to obtain node embeddings based on proximity and structural similarity, and then use the k-means algorithm to perform cluster analysis to obtain node community labels and node role labels; use the mataplotlib library for visualization and draw nodes, divide nodes with the same type of community labels into the same subgraph, and give nodes with the same type of role labels the same shape; by analyzing the community labels and author data tables, explore the topic association patterns in the current environment; by analyzing the node role labels, explore the talent distribution in the current academic environment.

Images