Social network alignment method and system based on multiple information fusion and graph optimization

Abstract

This invention provides a social network alignment method and system based on multi-information fusion and graph optimization, belonging to the field of internet social network technology. Based on constructed social network user graph data, it obtains user structural information, user behavioral information, and user attribute information. According to the constructed optimization model, it obtains the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavioral information embedding matrix, and attribute information embedding matrix, respectively, and calculates the three node embedding similarity matrices of the source and target networks. Based on a gating mechanism, it obtains the fused embedding similarity matrix and acquires the user alignment nodes. This invention extracts the semantic features of users, reduces the sparsity of the network structure, eliminates the need for pre-training data, and is more suitable for the needs of large-scale networks. By learning user attribute features and network topology features, it reduces the interference of structural noise and attribute noise, improving the accuracy of cross-social network user alignment.

Description

Technical Field

[0001] This invention relates to the field of internet social network alignment technology, specifically to a social network alignment method and system based on multi-information fusion and graph optimization. Background Technology

[0002] With the rapid development of the internet, social networks have experienced explosive growth. To meet their diverse needs, users often register accounts on multiple social networks. The accounts of the same user on different social networks can act as bridges between these networks, providing raw data for cross-social network research tasks.

[0003] Accounts belonging to the same user across different social networks are called aligned users or anchor links, and the process of discovering such accounts is called network alignment. As an upstream task in many cross-social network research projects, social network alignment has attracted significant attention from scholars both domestically and internationally in recent years. Social network alignment has important implications for both industrial applications and cybersecurity.

[0004] (1) Industrial Applications. Social networks are important platforms for business activities, and social network alignment will unleash greater value from them. For example, social network alignment can correlate informational social networks with cold-start social networks, helping cold-start networks to acquire user preferences and effectively conduct cross-domain recommendations; social network alignment establishes pathways for the spread of influence among multiple social networks, providing the possibility of revealing more authentic user influence. Based on social network alignment, the fusion of data from different sources on the same user or group can enrich user or group profiles, uncover user or group preferences, and thus conduct business activities in a more targeted manner.

[0005] (2) In terms of cybersecurity. User identity and true intent analysis are fundamental prerequisites for maintaining national security in cyberspace, and social network alignment is a key link in user identity analysis. In real life, some users do not use their real names on certain online social networks, and the content they publish on anonymous networks may differ from that on real-name networks. For anonymous users, social network alignment methods can be used to match them with real-name networks, obtain the user's real identity, and uncover their true intentions and behavioral patterns, thereby providing targeted, positive, and correct guidance to maintain and protect cybersecurity.

[0006] Based on the types of data used in alignment methods, cross-social network alignment can be mainly divided into three categories: attribute-based social network alignment, account behavior-based social network alignment, and network structure-based social network alignment. Attribute-based alignment assumes that user profile attributes across different social networks typically exhibit similarities; for example, a user's username, address, and avatar may be identical or similar across different networks. However, real-world datasets often suffer from issues such as missing, inconsistent, and heterogeneous user attributes. Furthermore, these methods do not consider the role of social network topology in alignment. Account behavior-based alignment utilizes information such as user posting, login / logout, and writing style, assuming similar user behavior across different social networks. These methods require large amounts of user data for training and often suffer from insufficient data volume. Network structure-based alignment assumes that the neighborhood structure of anchor links is similar. This approach includes spectral-based alignment, such as the FINAL and REGAL models, and network representation learning-based alignment, such as the DeepLink and GAlign models. These methods do not require user attribute information or posting behavior data, providing a solution for large-scale network alignment. However, such solutions require a high degree of consistency in network topology and are susceptible to noise interference.

[0007] In summary, the existing technical problems are: existing network alignment solutions are susceptible to structural noise and attribute noise, resulting in low alignment accuracy, and the high model complexity makes them unsuitable for large-scale network alignment. Summary of the Invention

[0008] The purpose of this invention is to provide a social network alignment method and system based on multi-information fusion and graph optimization, so as to solve at least one of the technical problems existing in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] On the one hand, this invention provides a social network alignment method based on multi-information fusion and graph optimization, comprising:

[0011] Based on the constructed social network user graph data, the structural information of users is obtained based on the centrality principle, the behavioral information of users is obtained based on the text of multi-source networks, and the attribute information of users is obtained based on the graph neural network model of multi-hop neighbor aggregation.

[0012] Based on the optimization model constructed using graph optimization theory, the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavior information embedding matrix, and attribute information embedding matrix are obtained respectively.

[0013] The embedding similarity matrix of the three nodes in the source network and the target network is calculated based on the transmission matrix;

[0014] The fusion embedding similarity matrix is ​​obtained based on the gating mechanism, and the user alignment node is obtained.

[0015] Preferably, constructing social network user graph data includes: constructing social network user data, including the social network topology graph structure, user-posted text information, and user homepage attributes; obtaining the structural information of network user nodes through feature vector centrality, and through B... e The RT model extracts behavioral information of network user nodes and extracts attribute information of network user nodes through a GCN network based on multi-hop neighbor aggregation. Based on a gating mechanism, the three types of user information are fused to obtain a user vector representation.

[0016] Preferably, constructing social network user structure information and obtaining user behavior information includes:

[0017] Let the source network be G. S =(V S E S F S The target network is G. T =(V T E T F T This paper constructs an undirected weighted graph using the social network relationships of users in the input data; extracts the feature vector centrality of users using the edge weight matrix W; and obtains the user's structural vector representation struct using logarithmic binning techniques. * (i);

[0018] Input all of the user's post text into the Bert model to obtain the user's text vector. User text vectors are formed by adding the shared semantics of the source and target networks and the global semantics of the network in which they reside, i.e.:

[0019] sharedEmb i =textEmb * (i)-globalEmb *

[0020] sharedEmb i Indicates shared semantics, gl0balEmb * This represents the global semantics of the source or target network. The global semantics can be derived from all text inputs to the current network into B. e Obtained from the RT model;

[0021] By training T eThe xtCNN network minimizes the distance between the shared vector of the central node and the shared vector of its neighboring nodes to obtain the final vector representation of user behavior information.

[0022] Preferably, the final user behavior information vector representation includes:

[0023] text i =ReLU(Conv1D(sharedEmb) i ,h)+b)

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] Where ReLU represents the activation function, Conv1D represents the TextCNN convolutional layer, and the kernel size is h×d, where h represents the kernel window size and d represents the word embedding dimension; text i It is the output of the i-node after passing through TextCNN. It is the output of TextCNN of all the neighbor nodes of i; and What is minimized is the difference between the shared vector of the central node and its neighboring nodes. What is minimized is the difference between the shared vectors of neighboring nodes in two networks; μ1 and μ2 control the importance of intra-network loss and inter-network loss.

[0030] Preferably, the attribute information of the network user nodes is obtained, including:

[0031] Vectorize the original attribute information of the nodes to obtain the initial attribute information f. i , will f i As input to a two-layer GCN network, the output of node i, which aggregates first-order neighbor nodes at each layer, is: The output of the aggregated second-order neighbor nodes is Obtain the first-order neighbor aggregation vector matrix of the node. and second-order neighbor aggregation vector Aggregation via fully connected neural networks:

[0032]

[0033]

[0034] in It is a learnable parameter matrix, b∈R dim It is a bias term, and the loss function L is obtained by constraining the neural network through node structure information. attr ;

[0035] Aggregate all hidden layer output vectors as attribute information for the end user:

[0036]

[0037] Where, θ (l) The importance of controlling the output of each GCN layer.

[0038] Preferably, the optimization equation for the transfer matrix from the source network node embedding matrix to the target network node embedding matrix, based on the optimal transmission theory, is as follows:

[0039] L=||Cost⊙Q|| F +λ1L topology +λ2L one-to-one

[0040] Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j ))

[0041]

[0042]

[0043] Where X and Y represent the structural embedding matrix, behavioral embedding matrix, and attribute embedding matrix of the source and target networks, respectively; Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j )) represents the sinkhorn distance between source network node i and target network node j, L topology L represents the graph structure constraint. one-to-one This represents alignment constraints; λ1 and λ2 control the importance of these two constraints. and Let I represent the adjacency matrix with self-loops of the source and target networks, and let I denote the identity matrix. and It is the Laplacian matrix of the source network graph and the target network graph, where and

[0044] Preferably, the node structure information, node behavior information, and node attribute information of the source network and the target network are transmitted in an optimized manner to obtain the corresponding transition matrix. Three similarity matrices can then be obtained using cosine similarity as the structural similarity matrix sim. struct Behavioral similarity matrix sim act and attribute similarity matrix sim attr Based on the gating mechanism, these three similarity matrices are fused:

[0045] hidden1=σ(W gate1 [sim struct , sim act , sim attr ]+b1)

[0046] hidden2=σ(W gate2 hidden1+b2)

[0047] gate=σ(W gate hidden2+b gate )

[0048] fusedSim=gate⊙[sim struct , sim act , sim attr ]

[0049]

[0050] Where hidden1 and hidden2 represent the hidden layer output vectors, W gate1 W gate2 W gate denoted as the learnable parameter matrix, gate as the calculated gate vector, fusedSim as the final fused node similarity matrix, Loss as the loss function of the neural network, α1, α2, and α3 control the importance of node structural similarity, node behavioral similarity, and node attribute similarity, respectively; σ is the activation function.

[0051] Secondly, the present invention provides a social network alignment system based on multi-information fusion and graph optimization, comprising:

[0052] The information acquisition module is used to acquire user structural information based on the centrality principle, acquire user behavioral information based on text from multi-source networks, and acquire user attribute information based on a graph neural network model of multi-hop neighbor aggregation, based on the constructed social network user graph data.

[0053] The matrix acquisition module is used to acquire the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavior information embedding matrix, and attribute information embedding matrix, respectively, based on the optimization model constructed according to graph optimization theory.

[0054] The similarity matrix acquisition module is used to calculate the embedding similarity matrix of three nodes in the source network and the target network based on the transmission matrix.

[0055] The alignment node acquisition module is used to obtain the fusion embedding similarity matrix based on the gating mechanism and to obtain the user alignment node.

[0056] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the social network alignment method based on multi-information fusion and graph optimization as described above.

[0057] Fourthly, the present invention provides a computer program product, including a computer program that, when run on one or more processors, is used to implement the social network alignment method based on multi-information fusion and graph optimization as described above.

[0058] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the social network alignment method based on multi-information fusion and graph optimization as described above.

[0059] The beneficial effects of this invention are: it can extract users' semantic features, reduce the sparsity of the network structure, eliminate the need for pre-training data, and is more suitable for the needs of large-scale networks; by learning user attribute features and network topology features, it reduces the interference of structural noise and attribute noise, and improves the accuracy of cross-social network user alignment.

[0060] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1This is a structural diagram of the social network alignment model based on multi-information fusion as described in an embodiment of the present invention.

[0063] Figure 2 This is a comparison chart of the running times of each model in the comparative experiment described in the embodiments of the present invention.

[0064] Figure 3 This is a schematic diagram illustrating the influence of each model and its components in the comparative experiment described in this embodiment of the invention. Detailed Implementation

[0065] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0066] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0067] Example 1

[0068] In this embodiment 1, a social network alignment system based on multi-information fusion and graph optimization is first provided. The system includes: an information acquisition module, used to acquire user structural information based on the centrality principle, user behavioral information based on text from a multi-source network, and user attribute information based on a graph neural network model of multi-hop neighbor aggregation, according to the constructed social network user graph data; a matrix acquisition module, used to acquire the transmission matrix from the source network node structure embedding matrix to the target network structure information embedding matrix, behavioral information embedding matrix, and attribute information embedding matrix, respectively, according to an optimization model constructed based on graph optimization theory; a similarity matrix acquisition module, used to calculate the three node embedding similarity matrices of the source network and the target network based on the transmission matrix; and an alignment node acquisition module, used to acquire the fused embedding similarity matrix based on a gating mechanism to acquire the user alignment node.

[0069] In this embodiment 1, the above-described system is used to implement a social network alignment method based on multi-information fusion and graph optimization, including: using an information acquisition module, based on the constructed social network user graph data, to acquire user structural information based on the centrality principle, user behavioral information based on text from multi-source networks, and user attribute information based on a graph neural network model of multi-hop neighbor aggregation; using a matrix acquisition module, based on an optimization model constructed based on graph optimization theory, to acquire the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavioral information embedding matrix, and attribute information embedding matrix, respectively; using a similarity matrix acquisition module, to calculate the three node embedding similarity matrices of the source network and the target network based on the transmission matrices; and using an alignment node acquisition module, to acquire the fused embedding similarity matrix based on a gating mechanism, and to acquire the user alignment node.

[0070] Constructing social network user graph data includes: constructing social network user data, including the social network topology graph structure, the text information posted by users, and the attributes of user homepages; obtaining the structural information of network user nodes through feature vector centrality, extracting the behavioral information of network user nodes through the BERT model, and extracting the attribute information of network user nodes through a GCN network based on multi-hop neighbor aggregation; and fusing the three types of user information based on a gating mechanism to obtain the user vector representation.

[0071] Constructing social network user structure information and obtaining user behavior information includes:

[0072] Let the source network be G. S =(V S E s F s The target network is G. T =(V T E T F T This paper constructs an undirected weighted graph using the social network relationships of users in the input data; extracts the feature vector centrality of users using the edge weight matrix W; and obtains the user's structural vector representation struct using logarithmic binning techniques. * (i);

[0073] Input all of the user's post text into the Bert model to obtain the user's text vector. User text vectors are formed by adding the shared semantics of the source and target networks and the global semantics of the network in which they reside, i.e.:

[0074] sharedEmb i =textEmb * (i)-globalEmb *

[0075] sharedEmb i Indicates shared semantics, globalEmb * It represents the global semantics of the source or target network. The global semantics can be obtained by inputting all the text of the current network into the BERT model.

[0076] By training the TextCNN network, the distance between the shared vector of the central node and the shared vector of the neighboring nodes is minimized, thus obtaining the final vector representation of user behavior information.

[0077] The final user behavior information vector representation is obtained, including:

[0078] text i =ReLU(Conv1D(sharedEmb) i ,h)+b)

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Where ReLU represents the activation function, Conv1D represents the TextCNN convolutional layer, and the kernel size is h×d, where h represents the kernel window size and d represents the word embedding dimension; text i It is the output of the i-node after passing through TextCNN. It is the output of TextCNN of all the neighbor nodes of i; and What is minimized is the difference between the shared vector of the central node and its neighboring nodes. What is minimized is the difference between the shared vectors of neighboring nodes in two networks; μ1 and μ2 control the importance of intra-network loss and inter-network loss.

[0085] Obtain the attribute information of network user nodes, including:

[0086] Vectorize the original attribute information of the nodes to obtain the initial attribute information f. i , will f i As input to a two-layer GCN network, the output of node i, which aggregates first-order neighbor nodes at each layer, is: The output of the aggregated second-order neighbor nodes is Obtain the first-order neighbor aggregation vector matrix of the node. and second-order neighbor aggregation vector Aggregation via fully connected neural networks:

[0087]

[0088]

[0089] in It is a learnable parameter matrix, b∈R dim It is a bias term, and the loss function L is obtained by constraining the neural network through node structure information. attr ;

[0090] Aggregate all hidden layer output vectors as attribute information for the end user:

[0091]

[0092] Where, θ (l) The importance of controlling the output of each GCN layer.

[0093] The optimization equation for the transition matrix from the source network node embedding matrix to the target network node embedding matrix, based on the optimal transmission theory, is as follows:

[0094] L=||Cost⊙Q|| F +λ1L topology +λ2L one-to-one

[0095] Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j ))

[0096]

[0097]

[0098] Where X and Y represent the structural embedding matrix, behavioral embedding matrix, and attribute embedding matrix of the source and target networks, respectively; Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j )) represents the sinkhorn distance between source network node i and target network node j, L topology L represents the graph structure constraint. one-to-one This represents alignment constraints; λ1 and λ2 control the importance of these two constraints. and Let I represent the adjacency matrix with self-loops of the source and target networks, and let I denote the identity matrix. and It is the Laplacian matrix of the source network graph and the target network graph, where and

[0099] For the node structure information, node behavior information, and node attribute information of the source network and the target network, optimal transmission is performed to obtain the corresponding transition matrices. Three similarity matrices, sim, can be obtained through cosine similarity to form the structural similarity matrix. struct Behavioral similarity matrix sim act and attribute similarity matrix sim attr Based on the gating mechanism, these three similarity matrices are fused:

[0100] hidden1=σ(W gate1 [sim struct , sim act , sim attr ]+b1)

[0101] hidden2=σ(W gate2 hidden1+b2)

[0102] gate=σ(W gate hidden2+b gate )

[0103] fusedSim=gate⊙[sim struct , sim act , sim attr ]

[0104]

[0105] Where hidden1 and hidden2 represent the hidden layer output vectors, W gate1 W gate2 W gate denoted as the learnable parameter matrix, gate as the calculated gate vector, fusedSim as the final fused node similarity matrix, Loss as the loss function of the neural network, α1, α2, and α3 control the importance of node structural similarity, node behavioral similarity, and node attribute similarity, respectively; σ is the activation function.

[0106] Example 2

[0107] This embodiment 2 provides a social network user alignment method based on multi-information fusion. The method includes: constructing social network user graph data; obtaining user structural information based on the centrality principle; obtaining user behavioral information based on text from multi-source networks; and obtaining user attribute information based on a graph neural network model using multi-hop neighbor aggregation. An optimization model is designed based on graph optimization theory to obtain the transfer matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavioral information embedding matrix, and attribute information embedding matrix, respectively. Based on these three transfer matrices, three node embedding similarity matrices of the source and target networks are calculated. These similarity matrices are fused using a gating mechanism to finally obtain the user-aligned nodes.

[0108] This involves constructing social network user data, including the social network topology, user-posted text information, and user profile attributes. The structural information (struct) of network user nodes is obtained through feature vector centrality. * (i) Extracting the behavioral information of network user nodes using the Bert model (sharedEmb) i The attribute information attrEmb of network user nodes is extracted through a GCN network based on multi-hop neighbor aggregation. * (i). Based on the gating mechanism, the user vector representation fusedEmb is obtained by fusing three types of user information. * (i). Here, *∈{S,T} represents the source network or the target network.

[0109] Constructing social network user structure information and obtaining user behavior information includes:

[0110] Let the source network be G. S =(V S E S F S The target network is G. T =(V T E T F T This paper constructs an undirected weighted graph using the social network relationships of users in the input data. The feature vector centrality of users is extracted using the edge weight matrix W, and the structural vector representation of users is obtained through logarithmic binning. * (i).

[0111] Input all of the user's post text into the Bert model to obtain the user's text vector. User text vectors are formed by adding the shared semantics of the source and target networks and the global semantics of the network in which they reside.

[0112] sharedEmb i =textEmb* (i)-globalEmb *

[0113] sharedEmb i Indicates shared semantics, globalEmb * This represents the global semantics of the source or target network. The global semantics can be obtained by inputting all text into the current network's BERT model. This is achieved through training T... e The xtCNN network minimizes the distance between the shared vector of the central node and the shared vector of its neighboring nodes to obtain the final vector representation of user behavior information.

[0114] text i =ReLU(Conv1D(sharedEmb) i ,h)+b)

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] Where ReLU represents the activation function, Conv1D represents the TextCNN convolutional layer, the kernel size is h×d, h represents the kernel window size, and d represents the word embedding dimension. i It is the output of the i-node after passing through TextCNN. It is the output of TextCNN for all neighboring nodes of i. and What is minimized is the difference between the shared vector of the central node and its neighboring nodes. What is minimized is the difference between the shared vectors of neighboring nodes in two networks. μ1 and μ2 control the importance of intra-network loss and inter-network loss, respectively.

[0121] Obtain the attribute information of network user nodes, including:

[0122] Vectorize the original attribute information of the nodes to obtain the initial attribute information f. i , will f i As input to a two-layer GCN network, the output of node i, which aggregates first-order neighbor nodes at each layer, is: The output of the aggregated second-order neighbor nodes is Obtain the first-order neighbor aggregation vector matrix of the node. and second-order neighbor aggregation vector Aggregation is achieved through a fully connected neural network.

[0123]

[0124]

[0125] in, It is a learnable parameter matrix, b∈R dim It is a bias term, and the loss function L is obtained by constraining the neural network through node structure information. attr Aggregate all hidden layer output vectors as attribute information for the end user.

[0126]

[0127] Where, θ (l) The importance of controlling the output of each GCN layer.

[0128] The optimization equation for the transition matrix from the source network node embedding matrix to the target network node embedding matrix, based on the optimal transmission theory, is as follows:

[0129] L=||Cost⊙Q|| F +λ1L topology +λ2L one-to-one

[0130] Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j ))

[0131]

[0132]

[0133] Where X and Y can represent the structural embedding matrix, behavioral embedding matrix, and attribute embedding matrix of the source and target networks, respectively. Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j )) represents the sinkhom distance between source network node i and target network node j, L topology L represents the graph structure constraint. one-to-one These represent alignment constraints. λ1 and λ2 control the importance of these two constraints. and Let I represent the adjacency matrix with self-loops of the source and target networks, and let I denote the identity matrix. and It is the Laplacian matrix of the source network graph and the target network graph, where and

[0134] For the node structure information, node behavior information, and node attribute information of the source network and the target network, optimal transmission is performed to obtain the corresponding transition matrices. Three similarity matrices, sim, can be obtained through cosine similarity to form the structural similarity matrix. struct Behavioral similarity matrix sim act and attribute similarity matrix sim attr Based on a gating mechanism, these three similarity matrices are fused:

[0135] hidden1=σ(W gate1 [sim struct Sim act , sim attr ]+b1)

[0136] hidden2=σ(W gate2 hidden1+b2)

[0137] gate=σ(W gate hidden2+b gate )

[0138] fusedSim=gate⊙[sim struct , sim act , sim attr ]

[0139]

[0140] Where hidden1 and hidden2 represent the hidden layer output vectors, W gate1 W gate2 W gate This represents the learnable parameter matrix, `gate` represents the calculated gate vector, `fusedSim` represents the final fused node similarity matrix, `Loss` is the loss function of the neural network, and `α1`, `α2`, and `α3` control the importance of node structural similarity, node behavioral similarity, and node attribute similarity, respectively. `σ` is the activation function, specifically the sigmoid function.

[0141] Example 3

[0142] To address the issues that existing network alignment solutions are susceptible to structural and attribute noise, resulting in low alignment accuracy, and that their high model complexity makes them unsuitable for large-scale network alignment, this embodiment 3 proposes a social network alignment scheme based on multi-information fusion, which mainly includes the following:

[0143] Multiple information acquisition from users. Structural information of network user nodes is obtained through feature vector centrality, and logarithmic binning is used to obtain user structure vector representations. Behavioral information of network user nodes is extracted using the BERT model, and a TextCNN network is trained using a loss function constructed from the text embeddings of neighboring nodes to obtain stable multi-source network shared behavioral vectors. Attribute information of network user nodes is extracted using a GCN network based on multi-hop neighbor aggregation, and the neural network is trained by constraining the node structure information.

[0144] Similarity calculation is based on optimal transmission and similarity aggregation based on a gating mechanism. For the node structure, behavior, and attribute information of both the source and target networks, optimal transmission is performed to obtain the corresponding transition matrices. Three similarity matrices are then obtained using cosine similarity. A gating mechanism is designed to aggregate these three similarity matrices to obtain the final user-aligned nodes.

[0145] For a social network user alignment method based on multi-information fusion: Nodes from different social networks are obtained through web crawling, and nodes belonging to the same user in two networks are denoted as anchor nodes. The user's friend relationships are obtained, and edges are added between user nodes with friend relationships to construct a graph structure, represented as G = (V, E, F), where V represents the node set, E represents the edge set, and F represents the node attribute. Existing network alignment methods often simply set the weight of network edges to 1. However, real social networks such as Weibo and Douban may have mutual following relationships, which are obviously closer than one-way following. Therefore, when constructing the graph, the weight of bidirectional nodes is set to 2, and the weight of unidirectional nodes is set to 1. This setting increases the influence of neighboring nodes with bidirectional friend relationships and weakens the influence of neighboring nodes with only unidirectional relationships, resulting in an edge weight matrix W∈R. |V|×|V| Noting that the neighborhood structures of anchor nodes in different social networks are usually consistent, this paper utilizes structural information to assist in social network alignment. For the graph structures of the source and target networks, the structural vectors of the graph nodes are obtained using eigenvector centrality that considers edge weights, as shown in the following formula:

[0146]

[0147] Among them, A i，j Let C represent the element in the i-th row and j-th position of the adjacency matrix. eigen (i) is a numerical score representing the feature centrality, C eigen (j) represents the centrality score of the neighboring nodes of node i. This is used to define the structural attribute characteristics of user i in the network. *(i) Using a logarithmic binning strategy, for the Katz centrality C of nodes in the network... eigen (i), 1 is allocated to struct * (i) At this point, `width` is the specified bin width. In this embodiment, `struct` will be used. * (i) represents the user's structure information.

[0148] For social network users, their posting information over a period of time is crawled as user behavior information. Considering that users' posting content and writing style are similar across different social networks, user behavior information can be used as auxiliary information for network alignment. First, all posts by user i are concatenated into a long sentence. *∈{S,T} represents the source network or target network, which is input into the BERT model to obtain the text embedding. However, text information across different social networks is often heterogeneous. For example, a user's Weibo post might be more about everyday life, while their Douban post might focus on movie sharing and reviews. To eliminate the potential impact of text heterogeneity on network alignment, a neural network is trained to obtain the shared text vector of anchor node (i, j) across different networks, defined as sharedEmb. i Therefore, user text embedding in a certain network can be represented as:

[0149] textEmb * (i) = sharedEmb i +globalEmb *

[0150] globalEmb * Represents the global semantic information of the source or target network. GlobalEmb is obtained by inputting all user text from a given network into the BERT model. * = Bert(text * The final user text shared vector is obtained as follows:

[0151] sharedEmb i =textEmb * (i)-globalEmb *

[0152] sharedEmb i The text shared vectors of the sharedEmb nodes have some similarity to those of the neighboring nodes. The neighboring nodes from the source and target networks are used to train the sharedEmb. i By minimizing the reconstruction error, a final stable user text vector representation is obtained:

[0153] text i =ReLU(Connv1D(sharedEmb) i ,h)+b)

[0154]

[0155]

[0156]

[0157]

[0158] in This represents the neighboring nodes of node i in the source or target network. This represents the number of neighboring nodes of node i in the network. Conv1D represents the TextCNN network, used for text feature extraction, with a convolutional kernel size of h×d, where h represents the kernel window size and d represents the word embedding dimension. It is the output of TextCNN for all neighboring nodes of i. and What is minimized is the difference between the shared vector of the central node and its neighboring nodes. What is minimized is the difference between the shared vectors of two network neighbor nodes.

[0159] Combining the three errors yields the final loss function:

[0160]

[0161] Where μ1 and μ2 are hyperparameters controlling the importance of the error. Using the neural network described above, a stable shared text vector `sharedEmb` can be obtained for a given user. i In this embodiment, it is regarded as user behavior information.

[0162] By crawling, we extract attribute information that exists on both user homepages in two networks, and vectorize it. The attribute information of user i is represented as f. i In this embodiment, a graph neural network is used to train attribute information. In traditional graph neural network structures, such as GCN and GAT models, the one-hop neighbors of node i are usually used to obtain the vector representation of the hidden layer. Here, l∈{1,...,k} represents the l-th layer of the graph neural network. However, this simple method of aggregating first-order neighbors cannot characterize the situation in real-world social networks. For example, two users may be friends on Weibo but not on Douban, even if they have similar interests and follow the same user. In this case, they are first-order neighbors on Weibo but second-order neighbors on Douban. Obviously, learning the vector representations of users on the two networks through the neural network will lead to discrepancies.

[0163] Therefore, in this embodiment, a GCN that aggregates one-hop and two-hop neighbors is used to learn the vector representation of nodes. For one-hop neighbor information aggregation, the hidden layer vector of node i in the l-th layer of the GCN is:

[0164]

[0165] Where j∈N1(i) represents all first-order neighbor nodes of node i, c i It is a normalized constant. These are learnable parameters, where ⊙ represents element-wise multiplication. Edge weights W are used to adjust the aggregate weights of neighbor node features. Node attribute information is used as input to the neural network, i.e. Similarly, for two-hop neighbor aggregation, the hidden layer vector of the user is learned using the following formula:

[0166]

[0167] Where α ij (l) Let represent the weights between node i and its second-hop neighbor nodes. Note that not all second-hop neighbors contribute positively to network alignment. The GCN model learns rich information about node proximity structure and node attributes; nodes with closer relationships to a node are more helpful in representing the node and can improve the accuracy of network alignment. Therefore, given a node i and its second-hop neighbor node j, their calculated weights are:

[0168]

[0169] Where Ws is a learnable parameter, the first-order neighbor aggregation vector matrix of the node is obtained. and second-order neighbor aggregation vector Then, aggregate using the following formula:

[0170]

[0171] in It is a learnable parameter matrix, b∈R dim This is a bias term. Ultimately, at each layer of the GCN, we can obtain a dimension of dim.(l) The node embedding vector. In this embodiment, the structural information vector of the node is used as a constraint term to learn the user's attribute vector representation, and its loss function is:

[0172]

[0173] While multi-level neighbor aggregation GCNs can capture more neighbor node information in a single layer, multi-layer GCNs can capture information at multiple scales through inter-layer information transfer and feature updates. Furthermore, the nonlinear transformations of each layer gradually enhance the nonlinear expressive power of node features. Therefore, this embodiment uses a two-layer GCN network. The node vector representations learned in each layer are helpful for network alignment, so this paper needs to aggregate the embedding vectors output from all layers. Considering the potential differences in semantic information between different networks, their embedding spaces will also differ, leading to anchor nodes not being close in the embedding space. Therefore, this embodiment shares the weight matrix between the source and target GCNs, ensuring consistency in the embedding space. In this embodiment, the embedding dimension dim for each layer is set. (l) =dim, for the source network and the target network, their embedding matrices will be obtained as follows:

[0174]

[0175] θ (l) This is a configurable hyperparameter. Source and target network embedding matrices. |V S |,|V T | represents the number of nodes in the source network and the target network, respectively. In this embodiment, attrEmb * (i) is the attribute information of node i.

[0176] In this embodiment, for the user vector alignment method based on optimal transmission theory: Optimal transmission theory provides an efficient method for comparing two probability distributions. Traditional graph alignment methods use a method of first calculating the similarity of embedding vectors and then directly comparing them, ignoring the semantic differences that may exist in the structural, behavioral, and attribute information of the two networks. If the node features in the two networks are inconsistent, the corresponding node embeddings are usually also inconsistent and cannot be directly compared. Furthermore, optimal transmission theory is more interpretable than traditional network alignment methods.

[0177] By inputting the source network and the target network into the neural network model proposed above, we can obtain the structure embedding vectors (structs) of the source network and the target network, respectively. * (i) Behavior embedding vector sharedEmb *(i) and attribute embedding vector attrEmb * (i) The goal of optimal transmission is to find a transmission matrix. This minimizes the cost of transmitting data from the source network node to the target network. The optimization equation for Q is:

[0178] argmin Q∈B f(Q) = ||Cost⊙Q|| F

[0179]

[0180] Here, Cost represents the cost matrix. Traditional methods use Euclidean distance to measure the geometric distance between two vectors, without considering the probability distribution information of the data. Sinkhorn distance, on the other hand, can measure the distance between two distributions and performs better in handling social network embedding alignment. The cost matrix is ​​defined for the three embedding matrices of the source and target networks as follows:

[0181] Cost(s i , t j ) = sinkhorn(X(s i ), Y(t) j ))

[0182] X∈{struct S sharedEmb S ,attrEmb S}

[0183] Y∈{struct T sharedEmb T ,attrEmb T}

[0184] in To represent nodes in two networks, the sinkhorn distance is calculated using an iterative update method, as follows:

[0185] (1) Calculate X S (s i ), Y T (t j Euclidean distance between )

[0186] (2) Initialize the two matrices Where ε is the regularization parameter.

[0187] (3) Iterate and update these two matrices for t rounds, using the following update formula:

[0188]

[0189]

[0190] (4) The final sinkhorn distance can be obtained by the following formula:

[0191]

[0192] During optimized transmission, the network structure may change, which is detrimental to network alignment. That is, if nodes are connected in the source network, they should also be connected in the corresponding second network after transmission. Therefore, a topology penalty term needs to be introduced:

[0193]

[0194] in and Let I represent the adjacency matrix with self-loops of the source and target networks, and let I denote the identity matrix. and It is the Laplacian matrix of the source network graph and the target network graph, where and Furthermore, our method only considers one-to-one alignment, so we need to add a constraint to ensure that a node in the source network can have at most one aligned node in the target network:

[0195]

[0196] Combining the above factors, the final optimization equation is obtained as follows:

[0197] L=||Cost⊙Q|| F +λ1L topology +λ2L one-to-one

[0198] Update L using gradient descent to minimize it:

[0199]

[0200] Here, λ1 and λ2 control the importance of topological constraints and one-to-one constraints. The transition matrix Q can be iteratively updated as follows:

[0201]

[0202] η is a hyperparameter used for gradient updates. After T rounds of iterative optimization, the final transition matrix Q from the source network to the target network is obtained. The source network embedding matrix is ​​then transferred to the target network space as follows:

[0203] map(X S ) = XS ·Q

[0204] The similarity matrix between the source network nodes and the target network nodes can be calculated using the cosine similarity formula:

[0205]

[0206] By optimizing the transmission of structural, behavioral, and attribute information of the source and target networks respectively, the structural similarity matrix sim can be obtained. struct Behavioral similarity matrix sim act and attribute similarity matrix sim attr In this embodiment, a gating mechanism is used to fuse these three similarity matrices of the nodes. The gating vector is calculated using a multilayer perceptron network as follows:

[0207] hidden1=σ(W gate1 [sim struct , sim act , sim attr ]+b1)

[0208] hidden2=σ(W gate2 hidden1+b2)

[0209] gate=σ(W gate hidden2+b gate )

[0210] Where hidden1 and hidden2 represent the hidden layer output vectors, W gate1 W gate2 W gate Let represent the learnable parameter matrix, and 'gate' represent the calculated gate vector. The final aggregate vector is:

[0211] fusedSim=gate⊙[sim struct , sim act , sim attr ]

[0212] The multilayer perceptron network is optimized by minimizing the distance between the aggregate vector and structural, behavioral, and attribute information. The loss function equation is as follows:

[0213]

[0214] α1, α2, and α3 control the importance of structural information loss, behavioral information loss, and attribute information loss, respectively. Through the learning of the above neural network, the user's aggregated similarity matrix fusedSim can be obtained from the source network and the target network. The maximum value in each row of the similarity matrix is ​​the aligned node obtained by the model in this embodiment.

[0215] Comparative experiment:

[0216] In this embodiment, the method is also verified through comparative experiments, the specific contents of which include:

[0217] The dataset used in the experiment includes:

[0218] 1. Douban Online and Douban Offline Datasets: These are two networks collected from the Douban social network, treating online users as nodes and friend relationships as edges. Douban Online contains online users, while Douban Offline contains users who appear together in offline social activities. Each dataset contains 1118 anchor nodes.

[0219] 2. Allmovie and Tmdb datasets: Allmovie is constructed from movies on the Rotten Tomatoes website. If two movies share at least one common actor, then an edge exists between the movie nodes. The Imdb network is similarly extracted from the Imdb online website. It contains 5176 anchor nodes.

[0220] 3. Facebook and Twitter datasets: These are two network structures scraped from the two major social networking sites, Facebook and Twitter, containing 1043 anchor nodes.

[0221] The detailed information of the dataset is shown in Table 1:

[0222] Table 1. Dataset Description

[0223]

[0224] In this embodiment, the evaluation metrics used in the experiment are as follows:

[0225] 1. From the perspective of prediction accuracy, the Acc@q evaluation method is used. This method measures how many of the top-q candidate nodes are true anchor nodes. The calculation formula is as follows:

[0226]

[0227] Where M represents the true aligned node set, sim q(u) represents the top-q elements in the u-th row of the alignment matrix sim, sorted by value from largest to smallest.

[0228] 2. From a ranking perspective, the Mean Average Precision (MAP) is used to evaluate the overall performance of the algorithm.

[0229]

[0230] Among them, rank i This represents the rank of the true aligned node in the anchor node set M within the alignment matrix obtained from the model output. A larger MAP result indicates higher evaluation accuracy of the algorithm.

[0231] The comparison algorithm used in the experiment is as follows:

[0232] 1. Big-Align: It transforms the network alignment problem into a bipartite graph problem, represents the corresponding matrix lookup task as an optimization problem, and provides an efficient solution algorithm.

[0233] 2. REGAL: This method proposes a cross-network matrix factorization approach to learn representations of multiple networks while maintaining structural similarity, and to match nodes between different networks based on the learned embeddings.

[0234] 3. CENALP: This method unifies network alignment and link prediction tasks. First, it utilizes a biased random walk strategy customized across networks, and then learns node embeddings by maximizing the co-occurrence probability of nodes during the walk.

[0235] 4. GAlign: This method uses a multi-order GCN model based on local and global network structure and attribute information to obtain the alignment matrix of the network.

[0236] 5. CONE: This method learns proximity within the network through node embedding and matches nodes across the network through embedding subspace alignment.

[0237] 5. Grad-Align: This method gradually discovers aligned node pairs through multi-layer embedding similarity and Tversky similarity of GNN.

[0238] Experimental setup: For the hyperparameters in this model, the node behavior information control weights are μ1 = 0.5, μ2 = 0.5; the gated loss control weights are α1 = 0.2, α2 = 0.2, α3 = 0.6; the neural network model output dimension is dim = 128; and the multi-layer GCN embedding aggregation parameters are... The nearest neighbor node count K=1, the optimal transmission update rounds T=50, the sinkhorn distance update rounds t=10, and the optimal transmission update parameter η=0.01. For the hyperparameters of the comparison models, the parameters with the best experimental results in the paper were used. Each model was tested 50 times, and the average of the results was taken as the final result to reduce random error. All experiments in this paper were conducted on an Intel(R) Xeon(R) Gold 5118 CPU@2.30GHz system, which included a 16GB Tesla V100 GPU.

[0239] Table 2 Experimental Results

[0240]

[0241] Experimental Results: To verify the performance of the proposed model on real datasets, the model was compared with several state-of-the-art unsupervised network alignment models. The experimental results are shown in Table 2. The results show that, across the three models used in the experiment, the proposed model outperforms the other baseline models in both accuracy and mean precision.

[0242] BigAlign and CONE only consider network alignment from the perspective of network topology, resulting in poor alignment performance on the Doubanonline-Douban offline and Allmovie-Imdb datasets. This is because these networks suffer from network sparsity in real social networks and structural semantic inconsistencies between different networks. Conversely, since both GAlign and Grad models use node attributes as input to the neural network, the consistency between node attributes allows these models to perform well on the Douban online-Douban offline and Allmovie-Imdb datasets. However, on the Facebook-Twitter dataset, which lacks node attributes, GAlign and Grad perform poorly in network alignment, while the CONE model, which only considers topology, performs better, indicating a strong structural consistency between the Facebook and Twitter networks on this dataset. The network alignment model proposed in this paper considers both network node attributes and the potential structural consistency between different networks, making the model more universal and performing well in alignment across various social networks.

[0243] Model Performance: To compare the average runtime complexity of the models, experiments were conducted on three datasets. The experimental results are as follows: Figure 2As shown, it can be seen that the running time of the model in this embodiment is comparable to that of lightweight models such as GAlign and Grad, and CENALP has the highest running time for all datasets. Figure 2 In the bar charts showing the datasets Douban, Allmovie-lmbd, and Facebook-Twitter on the horizontal axis, the bars from left to right represent the model of this embodiment, BigAlign, REGAL, CENALP, Galign, CONE, and Grad, respectively. While BigAlign has the lowest runtime across all datasets, its alignment accuracy is also the worst. Therefore, the model proposed in this embodiment demonstrates good performance in both time efficiency and algorithm accuracy.

[0244] Ablation experiment: To verify the function of each part of the model, the model settings were changed to obtain the following model:

[0245] 1. Model 1: Complete model, which includes fused node structure information, behavior information and attribute information.

[0246] 2. Model 2: Model lacking structural information

[0247] 3. Model 3: Model lacking behavioral information

[0248] 4. Model 4: Model lacking attribute information

[0249] Because the Facebook-Twitter dataset lacks attribute information, this embodiment conducts experiments on the Douban and Allmovie-Imdb datasets, and the experimental results are shown in the figure below. Figure 3 As shown. Figure 3 In the bar chart for each model, the left bars represent the Douban dataset, and the right bars represent the Allmovie-Imdb dataset.

[0250] Depend on Figure 3 The results show that structural information has little impact on node alignment in real social networks, while behavioral information and node attributes have a greater impact on alignment performance. In other words, structural consistency is usually not satisfied in real social networks, so structure contributes relatively little to alignment, while richer textual and attribute information contributes more to node alignment.

[0251] Example 4

[0252] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When executed by a processor, the computer instructions implement the social network alignment method based on multi-information fusion and graph optimization as described above. The method includes:

[0253] Based on the constructed social network user graph data, the structural information of users is obtained based on the centrality principle, the behavioral information of users is obtained based on the text of multi-source networks, and the attribute information of users is obtained based on the graph neural network model of multi-hop neighbor aggregation.

[0254] Based on the optimization model constructed using graph optimization theory, the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavior information embedding matrix, and attribute information embedding matrix are obtained respectively.

[0255] The embedding similarity matrix of the three nodes in the source network and the target network is calculated based on the transmission matrix;

[0256] The fusion embedding similarity matrix is ​​obtained based on the gating mechanism, and the user alignment node is obtained.

[0257] Example 5

[0258] This embodiment 5 provides a computer program product, including a computer program that, when run on one or more processors, implements the social network alignment method based on multi-information fusion and graph optimization as described above. The method includes:

[0259] Based on the constructed social network user graph data, the structural information of users is obtained based on the centrality principle, the behavioral information of users is obtained based on the text of multi-source networks, and the attribute information of users is obtained based on the graph neural network model of multi-hop neighbor aggregation.

[0260] Based on the optimization model constructed using graph optimization theory, the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavior information embedding matrix, and attribute information embedding matrix are obtained respectively.

[0261] The embedding similarity matrix of the three nodes in the source network and the target network is calculated based on the transmission matrix;

[0262] The fusion embedding similarity matrix is ​​obtained based on the gating mechanism, and the user alignment node is obtained.

[0263] Example 6

[0264] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the social network alignment method based on multi-information fusion and graph optimization as described above. The method includes:

[0265] Based on the constructed social network user graph data, the structural information of users is obtained based on the centrality principle, the behavioral information of users is obtained based on the text of multi-source networks, and the attribute information of users is obtained based on the graph neural network model of multi-hop neighbor aggregation.

[0266] Based on the optimization model constructed using graph optimization theory, the transmission matrices from the source network node structure embedding matrix to the target network structure information embedding matrix, behavior information embedding matrix, and attribute information embedding matrix are obtained respectively.

[0267] The embedding similarity matrix of the three nodes in the source network and the target network is calculated based on the transmission matrix;

[0268] The fusion embedding similarity matrix is ​​obtained based on the gating mechanism, and the user alignment node is obtained.

[0269] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A social network alignment method based on multiple information fusion and graph optimization, characterized in that, Comprise: According to the constructed social network user graph data, the structural information of the user is obtained based on the centrality principle, the behavior information of the user is obtained based on the text of the multi-source network, and the attribute information of the user is obtained based on the multi-hop neighbor aggregation graph neural network model; According to the optimization model constructed based on the graph optimization theory; Respectively obtain the transmission matrix of the source network node structure embedding matrix to the target network structure information embedding matrix, the behavior information embedding matrix and the attribute information embedding matrix; Based on the transmission matrix, three node embedding similarity matrices of the source network and the target network are calculated; For the source network and target network node structure information, node behavior information and node attribute information, the optimal transmission is made to obtain the corresponding transition matrix, and the three similar shape matrices are obtained as the structure similarity matrix , the behavior similarity matrix , and the attribute similarity matrix ; Based on the gating mechanism, the three similarity matrices are fused: ； ； ； ； ； wherein, denotes the hidden layer output vector, denotes the learnable parameter matrix, denotes the computed gating vector, denotes the final obtained fused node similarity matrix, is the loss function of the neural network, respectively control the importance of node structure similarity, node behavior similarity and node attribute similarity; is an activation function; Based on the gating mechanism, the fused embedding similarity matrix is obtained, and the user alignment node is obtained.

2. The social network alignment method based on multiple information fusion and graph optimization according to claim 1, characterized in that, Constructing social network user graph data, comprising: constructing social network user data, including social network topology graph structure, user published text information and user homepage attribute; the structural information of the network user node is obtained through the feature vector centrality, the behavior information of the network user node is extracted through the Bert model, and the attribute information of the network user node is extracted through the GCN network based on multi-hop neighbor aggregation; the user vector representation is obtained by fusing the three kinds of information of the user based on the gating mechanism. 3.The social network alignment method based on multiple information fusion and graph optimization according to claim 2, characterized in that, Constructing social network user structure information and obtaining user behavior information, comprising: The source network is , the target network is , the social network relationship of the user in the input data is used to construct an undirected weighted graph, the edge weight matrix W is used to extract the feature vector centrality of the user, and the logarithmic binning technology is used to obtain the structural vector representation of the user ; Input all the post text of the user into the Bert model to obtain the user text vector The user text vector is composed of the shared semantics of the source network and the target network and the global semantics in the network, that is: ； wherein represents a shared semantic, represents a global semantic of the source network or the target network, and the global semantic can be obtained by inputting all text of the current network into a Bert model. By training the TextCNN network, the distance between the shared vector of the central node and the shared vector of the neighbor node is minimized to obtain the final user behavior information vector representation.

4. The social network alignment method based on multiple information fusion and graph optimization of claim 3, wherein, Obtaining the final user behavior information vector representation, comprising: ； ； ； ； ； ； where, represents the activation function, represents the TextCNN convolution layer, and the convolution kernel size is , represents the convolution kernel window size, represents the word embedding dimension; is the output of the i node through the TextCNN, is the output of all neighbor nodes of i through the TextCNN; and minimized is the difference between the center node shared vector and the neighbor node, minimized is the difference between the neighbor node shared vectors of the two networks; controlled is the importance of the intra-network loss and the inter-network loss.

5. The social network alignment method based on multiple information fusion and graph optimization according to claim 4, characterized in that, Obtaining the attribute information of the network user node, comprising: Vectorize the original attribute information of the nodes to obtain the initial attribute information. ,Will As input to a two-layer GCN network, the output of node i, which aggregates first-order neighbor nodes at each layer, is: The output of the aggregated second-order neighbor nodes is This yields the first-order neighbor aggregation vector matrix of the node. and second-order neighbor aggregation vector Aggregation through a fully connected neural network: ； ； wherein is a learnable parameter matrix, is a bias term, the loss function is constrained by the node structure information ; Aggregating all hidden layer output vectors as the final user attribute information: ； wherein, Controlling the importance of each layer GCN output.

6. The social network alignment method based on multiple information fusion and graph optimization of claim 5, wherein, Based on the optimization transmission theory, the transfer matrix optimization equation for transferring the source network node embedding matrix to the target network node embedding matrix is designed as: ； ； ； ； where X and Y represent the structure embedding matrix, the behavior embedding matrix, and the attribute embedding matrix of the source network and the target network; denotes the sinkhorn distance between the source network node i and the target network node j, denotes the graph structure constraint, denotes the alignment constraint; controls the importance of the two constraints, and denotes the adjacency matrix with self-loops of the source network and the target network, I denotes the identity matrix, and is the Laplacian matrix of the source network graph and the target network graph, where and .

7. A social network alignment system based on multiple information fusion and graph optimization, characterized in that, Comprise: The information acquisition module is used for obtaining the structural information of the user based on the centrality principle according to the constructed social network user graph data, obtaining the behavior information of the user based on the text of the multi-source network, and obtaining the attribute information of the user based on the multi-hop neighbor aggregation graph neural network model; The matrix acquisition module is used for constructing the optimization model based on the graph optimization theory; Respectively obtain the transmission matrix of the source network node structure embedding matrix to the target network structure information embedding matrix, the behavior information embedding matrix and the attribute information embedding matrix; The similarity matrix acquisition module is used for calculating three node embedding similarity matrices of the source network and the target network based on the transmission matrix; For the node structure information, node behavior information and node attribute information of source network and target network, the optimal transmission is made respectively to obtain the corresponding transfer matrix, and three similar shape matrices are obtained as the structure similarity matrix , the behavior similarity matrix and the attribute similarity matrix ; Based on the gating mechanism, the three similarity matrices are fused: ； ； ； ； ； wherein, denotes the hidden layer output vector, denotes the learnable parameter matrix, denotes the computed gating vector, denotes the final obtained fused node similarity matrix, is a loss function of the neural network, respectively control the importance of node structure similarity, node behavior similarity and node attribute similarity; is an activation function. The alignment node acquisition module is used for obtaining the fused embedding similarity matrix based on the gating mechanism, and obtaining the user alignment node.

8. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is used for storing computer instructions, and the computer instructions are executed by the processor to realize the social network alignment method based on multiple information fusion and graph optimization in any one of claims 1-6.

9. An electronic device, comprising: Comprise: A processor, a memory and a computer program; wherein the processor is connected with the memory, the computer program is stored in the memory, when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the instructions for realizing the method for aligning social network based on multiple information fusion and graph optimization according to any one of claims 1-6.

Images