Dynamic social network alignment method based on longitudinal federation and canonical correlation analysis

Abstract

The invention belongs to the field of social network analysis, and particularly relates to a dynamic social network alignment method based on longitudinal federation and canonical correlation analysis, which comprises the following steps: simulating a user space-time relationship through a dynamic space-time diagram self-encoding memory model and an attention mechanism to obtain a user relationship matrix; constructing a user attribute matrix, and fusing the user relation matrix to obtain a user matrix; and through a training model based on federal learning, inputting the user feature matrixes of the platform X and the platform Y into the model for training to obtain a prediction result of cross-domain user alignment. According to the cross-platform user alignment prediction method, cross-platform user alignment prediction is carried out by aiming at the time sequence characteristics of the user relationship, combining the dynamic relationship expression updating and modeling and fusing other non-time sequence characteristics, the problems of cross-domain data privacy leakage, social network dynamics and the like can be effectively solved through the method, and finally accurate cross-platform network user alignment is realized.

Description

Technical Field

[0001] The present invention belongs to the field of social network analysis, and in particular relates to a dynamic social network alignment method based on vertical federation and canonical correlation analysis. Background Art

[0002] Social network alignment is to align the same user on different social networks, also known as anchor link prediction. It is widely used in cross-domain recommendation, link prediction and network fusion. In today's information society, social networks have become an important part of people's daily lives. People participate in multiple social media to meet their different spiritual needs. However, with the continuous emergence of various social network platforms, how to effectively integrate users and their relationship information on these platforms to achieve cross-platform information sharing and user behavior analysis has become a major challenge in current research. The core of this challenge lies in social network alignment, that is, matching and associating user information from different social network platforms.

[0003] For social network alignment, existing research methods can be roughly divided into two categories: one is the network alignment method based on graph neural network, which mainly uses the graph structure information of social networks and the characteristic information of users to align users. The other is the heterogeneous social network alignment method based on network representation learning, which uses the structural information in heterogeneous social networks and the complex interaction relationship between users to align users. However, the current user alignment still has the following shortcomings:

[0004] 1. The dynamics of social networks. In the real world, the topology of social networks is always changing over time. This dynamic change makes user behavior and relationship links extremely complex and difficult to accurately capture and predict. The frequent changes in user relationships increase the difficulty of alignment algorithms. Therefore, it is necessary to explore new technologies and methods to effectively respond to network dynamics and improve the accuracy of social network alignment.

[0005] 2. Sparsity and imbalance of cross-platform network structure. Although cross-domain networks may describe the same or similar users and relationships, due to the natural distribution of data, different collection methods and network construction strategies, the connections between user nodes are relatively sparse, resulting in unstable alignment results and low accuracy. At the same time, cross-domain network data is unbalanced, which increases the risk of privacy leakage for data-disadvantaged parties. Therefore, it is an important issue to strengthen data security and alleviate network sparsity in the case of data imbalance.

[0006] 3. Cross-domain user data leakage and distribution differences. When implementing cross-platform user alignment, centralized learning has the risk of user privacy data leakage. At the same time, different social platforms have different functions, user groups, and network structures. User data distribution may be biased, which in turn leads to differences in the distribution of user representations. Therefore, how to solve this difference on the basis of data security and achieve accurate cross-platform user alignment has become a key issue that needs to be solved urgently. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a dynamic social network alignment method based on vertical federation and canonical correlation analysis, comprising the following steps:

[0008] S1. Use the dynamic spatiotemporal graph autoencoder memory model and attention mechanism to simulate the user spatiotemporal relationship and obtain the user relationship matrix.

[0009] The dynamic spatiotemporal graph autoencoding memory model is composed of a GAE network and an LSTM unit;

[0010] S2, construct a user attribute matrix and merge it with the user relationship matrix to obtain a user matrix;

[0011] S3. Through the training model based on federated learning, the user feature matrix of platforms X and Y is input into the model for training to obtain the prediction results of cross-domain user alignment.

[0012] Beneficial effects of the present invention:

[0013] The present invention can effectively simulate the changes in user relationships over time through a dynamic spatiotemporal graph self-encoding memory model; the present invention combines network structure enhancement and privacy protection technology to reflect user characteristics from multiple angles, thereby improving the comprehensiveness and accuracy of user characteristics; the present invention is carried out under a vertical federated learning framework, combined with typical correlation analysis technology, effectively avoiding the risk of data privacy leakage across social platforms, achieving the capture of the correlation of user representations across platforms, and more accurately achieving user alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the process of the present invention;

[0015] Figure 2 This is a schematic diagram of the dynamic spatiotemporal graph self-encoding memory model process;

[0016] Figure 3 Embedding diagram for multi-level attributes;

[0017] Figure 4 Schematic diagram of the training model process for federated learning. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] A dynamic social network alignment method based on vertical federation and canonical correlation analysis, the process of the present invention is as follows Figure 1 As shown, the specific steps include:

[0020] S1. Use the dynamic spatiotemporal graph autoencoder memory model and attention mechanism to simulate the user spatiotemporal relationship and obtain the user relationship matrix.

[0021] The dynamic spatiotemporal graph autoencoding memory model is composed of LSTM units of the GAE network;

[0022] S2, construct a user attribute matrix and merge it with the user relationship matrix to obtain a user matrix;

[0023] S3. Through the training model based on federated learning, the user feature matrix of platforms X and Y is input into the model for training to obtain the prediction results of cross-domain user alignment.

[0024] The user relationship matrix is ​​obtained by simulating the user's spatiotemporal relationship through the dynamic spatiotemporal graph self-encoding memory model and the attention mechanism, as shown in Figure 2 As shown, specifically including:

[0025] S11. Obtaining the binary set G of user relationship-time series features t , specifically expressed as,

[0026] G t ={(V,E1),(V,E2),...,(V,E T )}={(V,E t )|t∈[0,T]}

[0027] Among them, the user behavior-time series feature set G t ={(V,E t )|t∈[0,T]} is defined as the set consisting of V and E. Among them:

[0028] V={ν1,ν2,···,ν n}

[0029] E={e ij =(v i ,v j )|v i ,v j ∈V}

[0030] Where V represents the user node set, E represents the set of n users and the set of undirected edges between users within time t, and each edge e ij ∈E represents user v i and v j The binary state of the links between words;

[0031] S12. Calculate the input static node embedding Z of the encoding part in the GAE network according to the dynamic user relationship set, which is specifically expressed as:

[0032]

[0033] Where A represents the adjacency matrix, W0 and W1 are the parameters to be learned, and D represents the degree matrix of the node;

[0034] S13, the adjacency matrix reconstructed by computing the encoded output according to the static node embedding Z Specifically expressed as:

[0035]

[0036] To get a good node embedding Z, the reconstructed adjacency matrix should be As similar as possible to the original adjacency matrix A. Therefore, GAE uses cross entropy as the loss function during training, which is specifically expressed as:

[0037]

[0038] Where y represents the value of an element in the adjacency matrix A (0 or 1), Represents the reconstructed adjacency matrix The value of the corresponding element in (between 0 and 1);

[0039] S14. For the user relationship-time series set G t , after passing through the GAE network, we get the static embedding set {Z t ,t∈[0,T]}, at time step t=0, the state S0 of the LSTM unit is initialized; from t=1 to t=T, the LSTM unit processes the embedding set Z of each time step in turn t , and updates its internal state S t :

[0040] S t =f(S t-1 ,Z t )

[0041] Among them, f is the function of the LSTM unit, which is responsible for the state S of the previous time step t-1and the embedding Z of the current time step t Update the current state S t ;

[0042] S141, the output f of the forget gate at the current time step t , specifically expressed as:

[0043] f t =sigmoid(W f ·[h t-1 ,Z t ]+b f )

[0044] Among them, W f represents the weight matrix of the forget gate, h t-1 represents the hidden state of the previous time step, b f Represents the bias term of the forget gate. The forget gate is used to determine which network information will be discarded from the cell state;

[0045] S142, the output i of the input gate of the current time step t and candidate cell states Specifically expressed as:

[0046] i t =sigmoid(W t ·[h t-1 ,Z t ]+b t )

[0047]

[0048] Among them, W i and W c represents the weight matrix of the input gate and candidate cell state, b i and b c Represents the bias term of the input gate and candidate cell state. The input gate determines which new network information in the current time step will be stored in the cell state;

[0049] S143, cell state C at the current time step t :

[0050]

[0051] Among them, C t-1 Represents the cell state at the previous time step. Use the output of the forget gate to discard some parts of the old cell state, and add the product of the output of the input gate and the candidate cell state to update the cell state;

[0052] S144, the output o of the output gate of the previous time step t , specifically expressed as:

[0053] o t =sigmoid(W O ·[h t-1 ,Z t ]+b O )

[0054] Among them, W o represents the weight matrix of the output gate, b o Represents the bias term of the output gate. The output gate determines which part of the network information of the next hidden state will be output;

[0055] S145, LSTM output h of the current time step t , specifically expressed as:

[0056] h t =o t *tanh(C t )

[0057] Filter through a sigmoid function, then pass the cell state through tanh and multiply it with the output of the output gate to get the required next hidden state h t ;

[0058] S146, according to the LSTM output h of the tth time step t Calculate the attention weight a at the tth time step t , specifically expressed as:

[0059] a t = sfotmax(tanh(W a ·h t +b a ))

[0060] Where W a represents the learnable weight matrix, b a represents the bias term;

[0061] S147, according to the attention weight a of the tth time step t The LSTM output h at the current time step t Calculate the user relationship embedding Q at the tth time step t , specifically expressed as:

[0062]

[0063] Construct a user attribute matrix and merge it with the user relationship matrix, such as Figure 3 As shown, the specific steps include:

[0064] S21. Construct user attribute set A:

[0065] A={A c ,A w ,A t}

[0066] The present invention divides the user's attribute information including user name, gender, geographic location, educational background and historical generated content into three levels: character level A c , word level A w and subject level A t ;

[0067] S22, character-level feature matrix:

[0068] S221. For character-level attribute sets Use the Bag of Words (BoW) model to vectorize attributes. User v i The corresponding character-level attributes are Contains characters, strings, and numbers. Construct vocabulary w = w1, w2, ..., w k ,…,w m , where k∈{1,2,…,m};

[0069] S222, the vocabulary w vector is represented as In the formula is the corresponding w in w k The number of attributes; finally, the attribute set A c Vectorized to

[0070]

[0071] Use an autoencoder to weight the count matrix X c To reduce the dimension, first, the autoencoder transforms the input vector through a deterministic function Mapping to a latent representation as follows:

[0072]

[0073] Where W is the weight matrix and b is the bias vector;

[0074] S223. Obtained potential representation Mapped to a reconstructed vector through the inverse mapping function g(·) as follows:

[0075]

[0076] Where W * is the weight matrix, b *is the bias vector. In order to minimize the average reconstruction loss, this section optimizes the parameters.

[0077]

[0078] After optimization, the character-level feature matrix is ​​obtained as follows:

[0079] P c =WX c +b

[0080] Where W and b are the weight matrix and the bias of the autoencoder respectively;

[0081] S23, word-level feature matrix:

[0082] For word-level attribute sets Use word2vec to capture the user's word-level attribute features. i The corresponding word-level attributes are Includes gender, geographic location, and educational background.

[0083] S231, attribute Split into m word sequences w i =w i1 ,w i2 ,…,w ik ,…,w im , where w ik Representation attributes The vocabulary w of the kth word in ;

[0084] S232, User v i The word-level attribute vector of It can be expressed as w i The word vectors of all words in are summed up, as shown in formula (23).

[0085]

[0086] S233, get word-level feature matrix

[0087] S24, topic-level feature matrix:

[0088] For the subject-level property set The topic features of user attributes are extracted using latent Dirichlet distribution. i Theme-level attributes of It consists of paragraphs or articles. The i-th topic-level attribute is the topic-level document w i , that is, the subject-level attribute set A t Corresponding document set w=w1,w2,…,wi ,…,w n .

[0089] The way LDA generates topics from a document is as follows: generate a document w according to the prior probability i ; Generate document w by sampling from Dirichlet distribution α i The topic distribution Generate topic z by sampling from Dirichlet distribution β ij Word distribution Multinomial distribution from words to topics Sample the generated document w i The topic z of the jth word in ij ; From the multinomial distribution of words Sampling generated words w ij .

[0090] Therefore, the joint distribution of all variables in the LDA model looks like this:

[0091]

[0092] Use Gibbs sampling to perform sampling and update after each sampling. and θ m,k The value is calculated and the cycle is repeated until convergence. After the overall sampling is completed, the statistics θ m,k Get the topic distribution θ of each document in the document set m . As shown below:

[0093]

[0094] Therefore, we get user v i The topic-level feature vector The topic-level attribute matrix for all users is

[0095] S25. User attribute matrix:

[0096] The attribute matrices of the above three levels are combined, and the fusion matrix P is processed using the perturbation matrix γ to obtain the user attribute matrix P′. Then the user relationship matrix Q is combined to obtain the user matrix U. As shown below:

[0097] U=Q+P′=Q+γP=Q+γ(P c +P w +P t ).

[0098] By establishing a training model based on federated learning, the user feature matrix of platforms X and Y is input into the model for training, and the prediction results of cross-domain user alignment are obtained, such as Figure 4As shown, the specific steps include:

[0099] S31. Model training:

[0100] The user matrix U calculated above X and U Y As the input of model training, according to the loss function L CCA , platforms X and Y are used as source and target domains for federated training. The formula is as follows:

[0101]

[0102] Among them, W X represents the projection matrix of X, W Y represents the projection matrix of Y, C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YT represents the covariance matrix of Y, λ represents the regularization parameter, Ω(Θ) represents the regularization term, and T represents the matrix transpose.

[0103] The goal of model training is two network parameters W X and W Y , so that the two networks have the maximum correlation after CCA calculation. The training steps are as follows:

[0104] S311, initialize model parameters. Platform X initializes the projection matrix Platform Y initialization projection matrix

[0105] S312, platform X calculates the typical variable matrix G X =U X W X , G X After encryption, it is sent to platform Y.

[0106] S313, platform Y calculates the typical variable matrix G Y =U Y W Y . Using the received encrypted G X and local G Y Securely jointly compute the encrypted cross-covariance matrix [[C XY ]] Calculate the encrypted residual [[d]] = [[G X -G Y ]]. Where d is the residual, which represents the difference between the two typical variable matrices, and [[·]] means encrypting the data.

[0107] S313, solve the encrypted gradient. Platform X and Y are based on the loss function L CCA, calculate the following encrypted gradient: The specific formula is:

[0108]

[0109] in, represents the partial derivative, L CCA Denotes the loss function L CCA , C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YY represents the covariance matrix of Y and T represents the matrix transpose.

[0110] S315: Upload the solved encrypted gradient to a trusted third-party server for decryption. The decrypted gradient is returned to platforms X and Y respectively. Update parameters using gradient descent: Platforms X and Y update their respective projection matrices W according to the decrypted gradient. X and W Y :

[0111]

[0112] Among them, η represents the learning rate, t represents the current training round, and t+1 represents the next training round.

[0113] Repeat S312 to S315 until the algorithm converges.

[0114] S33, Model prediction:

[0115] The third-party server can calculate the maximum correlation according to the following formula, and then verify it according to the cosine similarity to predict the potential anchor user pairs.

[0116]

[0117] Among them, ρ represents the correlation coefficient, maxcorr() represents the maximum correlation, U X represents the user embedding matrix of platform X, U Y represents the user embedding matrix of platform Y, max represents maximization, W X represents the projection matrix of X, W Y represents the projection matrix of Y, C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YY represents the covariance matrix of Y and T represents the matrix transpose.

[0118] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic social network alignment method based on vertical federation and canonical correlation analysis, characterized in that: include: S1. Use the dynamic spatiotemporal graph autoencoder memory model and attention mechanism to simulate the user spatiotemporal relationship and obtain the user relationship matrix. The dynamic spatiotemporal graph autoencoding memory model is composed of a GAE network and an LSTM unit; S2, construct a user attribute matrix and merge it with the user relationship matrix to obtain a user matrix; S3. Through the training model based on federated learning, the user feature matrix of platforms X and Y is input into the model for training to obtain the prediction results of cross-domain user alignment.

2. A dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 1, characterized in that: The dynamic spatiotemporal graph autoencoding memory model and attention mechanism are used to simulate the spatiotemporal relationship of users, including: S11. Obtain user relationship-time series set G t ={(V,E t )|t∈[0,T]}; where V={v1,v2,···,v n } represents a node set, E t represents the set of undirected edges between users, and T represents the time series; S12. Calculate the input static node embedding of the encoding part in the GAE network based on the user relationship-time series set in, represents the symmetric normalized adjacency matrix, ReLU() represents the activation function, W0 represents the initial weight matrix, and W1 represents the second layer weight matrix; S13, calculate the adjacency matrix reconstructed from the decoded output according to the static node embedding Z Among them, sigmoid() means converting the score calculated by the decoder into a probability value, and T means matrix transpose; S14. For the user relationship-time series set G t , after passing through the GAE network, we get the static embedding set {Z t ,t∈[0,T]}, at time step t=0, the state S0 of the LSTM unit is initialized; from t=1 to t=T, the LSTM unit processes the embedding set Z of each time step in turn t , and updates its internal state S t : S t =f(S t-1 ,Z t ) Among them, f is the function of the LSTM unit, which is responsible for the state S of the previous time step t-1 and the embedding Z of the current time step t Update the current state S t ; S15, the attention mechanism is introduced to better capture the importance of different time steps, and the LSTM outputs of all time steps are combined to obtain the user relationship matrix Q: Among them, a t represents the attention weight at the tth time step, h t Represents the LSTM output at the current time step.

3. A dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 2, characterized in that: The main process of step S14 includes: S141, LSTM first step through the forget gate to determine which network information will be discarded from the cell state: f t =sigmoid(W f ·[h t-1 ,Z t ]+b f ) Among them, h t-1 represents the hidden state of the previous time step, Z t represents the input of the current time step, f t Represents the output of the forget gate at the current time step, W f represents the weight matrix of the forget gate, b f represents the bias term of the forget gate, [·] represents the vector concatenation operation; S142, LSTM uses the input gate to determine which new network information in the current time step will be stored in the cell state, including: using the sigmoid function to determine which values ​​will be updated and creating a candidate cell state By applying the tanh function: i t =sigmoid(W t ·[h t-1 ,Z t ]+b t ) Among them, i t represents the output of the input gate at the current time step, represents the candidate cell state at the current time step, W t and W c represents the weight matrix of the input gate and candidate cell state, b t and b c A bias term representing the input gate and candidate cell state; S143, LSTM uses the output of the forget gate to discard some parts of the old cell state, and adds the product of the output of the input gate and the candidate cell state to update the cell state: Among them, C t represents the cell state at the current time step, C t-1 represents the cell state at the previous time step; S144, LSTM determines which part of the network information in the next hidden state will be output through the output gate. This output will be based on the cell state o t , but it will be filtered through a sigmoid function, and then the cell state o t Through the tanh function and multiplying it with the output of the output gate, we get the required next hidden state h t ; o t =sigmoid(W O ·[h t-1 ,Z t ]+b O ) h t =o t *tanh(C t ) Among them, W O represents the weight matrix of the output gate, b O represents the bias of the output gate, h t Represents the LSTM output of the current time step; S145, update its internal state S t =[C t ,h t ].

4. The dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 1, characterized in that: Construct a user attribute matrix and integrate it with the user relationship matrix, including: S21. Construct user attribute set A = {A c ,A w ,A t The user attributes include username, gender, geographic location, educational background and user-generated content on the two social platforms, and are divided into three levels: character level A c , word level A w and subject level A t ; S22. Construct character-level feature matrix: For character-level attribute sets Use the bag-of-words vector model to vectorize attributes and obtain a count-weighted matrix in, Represents user v i The corresponding character-level attributes, T represents the matrix transpose, Represents character-level attributes Bag-of-words vectors; Use an autoencoder to weight the count matrix X c Perform dimensionality reduction to obtain the character-level feature matrix in, Represents user v i The character-level attribute vector of ; S23. Construct word-level feature matrix: For word-level attribute sets Use word2vec to capture the user's word-level attribute features and obtain the word-level feature matrix in, Represents user v i The corresponding word-level attributes, Represents user v i The word-level attribute vector of ; S24. Construct the topic-level feature matrix: For the subject-level property set The topic features of user attributes are extracted using latent Dirichlet distribution, and the topic-level attribute matrix of all users is obtained as in, Represents user v i The subject-level attributes of Represents user v i The topic-level feature vector of S25. Build a user matrix: The character-level feature matrix, word-level feature matrix, and topic-level feature matrix are integrated to obtain the integrated matrix P, and the perturbation matrix γ is used to process the integrated matrix P to obtain the user attribute matrix P′, and then the user relationship matrix Q is integrated to obtain the user matrix U=Q+P′=Q+γP=Q+γ(P c +P w +P t ), where P c represents the character-level feature matrix, P w represents the word-level feature matrix, P t Represents the topic-level feature matrix.

5. The dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 1, characterized in that: The user feature matrix of platforms X and Y is input into the model for training to obtain the prediction results of cross-domain user alignment, including: S31, the user matrix U X and U Y As the input of model training, according to the loss function L CCA , use platforms X and Y as source and target domains for federated training until the algorithm converges; S32. The third-party server calculates the maximum network correlation based on canonical correlation analysis CCA, and then calculates the correlation coefficient between users based on cosine similarity to predict potential anchor user pairs, i.e., aligned users.

6. A dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 5, characterized in that: The loss function L CCA ,include: Among them, W X represents the projection matrix of X, W Y represents the projection matrix of Y, C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YY represents the covariance matrix of Y, λ represents the regularization parameter, Ω(Θ) represents the regularization term, and T represents the matrix transpose.

7. The dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 5, characterized in that: Through the training model based on federated learning, the user matrix U X and U Y As the input of model training, according to the loss function L CCA , using platforms X and Y as source and target domains for federated training, including: S311, Initialize model parameters: Platform X initializes projection matrix Platform Y initialization projection matrix in, Represents the projection matrix W X The i-th projection vector of Represents the projection matrix W Y The i-th projection vector of ; S312, platform X calculates the typical variable matrix G X =U X W X , G X After encryption, it is sent to platform Y; among them, U X represents the user matrix of platform X; S313, platform Y calculates the typical variable matrix G Y =U Y W Y , using the received encrypted G X and local G Y Securely jointly compute the encrypted cross-covariance matrix [[C XY ]]; Calculate the encrypted residual [[d]] = [[G X -G Y ]]; where d represents the residual used to represent the difference between two typical variable matrices, [[]] represents encryption of data, and U Y represents the user matrix of platform Y; S313, solve the encrypted gradient: platform X and Y according to the loss function L CCA , calculate the encrypted gradient: in, represents the partial derivative, L CCA Denotes the loss function L CCA , C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YY represents the covariance matrix of Y, and T represents the matrix transpose; S315: Upload the solved encrypted gradient to a trusted third-party server for decryption. The decrypted gradient is returned to platforms X and Y respectively. Platforms X and Y use gradient descent to update their respective projection matrices W according to the decrypted gradient. X and W Y : Among them, η represents the learning rate, t represents the current training round, and t+1 represents the next training round.

8. The dynamic social network alignment method based on vertical federation and canonical correlation analysis according to claim 5, characterized in that: The third-party server calculates the maximum network correlation based on canonical correlation analysis (CCA), and then calculates the correlation coefficient between users based on cosine similarity to predict potential anchor user pairs, including: Among them, ρ represents the correlation coefficient, maxcorr() represents the maximum correlation, U X represents the user embedding matrix of platform X, U Y represents the user embedding matrix of platform Y, max represents maximization, W X represents the projection matrix of X, W Y represents the projection matrix of Y, C XY represents the cross covariance matrix between X and Y, C XX represents the covariance matrix of X, C YY represents the covariance matrix of Y and T represents the matrix transpose.

Images