A social network alignment method

Abstract

The present invention belongs to the field of Internet technology, and specifically relates to a social network alignment method. The method collects network user data, forms a first social network graph and a second social network graph according to user relationships, selects a portion of users to form user pairs to form a third social network graph, calculates an adjacency matrix, inputs the adjacency matrix into an encoder to obtain a first user node representation matrix; inputs the first user node representation matrix and the adjacency matrix into a denoising model to obtain a covering matrix; obtains a covering adjacency matrix based on the covering matrix and the adjacency matrix, inputs the covering adjacency matrix into an encoder to obtain a second user node representation matrix; obtains a difference between a first similarity and a second similarity based on the second user node representation matrix, and updates the learning parameters of the denoising model; outputs the updated covering adjacency matrix, and aligns users in the network graph based on the covering adjacency matrix. The present invention can remove noise generated by the existence of identical users in a social network, thereby improving the effect of social network alignment.

Description

Technical Field

[0001] The present invention belongs to the field of Internet technology, and in particular relates to a social network alignment method. Background Art

[0002] In the internet age, users have a wide variety of data across various online service platforms. This data holds great value for mining, but it also presents numerous challenges. For example, data from different platforms is not interoperable, making data mining difficult. Therefore, the problem of identifying common users across multiple social networking platforms is defined as network alignment. Network alignment can connect users across different platforms, thereby mining valuable data.

[0003] Based on the structured data characteristics of the network, users can be regarded as nodes and the relationships between users as edges. Therefore, most social network alignment methods are based on graph convolutional neural networks and graph structures to learn and find the same nodes in two different networks. However, these alignment methods are explicit or implicit. During the alignment process, it is assumed that the two social networks used are non-overlapping. However, in reality, the original different social networks inevitably have the same users. Therefore, the existing technology will produce low alignment effects due to the noise generated by these same users when aligning social networks. Summary of the Invention

[0004] In order to solve the problem of low network alignment effect caused by the noise inevitably generated by the presence of the same user in different social networks during social network alignment, the present invention proposes a social network alignment method, which specifically includes the following steps:

[0005] S1: Collect social network user data from different social platforms and form a first social network graph and a second social network graph according to user relationships;

[0006] S2: Select a portion of known identical users from the first social network graph and the second social network graph to form user pairs, and merge the user pairs to form a third social network graph;

[0007] S3: Calculate the adjacency matrix of the third social network graph, and input the adjacency matrix of the third social network graph into the encoder to obtain the first user node representation matrix of all users;

[0008] S4: Inputting the first user node representation matrix and the adjacency matrix into the denoising model to calculate a covering matrix of the third social network graph;

[0009] S5: Calculate the covered adjacency matrix of all users in the third social network graph according to the product of the largest K values ​​in the covered matrix and the adjacency matrix of the third social network graph;

[0010] S6: Inputting the masked adjacency matrix into the encoder to obtain a second user node representation matrix for all users;

[0011] S7: Determine, based on the second user node representation matrix, a first similarity between another portion of the known identical users in the third social network graph, and a second similarity between the second user node representation matrix and a set of user pairs consisting of a portion of the users;

[0012] S8: using the difference between the first similarity and the second similarity to update the learning parameters of the denoising model;

[0013] S9: Repeat S4-S8. When the difference is less than the set threshold, output the updated masked adjacency matrix, and merge and align the unknown users in the third social network graph according to the masked adjacency matrix.

[0014] The present invention has at least the following beneficial effects:

[0015] The present invention designs a social network alignment method. This method obtains a portion of known identical users from a first social network graph and a second social network graph, merges a portion of the known identical users to form a third social network graph, calculates an adjacency matrix of the third social network graph, obtains a first node representation matrix through the adjacency matrix, inputs the first node representation matrix and the adjacency matrix into a denoising model to obtain a covering matrix, masks the adjacency matrix according to the covering matrix to obtain a covered adjacency matrix, merges and aligns (denoises) unknown users in the third social network graph according to the covered adjacency matrix to obtain a denoised third social network graph. Simultaneously, the present invention calculates a second node representation matrix based on the covered adjacency matrix, and updates the learning parameters of the denoising model by calculating the difference between the first similarity between another portion of the known identical users in the third social network graph in the second node representation matrix and the second similarity between the second user node representation matrix and a set of user pairs consisting of the portion of users, thereby improving the denoising effect of the model and accelerating the denoising time. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method of the present invention;

[0017] Figure 2 Schematic diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0019] The present invention proposes a social network alignment method, such as Figure 1 and Figure 2 As shown, the specific steps include:

[0020] Social network user data of different social platforms are collected, and a first social network graph and a second social network graph are formed according to user relationships, wherein the user relationship is one of following, friendship, communication, etc. between users.

[0021] A portion of known identical users in the first social network graph and the second social network graph are selected to form user pairs, and the user pairs are merged to form a third social network graph, wherein the known identical users are limited users obtained based on social network user data when aligning the two social network graphs. Identical users can be represented by having the same identity information, the same ID, the same age, or the same birthday.

[0022] Furthermore, the formation process of the third social network graph is as follows:

[0023]

[0024] Where map represents the mapping function, node represents any user in the first social network graph and the second social network graph, otherwise represents the user that does not belong to A in the first social network graph and the second social network graph, G merage Represents the third social network graph, and A represents a user pair.

[0025] An adjacency matrix of the third social network graph is calculated, and the adjacency matrix of the third social network graph is input into an encoder to obtain a first user node representation matrix of all users, wherein the encoder is a network alignment encoder.

[0026] Preferably, calculating the adjacency matrix of the third social network graph specifically includes:

[0027]

[0028] Among them, N e (j) represents the set of users who have relationships with user j in the third social network graph, A merage represents the adjacency matrix of the third social network graph, and i and j represent any two different users in the third social network graph.

[0029] The first user node representation matrix and the adjacency matrix are input into the denoising model to calculate the covering matrix of the third social network graph, wherein the graph convolutional neural network model and the regularization model.

[0030] Preferably, calculating the covering matrix of the third social network graph specifically includes:

[0031] The first user node representation matrix is ​​used as the original node features of the graph convolutional neural network model, and the adjacency matrix of the third social network graph is used as the input of the graph convolutional neural network model to calculate the node features of the third social network graph:

[0032]

[0033] Among them, H (l) represents the node features of the lth layer in the graph convolutional neural network, σ represents the nonlinear activation function, yes The degree matrix of Indicates A merage Wave, A merage Represents the adjacency matrix of the third social network graph, I is the identity matrix, when l=n-1, H (n) Represents the node features of the third social network graph. When l = 0, H (0) =X merage is the original feature of the node, W represents the learning parameter in the graph convolutional neural network, and n represents the number of layers of the graph convolutional neural network. W represents the learning parameter in the graph convolutional neural network, and the initial value is randomly set by those skilled in the art. The value of W in each layer of the graph convolutional neural network is similar to a constant.

[0034] The node features of the third social network graph are regularized using a regularization model, and the regularized node features are decoded to obtain the covering matrix of the third social network graph:

[0035] Z=Normalize(H (n) )

[0036] Mask=σ(Z@Z T )

[0037] Normalize represents the regularization function, Z represents the node features after regularization of the node features of the third social network graph, Z T represents the transpose of Z, @ represents matrix multiplication, Mask represents the masking matrix of the third social network graph, and σ represents a nonlinear activation function.

[0038] According to the product of the largest K values ​​in the mask matrix and the adjacency matrix of the third social network graph, the mask adjacency matrix of all users in the third social network graph is calculated. For example, the largest K values ​​in the mask matrix Mask are selected to form topK(Mask), and topK(Mask)*A is used to calculate the mask adjacency matrix of all users in the third social network graph. merage Calculate the third user graph network structure G merage The masked adjacency matrix A mask .

[0039] The masked adjacency matrix is ​​input into the encoder to obtain a second user node representation matrix for all users, wherein the encoder is a network alignment encoder.

[0040] Based on the second user node representation matrix, a first similarity between another part of the known identical users in the third social network graph and a second similarity between the second user node representation matrix and a set of user pairs consisting of a part of the users are derived. Preferably, node representations of another part of the known identical users are selected from the second user node representation matrix, the first similarity is derived by calculating the similarity of the node representations between the another part of the known identical users, node representations of user pairs are selected from the second user node representation matrix, and a set of user pair node representations is formed, and the second similarity is derived by calculating the similarity between the set of user pair node representations and the second user node representation matrix, including:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] Wherein, node is any user in the first social network graph and the second social network graph, Y1 and Y2 represent any two users in another part of the known same users, sim represents the cosine similarity function, The node representation set of another part of the known same users, Indicates that user Y1 represents the matrix X in the second node mask The nodes in represent, Indicates that user Y2 represents the matrix X in the second node mask The node in sim1 represents and The first similarity, Represents all user pairs in the second user node representation matrix Xmask The node in sim2 represents X mask and The second similarity, It represents the difference between the first similarity and the second similarity, A represents the user pair, and B represents another part of the known identical users.

[0047] The difference between the first similarity and the second similarity is used to update the learning parameters of the denoising model.

[0048] Repeat S4-S8, and when the difference is less than the set threshold, output the updated masked adjacency matrix, and merge and align the unknown users in the third social network graph according to the masked adjacency matrix.

Claims

1. A social network alignment method, characterized in that: The specific steps include: S1: Collect social network user data from different social platforms and form a first social network graph and a second social network graph according to user relationships; S2: Select a portion of known identical users from the first social network graph and the second social network graph to form user pairs, and merge the user pairs to form a third social network graph; S3: Calculate the adjacency matrix of the third social network graph, and input the adjacency matrix of the third social network graph into the encoder to obtain the first user node representation matrix of all users; S4: Inputting the first user node representation matrix and the adjacency matrix into the denoising model to calculate a covering matrix of the third social network graph; S5: Calculate the covered adjacency matrix of all users in the third social network graph according to the product of the largest K values ​​in the covered matrix and the adjacency matrix of the third social network graph; S6: Inputting the masked adjacency matrix into the encoder to obtain a second user node representation matrix for all users; S7: Determine, based on the second user node representation matrix, a first similarity between another portion of the known identical users in the third social network graph, and a second similarity between the second user node representation matrix and a set of user pairs consisting of a portion of the users; S8: using the difference between the first similarity and the second similarity to update the learning parameters of the denoising model; S9: Repeat S4-S8. When the difference is less than the set threshold, output the updated masked adjacency matrix, and merge and align the unknown users in the third social network graph according to the masked adjacency matrix.

2. A social network alignment method according to claim 1, characterized in that: The S2 includes: Where map represents the mapping function, node represents any user in the first social network graph and the second social network graph, otherwise represents the user that does not belong to A in the first social network graph and the second social network graph, G merage Represents the third social network graph, and A represents the user in the user pair.

3. A social network alignment method according to claim 1, characterized in that: Calculating the adjacency matrix of the third social network graph includes: Among them, N e (j) represents the set of users who have relationships with user j in the third social network graph, A merage represents the adjacency matrix of the third social network graph, and i and j represent any two different users in the third social network graph.

4. A social network alignment method according to claim 1, characterized in that: The denoising model includes: a graph convolutional neural network model and a regularization model.

5. A social network alignment method according to claim 4, characterized in that: The calculating and obtaining the covering matrix of the third social network graph includes: S41: Using the first user node representation matrix as the original node features of the graph convolutional neural network, and using the adjacency matrix of the third social network graph as the input of the graph convolutional neural network, to calculate the node features of the third social network graph; S42: Regularizing the node features of the third social network graph using a regularization model, and decoding the regularized node features to obtain a covering matrix of the third social network graph.

6. A social network alignment method according to claim 5, characterized in that: The S41 includes: Among them, H (l) represents the node features of the lth layer in the graph convolutional neural network, σ represents the nonlinear activation function, yes The degree matrix of Indicates A merage Wave, A merage Represents the adjacency matrix of the third social network graph, I is the identity matrix, when l=n-1, H (n) Represents the node features of the third social network graph. When l = 0, H (0) =X merage is the original feature of the node, W represents the learning parameter in the graph convolutional neural network, and n represents the number of layers of the graph convolutional neural network.

7. A social network alignment method according to claim 6, characterized in that: The S42 includes: Z=Normalize(H (n) ) Mask=σ(Z@Z T ) Normalize represents the regularization function, Z represents the node features after regularization of the node features of the third social network graph, Z T represents the transpose of Z, @ represents matrix multiplication, Mask represents the masking matrix of the third social network graph, and σ represents a nonlinear activation function.

8. The social network alignment method according to claim 1, characterized in that: The S7 includes: S71: Selecting node representations of another part of the known same users from the second user node representation matrix, and calculating the similarity of the node representations between the other part of the known same users to obtain a first similarity; S72: Selecting node representations of user pairs from the second user node representation matrix and forming a user pair node representation set, and calculating the similarity between the user pair node representation set and the second user node representation matrix to obtain a second similarity.

Images