Heterogeneous network user anchor link prediction method and electronic device based on type perception

By using self-attention graph neural network for type perception in heterogeneous networks, the similarity of type-aware vectors and fusion vectors of nodes is solved, and a more accurate heterogeneous network anchor link is achieved. predict.

CN114625978BActive Publication Date: 2024-11-08NAT COMP NETWORK & INFORMATION SECURITY MANAGEMENT CENT +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011454130.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-12-10
Publication Date
2024-11-08
Estimated Expiration
2040-12-10

AI Technical Summary

Technical Problem

The prior art ignores local type information in heterogeneous networks, resulting in poor prediction of heterogeneous network anchor links, especially when there are inconsistent node types in different networks.

Method used

Using a type perception method based on a self-attention graph neural network, the multiple embedded representations of hierarchical learning nodes are used to predict heterogeneous network anchor links using the similarity of type perception vectors, and the similarity is calculated based on the Euclidean distance or cosine distance, and the similarity is judged. Detect whether the network and the target network are anchor links.

Benefits of technology

Effectively aligning the account information of the same natural person in different networks improves the comprehensiveness of user portraits, solves the problems of modeling of multiple types of information and type neighborhood alignment noise in heterogeneous networks, and improves the accuracy of anchor link prediction .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114625978B_ABST
    Figure CN114625978B_ABST
Patent Text Reader

Abstract

The present invention provides a heterogeneous network user anchor link prediction method and electronic device based on type perception, including collecting network information of a network to be detected and a target network, constructing a heterogeneous network to be detected and a target heterogeneous network; obtaining the initial feature vector of each dimension of each node, and obtaining the initial feature vector representation of each node; inputting the initial feature vector representation into a self-attention graph neural network for learning, and calculating the type perception vector and type fusion vector of each node; judging whether the network to be detected and the target network are anchor links by the similarity of the type perception vector and type fusion vector belonging to the same type between each node pair in the heterogeneous network to be detected and the target heterogeneous network. The present invention adopts an attention mechanism to capture the influence between nodes and type information, and uses a graph attention network to learn type perception vectors and type fusion vectors, so as to solve the problem of modeling multiple types of information in heterogeneous networks and achieve better anchor link prediction effects.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computers, and in particular relates to a type-aware based heterogeneous network user anchor link prediction method and an electronic device. Background Art

[0002] With the development of the Internet, the role of network platforms has gradually been subdivided. For example, users share their feelings with friends on Twitter and share videos on YouTube. Therefore, in order to meet the needs, more and more users have registered accounts on multiple platforms. However, each account only reflects the characteristics of the user in a certain aspect. In order to more comprehensively characterize the user, it is necessary to integrate the information of different accounts of the same user, that is, to identify the accounts of the same user on different network platforms, which is called anchor link prediction.

[0003] With the rise of network embedding, anchor link prediction based on embedding has become a trend. The core of existing anchor link prediction methods is: embedding and alignment. The goal of the embedding part is to obtain the vector representation of nodes (accounts) in each network based on the network embedding method. The alignment part obtains potential anchor links by estimating the similarity between the embedding representation vectors of nodes in different networks. However, in heterogeneous networks, there are few anchor link prediction methods based on network embedding. Most methods fuse the information of various types of nodes related to user nodes from a global perspective to obtain the embedding vector of the user node (called the fusion vector). Then, based on the similarity between the embedding vectors of the node pairs, it is judged whether they are anchor links.

[0004] But if Figure 1 As shown, the prior art ignores the impact of local type information (information of each type of node associated with the user node) on user node alignment when fusing vectors. This impact is more obvious when there are inconsistent node types in different heterogeneous networks. Summary of the invention

[0005] The purpose of the present invention is to provide a type-aware heterogeneous network user anchor link prediction method, which learns multiple embedded representations of nodes through a hierarchical graph neural network, uses the similarity of type-aware vectors to predict anchor links in heterogeneous networks, and better aligns the account information of the same natural person in different networks, thereby more comprehensively portraying user portraits.

[0006] The technical solution of the present invention is as follows:

[0007] A method for predicting anchor links of heterogeneous network users based on type perception, the steps of which include:

[0008] 1) Collecting network information of the network to be detected and the target network, respectively constructing the heterogeneous network to be detected and the target heterogeneous network, wherein each node in the heterogeneous network to be detected and the target heterogeneous network contains partial attribute information of the corresponding network to be detected and the target network;

[0009] 2) Obtaining the initial feature vectors of each dimension of each node in the heterogeneous network to be detected and the target heterogeneous network, and obtaining the initial feature vector representation of each node;

[0010] 3) Input the initial feature vector representation into a self-attention graph neural network for learning, and calculate the type perception vector and type fusion vector of each node according to the weight of each type perception vector of the corresponding neighboring nodes of each node;

[0011] 4) By determining the similarity of the type perception vectors and type fusion vectors of the same type between each pair of nodes in the heterogeneous network to be detected and the target heterogeneous network, it is determined whether the network to be detected and the target network are anchor links.

[0012] Furthermore, the initial feature vector representation of each node is obtained through the following strategy:

[0013] 1) For nodes containing text information, methods for obtaining initial feature vector representation include: Word2vec method;

[0014] 2) For nodes containing image information, methods for obtaining initial feature vector representation include: resnet method;

[0015] 3) For nodes whose text information or image information is unclear, methods for obtaining initial feature vector representation include: random allocation method.

[0016] Furthermore, the type-aware vector is obtained by the following steps:

[0017] 1) Compute node v i Each corresponding neighboring node v j The type-aware vector weight of type r in the

[0018] 2) Normalize the type perception vector weight of type r to obtain node v i and v j The attention coefficient of type r;

[0019] 3) By calculating the neighboring node v j Weighted summation, get v of type r i Type-aware vectors

[0020] Furthermore, each type of perception vector is weightedly summed with the corresponding type of perception vector plus the attention weight to obtain a type fusion vector.

[0021] Furthermore, the similarity between the type-aware vector and the type-fusion vector is calculated by using the Euclidean distance or cosine distance between the node pairs. Furthermore, the following steps are used to determine whether the network to be detected and the target network are anchor links:

[0022] 1) The similarity of each type of perception vector and type fusion vector is weighted and summed with the corresponding set weight to obtain a similarity score;

[0023] 2) Based on the similarity score and a threshold, determine whether the network to be detected and the target network are anchor links.

[0024] Furthermore, when learning based on the self-attention graph neural network, the loss function is in, is the heterogeneous network node to be detected, Target heterogeneous network node, a represents the node type, represents a positive sample, Denotes negative samples, ω, λ r , ξ is a hyperparameter and ω+∑λ r =1, r, R are type-aware vector types, Divided into nodes Contains the type of the fused vector, Divided into nodes Contains type-aware vectors of type r, d(.,.) is the distance between two vectors, is the fusion loss, is the type of loss.

[0025] Furthermore, the type of loss includes: document loss or geographical location loss.

[0026] A storage medium stores a computer program, wherein the computer program is configured to execute the method described above when running.

[0027] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer to execute the method described above.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1) The present invention mainly uses the attention mechanism to capture the influence between nodes and type information, and can improve the time efficiency of the model through parallel computing;

[0030] 2) The model of the present invention is not a traditional machine learning method and needs to maintain a single network structure. The learned embedding representation is designed for the task of anchor link prediction, so it works better in the task of anchor link prediction;

[0031] 3) Using a multi-layer graph attention network to learn type-aware embedding vectors and type fusion vectors, it effectively solves the problem of how to model multiple types of information in heterogeneous networks;

[0032] 4) In the type-aware anchor linking process, the type-aware vector enables good alignment between neighbors of the same type, solving the noise problem caused by aligning neighbors of different types, thereby achieving better anchor link prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is an example diagram of anchor link prediction in a heterogeneous network using existing technology.

[0034] Figure 2 Schematic diagram of n-tuple vector representation learning of the present invention.

[0035] Figure 3 It is a schematic diagram of type-aware anchor link prediction of the present invention. DETAILED DESCRIPTION

[0036] In order to make the above features and advantages of the present invention more clearly understood, examples are given below and detailed description is given in conjunction with the accompanying drawings.

[0037] The type-aware anchor link prediction method for heterogeneous networks provided by the present invention is suitable for user alignment tasks in real scenarios. The main idea of ​​the method is to model the network platform in the real scenario as a heterogeneous network, and use the characteristics of the heterogeneous network containing multiple node types to learn a fusion vector and multiple type-aware vectors that fuse multiple types of information for each node in the heterogeneous network, which together constitute an n-tuple vector representation. Finally, by calculating the similarity of the fusion vector and the similarity of the type-aware vector, the purpose of network alignment is achieved.

[0038] The method of the present invention mainly includes three processes: heterogeneous network modeling, n-tuple representation learning and type-aware anchor link prediction. First, a heterogeneous network containing multiple types of nodes is constructed based on the information on the network. Before constructing a heterogeneous network, it is necessary to first determine what types of nodes are contained in the real network platform to be modeled, because different types of networks contain different types of information, so the types of nodes in the constructed heterogeneous network are also different, such as posting nodes and location nodes in social networks, paper nodes and conference nodes in citation networks, and the relationship between them. After determining the node types, the heterogeneous network corresponding to the real network platform can be modeled based on the nodes and the relationship between the nodes.

[0039] Please refer to Figure 2 The n-tuple vector representation learning diagram, the present invention uses set V to represent the node set in the network. Each node in the heterogeneous network carries some attribute features, such as post content, pictures, etc. In order to make this information available to the subsequent model, it is necessary to convert this information into a vector form that can be understood by the computer, that is, according to the node v i (v i ∈V) carries some attribute features to extract the initial feature vector of the node Specifically, for nodes containing text information, such as post content, geographic location description information, etc., Word2vec is used to encode the text information to obtain the text feature vector of the node. For nodes containing image information, resnet is used to extract image feature vectors For nodes with unclear text or image information, random allocation is used to obtain the initial feature vector.

[0040]

[0041] For the n-tuple vector representation of the node, a self-attention-based graph attention network (Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, Yoshua Bengio: Graph Attention Networks. ICLR (2018)) is constructed to integrate the user's domain neighborhood information of the same type and calculate the node v i Each node v in the neighborhood j The weight of the requested type awareness vector (self-attention)

[0042]

[0043] Where r(r∈R) represents the node neighborhood type, and the vector Represents the weight vector, matrix W r represents the weight matrix, .T represents the matrix transpose operation, and || represents the concatenation operation.

[0044] Then, normalize the weights to get node v i and v j The attention coefficient between:

[0045]

[0046] in Represents node v i The user's type perception vector can be obtained by weighted summing of the neighborhood nodes. t represents the subscript of the node:

[0047]

[0048] Where σ represents the activation function. Because there is more than one type of node in the network, there will be multiple type perception vectors for each node. The second layer of GAT aims to fuse the different type perception vectors of user nodes to obtain the type fusion embedding vector The present invention learns a weight for each node type in the network:

[0049]

[0050] The weight matrix W f and Project the initialization vector and type-aware vector to high-dimensional space respectively. i,r After normalization, we get:

[0051]

[0052] Where a represents the type of node. The perception vectors of different types in the network are weighted and summed by their corresponding type weights, and finally a type-fused embedding vector, referred to as the fusion vector, is obtained.

[0053]

[0054] Through the above steps, multiple type perception vectors and a type fusion vector are obtained for each node in the source network and the target network, respectively. and To express.

[0055] In the type-aware anchor link prediction stage, since the type information of the node affects the prediction result, the model of the present invention not only aligns the fusion vector, but also aligns each type-aware vector. Specifically, according to the n-tuple vector representation part of each node fusion vector learned, by calculating the similarity (Euclidean distance, cosine distance) of the fusion vectors between the node pairs in the two networks, through triplet-loss, the distance between the node pairs belonging to the anchor link is as small as possible, and the distance between the node pairs that do not belong to the anchor link is as large as possible. Similarly, in the stage of aligning each type-aware phase, the present invention calculates the similarity of the type-aware vectors of the same type between the node pairs in the two networks respectively, so as to align each type-aware vector of the node pairs in the two networks. The loss function of the model is as follows:

[0056]

[0057] Where d(·,·) represents the distance between two vectors, which is usually calculated using cosine distance or Euclidean distance. represents a positive sample, Denotes negative samples, ω, λ r , ξ is a hyperparameter, and ω+∑λ r =1. F =R S ∪R T , where R S and R T Represent the source network and target network node type sets respectively. In the above loss function L, the first term represents the fusion loss, and the second term represents the type loss, including posting loss, geographic location loss, etc.

[0058] Please refer to Figure 3 , is a schematic diagram of type-aware anchor link prediction according to an embodiment of the present invention. Assuming that a node in a given source network and nodes in the target network The specific process of calculating whether the link between two nodes is an anchor link is as follows:

[0059] (1) Based on the nodes in the source network and target network nodes The attribute information carried by them, learning their initial feature vector and (dimension D), and their neighborhood information in their respective networks and

[0060] (2) The initial eigenvector of And its neighborhood information Input the first layer of graph attention network and calculate and its neighbors The self-attention weight.

[0061] (3) Use the Self-Attention weight calculated in the previous step to perform a dot product with the initial feature vector of each node in the neighborhood of type r to obtain a weighted r-type perception vector with a dimension of m.

[0062] (4) Use the methods in steps (2) and (3) to obtain the node Multiple type-aware vectors.

[0063] (5) Use the methods in steps (2), (3), and (4) to obtain the nodes in the target network. Multiple type-aware vectors of .

[0064] (6) The node The initial feature vector and multiple type-aware vectors are input into the second-layer graph neural network to calculate the additional attention weights of the initial features and each type-aware vector.

[0065] (7) Perform a dot product between the additional attention weight calculated in the previous step and each type perception vector to obtain a weighted type fusion vector with a dimension of m.

[0066] (8) Using a method similar to steps (6) and (7), we can obtain the node The type of the fused vector.

[0067] (9) Computing Node Pairs Type Euclidean distance between fusion vectors.

[0068] (10) Computing Node Pairs Euclidean distance between type-aware vectors of the same type.

[0069] (11) Based on the positive samples (known anchor links), an equal number of negative samples (node ​​pairs that are known not to be anchor links) are selected.

[0070] (12) The loss function of the given two pairs of nodes is calculated by triple-loss, where the triplet-loss of the type-aware vector accounts for 0.6 and the triplet-loss of the type-fusion vector accounts for 0.4, thereby judging the training of the entire model.

[0071] (13) In the testing phase, after learning the n-tuple vector representation of the node, the Euclidean distance of the node to multiple type-aware vectors and type-fusion vectors is calculated, and the weighted sum is performed according to the set weights to obtain the similarity score of each pair of nodes; based on the similarity score, it is determined whether they are anchor links, where the threshold is set to 0.5.

[0072] Finally, the present invention and other anchor link prediction algorithms perform anchor link prediction on Twitter-Foursquare and Aminer-Mag datasets respectively, and the results are shown in Table 1.

[0073]

[0074] Table 1

[0075] Among them, MAG (Tan, S.; Guan, Z.; Cai, D.; Qin, network embedding.In IJCAI 2016,1774–1780.), Deep Link (Zhou, F.; Liu, L.; Zhang, K.; Trajcevski, G.; Wu, J.; and Zhong, T. 2018. Deeplink: A deep learning approach for user identity linkage. In INFOCOM 2018,1313–1321.), HAN (Wang, 2019,2022–2032.), PME (Chen, H.; Yin, H.; Wang, W.; Wang, H.; Nguyen, QVH; and Li, X. 2018. PME: projected metric embedding on heterogeneous networks for link prediction. In KDD 2018,1177–1186.), HHNE(Wang,X.; Zhang,Y.; and Shi,C.2019.Hyperbolic heterogeneous information network embedding.In AAAI 2019,5337–5344.) is the method of others; P@K (the accuracy of Top K) and MAP@K (the average accuracy of all nodes in Top K) are evaluation indicators. In this paper, K takes the values ​​of 1 and 30.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. A person skilled in the art may modify or make equivalent substitutions for the technical solutions of the present invention without departing from the spirit and scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for predicting anchor links of heterogeneous network users based on type perception, the steps of which include: 1) Collecting network information of the network to be detected and the target network, respectively constructing the heterogeneous network to be detected and the target heterogeneous network, wherein each node in the heterogeneous network to be detected and the target heterogeneous network contains partial attribute information of the corresponding network to be detected and the target network; 2) Obtaining the initial feature vectors of each dimension of each node in the heterogeneous network to be detected and the target heterogeneous network, and obtaining the initial feature vector representation of each node; 3) The initial feature vector representation is input into a self-attention graph neural network for learning, and the type perception vector and type fusion vector of each node are calculated according to the weight of each type perception vector of the corresponding neighboring nodes of each node; Among them, when learning based on the self-attention graph neural network, the loss function is is the heterogeneous network node to be detected, Target heterogeneous network node, a represents the node type, represents a positive sample, represents negative samples, ω,λ r ,ξ is a hyperparameter andω+∑λ r =1, r, R are type-aware vector types, Divided into nodes Contains the type of fused vector, Divided into nodes Contains type-aware vectors of type r, d(.,.) is the distance between two vectors, is the fusion loss, is the type loss; the type-aware vector is obtained by the following steps: Compute Node i Each corresponding neighboring node j The type-aware vector weight of type r in the Normalize the type perception vector weight of type r to get the node υ i and j The attention coefficient of type r; By j Weighted summation, we get υ of type r i Type-aware vectors 4) By determining the similarity of the type perception vectors and type fusion vectors of the same type between each pair of nodes in the heterogeneous network to be detected and the target heterogeneous network, it is determined whether the network to be detected and the target network are anchor links.

2. The method according to claim 1, characterized in that The initial feature vector representation of each node is obtained through the following strategy: 1) For nodes containing text information, methods for obtaining initial feature vector representation include: Word2vec method; 2) For nodes containing image information, methods for obtaining initial feature vector representation include: resnet method; 3) For nodes whose text information or image information is unclear, methods for obtaining initial feature vector representation include: random allocation method.

3. The method according to claim 1, characterized in that The type fusion vector is obtained by weighted summing up each type of perception vector and the corresponding type of perception vector with the added attention weight.

4. The method according to claim 1, characterized in that The similarity between the type-aware vector and the type-fused vector is calculated using the Euclidean distance or cosine distance between node pairs.

5. The method according to claim 1, characterized in that Use the following steps to determine whether the network to be detected and the target network are anchor links: 1) The similarity of each type of perception vector and type fusion vector is weighted and summed with the corresponding set weight to obtain a similarity score; 2) Based on the similarity score and a threshold, determine whether the network to be detected and the target network are anchor links.

6. The method according to claim 1, characterized in that The types of losses include: loss of posting or loss of geographical location.

7. A storage medium storing a computer program, wherein: The computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

8. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.