A multi-feature fusion method for user identity association across social networks

Through the multi-feature fusion method, combined with BERT and SDNE graph embedding, a cross-social network user identity association model is built, which solves the challenges of massive user data processing and privacy protection, improves the robustness and accuracy of identity association, and achieves efficient User identity association.

CN114581254BActive Publication Date: 2024-12-03Chinese People's Liberation Army Cyberspace Force Information Engineering University
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210187864.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-28
Publication Date
2024-12-03
Estimated Expiration
2042-02-28

AI Technical Summary

Technical Problem

The research on user identity association across social networks faces the problem of processing massive user data and privacy protection challenges, which leads to insufficient robustness and accuracy of identity association.

Method used

The multi-feature fusion method is adopted to build an identity correlation analysis model through user profile information, user behavior and user relationships, and the feature vector is extracted using BERT and SDNE graph embedding methods, and binary classification is combined with machine learning models to determine whether the user is the same natural person.

Benefits of technology

It improves the robustness and accuracy of user identity associations across social networks, realizes efficient data fusion and user identity associations, and has good application prospects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114581254B_ABST
    Figure CN114581254B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of social network data analysis, and particularly relates to a multi-feature fused cross-social network user identity association method, which obtains original user data from multiple social network platforms, and calculates the similarity of corresponding indicators from dimensions such as user profile information, user behavior information, and user relationship information; the similarity under the three indicator dimensions is used as the input of the trained user identity association model, and the probability that the cross-platform users are the same natural person is output through the user identity association model, and the degree of association between the cross-platform users is analyzed based on the output probability. The present invention constructs a multi-feature fused identity association model based on user attributes, user behavior, and user relationships, improves the robustness and accuracy of the cross-social network user identity association scheme, and facilitates data analysis and application in actual scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of social network data analysis, and in particular relates to a multi-feature fusion cross-social network user identity association method. Background Art

[0002] As the main medium for obtaining and transmitting messages, social networking platforms have penetrated into people's lives and work. An Internet user generally has accounts on multiple social networking platforms. Cross-social network user identity association aims to dig out the correspondence between different identities of the same user on multiple social platforms. It can integrate user data scattered on different social networking platforms, assist in portraying a more comprehensive user portrait, and provide richer data support for complex social network analysis services. For example, based on user identity association relationships, in the field of network security, it can provide support for identifying fake accounts and illegal accounts, and in the field of recommendation systems, it can help commercial recommendation systems provide users with more accurate personalized services.

[0003] Identity association technology has a wide range of application value and has received increasing attention. However, in general, the research on network identity association technology is still in the exploratory stage, and the research on user identity association across social networks faces many challenges. The user data of social network platforms is massive, and it is a difficult problem to integrate the effective user information in large-scale complex networks to achieve efficient identity association. In addition, as privacy protection is increasingly valued, it is difficult to obtain all user identity attribute information and to confirm the authenticity of the information, which brings great challenges to user identity association. Summary of the invention

[0004] To this end, the present invention provides a multi-feature fused cross-social network user identity association method, which constructs a multi-feature fused identity association analysis model based on user profile information, user behavior, and user relationship, improves the robustness and accuracy of the cross-social network user identity association solution, and facilitates application in data analysis and application services in actual scenarios.

[0005] According to the design scheme provided by the present invention, a multi-feature fusion cross-social network user identity association method includes the following contents:

[0006] Obtain relevant data of user V from different social network platforms G, including user profile information P, user behavior information C and user relationship information E, etc.;

[0007] In the user profile dimension, important attributes and general attributes are defined, and the G from the social network platform is calculated in different situations. A and G B Users in and The similarity between

[0008] In the user behavior dimension, the semantic features of user posts are extracted to generate user behavior vectors, and then the user behavior vectors from the social network platform G are calculated. A and G B Users in and The similarity between

[0009] In the user relationship dimension, we expand the friend relationships in different social networks, generate user node vectors that integrate global and local relationship features, and then calculate the user node vectors from the social network platform G A and G B Users in and The similarity between

[0010] Fusion of user similarities in different dimensions determination and Are they the same natural person? Output and is the degree of association of the same natural person ij .

[0011] As a method for associating user identities across social networks of the present invention, further, in the similarity calculation of the user profile information dimension, important attributes and general attributes are defined, and the similarity is calculated in different situations. Attributes that can uniquely identify the user's identity (such as mobile phone number, email address, etc.) are defined as important attributes; other attributes that identify the user's identity are defined as general attributes (such as user name, etc.). First, based on the important attributes, it is matched whether the cross-platform users are the same natural person. For the case of a complete match, the maximum similarity output is the similarity under this dimension. For the case of an incomplete match, the similarity under this dimension is output by measuring the edit distance of the general attribute character sequence.

[0012] As a user identity association method across social networks of the present invention, further, in the calculation of similarity in the user behavior dimension, the characteristics of user behavior can be reflected by extracting the semantic information of user posts. User posts can usually reflect the changes in user consciousness and behavior. BERT is used to extract the semantic information of cross-platform user posts and generate user behavior vectors, and the cosine similarity between the generated vectors is calculated as the similarity in the user behavior dimension.

[0013] As a user identity association method across social networks of the present invention, further, in extracting semantic information of user posts and generating user behavior vectors, according to the user posting time, starting from the most recent posting behavior, n posts of the user are selected in time series, and each post is trained using BERT to extract semantic information, and the user behavior vector is generated by using the first layer plus the last layer of the BERT network to take the average pooling method, and the user behavior vector for calculating similarity is formed by splicing the user's n posting vectors.

[0014] As a user identity association method across social networks of the present invention, further, in the similarity calculation of the user relationship dimension, the SDNE graph embedding method is used to obtain the vectorized representation of the user nodes in the cross-platform social network, and the cosine similarity between the generated user node vectors is calculated as the similarity under the user relationship dimension.

[0015] As a user identity association method across social networks of the present invention, further, in obtaining a vectorized representation of user nodes using the SDNE graph embedding method, first, the association relationship between users is expanded based on the association relationship between users within the platform and the prior association relationship between different platforms (i.e., a predetermined user relationship belonging to the same natural person); then, the SDNE graph embedding method is used to obtain a vectorized representation of the user node, the user node adjacency matrix is ​​used as the input of the SDNE model, and the embedding vector is obtained by dimensionality reduction compression through an automatic encoder, and the local and global similarities between user input vectors are jointly optimized by unsupervised reconstruction of the node neighborhood feature vector and supervised Laplace eigenmap.

[0016] As a user identity association method across social networks of the present invention, further, the user similarities under different dimensions are integrated determination and Are they the same natural person? Output and is the degree of association of the same natural person ij First, obtain several user pairs belonging to the same natural person and user pairs not belonging to the same natural person in different social networks; then calculate the similarity of each user's profile information, user behavior, user relationship and other dimensions to form a similarity vector; next, use the similarity vectors belonging to the same natural person as positive samples, and vice versa as negative samples. Use positive and negative samples as inputs to the machine learning binary classification model to train a user identity association model. When performing cross-social network user identity association, the social network G A Medium User and social network G B Medium User Similarity of profile information User behavior similarity User relationship similarity In the input user identity association model, it is determined whether two users are the same natural person, and the output probability of the model determining that they are the same natural person is used as the user association degree.

[0017] Beneficial effects of the present invention:

[0018] The present invention integrates multiple dimensional features of user profile information, user behavior, and user relationship to construct an identity association model to realize user identity association analysis across social network platforms. By integrating the three dimensional features, user information can be fully utilized to calculate the similarity under the corresponding dimension, thereby improving the robustness and accuracy of the identity association analysis model; by combining important attributes and general attributes to calculate the similarity of user identity attributes, by drawing on the idea of ​​Sentence-BERT to generate user posting vectors and then calculating the similarity of user behavior, by extending the friend relationship and using SDNE for graph embedding to calculate the user relationship similarity, user profile information, behavior and relationship data can be fully utilized to associate user identities, and efficient fusion association can be achieved, which has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of the correlation analysis process of multi-feature fusion in the embodiment;

[0020] Figure 2 This is a schematic diagram of the user profile information similarity calculation process in the embodiment;

[0021] Figure 3 This is a schematic diagram of the user behavior similarity calculation process in the embodiment;

[0022] Figure 4 This is a schematic diagram of a cross-platform social network user vector representation process in an embodiment; DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.

[0024] The present invention provides a method for associating user identities across social networks by integrating multiple features. Figure 1 As shown, it contains the following contents:

[0025] S101, obtaining user data in the social network platform, including user profile information, user behavior information, user relationship information, etc.;

[0026] The original user data of multiple social networking platforms collected in this example is stored in a database (such as MongoDB) and preprocessed, including data cleaning and data alignment. Furthermore, the preprocessed user data is structured and stored through a graph database (such as Neo4j), wherein a node label is set in each node to indicate the name of the social platform where the node user is located, and an association label is set between the nodes to indicate the cross-platform association of the node user.

[0027] S102, in the user profile information dimension, define important attributes and general attributes, and calculate user similarity according to different situations;

[0028] In this example, the attributes that can uniquely identify the user's identity are defined as important attributes, and other identity attributes are general attributes. For example, mobile phone number and email address are important attributes, and user name is a general attribute. The similarity calculation process of user profile information is as follows: Figure 2 As shown. First, determine whether there are important attributes. If there are important attributes, determine whether the corresponding attribute values ​​are completely matched. If they are completely matched, output the similarity of 100%. If the important attributes do not match completely, continue to determine whether there are general attributes; if there are no important attributes, then determine whether there are general attributes. If not, the similarity based on the user identity attribute is recorded as missing. If it exists, use Jaro-Winkler Distance to calculate the user similarity. Jaro-Winkler Distance is a method for calculating the similarity between two strings. The calculation formula is:

[0029] Sim_jw(s1,s2)=d j (s1,s2)+(l×p(1-d j (s1,s2)))

[0030] Among them, s1 and s2 represent two user name strings; l represents the same length starting from the first character, with a maximum value of 4; p is the adjustment factor, which should be less than or equal to 0.25 and is usually set to 0.1; d j (s1,s2) is the Jaro Distance, calculated by the formula

[0031]

[0032] Among them, |s1| and |s2| represent the lengths of the two strings respectively, m represents the number of common strings, and t represents the number of transpositions.

[0033] Jaro-Winkler Distance is an extension of Jaro Distance. It gives higher similarity to strings that are identical from the beginning, taking into account string matching and transposition. It is suitable for calculating the similarity between shorter strings. The higher the final score, the greater the similarity. 0 means no similarity, and 1 means a perfect match.

[0034] S103: In the user behavior dimension, the semantic information of the user's post is extracted to generate a user behavior vector, and then the cosine similarity between the generated vectors of the user pairs is calculated as the similarity in the user behavior dimension.

[0035] User posts usually reflect their changes in consciousness and behavior patterns, and are difficult to forge, so they can be used for identity association. Generally speaking, cross-platform users belonging to the same natural person have similarities in the content of their posts. By extracting semantic information from user posts, the characteristics of the user's post content can be reflected. BERT (Bidirectional Encoder Representation from Transformers) is a pre-trained language representation model that has achieved excellent results in natural language processing tasks. Directly calculating semantic similarity using sentence vectors generated by BERT is not ideal, and the sentence vectors obtained from two similar sentences may be very different. Sentence-BERT uses the twin BERT network to generate sentence vectors separately, and then calculates the cosine similarity between them, and pre-trains the similarity model using the mean square error as the loss function, achieving good results. In the embodiment of this case, the idea of ​​Sentence-BERT can be used to design a similarity calculation model for user post vectors. The specific process is as follows: Figure 3 shown.

[0036] Each account contains multiple posts. First, we select n posts in chronological order, starting from the most recent post. Second, we train each post using BERT. BERT is a model that has been trained with a large number of data sets in advance. It has strong generalization capabilities and includes all preprocessing processes (such as word segmentation and vectorization). Therefore, when using it, you only need to fine-tune and provide text.

[0037] Then, the first layer plus the last layer are averaged and Pooling is used to generate a fixed-length sentence vector. There are two main ways for BERT to generate sentence vectors: taking CLS (feature position vector) to represent the sentence vector or doing a Pooling on all tokens (for Chinese, token refers to characters). Practical experience shows that the sentence vector of the first layer plus the last layer of Pooling is conducive to the calculation of semantic similarity. Next, the n vectors of each account are spliced ​​together to form the user post vector, for example Figure 3u and v in .

[0038] Next, we use cosine similarity to calculate the similarity between the user post vectors. Cosine similarity uses the cosine value of the angle between two vectors in the vector space to measure the size of the difference between two individuals. The calculation formula is:

[0039]

[0040] The closer the cosine value is to 1, the closer the angle is to 0 degrees, that is, the more similar the two vectors are. Where x and y represent two vectors.

[0041] S104. In the user relationship dimension, the friendship relationships between users across social networks are expanded, and a graph embedding algorithm is used to generate a user node vector that integrates global and local relationship features, and then the cosine similarity between the generated vectors of user pairs is calculated as the similarity in the user relationship dimension.

[0042] The association relationship between users connects independent user individuals to form a social network. In order to realize user identity association based on user relationship, it is necessary to convert user node information into vector form for subsequent association analysis. Graph Embedding (NE) represents nodes as low-dimensional and dense vectors, which contain network topology information and provide efficient and accurate node features for social network analysis. SDNE (Structural Deep Network Embedding) is a graph embedding method based on deep learning. Compared with graph embedding methods based on random walks (such as Deepwalk, Node2vec), it has stronger stability in practice and relatively low time complexity. SDNE uses an autoencoder structure to simultaneously optimize first-order and second-order similarities. The learned vector representation can retain global and local structures and is robust to sparse networks.

[0043] In this embodiment, the SDNE method can be used to obtain the vector representation of user nodes in the cross-platform social network. The specific steps are as follows: Figure 4 As shown in Figure 2, in the two social platforms, there is a correlation between users within the platform ( Figure 4 Dark solid line in the first step), and there is also a priori relationship between different platforms ( Figure 4 The dark dashed line in the first step) is the relationship between the predetermined pairs of users belonging to the same natural person. For example, A and A', C and C', D and D' in the two platforms all belong to the same natural person.

[0044] Some of the relationships between friends in the original data obtained may be missing. Before generating user vectors, the first step is to expand the friend relationship. Assuming that two users have a friend relationship in platform 1, then the friend relationship also exists in platform 2. For example, in platform 1, there is a friend relationship between C and D, so there should also be a friend relationship between C' and D' in platform 2; in platform 2, there is a friend relationship between A' and D', so there should also be a friend relationship between A and D in platform 1. Figure 4 The bold solid line in the first step is the extended friend relationship. In the second step, SDNE is used to generate user node vectors in the social network. The adjacency vector of the user node in the social network is used as the input of the SDNE model, and the input vector is reduced to a compressed embedding vector through the autoencoder. Then, the node neighborhood features and Laplace eigenmap are reconstructed through an unsupervised autoencoder to jointly optimize the second-order similarity and first-order similarity. The second-order similarity refers to the similarity between two nodes that have common neighbors but are not directly connected, which reflects the local characteristics of the social network; the first-order similarity refers to the similarity of the embeddings of user nodes with edge connections, which mainly reflects the global characteristics of the social network. For example, Figure 4 The adjacency matrix of the D nodes in platform 1 is used as the input vector, and an embedding vector is obtained through a deep autoencoder, and then unsupervised and supervised second-order and first-order similarity optimization is performed. In the unsupervised module, the node neighborhood features are reconstructed by the decoder, and the optimization goal is to minimize the mean square error between the original vector and the reconstructed feature vector. In the supervised module, the calculation of the distance between the two embedding vectors is optimized through the Laplace feature mapping idea. When similar nodes are mapped far away from each other in the embedding space, the feature mapping will be penalized. Finally, the cosine similarity is used to calculate the similarity between the final generated user node embedding vectors.

[0045] S105. Integrate user profiles, user behaviors, and user relationship dimensions to perform user identity association. Construct an experimental data set, including positive samples (belonging to the same natural person) and negative samples (not belonging to the same natural person), to verify the effectiveness of the method. Generate 10,000 account pairs based on 100 account information from each of the two social networking platforms in the data set, of which 20 are positive samples (two accounts belong to the same person) and 80 are negative samples (two accounts do not belong to the same person). The username attributes of the accounts all exist, but the profile attributes are incomplete. The 10,000 account pairs are divided into training and test sets at a ratio of 8:2.

[0046] Calculate the similarity of positive and negative samples in user profile information, user behavior, and user relationship dimensions, input 80% of the training samples into the machine learning algorithm, and train to obtain the user identity association model. In this example, the decision tree is selected as the machine learning classification algorithm. The algorithm is implemented using the corresponding module in the Scikit-Learn third-party library and set to default parameters.

[0047] The user identity association model is evaluated on a 20% test set using accuracy, precision, and recall.

[0048]

[0049]

[0050]

[0051] Among them, TP / TN represents the number of correctly predicted positive samples / negative samples, and FP / FN represents the number of incorrectly predicted positive samples / negative samples.

[0052] The experimental results of this embodiment are as follows:

[0053] The accuracy of the test set is 94.37%, the precision is 96.23%, the recall is 92.57%, and the test time is 0.000997 seconds. This model can not only determine whether a user pair is the same natural person, but also use the probability that the user pair is the same natural person as the degree of association.

[0054] The experimental results show that the multi-feature fusion cross-social network user identity association method proposed in the present invention can achieve higher accuracy and efficiency.

[0055] In the table, bold fonts indicate the best results, and running time refers to the time consumed to train the classifier by inputting three features: user profile attribute similarity, user behavior similarity, and user relationship similarity.

[0056] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0057] Based on the above method and / or system, an embodiment of the present invention also provides a server, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0058] Based on the above method and / or system, an embodiment of the present invention further provides a computer-readable medium having a computer program stored thereon, wherein the above method is implemented when the program is executed by a processor.

[0059] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.

[0060] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0061] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A multi-feature fusion method for associating user identities across social networks. It is characterized in that Contains the following: Obtain relevant data of user V from different social network platforms G, including user profile information P, user behavior information C and user relationship information E; In the user profile dimension, the attributes that uniquely identify the user are defined as important attributes, and other attributes that identify the user are defined as general attributes. A and G B Users in and Similarity between In the user behavior dimension, the semantic features of user posts are extracted to generate user behavior vectors, and then the user behavior vectors from the social network platform G are calculated. A and G B Users in and Similarity between In the user relationship dimension, we expand the friend relationships in different social networks, generate user node vectors that integrate global and local relationship features, and then calculate the user node vectors from the social network platform G A and G B Users in and Similarity between Among them, in the calculation of similarity in the user relationship dimension, the SDNE graph embedding method is used to obtain the vectorized representation of user nodes in the cross-platform social network, and the cosine similarity between the generated user node vectors is calculated as the similarity in the user relationship dimension; and in the use of the SDNE graph embedding method to obtain the vectorized representation of the user node, first, the association relationship between users is expanded based on the association relationship between users within the platform and the prior association relationship between different platforms; then, the SDNE graph embedding method is used to obtain the vectorized representation of the user node, and the user node adjacency matrix is ​​used as the input of the SDNE model, and the embedding vector is obtained by dimensionality reduction compression through the automatic encoder, and the local and global similarities between the user input vectors are jointly optimized by unsupervised reconstruction of the node neighborhood feature vector and supervised Laplace eigenmap; Fusion of user similarities in different dimensions determination and Are they the same natural person? Output and is the degree of association of the same natural person ij .

2. According to claim 1, a method for associating user identities across social networks by integrating multiple features, It is characterized in that In the similarity calculation of user profile information dimension, we first match whether the cross-platform users are the same natural person based on important attributes. For the case of complete match, the maximum similarity output is the similarity under this dimension. For the case of incomplete match, the similarity under this dimension is output by measuring the edit distance of the general attribute character sequence.

3. According to the multi-feature fusion cross-social network user identity association method of claim 1, It is characterized in that In the similarity calculation of the user behavior dimension, BERT is used to extract the semantic information of cross-platform user posts and generate user behavior vectors. The cosine similarity between the generated vectors is calculated as the similarity under the user behavior dimension.

4. According to claim 3, a method for associating user identities across social networks by integrating multiple features, It is characterized in that In extracting the semantic information of user posts and generating user behavior vectors, based on the user's posting time, starting from the most recent posting behavior, n posts of the user are selected in time series, and BERT is used to train each post to extract semantic information. The user behavior vector is generated by taking the average pooling method of the first layer and the last layer of the BERT network. The user's n posting vectors are spliced ​​to form a user behavior vector for calculating similarity.

5. According to the multi-feature fusion cross-social network user identity association method of claim 1, It is characterized in that Comprehensive user similarity under different dimensions determination and Are they the same natural person? Output and is the degree of association of the same natural person ij ,First, obtain several user pairs belonging to the same natural person and user pairs not belonging to the same natural person in different social networks; ,then calculate the similarity of each user’s profile information, user behavior and user relationship dimensions to form a ,similarity vector; Next, the similarity vectors belonging to the same natural person are taken as positive samples, and vice versa as negative samples. The positive and negative samples are used as the input of the machine learning binary classification model to train the user identity association model. When associating user identities across social networks, the social network G A Medium User and social network G B Medium User Similarity of profile information User behavior similarity User relationship similarity In the input user identity association model, it is determined whether two users are the same natural person, and the output probability of the model determining that they are the same natural person is used as the user association degree.

Citation Information

Patent Citations

  • Information pushing method and device based on graph network

    CN112069398A

  • Cross-social network user identity recognition method based on joint embedded learning model

    CN113434782A

  • Cross-space target virtual identity association method based on multilayer attribute analysis

    CN113779520A