An identity correlation method and device based on SDNE embedding and deep learning

Abstract

The application provides an identity correlation method and device based on SDNE embedding and deep learning, and relates to the technical field of identity correlation; comprising: constructing a first data set and a second data set; based on the basic data of the correlated users in any social network, updating the user link relationship between the correlated users in another social network by using a network expansion method; inputting the updated first data set and second data set into an SDNE embedding network to obtain a first low-dimensional embedding result set and a second low-dimensional embedding result set; inputting the first low-dimensional embedding result set and the second low-dimensional embedding result set into a deep learning identity correlation model respectively to obtain a first vector set and a second vector set; and performing identity correlation by using a user alignment method based on the first vector set and the second vector set. The present application performs cross-social network user identity correlation by combining user name information and network topology structure information, thereby improving the accuracy of identity correlation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of identity association, in particular to an identity association method and device based on SDNE embedding and deep learning. BACKGROUND

[0002] User identity association refers to identifying that the user accounts on different online social network platforms belong to the same natural person by analyzing the user identity information on different online social network platforms. With the popularity of the Internet and the rise of online social media, the number of personal accounts and personal profiles created by users on different platforms is increasing, and user identity association has become an important research field. By associating the behavior and related features of users on different platforms, a user's online activities and interests can be more comprehensively understood; social phenomena such as social network structure, information dissemination and public opinion evolution can be better understood and studied; and social institutions and enterprises can more accurately identify and target user groups.

[0003] At present, the main cross-social network user identity association technology is achieved by user attribute information, user behavior information and network topology structure. The cross-social network identity association method based on user attribute information or user behavior information is not robust enough; the cross-social network identity association method based on network topology cannot fully capture information, resulting in inaccurate association results. At the same time, based on the huge number of users in social networks, the existing cross-social network user identity association methods are difficult to achieve both high association accuracy and high association efficiency. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide an identity association method and device based on SDNE embedding and deep learning, which can quickly and accurately associate user identities between different social networks to solve the above problems in the prior art.

[0005] In a first aspect, an identity association method based on SDNE embedding and deep learning is provided, which can include:

[0006] A first data set and a second data set are constructed; wherein the first data set includes basic data of all users in a first social network; the second data set includes basic data of all users in a second social network; the all users include: associated users and unassociated users; the basic data includes: a user name and a user link relationship; the associated user is a user who has associated the same user's user name in the first social network with the user name in the second social network; the user link relationship is a follow relationship between different users in the same social network;

[0007] Based on the basic data of the associated users in any social network, the user link relationship between the associated users in another social network is updated by using a network expansion method, to obtain an updated first data set and an updated second data set;

[0008] The updated first data set and the updated second data set are respectively input into a pre-constructed SDNE embedding network, to obtain a first low-dimensional embedding result set of all users in the first social network and a second low-dimensional embedding result set of all users in the second social network;

[0009] The first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model, to obtain a first vector set and a second vector set; the first vector set includes user vectors of all unassociated users in the first social network; and the second vector set includes user vectors of all unassociated users in the second social network;

[0010] Based on the first vector set and the second vector set, a user alignment method is used to associate the unassociated users in the first social network with the unassociated users in the second social network.

[0011] In an optional implementation, based on the basic data of the associated users in any social network, the user link relationship between the associated users in another social network is updated by using a network expansion method, to obtain an updated first data set and an updated second data set, which includes:

[0012] For any two associated users having a user link relationship in any social network, it is determined whether the two associated users also have a user link relationship in another social network;

[0013] If not, a user link relationship is added for the two associated users in the other social network, to obtain an updated user link relationship;

[0014] Based on the obtained updated user link relationship, a data set corresponding to the social network is updated, to obtain an updated first data set and an updated second data set.

[0015] In an optional implementation, the first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model, to obtain a first vector set and a second vector set, which includes:

[0016] Low-dimensional embedding results corresponding to all associated users in the first social network in the first low-dimensional embedding result set are extracted, and low-dimensional embedding results corresponding to all associated users in the second social network in the second low-dimensional embedding result set are extracted;

[0017] construct a dataset based on the extracted low-dimensional embedding results; wherein the dataset comprises a training dataset and a testing dataset;

[0018] train a pre-constructed deep learning identity correlation model using the training dataset to obtain a trained deep learning identity correlation model;

[0019] test the trained deep learning identity correlation model using the testing dataset to obtain a tested deep learning identity correlation model;

[0020] input the low-dimensional embedding results of all uncorrelated users in the first social network in the first low-dimensional embedding result set into the tested deep learning identity correlation model to obtain a first vector set;

[0021] input the low-dimensional embedding results of all uncorrelated users in the second social network in the second low-dimensional embedding result set into the tested deep learning identity correlation model to obtain a second vector set.

[0022] In an optional implementation, based on the first vector set and the second vector set, a user alignment method is used to correlate the identities of the uncorrelated users in the first social network with the uncorrelated users in the second social network, comprising:

[0023] for any user vector in the first vector set, calculate the cosine similarity of the user vector with all user vectors in the second vector set to obtain a first similarity calculation result;

[0024] according to the first similarity calculation result, arrange all user vectors in the second vector set in descending order to obtain a first arrangement result;

[0025] based on the first arrangement result, select the top n user vectors in the second vector set; wherein n represents the number of selected user vectors in the second vector set;

[0026] calculate the edit distance of the usernames corresponding to the selected user vectors in the second vector set and the usernames corresponding to the user vectors in the first vector set to obtain an edit distance calculation result; wherein the edit distance calculation result is the number of edits required to change the usernames corresponding to the selected user vectors in the second vector set to the usernames corresponding to the user vectors in the first vector set;

[0027] according to the edit distance calculation result, correlate the identities of the usernames corresponding to the selected user vectors in the second vector set with the usernames corresponding to the user vectors in the first vector set.

[0028] In an optional implementation, according to the edit distance calculation result, identity association is performed between the username corresponding to the user vector in the selected second vector set and the username corresponding to the user vector in the first vector set, including:

[0029] The edit distance calculation result is multiplied by a weight coefficient to obtain a weighted edit distance calculation result.

[0030] The weighted edit distance calculation result is added to the first similarity calculation result to obtain a second similarity calculation result.

[0031] According to the second similarity calculation result, the user vectors in the selected second vector set are arranged in descending order to obtain a second arrangement result.

[0032] The user vector at the first position in the second arrangement result is selected as the user vector closest to the user vector in the first vector set.

[0033] The second similarity calculation result corresponding to the selected closest user vector is compared with a preset threshold.

[0034] If the second similarity calculation result corresponding to the closest user vector is not less than the preset threshold, identity association is performed between the username corresponding to the closest user vector and the username corresponding to the user vector in the first vector set.

[0035] In an optional implementation, the user corresponding to the closest user vector is an unassociated user in the second social network, and the user corresponding to the user vector in the first vector set is an unassociated user in the first social network.

[0036] Identity association between the username corresponding to the closest user vector and the username corresponding to the user vector in the first vector set includes:

[0037] The closest user vector corresponding to the unassociated user in the second social network is determined, and the unassociated user in the first social network corresponding to the user vector in the first vector set is determined.

[0038] Identity association is performed between the username of the unassociated user in the second social network and the username of the unassociated user in the first social network.

[0039] In an optional implementation, the basic data of the associated user further includes a corresponding relationship between the usernames in different social networks.

[0040] After the unassociated user in the first social network and the unassociated user in the second social network are identity associated by using a user alignment method based on the first vector set and the second vector set, the method further includes:

[0041] transforming an unassociated user in the first social network and the second social network that is associated with an identity into an associated user;

[0042] adding a corresponding relationship of a username of the associated user between different social networks in the basic data of the associated user;

[0043] updating the first data set and the second data set based on the associated user after adding the corresponding relationship of the username between different social networks.

[0044] In a second aspect, an identity association device based on SDNE embedding and deep learning is provided, which can include:

[0045] a construction unit configured to construct a first data set and a second data set; wherein the first data set includes basic data of all users in a first social network; the second data set includes basic data of all users in a second social network; the all users include associated users and unassociated users; the basic data includes a username and a user link relationship; the associated users are users who have associated a username of a same user in the first social network with a username in the second social network; the user link relationship is a follow relationship between different users in a same social network;

[0046] an updating unit configured to update a user link relationship between associated users in another social network based on the basic data of the associated users in any social network by using a network expansion method, to obtain an updated first data set and an updated second data set;

[0047] an SDNE embedding unit configured to input the updated first data set and the updated second data set into a pre-constructed SDNE embedding network respectively, to obtain a first low-dimensional embedding result set of all users in the first social network and a second low-dimensional embedding result set of all users in the second social network;

[0048] an identity association unit configured to input the first low-dimensional embedding result set and the second low-dimensional embedding result set into a pre-constructed deep learning identity association model respectively, to obtain a first vector set and a second vector set; wherein the first vector set includes user vectors of all unassociated users in the first social network; the second vector set includes user vectors of all unassociated users in the second social network; and based on the first vector set and the second vector set, a user alignment method is used to associate the unassociated users in the first social network with the unassociated users in the second social network.

[0049] In a third aspect, an electronic device is provided, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus.

[0050] The memory is configured to store a computer program.

[0051] The processor is configured to execute the program stored in the memory to implement the method steps of any one of the first aspect.

[0052] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the method steps of any one of the first aspect.

[0053] The present application combines user name information and network topology information to perform cross-social network user identity association, makes up for the deficiency caused by using single user attribute information, combines the SDNE algorithm to express the local and global structure information of the network, improves the generalization of identity association, and expands the user relationship of the captured social network information before identity association to discover more potential user relationships and improve the accuracy of identity association. The present application can be applied to real social network user identity association, has high identity association accuracy, and has higher association efficiency than existing identity association methods. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0055] Figure 1 A flow chart of an identity association method based on SDNE embedding and deep learning provided by the embodiments of the present application;

[0056] Figure 2 A social network topology diagram provided by the embodiments of the present application;

[0057] Figure 3 An SDNE embedding structure diagram provided by the embodiments of the present application;

[0058] Figure 4 An architecture diagram of an identity association method based on SDNE embedding and deep learning provided by the embodiments of the present application;

[0059] Figure 5A Twitter-FourSquare dataset prediction ranking F1 value comparison chart provided by an embodiment of the present application;

[0060] Figure 6 A Lastfm-MySpace dataset prediction ranking F1 value comparison chart provided by an embodiment of the present application;

[0061] Figure 7 A Livejournal-MySpace dataset prediction ranking F1 value comparison chart provided by an embodiment of the present application;

[0062] Figure 8 A structure schematic diagram of an identity association device based on SDNE embedding and deep learning provided by an embodiment of the present application;

[0063] Figure 9 A structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, and not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0065] For the convenience of understanding, the terms involved in the embodiments of the present application are explained as follows:

[0066] Anchor link, which means the link for associating and identifying the identities or accounts of two or more users in multiple social networks or social platforms as the same user entity.

[0067] Anchor user (or anchor node), which refers to a user participating in two social networks at the same time. Other users not participating in two social networks at the same time are non-anchor users (or non-anchor nodes).

[0068] Identity association, which means judging whether the users from two different social networks belong to the same natural person.

[0069] The identity association method based on SDNE embedding and deep learning provided in the embodiments of the present application can be applied in a server or a terminal with strong computing capability. The server can be a physical server, a server cluster composed of multiple physical servers or a distributed system, or a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (CDN), and big data and artificial intelligence platform. The terminal can be a user equipment (UE) such as a mobile phone, a smart phone, a notebook computer, a digital broadcast receiver, a personal digital assistant (PDA), a tablet computer (PAD), a handheld device, a vehicle-mounted device, a wearable device, a computing device or other processing device connected to a wireless modem, a mobile station (MS), a mobile terminal (Mobile Terminal), etc. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0070] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings of the specification, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0071] Figure 2 A flowchart of an identity association method based on SDNE embedding and deep learning provided in the embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, the method can include the following steps. Figure 2

[0072] In step S110, a first data set and a second data set are constructed.

[0073] In the embodiments of the present application, the first data set includes basic data of all users in a first social network; and the second data set includes basic data of all users in a second social network. Specifically, the first social network and the second social network are two different social networks or social platforms, which can be selected from any two of microblog, Facebook, Twitter, FourSquare, Lastfm, MySpace and Livejournal; or can be other social networks with non-repeated usernames.

[0074] In the embodiments of the present application, all users include associated users and unassociated users; and the associated users are users who have associated the username of the same user in the first social network with the username in the second social network.​

[0075] In the embodiments of the present application, the basic data of the associated users includes a user name, a user link relationship and a user name correspondence relationship between different social networks; and the basic data packet of the unassociated users includes a user name and a user link relationship. The user link relationship is a follow relationship between different users in the same social network.

[0076] In the embodiments of the present application, after obtaining the basic data of all users in each social network, the obtained basic data of all users in each social network needs to be cleaned and normalized; then, based on the normalized basic data of all users in the first social network, a first data set is constructed; and based on the normalized basic data of all users in the second social network, a second data set is constructed.

[0077] In step S120, based on the basic data of the associated users in any social network, a network expansion method is used to update the user link relationship between the associated users in another social network, to obtain an updated first data set and an updated second data set.

[0078] In the embodiments of the present application, based on the basic data of the associated users in any social network, a network expansion method is used to update the user link relationship between the associated users in another social network, to obtain an updated first data set and an updated second data set, which includes:

[0079] For any two associated users with a user link relationship in any social network, it is determined whether the two associated users also have a user link relationship in another social network;

[0080] If not, a user link relationship is added between the two associated users in another social network, to obtain an updated user link relationship; based on the obtained updated user link relationship, the data set corresponding to the social network is updated, to obtain an updated first data set and an updated second data set.

[0081] In actual application, because the data in the data set collected in the real social network may not be comprehensive in network topology structure, in the embodiments of the present application, the first data set and the second data set are network expanded, the potential user relationship is identified and the first data set and the second data set are updated, to obtain a comprehensive network topology structure of the first social network and the second social network, and to increase the user association accuracy. The network expansion is to find the common friends of the anchor users in the two social networks through the known anchor link, and to complete the missing friend relationship.

[0082] In the embodiments of the present application, a social network is abstracted into an undirected and unweighted graph, and a social network is represented by G=(V, E), where V={v1, v2,... vN} is a set of user nodes in the social network, n} is a set of user nodes in the social network, is a set of edges in the social network. When there is a user link relationship (i.e., a follow relationship) between two user nodes in the social network, there is an edge between the two user nodes.

[0083] In the embodiments of the present application, two social networks are taken as examples, which are a first social network G S =(V S , E S ) and a second social network G T =(V T , E T ). For convenience of distinction, the first social network is represented by S network, and the second social network is represented by T network.

[0084] As shown in Figure 2 , the S network includes an i user (i.e., v i ), a j user (i.e., v j ), an m user (i.e., v m ), an a user (i.e., v a ), and a p user (i.e., v p ); and the T network includes a k user (i.e., v k ), an h user (i.e., v h ), an l user (i.e., v l ), an n user (i.e., v n ), and a c user (i.e., v c ). When there is an anchor link between any user in the S network and any user in the T network, it indicates that the user in the S network and the user in the T network are the same natural person, and the user is an associated user at this time.

[0085] According to Figure 2 , there is an anchor link between the i user (i.e., v i ) in the S network and the k user (i.e., v k ) in the T network, which indicates that the i user in the S network and the k user in the T network are the same natural person, and both the i user and the k user are associated users; similarly, there is an anchor link between the j user (i.e., v j ) in the S network and the l user (i.e., v l ) in the T network, which indicates that the j user in the S network and the l user in the T network are the same natural person, and both the j user and the l user are associated users. Meanwhile, according to Figure 2It can be concluded that the user link relationship exists between the k user and the l user in the T network (i.e., the k user and the l user are also social network friends).

[0086] Therefore, based on the above judgment, the user link relationship in the T network can be extended, and a user link relationship between the k user and the l user is added, indicating that they are social network friends. The network extension is specifically expressed as formula (1).

[0087]

[0088] wherein, V S denotes the S network; V T denotes the T network; denotes the i user in the S network; denotes the k user in the T network; denotes an anchor link, at this time, and denote the same user in the S network and the T network, i.e., the i user in the S network and the k user in the T network are the same natural person; denotes an anchor link, at this time and denote the same user in the S network and the T network, i.e., the j user in the S network and the l user in the T network are the same natural person; denotes the user link relationship between the i user in the S network and the j user in the S network, i.e., the i user in the S network and the j user in the S network have a follow relationship or have a friend relationship; denotes the user link between the k user in the T network and the l user in the T network.

[0089] The present application supplements the missing user link relationship in the first social network and the second social network, thereby more accurately reflecting the link relationship between users in the same social network. This helps to improve the reliability of the social network representation and provides a more accurate basis for subsequent data analysis and application. Through a reasonable network expansion strategy, more comprehensive and complete network information can be obtained, thereby better understanding and analyzing the relationship between nodes in the social network.

[0090] In step S130, the updated first data set and the updated second data set are respectively input into the pre-constructed SDNE embedding network to obtain a first low-dimensional embedding result set of all users in the first social network and a second low-dimensional embedding result set of all users in the second social network.

[0091] In the embodiment of the present application, the updated first data set and the second data set are input into the pre-constructed SDNE embedding network; in the SDNE embedding network, firstly, the network topology graph of the first social network and the network topology graph of the second social network are constructed based on the updated first data set and the second data set respectively; secondly, the network topology graph of the first social network and the network topology graph of the second social network are trained by using the encoder and the decoder respectively; finally, the low-dimensional embedding results of the users of the first social network and the low-dimensional embedding results of all users of the second social network are extracted in the encoder part respectively as the final output of the SDNE embedding network.

[0092] In the embodiment of the present application, the SDNE embedding network includes a self-encoder, a decoder, a reconstruction loss function and a regularization loss function, which are used to capture the implicit information and the association relationship in the social network structure by learning the low-dimensional representation of the social network; wherein the input of the self-encoder is the adjacency matrix of the social network, the self-encoder maps the input adjacency matrix to the node embedding vector in the low-dimensional space through an encoder, and then reconstructs the node embedding vector into the adjacency matrix through a decoder; finally, the low-dimensional representation of the nodes in the social network is learned by minimizing the reconstruction error.

[0093] The construction process of the SDNE embedding network in the embodiment of the present application is as follows:

[0094] Firstly, in order to maintain the similarity between nodes, the first-order and second-order similarities are introduced into the SDNE embedding network.

[0095] Among them, the first-order similarity represents the direct link relationship between users in the same social network, and the loss function of the first-order similarity is defined as formula (2) as follows:

[0096]

[0097] Among them, s i，j represents the element value in the adjacency matrix of the social network, when there is a direct link relationship between the users (i.e. nodes) of the social network, s i，j = 1, otherwise s i，j = 0; is the embedding vector of user i, is the low-dimensional embedding result of user j, and user i and user j are a pair of adjacent users. tr(·) represents the trace of the adjacency matrix of the social network, L is the Laplace matrix, and Y is the hidden layer space vector matrix processed by the encoder. Minimizing the loss function L 1st of the first-order similarity makes the embedding vectors corresponding to the two adjacent users in the social network topology graph close in the hidden space.

[0098] The second-order similarity represents the direct and indirect link relationship between users in the same social network, i.e., the neighbor users shared by the users. The loss function of the second-order similarity is defined as formula (3) as follows:

[0099]

[0100] wherein, represents the user node after the training of the SDNE embedding network, i.e., the user embedding vector; x i represents the user node in the social network, x i is input into the SDNE embedding network in the form of an adjacency matrix vector; N represents the number of users in the social network. Since the user node is input into the SDNE embedding network in the form of an adjacency matrix vector, N also represents the number of user vectors in the social network.

[0101] Secondly, to solve the problem that the non-zero elements in the social network adjacency matrix are far less than the zero elements due to the sparsity of the social network, the loss function is improved as follows:

[0102]

[0103] wherein, b i represents the error weight; represents the user node set after the training, X represents the user node set in the social network, B represents the error weight set, F represents the Frobenius norm; and is the Hadamard product, which means that the corresponding elements are multiplied. The zero in the social network adjacency matrix corresponds to b=1, and the non-zero element corresponds to b>1. The weighted loss function is used, and the non-zero element has a higher penalty coefficient.

[0104] Then, the formula of the encoder is as follows:

[0105]

[0106]

[0107] wherein, represents the first hidden layer of the encoder; and is a nonlinear function; W (k) is the parameter matrix of the kth hidden layer, represents any one of the 2nd to Kth hidden layers; and K represents the last hidden layer; W (1) is the parameter matrix of the 1st layer; and b (k) is the bias term of the kth layer; as Figure 3 shown, from x i to is the encoder, from to is a decoder, is a low-dimensional embedding result of user i; x j to is an encoder, from to is a decoder, is a low-dimensional embedding result of user j. The low-dimensional embedding result of the user is represented by a vector, that is, the obtained low-dimensional embedding result of the user is actually a low-dimensional embedding result.

[0108] Finally, a regularization is added to prevent overfitting, and the loss function of the SDNE embedding network of the embodiment of the application is obtained as formula (7) as follows:

[0109]

[0110] wherein vL reg is a regularization term, a is a parameter for controlling the first-order loss function, v is a parameter for controlling the regularization term; L 2nd is a second-order loss function.

[0111] The SDNE embedding network of the embodiment of the application considers the local and global information of the network structure. It uses two loss functions to optimize the network embedding: reconstruction loss and regularization loss. The reconstruction loss is used to measure the reconstruction error between the encoder and the decoder. The regularization loss is used to maintain the smoothness and continuity of the network embedding to prevent overfitting and instability of the embedding.

[0112] The training process of the SDNE embedding network of the embodiment of the application learns the embedding representation of the network nodes by minimizing the reconstruction loss and the regularization loss. Once the model training is completed, the low-dimensional embedding vector of the node can be extracted using the encoder part.

[0113] In the embodiment of the application, the first low-dimensional embedding result set includes the low-dimensional embedding results of all unassociated users and all associated users in the first social network; and the second low-dimensional embedding result set includes the low-dimensional embedding results of all unassociated users and all associated users in the second social network.

[0114] In step S140, the first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model to obtain a first vector set and a second vector set.

[0115] In the embodiment of the application, the first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model to obtain a first vector set and a second vector set, including:

[0116] extracting low-dimensional embedding results corresponding to all associated users in the first social network from the first low-dimensional embedding result set, and extracting low-dimensional embedding results corresponding to all associated users in the second social network from the second low-dimensional embedding result set;

[0117] constructing a data set based on the extracted low-dimensional embedding results; wherein the data set comprises a training data set and a test data set;

[0118] training a pre-constructed deep learning identity correlation model using the training data set to obtain a trained deep learning identity correlation model;

[0119] testing the trained deep learning identity correlation model using the test data set to obtain a tested deep learning identity correlation model;

[0120] inputting low-dimensional embedding results corresponding to all unassociated users in the first social network in the first low-dimensional embedding result set into the tested deep learning identity correlation model to obtain a first vector set;

[0121] inputting low-dimensional embedding results corresponding to all unassociated users in the second social network in the second low-dimensional embedding result set into the tested deep learning identity correlation model to obtain a second vector set.

[0122] In the embodiments of the present application, the first vector set comprises user vectors of all unassociated users in the first social network; and the second vector set comprises user vectors of all unassociated users in the second social network.

[0123] Specifically, the deep learning identity correlation model uses a multi-layer perceptron (MLP) to learn a mapping function between the two social networks based on the first low-dimensional embedding result set and the second low-dimensional embedding result set, and uses a gradient descent method as an optimization algorithm to iteratively update parameters, while optimizing the mapping function by minimizing the following loss function formula (8):

[0124]

[0125] wherein cos(·) represents the cosine similarity between two user vectors, Φ(·) represents the mapping function, L(·) represents the loss function, and the range of the loss function is 0-2, the smaller the value, the higher the similarity, indicating that the function Φ mapping is more accurate.

[0126] In the embodiment of the present application, two-thirds of the data in the data set are taken as the training data set, and one-third of the data in the data set are taken as the test data set; the deep learning identity correlation model is trained by using the training data set, and the related mapping function and parameters are saved; the test data set is input into the trained deep learning identity correlation model for testing to determine the accuracy of the trained deep learning identity correlation model; if the accuracy meets the preset threshold, it is determined that the trained deep learning identity correlation model passes the test; and the deep learning identity correlation model that passes the test (with the corresponding saved related mapping function and parameters as the function and parameters of the model) is applied to the identity correlation of users of different social networks.

[0127] In step S150, based on the first vector set and the second vector set, the user alignment method is used to perform identity correlation between the uncorrelated users in the first social network and the uncorrelated users in the second social network.

[0128] In the embodiment of the present application, after obtaining the first vector set and the second vector set, one is arbitrarily selected from the first vector set as a to-be-matched user vector; similarity calculation is performed on the to-be-matched user vector and all user vectors in the second vector set, and the edit distance of the corresponding usernames of the user vectors is fused as a constraint to screen n user vectors that match the to-be-matched user vector from all user vectors in the second vector set; the n user vectors are sorted in descending order to determine the user vector in the second vector set that has the highest matching degree with the to-be-matched user vector; and identity correlation is performed between the determined username corresponding to the user vector and the username of the to-be-matched user vector. Wherein, n represents the number, which is a positive integer, and can be selected as 100.

[0129] Specifically, based on the first vector set and the second vector set, the user alignment method is used to perform identity correlation between the uncorrelated users in the first social network and the uncorrelated users in the second social network, including:

[0130] For any user vector in the first vector set, the cosine similarity of the user vector and all user vectors in the second vector set is calculated to obtain a first similarity calculation result; and all user vectors in the second vector set are arranged in descending order according to the first similarity calculation result to obtain a first arrangement result;

[0131] Based on the first arrangement result, the top n user vectors in the second vector set are selected; wherein n represents the number of selected user vectors in the second vector set; and the edit distance of the usernames corresponding to the selected user vectors in the second vector set and the usernames corresponding to the user vectors in the first vector set is calculated to obtain an edit distance calculation result;

[0132] According to the edit distance calculation result, identity correlation is performed between the usernames corresponding to the selected user vectors in the second vector set and the usernames corresponding to the user vectors in the first vector set.

[0133] Specifically, the edit distance calculation result is the number of edits for transforming the username corresponding to the user vector in the second vector set to the username corresponding to the user vector in the first vector set.

[0134] In the embodiment of the present application, according to the edit distance calculation result, the username corresponding to the selected user vector in the second vector set is identity-associated with the username corresponding to the user vector in the first vector set, comprising:

[0135] The edit distance calculation result is multiplied by the weight coefficient to obtain a weighted edit distance calculation result; the weighted edit distance calculation result is added to the first similarity calculation result to obtain a second similarity calculation result.

[0136] According to the second similarity calculation result, the selected user vector in the second vector set is arranged in descending order to obtain a second arrangement result; the user vector located at the first item of the second arrangement result is selected as the user vector closest to the user vector in the first vector set.

[0137] The second similarity calculation result corresponding to the selected closest user vector is compared with a preset threshold value; if the second similarity calculation result corresponding to the closest user vector is not less than the preset threshold value, the username corresponding to the closest user vector is identity-associated with the username corresponding to the user vector in the first vector set.

[0138] In the embodiment of the present application, the user corresponding to the closest user vector is an unassociated user in the second social network; the user corresponding to the user vector in the first vector set is an unassociated user in the first social network. That is, the present application actually selects the closest user from all unassociated users in the second social network for each unassociated user in the first social network, and then identity-associates them.

[0139] In the embodiment of the present application, identity-associating the username corresponding to the closest user vector with the username corresponding to the user vector in the first vector set comprises:

[0140] The unassociated user in the second social network corresponding to the closest user vector is determined, and the unassociated user in the first social network corresponding to the user vector in the first vector set is determined; the username of the unassociated user in the second social network is identity-associated with the username of the unassociated user in the first social network.

[0141] In the embodiments of the present application, the identity association is actually to determine whether the users from the two different social networks belong to the same natural person; therefore, the identity association problem of the unassociated users in the first social network and the unassociated users in the second social network is converted into a binary classification problem, and the function f is constructed, wherein if the i-th user in the S network and the k-th user in the T network belong to the same natural person, 1 is output; otherwise, 0 is output; as shown in formula (9).

[0142]

[0143] Specifically, the steps of the user alignment method can be as follows: for any one vector set obtained by the deep learning identity association model, the similarity between any one user vector in the vector set and all user vectors of another social network is calculated; the results are ranked in descending order of similarity, and the first n items are retained; according to the first n items of the retained results, an optimal vector set is screened from the two vector sets; for any one user vector in the optimal vector set, the similarity between any one user vector in the optimal vector set and all user vectors of another social network is calculated; the first n user vectors are ranked in descending order of similarity, and the first n user vectors are retained; for the first n user vectors, the edit distance of the user name is superimposed as a further constraint, as shown in formula (10), to obtain the final similarity score; whether the users in the two social networks belong to the same natural person is determined based on the final similarity score.

[0144] Specifically, the final similarity score calculation formula is as follows:

[0145] S' = S + γS.name (10)

[0146] Wherein, S' is the final similarity score, S is the cosine similarity between the two user vectors, S.name is the edit distance of the user names corresponding to the two user vectors, and γ is the weight coefficient.

[0147] For example, for the user vector of the i-th user in the first vector set, the similarity between all user vectors of the second social network and the user vector of the i-th user is calculated; and the first n items are retained in descending order of similarity; assuming that n is 50; for the user vector of the k-th user in the second vector set, the similarity between all user vectors of the first social network and the user vector of the k-th user is calculated; and the first n items are retained in descending order of similarity; according to the first n similarity calculation results of the i-th user and the first n similarity calculation results of the k-th user, it is determined which of the first vector set and the second vector set is the optimal vector set. Specifically, the first n similarity calculation results and the corresponding user similarity are selected as the optimal vector set.

[0148] If the first vector set is the optimal vector set, for any one user vector in the first vector set, the similarity between the user vector and the user vector of all unassociated users of the second social network is calculated; the user vectors of the top n unassociated users of the second social network are ranked in descending order of similarity; for the top n user vectors, the edit distance of the user name is superimposed as a further constraint to obtain the final similarity score (i.e. the second similarity calculation result); the final similarity scores of the top n user vectors are ranked in descending order, and the final similarity score of the user vector in the first position is compared with the threshold value δ, if the final similarity score of the user vector in the first position is greater than the threshold value δ (i.e. the similarity score of the most similar user vector exceeds the threshold value δ), it is determined that the user is associated, i.e. the users in the two social networks belong to the same natural person.

[0149] In the embodiment of the present application, after the user alignment method is used to perform identity association between the unassociated users in the first social network and the unassociated users in the second social network based on the first vector set and the second vector set, the method further comprises:

[0150] Converting the unassociated users in the first social network and the second social network that are associated into associated users; adding the corresponding relationship of the usernames of the associated users between different social networks in the basic data of the associated users; updating the first data set and the second data set based on the associated users after adding the corresponding relationship of the usernames between different social networks.

[0151] As shown in Figure 4 In one embodiment of the present application, the identity association method based on SDNE embedding and deep learning can include the following steps:

[0152] Step 1, data preprocessing: collecting the original data set in the real social networks S and T, and performing data preprocessing on the original data set to obtain the S network data set and the T network data set. The S network data set is an adjacency matrix containing all users of the S network; the T network data set is an adjacency matrix containing all users of the T network.

[0153] Step 2, network expansion: updating the user link relationship between the associated users in another social network according to the user link relationship between the associated users in any social network.

[0154] Step 2.1, determine the associated users in the S network data set and the T network data set (i.e. the users with anchor links in the S network and the T network).

[0155] Step 2.2, for any two associated users with a user link relationship, determine whether the two associated users also have a user link relationship in another social network: if not, add a user link relationship between the two associated users in another social network, and obtain an updated user link relationship.

[0156] Step 2.3, based on the obtained updated user link relationship, update the data set corresponding to the social network.

[0157] Step 3, input the updated first data set and the updated second data set (i.e. the adjacency matrix of all users corresponding to two social networks) into the pre-constructed SDNE embedding network, and output the low-dimensional embedding results of all users in the S network and the low-dimensional embedding results of all users in the T network.

[0158] Step 4, train the pre-constructed deep learning identity correlation model, and output the user vectors of all unassociated users using the trained deep learning identity correlation model. The deep learning identity correlation model is composed of an input layer, a hidden layer and an output layer.

[0159] Step 4.1, use the low-dimensional embedding results of all users with anchor links in two social networks (i.e. associated users in the S network and associated users in the T network) to perform supervised deep neural network training on the pre-constructed deep learning identity correlation model.

[0160] Step 4.1.1, use two-thirds of the low-dimensional embedding results of all users with anchor links in two social networks (i.e. associated users in the S network and associated users in the T network) as training data to train the pre-constructed deep learning identity correlation model;

[0161] Step 4.1.2, use the remaining one-third of the low-dimensional embedding results of all users with anchor links in two social networks (i.e. associated users in the S network and associated users in the T network) as test data to test the trained deep learning identity correlation model.

[0162] Step 4.2, use the low-dimensional embedding results of all unassociated users in the S network as input to the tested deep learning identity correlation model, and obtain the user vectors of all unassociated users in the S network; wherein the user vectors of all unassociated users in the S network include x1, x2, …, x i .

[0163] Step 4.3, use the low-dimensional embedding results of all unassociated users in the T network as input to the tested deep learning identity correlation model, and obtain the user vectors of all unassociated users in the T network; wherein the user vectors of all unassociated users in the T network include y1, y2, …, y i.

[0164] Step 5, for the user vector x1 of the unassociated user in the S network, cosine similarity of x1 and user vectors (y1, y2, …, y i ) of all unassociated users in the T network is calculated, and the y user vectors are arranged in descending order according to the calculated cosine similarity; the y user vectors ranked in the top 100 in terms of cosine similarity with the x1 user vector are taken; all unassociated users (x2, …, x i ) in the S network are traversed to obtain the y user vectors ranked in the top 100 in terms of cosine similarity with the user vector of each unassociated user in the S network in the T network.

[0165] Figure 4 In the formula, S1 st is the user vector ranked in the first place (i.e., the y user vector with the highest cosine similarity) among the y user vectors ranked in the top 100 in terms of cosine similarity; S2 nd is the user vector ranked in the second place among the y user vectors ranked in the top 100 in terms of cosine similarity; and S100th is the user vector ranked in the one hundredth place (i.e., the y user vector with the lowest cosine similarity) among the y user vectors ranked in the top 100 in terms of cosine similarity. Figure 4 In the formula, S1 st : x1-y il indicates the user vector y il with the highest cosine similarity with the x1 user vector in the T network; S1 st : x2-y i2 indicates the user vector y i2 of the unassociated user with the highest cosine similarity with the x2 user vector in the T network; S1 st : x n -y in indicates the user vector y n with the highest cosine similarity with the x in user vector in the T network.

[0166] Step 6, the edit distance between the user name corresponding to the x1 user vector and the user name corresponding to the y user vector ranked in the top 100 in terms of cosine similarity with the x1 user vector is calculated by using the final similarity score calculation formula (i.e., formula 10); and the 100 y user vectors are arranged in descending order based on the edit distance.

[0167] Step 7, the user corresponding to the y user vector ranked in the first place is associated with the user corresponding to the x1 user vector to build an anchor link, and the corresponding user names are identity-associated.

[0168] The technical effects of the present application will be further described below in combination with simulation experiments.

[0169] The experimental running environment of the simulation experiment of the present application is Ubuntu 22.04 operating system, the CPU processor is Intel Core i7 13700K, the memory is 64GB DDR5 4800MHz, the software used is Python3.6, tensorflow1.4 and JupyterLab3.

[0170] The simulation experiment data sets of the present application are all from users, user links and anchor links in real social networks. Table 1 below is the basic information of the data sets used in the experiment. Among them, the Twitter-FourSquare data set contains part of the user nodes, corresponding user relationships and part of the aligned user anchor links of two social networks (Twitter and Foursquare). The Lastfm-MySpace data set contains user profiles and undirected links between users in MySpace; the Lastfm data set contains part of the users and the user's attention links. The Livejournal-MySpace data set contains part of the users and the corresponding user links. MySpace contains user profiles and undirected links between users. The data in Table 1 is only contained in the data set, and does not represent the actual number of users in each social network.

[0171] Table 1 Basic information of data sets used in the experiment

[0172]

[0173] In Table 1, when the social network platform is Twitter, the number of user nodes is 5552, and the number of user links is 164920, it means that the number of users in Twitter is 5552, and the number of user links existing in Twitter is 164920; when the number of anchor links corresponding to Twitter and FourSquare is 3148, it means that the number of associated users in Twitter and FourSquare is 3148.

[0174] Since the number of user nodes and user links in the Lastfm-MySpace data set and the Livejournal-MySpace data set is huge, and in order to further evaluate the performance of the algorithm, the present application uses these two pairs of social networks to construct artificial synthetic sub-networks and evaluate the performance of the algorithm. Given a pair of real social networks (for example, Livejournal-MySpace), the present application takes out all the anchor links and randomly selects 50% of them as non-anchor nodes. Therefore, the present application obtains two sub-networks Livejournal' and MySpace', which have ​There are 10 non-anchor nodes, arranged according to alignment. Then, for each node in the Livejournal network... There are several anchor nodes, and a random probability ρ is used to determine whether to add an edge to both networks. If both networks of an anchor node need to have an edge added, another parameter μ is used to control the probability that the added edge will connect to the exact same node in both networks.

[0175] The parameter μ is mainly used to adjust the similarity between Livejournal' and MySpace'. To make the artificially synthesized Livejournal' and MySpace' networks more realistic, some additional nodes and edges are randomly added. The additional nodes are the remaining nodes not included in the Livejournal' and MySpace' networks. The additional edges are added to the Livejournal' and MySpace' networks based on a random selection among the remaining nodes.

[0176] This application uses traditional precision (P), recall (R), and F1-measure values ​​to evaluate the performance of the identity association algorithm. The specific calculation method is as follows:

[0177] Precision P refers to the proportion of all correctly discovered anchor users to the actual number of anchor users, as shown in formula (11):

[0178]

[0179] Recall rate R refers to the proportion of all correctly discovered anchor users to the total number of predicted anchor users, as shown in formula (12):

[0180]

[0181] The F1-measure value is the weighted harmonic mean of precision P and recall R, as shown in formula (13):

[0182]

[0183] Where TP represents the number of correctly predicted anchor links, FP represents the number of incorrectly predicted anchor links, TN represents the number of correctly predicted non-anchor links, and FN represents the number of incorrectly predicted non-anchor links.

[0184] This simulation experiment uses the Deeplink algorithm as the benchmark algorithm and compares it with the identity association method based on SDNE embedding and deep learning proposed in this application. The performance of the algorithm is evaluated using a real social network dataset and two artificially synthesized social network datasets. The experimental results are shown in Table 2 below.

[0185] Table 2 Comparison with baseline algorithm results

[0186]

[0187] Among them, the identity correlation method based on SDNE embedding and deep learning proposed in the present application performs best on the Twitter-FourSquare dataset, with an accuracy rate of more than 30% higher than the benchmark algorithm, followed by the artificially synthesized Lastfm-MySpace dataset, with an accuracy rate of more than 15% higher than the benchmark algorithm, and finally the artificially synthesized Livejournal-MySpace dataset, with an accuracy rate of 13% higher than the benchmark algorithm.

[0188] Further, in order to analyze the reason for the large difference in experimental results that occurs using different datasets, the network expanded dataset is counted, and Table 3 below is the number of edges added to the three datasets used in the experiment after network expansion.

[0189] Table 3 Edge addition statistics

[0190]

[0191]

[0192] According to Table 3, it can be found that the number of edges added after network expansion of the Twitter-FourSquare dataset is much larger than that of the Lastfm-MySpace dataset and the Livejournal-MySpace dataset. First, the reason for this situation is mainly that the Twitter-FourSquare dataset has a large data base originally, and after network expansion, it obtains more abundant topological relationships, while the Lastfm-MySpace dataset and the Livejournal-MySpace dataset are artificially synthesized according to the anchor links of the original Lastfm-MySpace dataset and the Livejournal-MySpace dataset. Since the anchor links of the original data are limited, the size of the artificially synthesized network is also limited, and after network expansion, it does not obtain a large number of potential topological relationships. The size of the data limits the accuracy rate of the identity correlation method based on SDNE embedding and deep learning proposed in the present application, and finally leads to the experimental results being improved compared with the benchmark algorithm, but the effect is slightly worse than that of the real dataset.

[0193] The identity correlation method based on SDNE embedding and deep learning of the present application is compared with the Deeplink algorithm, the e-Deeplink algorithm, and the SDNE-UIL algorithm in identity correlation experiments; in the Deeplink algorithm, the embedding part uses the traditional random walk method, the e-Deeplink is the Deeplink algorithm with a network expansion module added, the SDNE-UIL is a deep learning identity correlation algorithm using SDNE for embedding expression, and the eSUIL represents the identity correlation method based on SDNE embedding and deep learning of the present application. The comparison results are shown in Table 4.

[0194] Table 4 Comparison of ablation experiment results

[0195]

[0196]

[0197] In Table 4, when only a single module is changed, the e-Deeplink has a 3%-5% improvement in P precision and a 3%-6% improvement in F1 value compared with the baseline algorithm Deeplink; the SDNE-UIL has a 3%-7% improvement in P precision and a 3%-6% improvement in F1 value compared with the baseline algorithm Deeplink; and the identity correlation method based on SDNE embedding and deep learning of the present application (the eSUIL algorithm) has a nearly 10% improvement or even a much better effect than adding a single module compared with the baseline algorithm Deeplink. The comparison results show that the network expansion and SDNE embedding modules both help to predict the final identity correlation result, and have a certain effect improvement compared with the baseline algorithm on different data sets. When the two modules are properly fused together, a better effect can be obtained.

[0198] Figure 5 The identity correlation method based on SDNE embedding and deep learning of the present application (i.e., the eSUIL algorithm) is compared with the Deeplink algorithm, the e-Deeplink algorithm, and the SDNE-UIL algorithm in the F1 value comparison (F1@k) between 1-30 rankings and 6 different k values on the Twitter-FourSquare data set. The eSUIL algorithm proposed in the present application achieves the highest accuracy in all rankings k.

[0199] Figure 6The identity correlation method based on SDNE embedding and deep learning (i.e., the eSUIL algorithm) of the present application is compared with the FI values (F1@k) of 6 different k values between 1-30 rankings of the Deeplink algorithm, the e-Deeplink algorithm, and the SDNE-UIL algorithm on the Lastfm-MySpace dataset. The proposed eSUIL algorithm of the present application achieves the highest accuracy in all rankings k.

[0200] Figure 7 The identity correlation method based on SDNE embedding and deep learning (i.e., the eSUIL algorithm) of the present application is compared with the FI values (F1@k) of 6 different k values between 1-30 rankings of the Deeplink algorithm, the e-Deeplink algorithm, and the SDNE-UIL algorithm on the Livejournal-MySpace dataset. The proposed eSUIL algorithm of the present application achieves the highest accuracy in all rankings k.

[0201] Corresponding to the above method, the present application also provides an identity correlation device based on SDNE embedding and deep learning, as shown in Figure 8 The identity correlation device based on SDNE embedding and deep learning includes:

[0202] The construction unit 810 is configured to construct a first data set and a second data set. The first data set includes basic data of all users in a first social network. The second data set includes basic data of all users in a second social network. The all users include: associated users and unassociated users. The basic data includes: a user name and a user link relationship. The associated users are users who have associated the same user's user name in the first social network with the user name in the second social network. The user link relationship is a follow relationship between different users in the same social network.

[0203] The update unit 820 is configured to update the user link relationship between the associated users in the other social network based on the basic data of the associated users in any social network by using a network expansion method, to obtain an updated first data set and an updated second data set.

[0204] The SDNE embedding unit 830 is configured to input the updated first data set and the updated second data set into a pre-constructed SDNE embedding network, respectively, to obtain a first low-dimensional embedding result set of all users in the first social network and a second low-dimensional embedding result set of all users in the second social network.

[0205] The identity association unit 840 is used to input the first low-dimensional embedding result set and the second low-dimensional embedding result set into a pre-built deep learning identity association model to obtain a first vector set and a second vector set; wherein, the first vector set includes user vectors of all unassociated users in the first social network; the second vector set includes user vectors of all unassociated users in the second social network; based on the first vector set and the second vector set, the user alignment method is used to associate the identities of unassociated users in the first social network with those in the second social network.

[0206] The functions of each functional unit in the identity association device based on SDNE embedding and deep learning provided in the above embodiments of this application can be implemented through the above method steps. Therefore, the specific working process and beneficial effects of each unit in the identity association device based on SDNE embedding and deep learning provided in the embodiments of this application will not be repeated here.

[0207] This application also provides an electronic device, such as... Figure 9 As shown, it includes a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940.

[0208] Memory 930 is used to store computer programs;

[0209] When the processor 910 executes the program stored in the memory 930, it performs the following steps:

[0210] Construct a first dataset and a second dataset; wherein, the first dataset includes basic data of all users in the first social network; the second dataset includes basic data of all users in the second social network; all users include: associated users and unassociated users; basic data includes: usernames and user link relationships; associated users are users whose usernames in the first social network and the second social network have been associated; user link relationships are the follow relationships between different users in the same social network;

[0211] Based on the basic data of users already associated in any social network, the user link relationships between users already associated in another social network are updated using the network expansion method, resulting in the updated first dataset and the updated second dataset.

[0212] The updated first dataset and the updated second dataset are respectively input into the pre-built SDNE embedding network to obtain the first low-dimensional embedding result set of all users in the first social network and the second low-dimensional embedding result set of all users in the second social network.

[0213] The first low-dimensional embedding result set and the second low-dimensional embedding result set are input into a pre-constructed deep learning identity correlation model respectively to obtain a first vector set and a second vector set; the first vector set includes user vectors of all uncorrelated users in the first social network; and the second vector set includes user vectors of all uncorrelated users in the second social network.

[0214] Based on the first vector set and the second vector set, a user alignment method is used to perform identity correlation between the uncorrelated users in the first social network and the uncorrelated users in the second social network.

[0215] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0216] The communication interface is used for communication between the electronic device and other devices.

[0217] The memory can include a Random Access Memory (RAM) and can also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0218] The processor described above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; can also be a Digital Signal Processing (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0219] Since the implementation manners and beneficial effects of the electronic device in the above-mentioned embodiments can be achieved by referring to the steps in the embodiments shown in the above-mentioned embodiments, the specific working process and beneficial effects of the electronic device provided by the embodiments of the present application will not be repeated here. Figure 1 The specific working process and beneficial effects of the electronic device provided by the embodiments of the present application will not be repeated here.

[0220] In a further example provided in the present application, a computer readable storage medium is provided, and the computer readable storage medium stores instructions which, when executed on a computer, cause the computer to perform the identity correlation method based on SDNE embedding and deep learning in any of the above examples.

[0221] In a further example provided in the present application, a computer program product containing instructions is provided, and the computer program product, when executed on a computer, causes the computer to perform the identity correlation method based on SDNE embedding and deep learning in any of the above examples.

[0222] Those skilled in the art should understand that the examples in the present application can be provided as a method, a system, or a computer program product. Therefore, the examples in the present application can be in the form of an entirely hardware example, an entirely software example, or an example combining software and hardware aspects. Moreover, the examples in the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0223] The examples in the present application are described with reference to the flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the examples in the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The function of one or more flows and / or blocks Figure 1 The function of one or more flows and / or blocks

[0224] These computer program instructions can also be stored in a computer readable memory capable of directing the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1 The function of one or more flows and / or blocks Figure 1 The function of one or more flows and / or blocks

[0225] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0226] Although preferred embodiments in the application have been described, those skilled in the art will recognize that additional modifications and variations to the preferred embodiments described herein can be made in light of the above teachings. The application therefore, is to be construed as including all such modifications and variations as falling within the scope of the application.

[0227] Obviously, numerous modifications and variations of the preferred embodiments are possible in light of the above teachings. It is therefore to be concluded that the application is intended to cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.​​

Claims

1. An identity association method based on SDNE embedding and deep learning, characterized in that, The method includes: Construct a first dataset and a second dataset; wherein, the first dataset includes basic data of all users in the first social network; the second dataset includes basic data of all users in the second social network; all users include: associated users and unassociated users; the basic data includes: usernames and user link relationships; associated users are users whose usernames in the first social network and the second social network have been associated; the user link relationships are the follow relationships between different users in the same social network; Based on the basic data of users already associated in any social network, the user link relationships between users already associated in another social network are updated using the network expansion method, resulting in the updated first dataset and the updated second dataset. The updated first dataset and the updated second dataset are respectively input into the pre-built SDNE embedding network to obtain the first low-dimensional embedding result set of all users in the first social network and the second low-dimensional embedding result set of all users in the second social network. The first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model to obtain a first vector set and a second vector set; wherein, the first vector set includes user vectors of all unassociated users in the first social network; and the second vector set includes user vectors of all unassociated users in the second social network. Based on the first vector set and the second vector set, the user alignment method is used to associate the identities of unassociated users in the first social network with those in the second social network.

2. The method as described in claim 1, characterized in that, Based on the basic data of users already associated in any social network, the user link relationships between users already associated in another social network are updated using network expansion methods, resulting in an updated first dataset and an updated second dataset, including: For any two linked users in any social network, determine whether the two linked users also have a user link relationship in another social network. If not, add a user link relationship for the two already associated users in the other social network to obtain the updated user link relationship; Based on the updated user link relationships obtained, the dataset corresponding to the social network is updated to obtain the updated first dataset and the updated second dataset.

3. The method as described in claim 1, characterized in that, The first low-dimensional embedding result set and the second low-dimensional embedding result set are respectively input into a pre-constructed deep learning identity association model to obtain a first vector set and a second vector set, including: Extract the low-dimensional embedding results of all associated users in the first social network corresponding to the first low-dimensional embedding result set, and extract the low-dimensional embedding results of all associated users in the second social network corresponding to the second low-dimensional embedding result set; Based on the extracted low-dimensional embedding results, a dataset is constructed; wherein the dataset includes a training dataset and a test dataset; The pre-built deep learning identity association model is trained using the training dataset to obtain the trained deep learning identity association model. The trained deep learning identity association model was tested using the test dataset to obtain the tested deep learning identity association model. The low-dimensional embedding results of all unassociated users in the first social network corresponding to the first low-dimensional embedding result set are input into the deep learning identity association model after testing to obtain the first vector set; The low-dimensional embedding results of all unassociated users in the second social network corresponding to the second low-dimensional embedding result set are input into the deep learning identity association model after testing to obtain the second vector set.

4. The method as described in claim 1, characterized in that, Based on the first vector set and the second vector set, the user alignment method is used to associate the identities of unassociated users in the first social network with unassociated users in the second social network, including: For any user vector in the first vector set, calculate the cosine similarity between the user vector and all user vectors in the second vector set to obtain the first similarity calculation result; Based on the first similarity calculation result, all user vectors in the second vector set are sorted in descending order to obtain the first sorting result; Based on the first ranking result, the top n user vectors in the second vector set are selected; where n represents the number of user vectors selected in the second vector set. Calculate the edit distance between the username corresponding to the user vector in the selected second vector set and the username corresponding to the user vector in the first vector set, and obtain the edit distance calculation result; wherein, the edit distance calculation result is the number of edits required to transform the username corresponding to the user vector in the second vector set into the username corresponding to the user vector in the first vector set; Based on the edit distance calculation result, the usernames corresponding to the user vectors in the selected second vector set are associated with the usernames corresponding to the user vectors in the first vector set.

5. The method as described in claim 4, characterized in that, Based on the edit distance calculation result, the usernames corresponding to the user vectors in the selected second vector set are associated with the usernames corresponding to the user vectors in the first vector set, including: Multiply the edit distance calculation result by the weight coefficient to obtain the weighted edit distance calculation result; The weighted edit distance calculation result is added to the first similarity calculation result to obtain the second similarity calculation result; Based on the second similarity calculation result, the user vectors in the selected second vector set are sorted in descending order to obtain the second sorting result; The user vector that is the first item in the second arrangement result is selected as the user vector that is closest to the user vector in the first vector set. The second similarity calculation result corresponding to the selected closest user vector is compared with a preset threshold. If the second similarity calculation result corresponding to the closest user vector is not less than the preset threshold, then the username corresponding to the closest user vector is associated with the username corresponding to the user vector in the first vector set.

6. The method as described in claim 5, characterized in that, The user corresponding to the closest user vector is an unassociated user in the second social network; the user corresponding to the user vector in the first vector set is an unassociated user in the first social network. Associating the username corresponding to the closest user vector with the username corresponding to the user vector in the first vector set includes: Determine the unassociated users in the second social network corresponding to the closest user vector, and determine the unassociated users in the first social network corresponding to the user vector in the first vector set; Associate the usernames of unassociated users in the second social network with the usernames of unassociated users in the first social network.

7. The method as described in claim 1, characterized in that, The basic data of the associated users also includes: the correspondence between usernames on different social networks; Based on the first vector set and the second vector set, after associating unrelated users in the first social network with unrelated users in the second social network using a user alignment method, the method further includes: Convert unlinked users who have their identities linked between the first and second social networks into linked users; Add the mapping relationship of the usernames of the associated users across different social networks to the basic data of the associated users; The first and second datasets are updated based on the associated users after adding the mapping relationship of usernames across different social networks.

8. An identity association device based on SDNE embedding and deep learning, characterized in that, The device includes: A construction unit is used to construct a first dataset and a second dataset; wherein, the first dataset includes basic data of all users in a first social network; the second dataset includes basic data of all users in a second social network; all users include: associated users and unassociated users; the basic data includes: usernames and user link relationships; associated users are users whose usernames in the first social network and the second social network have been associated; the user link relationships are the follow relationships between different users in the same social network; The update unit is used to update the user link relationships between users already associated in another social network based on the basic data of users already associated in any social network, using the network expansion method, to obtain the updated first dataset and the updated second dataset. The SDNE embedding unit is used to input the updated first dataset and the updated second dataset into the pre-built SDNE embedding network to obtain the first low-dimensional embedding result set of all users in the first social network and the second low-dimensional embedding result set of all users in the second social network. An identity association unit is used to input the first low-dimensional embedding result set and the second low-dimensional embedding result set into a pre-constructed deep learning identity association model to obtain a first vector set and a second vector set; wherein, the first vector set includes user vectors of all unassociated users in the first social network; the second vector set includes user vectors of all unassociated users in the second social network; based on the first vector set and the second vector set, a user alignment method is used to associate the identities of unassociated users in the first social network with those in the second social network.

9. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images