CN111931903A - Network alignment method based on double-layer graph attention neural network - Google Patents

Network alignment method based on double-layer graph attention neural network Download PDF

Info

Publication number
CN111931903A
CN111931903A CN202010654776.0A CN202010654776A CN111931903A CN 111931903 A CN111931903 A CN 111931903A CN 202010654776 A CN202010654776 A CN 202010654776A CN 111931903 A CN111931903 A CN 111931903A
Authority
CN
China
Prior art keywords
user
node
vector
network
social network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010654776.0A
Other languages
Chinese (zh)
Other versions
CN111931903B (en
Inventor
卢美莲
戴银龙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Posts and Telecommunications
Original Assignee
Beijing University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing University of Posts and Telecommunications filed Critical Beijing University of Posts and Telecommunications
Priority to CN202010654776.0A priority Critical patent/CN111931903B/en
Publication of CN111931903A publication Critical patent/CN111931903A/en
Application granted granted Critical
Publication of CN111931903B publication Critical patent/CN111931903B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/01—Social networking
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Business, Economics & Management (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Medical Informatics (AREA)
  • Economics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)

Abstract

The invention provides a network alignment method based on a double-layer graph attention neural network, which comprises two stages of network embedded representation and embedded vector space alignment. In a network embedding representation stage, a double-layer graph attention neural network is proposed to carry out network representation learning so as to extract an embedding vector of a user in a social network; in the embedding vector space alignment stage, a classification model is built by using the obtained embedding vector of the social network user nodes and part of known anchor link sets to predict the anchor links among different social networks, and a bidirectional embedding vector space alignment strategy is provided to meet the one-to-one matching constraint of user entities among different social networks. Through the arrangement, the method can effectively capture different influence weights among the user in the social network, the neighbor users and among various characteristics, so that the accurate representation of the user in the social network is learned, and the accuracy of anchor link prediction among different social networks is improved.

Description

Network alignment method based on double-layer graph attention neural network
Technical Field
The invention relates to the technical field of data mining and machine learning, in particular to a network alignment method based on a double-layer graph attention neural network.
Background
With the rapid development of the internet and mobile devices, online social networks have become an indispensable hot platform for people to share and communicate information. Because different social networking platforms provide different services, a person will typically register accounts on multiple social networking platforms at the same time to meet their different needs. These users shared by different social networking platforms naturally form anchor links connecting different social networks, facilitating information interaction between different social networks. Mining information interaction across multiple social fields can be effectively applied to multiple downstream social network applications such as cross-domain link prediction, cross-domain recommendation, cross-domain information diffusion and the like. However, in today's society, these social networking platforms are typically maintained separately by different companies, with some information isolation from each other. Therefore, aligning accounts belonging to the same user in different social platforms has become an urgent research topic to be solved. Current research on network alignment methods can be mainly divided into two categories: unsupervised-based network alignment methods and supervised-based network alignment methods. Wherein:
(1) the network alignment method based on unsupervised comprises the following steps: unsupervised-based network alignment models attempt to align user accounts between different social networks without known anchor links. In this type of method, researchers typically measure user similarity between different social networks according to the rarity of the user's user name in the social network and the consistency of the neighborhood structure, and then predict the anchor link using a greedy method or a method that minimizes the structural consistency of the two social networks.
(2) The network alignment method based on supervision comprises the following steps: the general idea of a supervised-based network alignment model is to translate the alignment problem between different social networks into a classification problem about anchor links, i.e. to determine whether any two users between different social networks have an anchor link relationship. Early researches constructed classification models by manually extracting certain characteristics of users in social networks, and although the method can solve the alignment problem of partial users in partial social network scenes to a certain extent, the method still has great limitations. Firstly, the method for manually extracting the user features is very complicated, and people usually cannot directly judge which features are effective, and the effective features of different social network scenes are possibly different; secondly, the social network platform often hides partial real information of the user for the purpose of protecting the privacy of the user, so that partial information loss often exists in manual extraction of user features, and the accuracy of the anchor link prediction task is affected.
In recent years, inspired by the widespread success of web presentation learning in a single social network analysis task, some researchers began to apply web presentation learning to a network alignment task between multiple social networks, and this type of approach attempted to learn a common embedded vector space for users in different social networks without manually extracting the effective features of the users in the social networks. These methods, while attempting to model a user's behavior in a social network from multiple aspects of the user's social structure and profile information, ignore the differences in the impact weights of their different neighboring user nodes when capturing their representations of user nodes, and the different impact weights of different attribute information on user information interactions.
In summary, in view of the importance of network alignment research on multiple social network analysis tasks and some limitations of existing research, the present invention aims to provide a network alignment method based on a two-layer graph attention neural network, which enables a model to learn an accurate representation of a user in a social network by combining a user-level attention mechanism and a feature-level attention mechanism, thereby improving the prediction accuracy of anchor links.
Disclosure of Invention
In view of this, the present invention provides a network alignment method based on a two-layer graph attention neural network, which can effectively capture different influence weights between a user and neighboring users and between various features in a social network while modeling a user social behavior by using information such as an attribute, a local social structure, and a global social structure of the user in the social network, so as to learn an accurate representation of the user in the social network, and improve accuracy of anchor link prediction between different social networks.
Based on the above object, the present invention provides a network alignment method based on a two-layer graph attention neural network, which is characterized by comprising:
basic definition: the social network abstraction is a directed graph G ═ (V, E, X), where V ═ V i1, …, N represents a set of user nodes in the social network, and N is the number of user nodes in the social network; e ═ Ei,j＝(vi,vj)|vi∈V,vje.V represents the set of relationships between users in a social network, ei,j＝(vi,vj) Representing a user viAnd user vjThere is an association between them; x ═ XiI 1, …, N represents a set of feature vectors for all users, for each user node viAll have a node feature vector xiCorrespondingly, the feature vector can be extracted from the aspects of personal data, behaviors and network social structure information of the user node, and the two networks to be aligned are named as a source social network and a target social network without loss of generality and are respectively named as GsAnd GtRepresents;
for any two users from different social networks
Figure BDA0002576285310000031
And
Figure BDA0002576285310000032
we use
Figure BDA0002576285310000033
Representing an anchor link relationship between the source social network and the target social network, wherein
Figure BDA0002576285310000034
And
Figure BDA0002576285310000035
the same user is respectively in different social networks GsAnd GtAccount (2) of (1); anchor links are one-to-one between two users in different social networksThe link relation is that the condition that two anchor links share the same user account of the same social network does not exist;
two different social networks GsAnd GtThe set formed by all the anchor link relations between the anchor links is called as an anchor link set and is used
Figure BDA0002576285310000036
Is shown in which
Figure BDA0002576285310000037
Representing a user account in the source social network,
Figure BDA0002576285310000038
representing a user account in a target social network; for two different social networks Gs＝(Vs,Es,Xs) And Gt＝(Vt,Et,Xt) Network alignment aims to discover a set of anchor-linked sets T between two social networks, where any element e 'in the set T'ijE T denotes two user accounts in different social networks
Figure BDA0002576285310000039
And
Figure BDA00025762853100000310
an anchor link between;
s1, a network preprocessing module: preprocessing a social network according to the input network type and the contained user attribute information, and constructing an initialized user node characteristic vector matrix;
s2, network embedded representation module: taking an initialized user node characteristic vector matrix obtained by a network preprocessing module and an adjacency matrix of a social network as input, and capturing a complex information interaction relation of a user in the social network through a double-layer graph attention neural network so as to learn potential information of user nodes in the social network and obtain an accurate user node embedded vector;
s3, embedding vector space alignment module: constructing a classification model according to the user node embedded vectors of the source social network and the target social network learned in the step S2 to predict anchor links, and adopting a bidirectional embedded vector space alignment strategy to meet the one-to-one matching constraint of user accounts among different social networks;
and S4, solving the intersection to complete network alignment.
Preferably, the step S2 includes the following: user viIs expressed as xiExtracting a plurality of feature vectors of the user according to the network type and transversely stacking the feature vectors to generate an initialized feature vector representation of the user
Figure BDA00025762853100000311
Wherein d represents the dimension of the user initialized feature vector, d ', d ", d'" appearing hereinafter represent different dimensions respectively, and the initialized feature vectors of all users in the social network are constructed into a state matrix X, wherein each row is the feature vector of a specific user node: x ═ X1,x2,…xN)T。
Preferably, the network type is a topological network, the feature vectors of the user are initialized randomly in a random matrix mode, and the weight parameters of the random matrix are learned through a training stage of a double-layer graph attention neural network model.
Preferably, if the network type is an attribute network, vectorizing the user attribute by using the following method: randomly initializing user data information such as a user name and the like in a word embedding mode to obtain a user name characteristic vector; mining the language style of the user from the long text information of the user by adopting a Doc2Vec model, and learning the text characteristic vector of the user; vector initialization is carried out on the user track information through spatial clustering to obtain a spatial feature vector of a user; and directly taking the user scoring and sign-in times as a characteristic dimension of the user, and carrying out vector initialization.
Preferably, the step S2 includes the following:
s2.1, embedding a user layer node into a representation submodule: the system is responsible for capturing different influence weights among users to perform weighted aggregation on local neighborhood information of the users in the social network, so that node embedded vectors of user levels are learned;
s2.2, embedding the feature layer nodes into a representation submodule: the method is used for learning the influence weight among different features of a user so as to capture the interactive relation among the features with finer granularity, thereby learning the node embedding vector of the user at the feature level;
s2.3, embedding a vector fusion submodule: the method is responsible for reserving and resetting user embedded vectors from different levels of user levels and feature levels so as to fuse node embedded vectors of multiple visual angles and improve the accuracy of network alignment tasks.
Preferably, S2.1 includes the following: using a learnable transformation matrix
Figure BDA0002576285310000041
Converting the input vector into a high-dimensional vector, namely:
Figure BDA0002576285310000042
according to the related knowledge of the attention neural network of the graph, for any two user nodes viAnd vjFirst, the relationship strength e between the two user nodes is calculatedij：
Figure BDA0002576285310000043
Wherein
Figure BDA0002576285310000044
And
Figure BDA0002576285310000045
representing a user node viAnd vjThe embedded vectors at the l-th layer,
Figure BDA0002576285310000046
represents the weight of the l-th layerThe parameter, "| |" is a concatenation operator, which represents that two vectors are transversely spliced, LeakyReLU (·) is an activation function of a neuron, and a user node v is calculatediWhen the neighborhood information is aggregated, the information contribution proportion from different neighbors adopts a softmax (·) function to normalize the user node and all the neighbor user nodes vk∈N(vi) The strength of the relationship between the two is calculated as follows:
Figure BDA0002576285310000051
aijis called user node viAnd vjAttention coefficient of (a) in betweenijThe larger the value of (v), the more closely the relationship between the two users is represented, and the user node v is obtained according to calculationiEach user node v attention coefficient with all its neighbor node users (including itself)iThe new embedded vector of (2) can be defined as follows:
Figure BDA0002576285310000052
wherein (-) is an activation function of a neuron, and a user-level node embedding vector h of each user in the social network can be obtained by performing linear aggregation of different influence weights on neighborhood information of the useriThe user-level vector matrix M ═ (h) is formed1,h2,…,hN)T。
Preferably, S2.2 includes the following: the user-level vector matrix M obtained in S2.1 is (h)1,h2,…,hN)TAs input, and taking into account the multidimensional attention between the features of any two user nodes in the social network, i.e. calculating an attention coefficient for each corresponding dimension of the two user node vectors,
suppose hiAnd hjDistribution represents two user nodes v in a social networkiAnd vjThe node of (2) is embedded into the vector, then two usersThe relationship between node-embedded vectors can be defined as:
f(hi,hj)＝W5·tanh(W4·hi+W3·hj+b2)+b1
wherein W3，W4And
Figure BDA0002576285310000053
is a parameter matrix, b1，
Figure BDA0002576285310000054
Is a bias term, tanh (-) is an activation function of a neuron,
using a feed-forward neural network to operate according to f (h)i,hj) Calculating the dependency relationship of any two user nodes on the characteristic level, and enabling betaijRepresenting a user node viAnd vjThe attention coefficient vector of the feature layer of [ beta ]ij]kSimilarly, in order to compare the attention coefficients of the features of the corresponding dimensions between different attention coefficient vectors, the attention coefficient vectors of all the neighbors of the user are normalized according to the corresponding feature dimensions by using the softmax (·) function, and then:
Figure BDA0002576285310000061
calculating attention coefficient [ beta ] between each dimension of any two usersij]kThese attention coefficients can then be combined into an attention coefficient vector β between the two users according to the corresponding feature dimensionsij＝([βij]1,[βij]2,…,[βij]d″) The dimensions of the attention vector and the user node vector are the same, each dimension [ β [ ]ij]kCorresponding to the weight of influence of each dimension of the user node vector, [ beta ]ij]kThe larger, the two user nodes v are representediAnd vjThe stronger the correlation degree of the characteristics of the kth dimension is, finally, for any user node v in the social networkiThe method is characterized in that the embedded vectors of the neighbor users are weighted and linearly aggregated according to the learned attention coefficient vectors among different user nodes, and the method is different from the aggregation mode of the user layer attention mechanism, wherein the aggregation function of the feature layer attention mechanism is to aggregate neighborhood information by adopting a mode of element-by-element multiplication:
Figure BDA0002576285310000062
wherein
Figure BDA0002576285310000068
Representing the element-by-element product between two vectors with the same shape to obtain a vector with the same shape, wherein the activation function of the neuron linearly aggregates the neighborhood information of the users according to the influence weights of different features to obtain the node embedded representation of the feature level of each user in the social network
Figure BDA0002576285310000063
Forming a feature level vector matrix
Figure BDA0002576285310000064
Figure BDA0002576285310000065
Preferably, S2.3 includes the following: embedding vector fusion submodule changes user level vector matrix M into (h)1,h2,…,hN)TAnd feature level vector matrix
Figure BDA0002576285310000066
As input, the weight parameters of node embedding vectors of different levels of the same user are learned automatically by using a gating mechanism to effectively retain and reset information representations from different levels,
for any user node v in social networkiThe module first calculates a user-level node embedding vector hiAnd feature level node embedded vectors
Figure BDA0002576285310000067
The weight relationship vector between them is calculated as follows:
Figure BDA0002576285310000071
wherein
Figure BDA0002576285310000072
And
Figure BDA0002576285310000073
is a parameter matrix of the gated neural network,
Figure BDA0002576285310000074
is a bias term, sigmoid (·) is an activation function of a neuron, and according to learning a weight relation vector F, user node embedded vectors from different layers can be selectively retained and reset, and the final user node embedded vector is expressed as follows:
Figure BDA0002576285310000075
wherein
Figure BDA0002576285310000076
Representing selective retention of node-embedded vectors at the user level,
Figure BDA0002576285310000077
represents the selective resetting of node-embedded vectors to the feature level, where 1-F is a vector operation, representing subtracting 1 from each dimension of the vector F,
embedding according to user level nodes of users in social networkThe in representation and the feature level node embedded representation are fused by using a gating mechanism, and the final node embedded representation z of each user in the social network can be obtainediThe component node embedding vector matrix Z ═ (Z)1,z2,…,zN)T；
For any given pair of user nodes v in a social networkiAnd vjThe node embedding vectors are respectively ziAnd zjThen the probability of an edge existing between the two nodes can be expressed as:
Figure BDA0002576285310000078
wherein
Figure BDA0002576285310000079
Is a sigmoid function;
in order to optimize the model parameters of the two-layer graph attention neural network, we need to define an objective function of the model whose goal is to maximize the probability of observable edge occurrences in the social network, i.e.:
Figure BDA00025762853100000710
to avoid trivial solution, for each edge e that can be observedi,j＝(vi,vj) We use a negative sampling technique to maximize the objective function, namely:
Figure BDA00025762853100000711
the first item models positive examples in the social network, the second item models negative examples by randomly generating edges associated with nodes through a negative sampling technology, and the sampled probability of each node satisfies
Figure BDA0002576285310000081
K denotes the number of edges of the negative example of the sample, dvRepresenting degrees of user node v; according to the objective function, parameters of the double-layer graph attention neural network model can be learned by adopting a back propagation optimization algorithm, so that node vector matrixes of the source social network and the target social network are obtained respectively
Figure BDA0002576285310000082
Figure BDA0002576285310000083
And
Figure BDA0002576285310000084
wherein | VsI represents the number of user nodes in the source social network, | VtI represents the number of user nodes in the target social network;
Figure BDA0002576285310000085
representing users in a source social network
Figure BDA0002576285310000086
The node of (a) embeds the vector,
Figure BDA0002576285310000087
representing users in a target social network
Figure BDA0002576285310000088
The node of (1) embeds a vector; zsAnd ZtAlso referred to as the embedded vector space for the source social network and the target social network.
Preferably, the step S3 includes the following: based on the step of S2, obtaining a node embedded vector matrix of the source social network
Figure BDA0002576285310000089
Embedding vector matrix with nodes of target social network
Figure BDA00025762853100000810
Figure BDA00025762853100000811
Each row of the node vector matrix represents a node embedded vector corresponding to one user in the social network, and the whole node vector matrix is also called an embedded vector space corresponding to the social network;
defining a mapping function M that maps user node vectors from one embedded vector space to another, assuming we now project the source social network to the target social network to find a target node that matches the source social network, given a partially known set of anchor links T as the supervisory information, the target function can be defined as:
Figure BDA00025762853100000812
wherein M iss→t() represents a mapping function from a source social network to a target social network, the mapping function being constructed using a multi-tiered perceptron, θ being a weighting parameter for the multi-tiered perceptron; the objective function aims to minimize the distance from a source user node to a target user node after mapping the source user node in a user pair with an anchor link relation to the target social network, construct a classification model to predict whether any two users between different social networks have an anchor link, and select the target user node closest to the source user node after projection to construct a candidate anchor link.
From the above, it can be seen that the complex interaction behavior of the user in the social network is modeled by using the graph neural network and the information of the user in the social network, such as the attributes, the local social structure and the global social structure, so that a more accurate user node embedding vector can be obtained;
the invention provides a double-layer attention neural network to learn attention coefficients between users at a user level and a characteristic level respectively, and capture differences of influence weights between different users from multiple visual angles, so that the learned user node embedded representation is more in line with the actual situation of the users in a social network;
the invention provides a bidirectional embedded vector space alignment strategy to predict anchor links among different social networks, so that users among different social networks are aligned and meet a one-to-one matching constraint relationship. Meanwhile, further confirmation of the bidirectional embedded vector space alignment strategy also improves the accuracy of anchor link prediction.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a method framework diagram of the present invention;
FIG. 2 is a fused representation of user node embedded representations from different perspectives of the present invention;
FIG. 3 is a schematic diagram of the spatial alignment strategy of the bi-directional embedded vector of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to specific embodiments and the accompanying drawings.
It is to be noted that technical terms or scientific terms used in the embodiments of the present invention should have the ordinary meanings as understood by those having ordinary skill in the art to which the present disclosure belongs, unless otherwise defined. The use of "first," "second," and similar terms in this disclosure is not intended to indicate any order, quantity, or importance, but rather is used to distinguish one element from another. The word "comprising" or "comprises", and the like, means that the element or item listed before the word covers the element or item listed after the word and its equivalents, but does not exclude other elements or items. The terms "connected" or "coupled" and the like are not restricted to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "upper", "lower", "left", "right", and the like are used merely to indicate relative positional relationships, and when the absolute position of the object being described is changed, the relative positional relationships may also be changed accordingly.
As shown in fig. 1 to 3, the embodiment:
first, the present invention describes the social network alignment problem as follows:
the social network abstraction is a directed graph G ═ (V, E, X), where V ═ V i1, …, N represents a set of user nodes in the social network, and N is the number of user nodes in the social network; e ═ Ei,j＝(vi,vj)|vi∈V,vje.V represents the set of relationships between users in a social network, ei,j＝(vi,vj) Representing a user viAnd user vjThere is an association between them; x ═ XiI 1, …, N represents a set of feature vectors for all users, for each user node viAll have a node feature vector xiCorrespondingly, the feature vector can be extracted from the aspects of personal data, behaviors, structural attributes and the like of the user node. For each edge e in the relationship seti,jLet us give wi,jRepresenting the weight of the edge. Wherein if two users have a link relationship in the social network, w i,j1, otherwise, wi,j0. Matrix array
Figure BDA0002576285310000101
Referred to as the adjacency matrix of graph G. Without loss of generality, we name the two networks to be aligned as the source social network and the target social network, and use G respectivelysAnd GtAnd (4) performing representation.
For any two users from different social networks
Figure BDA0002576285310000102
And
Figure BDA0002576285310000103
we use
Figure BDA0002576285310000104
Represents the association relationship, w'i,jRepresenting the relationship weight. If it is not
Figure BDA0002576285310000105
And
Figure BDA0002576285310000106
are two different accounts of the same user in the source and target social networks, respectively, we then let w'i,j1 denotes that there is an anchor link relationship between the two users, otherwise, w'i,j0. The goal of network alignment is to find a set of anchor links between different social networks
Figure BDA0002576285310000107
Figure BDA0002576285310000108
Wherein
Figure BDA0002576285310000109
Representing user nodes
Figure BDA00025762853100001010
From the source social network, the user may select,
Figure BDA00025762853100001011
representing user nodes
Figure BDA00025762853100001012
From target social network, w'i,jTwo user nodes are denoted by 1
Figure BDA00025762853100001013
And
Figure BDA00025762853100001014
belonging to the same user entity in the real world.
Secondly, referring to fig. 1, the present invention provides a network alignment method framework, which mainly comprises a network preprocessing module, a network embedded representation module and an embedded vector space alignment module. Wherein
A network preprocessing module: for the social network G ═ (V, E), we first perform preprocessing work on the network according to the type of the input network and the contained information, thereby constructing an initialized user vector matrix. Common network types can be divided into a topological network and an attribute network, and for the topological network, the invention adopts an embedding layer to randomly initialize a node embedding vector x of a useriLearning the weight parameters of the embedding layer through a training stage of the double-layer graph attention neural network model; for the attribute network, the user attribute is usually vectorized in different ways according to the contained information. For short text attributes similar to the user name, an embedding layer can be adopted for random initialization; for the long text attribute similar to the user comment, a theme model and other methods are usually adopted to learn the theme preference of the user; regarding the check-in information of the user, the access preference of the user to merchants in the same area is considered to be similar, and the check-in information of the user is initialized by adopting a spatial clustering method; the numerical attributes such as the user scoring times and the check-in times can be directly used as one dimension of the user attributes. After extracting the attribute vectors of the various attributes of the user, we stack these attribute vectors horizontally to generate the initialization vector of the user in the attribute network. Thus, for each user in the social network, we ultimately generate an initialized representation of the user
Figure BDA0002576285310000111
Where d represents the dimension of the user initialization vector. Constructing a state matrix X from the initialization vectors of all users in the social network, wherein each row is a feature vector of a specific user node: x ═ X1,x2,…xN)T。
Network embedded representation module: the module sets the initialized user node characteristic vector matrix X (X) obtained by the network preprocessing module as (X)1,x2,…xN)TAdjacency matrix with social network
Figure BDA0002576285310000114
As an input, capturing complex information interaction relation of a user in a social network through a two-layer graph attention machine neural network to learn potential information representation z of a user node in the social networki. The module may be further subdivided into three sub-modules: the system comprises a user layer node embedded expression submodule, a characteristic layer node embedded expression submodule and an embedded vector fusion submodule.
1. User layer node embedded representation submodule
The user layer node embedded representation submodule is responsible for capturing different influence weights among users to conduct weighted aggregation on local neighborhood information of the users in the social network, and therefore node embedded representation of the user level is learned. In order to ensure that the node vector has enough information representation capability, the invention firstly utilizes a learnable transformation matrix
Figure BDA0002576285310000112
Converting the input vector into a high-dimensional vector, namely:
Figure BDA0002576285310000113
for any two user nodes viAnd vjFirst, the strength e of the relationship between the two users is calculatedij：
Figure BDA0002576285310000121
Wherein
Figure BDA0002576285310000122
And
Figure BDA0002576285310000123
representing a user node viAnd vjIn the vector representation at the l-th layer,
Figure BDA0002576285310000124
representing the weight parameter of the l-th layer, "|" is a concatenation operator representing the horizontal concatenation of two vectors, and LeakyReLU (·) is the activation function of neurons. For calculating the user node viWhen the neighborhood information is aggregated, the information contribution proportion from different neighbors adopts the softmax (·) function to carry out the aggregation on the user node and all the neighbor user nodes vk∈N(vi) The relationship strength of (2) is normalized, and the calculation mode is as follows:
Figure BDA0002576285310000125
aijis called a user node viAnd vjAttention coefficient of (a) in betweenijThe larger the value of (a), the more closely the relationship between the two users is represented. According to the user node v obtained by calculationiEach user node v attention coefficient with all its neighbor node users (including itself)iThe new potential information representation of (a) may be defined as follows:
Figure BDA0002576285310000126
where (-) is the activation function of the neuron. According to the linear aggregation of different influence weights on the neighborhood information of the users, the user-level node embedded representation h of each user in the social network can be obtainediThe user-level vector matrix M ═ (h) is formed1,h2,…,hN)T。
2. Feature level node embedded representation submodule
The feature layer node embedded representation submodule is responsible for learning influence weights among different features of a user so as to capture an interactive relation among features with finer granularity, and therefore node embedded representation of the user at a feature level is learned. Feature level node embedded representationThe user level vector matrix M obtained in the previous stage of the module is (h)1,h2,…,hN)TAs input, and taking into account the multidimensional attention of features between any two nodes in the social network, an attention coefficient is calculated for each corresponding dimension of the two user node vectors.
Suppose hiAnd hjDistribution represents two user nodes v in a social networkiAnd vjThe relationship between two user node embedding vectors can be defined as:
f(hi,hj)＝W5·tanh(W4·hi+W3·hj+b2)+b1
wherein W3，W4And
Figure BDA0002576285310000131
is a parameter matrix, b1，
Figure BDA0002576285310000132
Is a bias term, and tanh (-) is the activation function of a neuron. Using a feed-forward neural network to operate according to f (h)i,hj) Calculating the dependency relationship of any two user nodes on the characteristic level. Let betaijRepresenting a user node viAnd vjThe attention coefficient vector of the feature layer of [ beta ]ij]kRepresenting the k-th dimension of the attention vector. Similarly, in order to compare the attention coefficients of the features of the corresponding dimensions between different attention coefficient vectors, the attention coefficient vectors of all the neighbors of the user are normalized according to the corresponding feature dimensions by using the softmax (·) function, and then:
Figure BDA0002576285310000133
calculating attention coefficient [ beta ] between each dimension of any two usersij]kThese attention coefficients can then be usedCombining the attention coefficient vectors beta between the two users according to the corresponding feature dimensionsij＝([βij]1,[βij]2,…,[βij]d). The dimensions of the attention vector and the user node vector are the same, each dimension [ β ]ij]kCorresponding to the weight of influence of each dimension of the user node vector, [ beta ]ij]kThe larger, the two user nodes v are representediAnd vjThe stronger the correlation of the features of the k-th dimension (c).
Finally, for any user node v in the social networkiAnd carrying out weighted linear aggregation on potential information representations of neighbor users according to the learned attention coefficient vectors among different user nodes. Different from the aggregation mode of the user layer attention mechanism, the aggregation function of the feature layer attention mechanism aggregates the neighborhood information in an element-by-element multiplication mode:
Figure BDA0002576285310000134
wherein
Figure BDA0002576285310000135
Represents the element-wise product between two vectors of the same shape, resulting in a vector of the same shape, (-) being the activation function of the neuron. According to the linear aggregation of the neighborhood information of the users according to the influence weights of different features, the node embedded representation of the feature level of each user in the social network can be obtained
Figure BDA0002576285310000141
Forming a feature level vector matrix
Figure BDA0002576285310000142
Figure BDA0002576285310000143
3. Embedded vector fusion submodule
The embedded vector fusion submodule is responsible for reserving and resetting user potential information from different levels of a user level and a characteristic level so as to fuse node embedded representation of multiple visual angles and improve the accuracy of a subsequent network alignment task. Embedding vector fusion submodule changes user level vector matrix M into (h)1,h2,…,hN)TAnd feature level vector matrix
Figure BDA0002576285310000144
As input, the weight parameters of different levels of node-embedded representations of the same user are learned automatically using a gating mechanism to efficiently retain and reset information representations from different levels.
As shown in FIG. 2, for any one user node v in the social networkiThe module first computes a user-level node-embedded representation hiAnd feature level node embedding representation
Figure BDA0002576285310000145
The weight relationship vector between them is calculated as follows:
Figure BDA0002576285310000146
wherein
Figure BDA0002576285310000147
And
Figure BDA0002576285310000148
is a parameter matrix of the gated neural network,
Figure BDA0002576285310000149
Figure BDA00025762853100001410
is a bias term, sigmoid (·) is the activation function of neurons. According to the learned weight relation vector F, the selection can be madeThe potential information of users from different layers is represented to be reserved and reset, and the final user node embedding is represented as follows:
Figure BDA00025762853100001411
wherein
Figure BDA00025762853100001412
Represents the selective retention of node-embedded representations at the user level,
Figure BDA00025762853100001413
represents the selective resetting of the node-embedded representation of the feature level, where 1-F is a vector operation, representing subtracting 1 from each dimension of the vector F.
According to the method, the user level node embedded representation and the characteristic level node embedded representation of the user in the social network are fused by using a gating mechanism, and the final node embedded representation z of each user in the social network can be obtainediThe component node embedding vector matrix Z ═ (Z)1,z2,…,zn)T。
For any given pair of user nodes v in a social networkiAnd vjThe node embedding vectors are respectively ziAnd zjThen the probability of an edge existing between the two nodes can be expressed as:
Figure BDA0002576285310000151
wherein
Figure BDA0002576285310000152
Is a sigmoid function;
in order to optimize the model parameters of the two-layer graph attention neural network, we need to define an objective function of the model whose goal is to maximize the probability of observable edge occurrences in the social network, i.e.:
Figure BDA0002576285310000153
to avoid trivial solution, for each edge e that can be observedi,j＝(vi,vj) We use a negative sampling technique to maximize the objective function, namely:
Figure BDA0002576285310000154
the first item models positive examples in the social network, the second item models negative examples by randomly generating edges associated with nodes through a negative sampling technology, and the sampled probability of each node satisfies
Figure BDA0002576285310000155
K denotes the number of edges of the negative example of the sample, dvRepresenting degrees of user node v; according to the objective function, parameters of the double-layer graph attention neural network model can be learned by adopting a back propagation optimization algorithm, so that node vector matrixes of the source social network and the target social network are obtained respectively
Figure BDA0002576285310000156
Figure BDA0002576285310000157
And
Figure BDA0002576285310000158
wherein | VsI represents the number of user nodes in the source social network, | VtI represents the number of user nodes in the target social network;
Figure BDA0002576285310000159
representing users in a source social network
Figure BDA00025762853100001510
The node of (a) embeds the vector,
Figure BDA00025762853100001511
representing users in a target social network
Figure BDA00025762853100001512
The node of (1) embeds a vector; zsAnd ZtAlso referred to as the embedded vector space for the source social network and the target social network.
An embedded vector space alignment module: based on the modules, node embedded vector matrixes of source social networks can be obtained respectively
Figure BDA0002576285310000161
Embedding vector matrix with nodes of target social network
Figure BDA0002576285310000162
Figure BDA0002576285310000163
Each row of the node vector matrix represents a node embedded representation corresponding to one user in the social network, and the whole node vector matrix is called an embedded vector space corresponding to the social network. In order to allow the two embedded vector spaces to be aligned effectively, we need to project the embedded vector space of the source social network and the embedded vector space of the target social network to a common vector space.
First, we define a mapping function M that maps the user node vector from one embedded vector space to another. Suppose we now project the source social network to the target social network to find a target node that matches the source social network. Given a partially known set of anchor links T as supervisory information, the objective function can be defined as:
Figure BDA0002576285310000164
wherein M iss→t() representing a transition from a source social network to a target social networkAnd mapping function, the invention adopts multilayer perceptron to construct mapping function, and theta is weight parameter of multilayer perceptron. The objective function is to minimize the distance from the target user node after mapping the source user node in the pair of users with anchor link relationship to the target social network, so as to construct a classification model to predict whether any two users between different social networks have anchor links. Since users typically have only one active account in different social networking platforms, target user nodes closest to the source user node after the projection are selected herein to construct candidate anchor links.
Since user alignment issues between different social networks typically satisfy one-to-one matching constraints, i.e., the same user entity, there is at most only one active account in different social network platforms. Whereas, as shown in fig. 3(a), a one-way embedded vector space mapping may result in a one-to-many matching relationship between social networks, which is contrary to the actual network scenario. Therefore, the invention provides a bidirectional embedded vector space alignment strategy to ensure that a network alignment task between two social networks meets a one-to-one matching constraint relationship. Referring to fig. 3, the present invention will describe a specific implementation flow of the module with reference to a specific example, and the steps are as follows:
step 1: constructing a multi-layered perceptron model projected from a source social network to a target social network based on a known anchor link set T
Figure BDA0002576285310000171
Learning a weight parameter θ by minimizing a distance from a source user node in an anchor link to a corresponding target user node after projection to a target social network1。
Step 2: based on learned multi-layer perceptron model
Figure BDA0002576285310000172
For each user node in the source social network
Figure BDA0002576285310000173
As shown in FIG. 3(a), first, the following steps are carried outIt projects to the target embedded vector space, then finds the target user node closest to the projected node in the target social network to form an anchor link with the source user node, and adds it to the candidate anchor link set
Figure BDA0002576285310000174
Figure BDA0002576285310000175
And step 3: constructing a multi-layered perceptron model projected from a target social network to a source social network based on a known anchor link set T
Figure BDA0002576285310000176
Learning a weight parameter θ by minimizing a distance from a target user node in an anchor link to a corresponding source user node after projection to the source social network2。
And 4, step 4: based on learned multi-layer perceptron model
Figure BDA0002576285310000177
For each user node in a target social network
Figure BDA0002576285310000178
As shown in FIG. 3(b), it is first projected into the source embedded vector space, then the source user node closest to the node it is projected is found in the source social network to form an anchor link with the target user node, and it is added to the set of candidate anchor links
Figure BDA0002576285310000179
Figure BDA00025762853100001710
And 5: taking a candidate anchor link set A1And A2Is used as the final predicted anchor link set A ═ A1∩A2。
Those of ordinary skill in the art will understand that: the discussion of any embodiment above is meant to be exemplary only, and is not intended to intimate that the scope of the disclosure, including the claims, is limited to these examples; within the idea of the invention, also features in the above embodiments or in different embodiments may be combined, steps may be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.
In addition, well known power/ground connections to Integrated Circuit (IC) chips and other components may or may not be shown within the provided figures for simplicity of illustration and discussion, and so as not to obscure the invention. Furthermore, devices may be shown in block diagram form in order to avoid obscuring the invention, and also in view of the fact that specifics with respect to implementation of such block diagram devices are highly dependent upon the platform within which the present invention is to be implemented (i.e., specifics should be well within purview of one skilled in the art). Where specific details (e.g., circuits) are set forth in order to describe example embodiments of the invention, it should be apparent to one skilled in the art that the invention can be practiced without, or with variation of, these specific details. Accordingly, the description is to be regarded as illustrative instead of restrictive.
While the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art in light of the foregoing description. For example, other memory architectures, such as Dynamic RAM (DRAM), may use the discussed embodiments.
The embodiments of the invention are intended to embrace all such alternatives, modifications and variances that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, substitutions, improvements and the like that may be made without departing from the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims (9)

1. A network alignment method based on a double-layer graph attention neural network is characterized by comprising the following steps:
basic definition: the social network abstraction is a directed graph G ═ (V, E, X), where V ═ Vi1, N represents a set of user nodes in the social network, and N is the number of user nodes in the social network; e ═ Ei，j＝(vi，vj)|vi∈V，vje.V represents the set of relationships between users in a social network, ei，j＝(vi，vj) Representing a user viAnd user vjThere is an association between them; x ═ XiI 1.. N } represents a set of feature vectors for all users, and v for each user nodeiAll have a node feature vector xiCorrespondingly, the feature vector can be extracted from the aspects of personal data, behaviors and network social structure information of the user node, and the two networks to be aligned are named as a source social network and a target social network without loss of generality and are respectively named as GsAnd GtRepresents;
for any two users from different social networks
Figure FDA0002576285300000011
And
Figure FDA0002576285300000012
we use
Figure FDA0002576285300000013
Representing an anchor link relationship between the source social network and the target social network, wherein
Figure FDA0002576285300000014
And
Figure FDA0002576285300000015
the same user is respectively in different social networks GsAnd GtAccount (2) of (1); the anchor link is two users in different social networksThe two anchor links share the same user account of the same social network;
two different social networks GsAnd GtThe set formed by all the anchor link relations between the anchor links is called as an anchor link set and is used
Figure FDA0002576285300000016
Is shown in which
Figure FDA0002576285300000017
Representing a user account in the source social network,
Figure FDA0002576285300000018
representing a user account in a target social network; for two different social networks Gs＝(Vs，Es，Xs) And Gt＝(Vt，Et，Xt) Network alignment aims to discover a set of anchor-linked sets T between two social networks, where any element e 'in the set T'ijE T denotes two user accounts in different social networks
Figure FDA0002576285300000019
And
Figure FDA00025762853000000110
an anchor link between;
s1, a network preprocessing module: preprocessing a social network according to the input network type and the contained user attribute information, and constructing an initialized user node characteristic vector matrix;
s2, network embedded representation module: taking an initialized user node characteristic vector matrix obtained by a network preprocessing module and an adjacency matrix of a social network as input, and capturing a complex information interaction relation of a user in the social network through a double-layer graph attention neural network so as to learn potential information of user nodes in the social network and obtain an accurate user node embedded vector;
s3, embedding vector space alignment module: constructing a classification model according to the user node embedded vectors of the source social network and the target social network learned in the step S2 to predict anchor links, and adopting a bidirectional embedded vector space alignment strategy to meet the one-to-one matching constraint of user accounts among different social networks;
and S4, solving the intersection to complete network alignment.
2. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 1, wherein the step of S2 includes the following steps: user viIs expressed as xiExtracting a plurality of feature vectors of the user according to the network type and transversely stacking the feature vectors to generate an initialized feature vector representation of the user
Figure FDA0002576285300000021
Wherein d represents the dimension of the user initialized feature vector, d ', d ", d'" appearing hereinafter represent different dimensions respectively, and the initialized feature vectors of all users in the social network are constructed into a state matrix X, wherein each row is the feature vector of a specific user node: x ═ X1，x2，...xN)T。
3. The method according to claim 2, wherein if the network type is a topological network, the eigenvector of the user is initialized randomly in a random matrix manner, and the weight parameter of the random matrix is learned through a training phase of the two-layer graph attention neural network model.
4. The method of claim 2, wherein if the network type is an attribute network, vectorizing the user attribute in the following manner is performed: randomly initializing user data information such as a user name and the like in a word embedding mode to obtain a user name characteristic vector; mining the language style of the user from the long text information of the user by adopting a Doc2Vec model, and learning the text characteristic vector of the user; vector initialization is carried out on the user track information through spatial clustering to obtain a spatial feature vector of a user; and directly taking the user scoring and sign-in times as a characteristic dimension of the user, and carrying out vector initialization.
5. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 2, wherein the step of S2 includes the following steps:
s2.1, embedding a user layer node into a representation submodule: the system is responsible for capturing different influence weights among users to perform weighted aggregation on local neighborhood information of the users in the social network, so that node embedded vectors of user levels are learned;
s2.2, embedding the feature layer nodes into a representation submodule: the method is used for learning the influence weight among different features of a user so as to capture the interactive relation among the features with finer granularity, thereby learning the node embedding vector of the user at the feature level;
s2.3, embedding a vector fusion submodule: the method is responsible for reserving and resetting user embedded vectors from different levels of user levels and feature levels so as to fuse node embedded vectors of multiple visual angles and improve the accuracy of network alignment tasks.
6. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 5, wherein S2.1 comprises the following contents: using a learnable transformation matrix
Figure FDA0002576285300000031
Figure FDA0002576285300000032
Converting the input vector into a high-dimensional vector, namely:
Figure FDA0002576285300000033
according to the related knowledge of the attention neural network of the graph, for any two user nodes viAnd vjFirst, the relationship strength e between the two user nodes is calculatedij：
Figure FDA0002576285300000034
Wherein
Figure FDA0002576285300000035
And
Figure FDA0002576285300000036
representing a user node viAnd vjEmbedding vectors at layer l, W2 (l)Representing the weight parameter of the l-th layer, "|" is a tandem operator and represents that two vectors are transversely spliced, LeakyReLU (·) is an activation function of a neuron, and a user node v is calculatediWhen the neighborhood information is aggregated, the information contribution proportion from different neighbors adopts a softmax (·) function to normalize the user node and all the neighbor user nodes vk∈N(vi) The strength of the relationship between the two is calculated as follows:
Figure FDA0002576285300000037
aijis called user node viAnd vjAttention coefficient of (a) in betweenijThe larger the value of (v), the more closely the relationship between the two users is represented, and the user node v is obtained according to calculationiEach user node v attention coefficient with all its neighbor node users (including itself)iThe new embedded vector of (2) can be defined as follows:
Figure FDA0002576285300000038
wherein (-) is an activation function of a neuron, and a user-level node embedding vector h of each user in the social network can be obtained by performing linear aggregation of different influence weights on neighborhood information of the useriThe user-level vector matrix M ═ (h) is formed1，h2，...，hN)T。
7. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 6, wherein S2.2 comprises the following contents: the user-level vector matrix M obtained in S2.1 is (h)1，h2，...，hN)TAs input, and taking into account the multidimensional attention between the features of any two user nodes in the social network, i.e. calculating an attention coefficient for each corresponding dimension of the two user node vectors,
suppose hiAnd hjDistribution represents two user nodes v in a social networkiAnd vjThe embedded vectors of two user nodes, the relationship between the embedded vectors of two user nodes can be defined as:
f(hi，hj)＝W5·tanh(W4·hi+W3·hj+b2)+b1
wherein W3，W4And
Figure FDA0002576285300000041
is a parameter matrix, b1，
Figure FDA0002576285300000042
Is a bias term, tanh (-) is an activation function of a neuron,
using a feed-forward neural network to operate according to f (h)i，hj) Calculating the dependency relationship of any two user nodes on the characteristic level, and enabling betaijRepresenting a user node viAnd vjThe attention coefficient vector of the feature layer of [ beta ]ij]kSimilarly, in order to compare the attention coefficients of the features of the corresponding dimensions between different attention coefficient vectors, the attention coefficient vectors of all the neighbors of the user are normalized according to the corresponding feature dimensions by using the softmax (·) function, and then:
Figure FDA0002576285300000043
calculating attention coefficient [ beta ] between each dimension of any two usersij]kThese attention coefficients can then be combined into an attention coefficient vector β between the two users according to the corresponding feature dimensionsij＝([βij]1，[βij]2，...，[βij]d″) The dimensions of the attention vector and the user node vector are the same, each dimension [ β [ ]ij]kCorresponding to the weight of influence of each dimension of the user node vector, [ beta ]ij]kThe larger, the two user nodes v are representediAnd vjThe stronger the correlation of the features of the k-th dimension of (a),
finally, for any user node v in the social networkiThe method is characterized in that the embedded vectors of the neighbor users are weighted and linearly aggregated according to the learned attention coefficient vectors among different user nodes, and the method is different from the aggregation mode of the user layer attention mechanism, wherein the aggregation function of the feature layer attention mechanism is to aggregate neighborhood information by adopting a mode of element-by-element multiplication:
Figure FDA0002576285300000051
wherein
Figure FDA0002576285300000052
Represents twoThe element-by-element product of the vectors with the same shape is used to obtain a vector with the same shape, wherein (DEG) the vector is an activation function of a neuron, and the node embedded representation of the feature level of each user in the social network can be obtained by carrying out linear aggregation on the neighborhood information of the users according to the influence weights of different features
Figure FDA0002576285300000053
Forming a feature level vector matrix
Figure FDA0002576285300000054
Figure FDA0002576285300000055
8. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 7, wherein S2.3 comprises the following contents: embedding vector fusion submodule changes user level vector matrix M into (h)1，h2，...，hN)TAnd feature level vector matrix
Figure FDA0002576285300000056
As input, the weight parameters of node embedding vectors of different levels of the same user are learned automatically by using a gating mechanism to effectively retain and reset information representations from different levels,
for any user node v in social networkiThe module first calculates a user-level node embedding vector hiAnd feature level node embedded vectors
Figure FDA0002576285300000057
The weight relationship vector between them is calculated as follows:
Figure FDA0002576285300000058
wherein
Figure FDA0002576285300000059
And
Figure FDA00025762853000000510
is a parameter matrix of the gated neural network,
Figure FDA00025762853000000511
is a bias term, sigmoid (·) is an activation function of a neuron, and according to learning a weight relation vector F, user node embedded vectors from different layers can be selectively retained and reset, and the final user node embedded vector is expressed as follows:
Figure FDA00025762853000000512
wherein
Figure FDA00025762853000000513
Representing selective retention of node-embedded vectors at the user level,
Figure FDA00025762853000000514
represents the selective resetting of node-embedded vectors to the feature level, where 1-F is a vector operation, representing subtracting 1 from each dimension of the vector F,
according to the method, the user level node embedded representation and the characteristic level node embedded representation of the user in the social network are fused by using a gating mechanism, and the final node embedded representation z of each user in the social network can be obtainediThe component node embedding vector matrix Z ═ (Z)1，z2，...，zN)T；
For any given pair of user nodes v in a social networkiAnd vjThe node embedding vectors are respectively ziAnd zjThen the two nodesThe probability of an edge existing between can be expressed as:
Figure FDA0002576285300000061
wherein
Figure FDA0002576285300000062
Is a sigmoid function;
in order to optimize the model parameters of the two-layer graph attention neural network, we need to define an objective function of the model whose goal is to maximize the probability of observable edge occurrences in the social network, i.e.:
Figure FDA0002576285300000063
to avoid trivial solution, for each edge e that can be observedi，j＝(vi，vj) We use a negative sampling technique to maximize the objective function, namely:
Figure FDA0002576285300000064
the first item models positive examples in the social network, the second item models negative examples by randomly generating edges associated with nodes through a negative sampling technology, and the sampled probability of each node satisfies
Figure FDA0002576285300000065
K denotes the number of edges of the negative example of the sample, dvRepresenting degrees of user node v; according to the objective function, parameters of the double-layer graph attention neural network model can be learned by adopting a back propagation optimization algorithm, so that node vector matrixes of the source social network and the target social network are obtained respectively
Figure FDA0002576285300000066
Figure FDA0002576285300000067
And
Figure FDA0002576285300000068
wherein | VsI represents the number of user nodes in the source social network, | VtI represents the number of user nodes in the target social network;
Figure FDA0002576285300000069
representing users in a source social network
Figure FDA00025762853000000610
The node of (a) embeds the vector,
Figure FDA00025762853000000611
representing users in a target social network
Figure FDA00025762853000000612
The node of (1) embeds a vector; zsAnd ZtAlso referred to as the embedded vector space for the source social network and the target social network.
9. The method for network alignment based on the double-layer graph attention neural network as claimed in claim 8, wherein the step of S3 includes the following steps: based on the step of S2, obtaining a node embedded vector matrix of the source social network
Figure FDA0002576285300000071
Embedding vector matrix with nodes of target social network
Figure FDA0002576285300000072
Wherein each row of the node vector matrix represents a node embedded vector corresponding to one user in the social network, and the whole node vector matrix is also called as the embedded vector corresponding to the social networkEntering a vector space;
defining a mapping function M that maps user node vectors from one embedded vector space to another, assuming we now project the source social network to the target social network to find a target node that matches the source social network, given a partially known set of anchor links T as the supervisory information, the target function can be defined as:
Figure FDA0002576285300000073
wherein M iss→t() represents a mapping function from a source social network to a target social network, the mapping function being constructed using a multi-tiered perceptron, θ being a weighting parameter for the multi-tiered perceptron; the objective function aims to minimize the distance from a source user node to a target user node after mapping the source user node in a user pair with an anchor link relation to the target social network, construct a classification model to predict whether any two users between different social networks have an anchor link, and select the target user node closest to the source user node after projection to construct a candidate anchor link.
CN202010654776.0A 2020-07-09 2020-07-09 Network alignment method based on double-layer graph attention neural network Active CN111931903B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010654776.0A CN111931903B (en) 2020-07-09 2020-07-09 Network alignment method based on double-layer graph attention neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010654776.0A CN111931903B (en) 2020-07-09 2020-07-09 Network alignment method based on double-layer graph attention neural network

Publications (2)

Publication Number Publication Date
CN111931903A true CN111931903A (en) 2020-11-13
CN111931903B CN111931903B (en) 2023-07-07

Family

ID=73312715

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010654776.0A Active CN111931903B (en) 2020-07-09 2020-07-09 Network alignment method based on double-layer graph attention neural network

Country Status (1)

Country Link
CN (1) CN111931903B (en)

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112381179A (en) * 2020-12-11 2021-02-19 杭州电子科技大学 Heterogeneous graph classification method based on double-layer attention mechanism
CN112396492A (en) * 2020-11-19 2021-02-23 天津大学 Conversation recommendation method based on graph attention network and bidirectional long-short term memory network
CN112395466A (en) * 2020-11-27 2021-02-23 上海交通大学 Fraud node identification method based on graph embedded representation and recurrent neural network
CN112446542A (en) * 2020-11-30 2021-03-05 西安电子科技大学 Social network link prediction method based on attention neural network
CN112507246A (en) * 2020-12-13 2021-03-16 天津大学 Social recommendation method fusing global and local social interest influence
CN112507247A (en) * 2020-12-15 2021-03-16 重庆邮电大学 Cross-social network user alignment method fusing user state information
CN112667920A (en) * 2020-12-29 2021-04-16 复旦大学 Text perception-based social influence prediction method, device and equipment
CN112800770A (en) * 2021-04-15 2021-05-14 南京樯图数据研究院有限公司 Entity alignment method based on heteromorphic graph attention network
CN112818257A (en) * 2021-02-19 2021-05-18 北京邮电大学 Account detection method, device and equipment based on graph neural network
CN112860810A (en) * 2021-02-05 2021-05-28 中国互联网络信息中心 Domain name multi-graph embedded representation method, device, electronic equipment and medium
CN113065045A (en) * 2021-04-20 2021-07-02 支付宝(杭州)信息技术有限公司 Method and device for carrying out crowd division and training multitask model on user
CN113095948A (en) * 2021-03-24 2021-07-09 西安交通大学 Multi-source heterogeneous network user alignment method based on graph neural network
CN113127752A (en) * 2021-03-18 2021-07-16 中国人民解放军战略支援部队信息工程大学 Social network account aligning method and system based on user naming habit mapping learning
CN113238885A (en) * 2021-05-08 2021-08-10 长安大学 Implicit deviation instruction prediction method and device based on graph attention network
CN113240098A (en) * 2021-06-16 2021-08-10 湖北工业大学 Fault prediction method and device based on hybrid gated neural network and storage medium
CN113407784A (en) * 2021-05-28 2021-09-17 桂林电子科技大学 Social network-based community dividing method, system and storage medium
CN113409157A (en) * 2021-05-19 2021-09-17 桂林电子科技大学 Cross-social network user alignment method and device
CN113628059A (en) * 2021-07-14 2021-11-09 武汉大学 Associated user identification method and device based on multilayer graph attention network
CN113779406A (en) * 2021-09-16 2021-12-10 浙江网商银行股份有限公司 Data processing method and device
CN113792937A (en) * 2021-09-29 2021-12-14 中国人民解放军国防科技大学 Social network influence prediction method and device based on graph neural network
CN113807012A (en) * 2021-09-14 2021-12-17 杭州莱宸科技有限公司 Water supply network division method based on connection strengthening
CN113901831A (en) * 2021-09-15 2022-01-07 昆明理工大学 Parallel sentence pair extraction method based on pre-training language model and bidirectional interaction attention
CN114662143A (en) * 2022-02-28 2022-06-24 北京交通大学 Sensitive link privacy protection method based on graph embedding
CN115063251A (en) * 2022-05-30 2022-09-16 华侨大学 Social communication propagation dynamic network representation method based on relationship strength and feedback mechanism
CN115861822A (en) * 2023-02-07 2023-03-28 海豚乐智科技(成都)有限责任公司 Target local point and global structured matching method and device
CN116049695A (en) * 2022-12-20 2023-05-02 中国科学院空天信息创新研究院 Group perception and standing analysis method, system and electronic equipment crossing social network
CN116566743A (en) * 2023-07-05 2023-08-08 北京理工大学 Account alignment method, equipment and storage medium
CN116776193A (en) * 2023-05-17 2023-09-19 广州大学 Method and device for associating virtual identities across social networks based on attention mechanism
CN117670572A (en) * 2024-02-02 2024-03-08 南京财经大学 Social behavior prediction method, system and product based on graph comparison learning

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180349477A1 (en) * 2017-06-06 2018-12-06 Facebook, Inc. Tensor-Based Deep Relevance Model for Search on Online Social Networks
CN109636658A (en) * 2019-01-17 2019-04-16 电子科技大学 A kind of social networks alignment schemes based on picture scroll product

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180349477A1 (en) * 2017-06-06 2018-12-06 Facebook, Inc. Tensor-Based Deep Relevance Model for Search on Online Social Networks
CN109636658A (en) * 2019-01-17 2019-04-16 电子科技大学 A kind of social networks alignment schemes based on picture scroll product

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
MEDIAN LU,ET AL.: "A Unified Link Prediction Framework for Predicting Arbitrary Relations in Heterogeneous Academic Networks", 《IEEE ACCESS》, pages 124967 - 124987 *

Cited By (45)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112396492A (en) * 2020-11-19 2021-02-23 天津大学 Conversation recommendation method based on graph attention network and bidirectional long-short term memory network
CN112395466A (en) * 2020-11-27 2021-02-23 上海交通大学 Fraud node identification method based on graph embedded representation and recurrent neural network
CN112395466B (en) * 2020-11-27 2023-05-12 上海交通大学 Fraud node identification method based on graph embedded representation and cyclic neural network
CN112446542A (en) * 2020-11-30 2021-03-05 西安电子科技大学 Social network link prediction method based on attention neural network
CN112446542B (en) * 2020-11-30 2023-04-07 山西大学 Social network link prediction method based on attention neural network
CN112381179B (en) * 2020-12-11 2024-02-23 杭州电子科技大学 Heterogeneous graph classification method based on double-layer attention mechanism
CN112381179A (en) * 2020-12-11 2021-02-19 杭州电子科技大学 Heterogeneous graph classification method based on double-layer attention mechanism
CN112507246A (en) * 2020-12-13 2021-03-16 天津大学 Social recommendation method fusing global and local social interest influence
CN112507246B (en) * 2020-12-13 2022-09-13 天津大学 Social recommendation method fusing global and local social interest influence
CN112507247A (en) * 2020-12-15 2021-03-16 重庆邮电大学 Cross-social network user alignment method fusing user state information
CN112507247B (en) * 2020-12-15 2022-09-23 重庆邮电大学 Cross-social network user alignment method fusing user state information
CN112667920A (en) * 2020-12-29 2021-04-16 复旦大学 Text perception-based social influence prediction method, device and equipment
CN112860810A (en) * 2021-02-05 2021-05-28 中国互联网络信息中心 Domain name multi-graph embedded representation method, device, electronic equipment and medium
CN112860810B (en) * 2021-02-05 2023-07-14 中国互联网络信息中心 Domain name multiple graph embedded representation method, device, electronic equipment and medium
CN112818257A (en) * 2021-02-19 2021-05-18 北京邮电大学 Account detection method, device and equipment based on graph neural network
CN113127752A (en) * 2021-03-18 2021-07-16 中国人民解放军战略支援部队信息工程大学 Social network account aligning method and system based on user naming habit mapping learning
CN113095948B (en) * 2021-03-24 2023-06-06 西安交通大学 Multi-source heterogeneous network user alignment method based on graph neural network
CN113095948A (en) * 2021-03-24 2021-07-09 西安交通大学 Multi-source heterogeneous network user alignment method based on graph neural network
CN112800770A (en) * 2021-04-15 2021-05-14 南京樯图数据研究院有限公司 Entity alignment method based on heteromorphic graph attention network
CN113065045A (en) * 2021-04-20 2021-07-02 支付宝(杭州)信息技术有限公司 Method and device for carrying out crowd division and training multitask model on user
CN113238885B (en) * 2021-05-08 2023-07-07 长安大学 Method and equipment for predicting implicit deviation instruction based on graph attention network
CN113238885A (en) * 2021-05-08 2021-08-10 长安大学 Implicit deviation instruction prediction method and device based on graph attention network
CN113409157A (en) * 2021-05-19 2021-09-17 桂林电子科技大学 Cross-social network user alignment method and device
CN113409157B (en) * 2021-05-19 2022-06-28 桂林电子科技大学 Cross-social network user alignment method and device
CN113407784B (en) * 2021-05-28 2022-08-12 桂林电子科技大学 Social network-based community dividing method, system and storage medium
CN113407784A (en) * 2021-05-28 2021-09-17 桂林电子科技大学 Social network-based community dividing method, system and storage medium
CN113240098A (en) * 2021-06-16 2021-08-10 湖北工业大学 Fault prediction method and device based on hybrid gated neural network and storage medium
CN113628059B (en) * 2021-07-14 2023-09-15 武汉大学 Associated user identification method and device based on multi-layer diagram attention network
CN113628059A (en) * 2021-07-14 2021-11-09 武汉大学 Associated user identification method and device based on multilayer graph attention network
CN113807012A (en) * 2021-09-14 2021-12-17 杭州莱宸科技有限公司 Water supply network division method based on connection strengthening
CN113901831A (en) * 2021-09-15 2022-01-07 昆明理工大学 Parallel sentence pair extraction method based on pre-training language model and bidirectional interaction attention
CN113901831B (en) * 2021-09-15 2024-04-26 昆明理工大学 Parallel sentence pair extraction method based on pre-training language model and bidirectional interaction attention
CN113779406A (en) * 2021-09-16 2021-12-10 浙江网商银行股份有限公司 Data processing method and device
CN113792937A (en) * 2021-09-29 2021-12-14 中国人民解放军国防科技大学 Social network influence prediction method and device based on graph neural network
CN114662143A (en) * 2022-02-28 2022-06-24 北京交通大学 Sensitive link privacy protection method based on graph embedding
CN114662143B (en) * 2022-02-28 2024-05-03 北京交通大学 Sensitive link privacy protection method based on graph embedding
CN115063251A (en) * 2022-05-30 2022-09-16 华侨大学 Social communication propagation dynamic network representation method based on relationship strength and feedback mechanism
CN116049695B (en) * 2022-12-20 2023-07-04 中国科学院空天信息创新研究院 Group perception and standing analysis method, system and electronic equipment crossing social network
CN116049695A (en) * 2022-12-20 2023-05-02 中国科学院空天信息创新研究院 Group perception and standing analysis method, system and electronic equipment crossing social network
CN115861822A (en) * 2023-02-07 2023-03-28 海豚乐智科技(成都)有限责任公司 Target local point and global structured matching method and device
CN116776193A (en) * 2023-05-17 2023-09-19 广州大学 Method and device for associating virtual identities across social networks based on attention mechanism
CN116566743A (en) * 2023-07-05 2023-08-08 北京理工大学 Account alignment method, equipment and storage medium
CN116566743B (en) * 2023-07-05 2023-09-08 北京理工大学 Account alignment method, equipment and storage medium
CN117670572A (en) * 2024-02-02 2024-03-08 南京财经大学 Social behavior prediction method, system and product based on graph comparison learning
CN117670572B (en) * 2024-02-02 2024-05-03 南京财经大学 Social behavior prediction method, system and product based on graph comparison learning

Also Published As

Publication number Publication date
CN111931903B (en) 2023-07-07

Similar Documents

Publication Publication Date Title
CN111931903B (en) Network alignment method based on double-layer graph attention neural network
CN111488734B (en) Emotional feature representation learning system and method based on global interaction and syntactic dependency
CN111462282B (en) Scene graph generation method
Li et al. Deep learning-based classification methods for remote sensing images in urban built-up areas
CN110427877B (en) Human body three-dimensional posture estimation method based on structural information
CN112925989B (en) Group discovery method and system of attribute network
CN108920678A (en) A kind of overlapping community discovery method based on spectral clustering with fuzzy set
CN111695415A (en) Construction method and identification method of image identification model and related equipment
CN113255895B (en) Structure diagram alignment method and multi-diagram joint data mining method based on diagram neural network representation learning
CN111325243B (en) Visual relationship detection method based on regional attention learning mechanism
CN113761250A (en) Model training method, merchant classification method and device
CN112200266A (en) Network training method and device based on graph structure data and node classification method
CN113065974A (en) Link prediction method based on dynamic network representation learning
CN112862147A (en) Comprehensive pipe rack operation and maintenance risk evaluation model and method based on BP neural network
CN116129286A (en) Method for classifying graphic neural network remote sensing images based on knowledge graph
Xie et al. Temporal‐enhanced graph convolution network for skeleton‐based action recognition
CN116385660A (en) Indoor single view scene semantic reconstruction method and system
Zou et al. Application of facial symmetrical characteristic to transfer learning
CN116010813A (en) Community detection method based on influence degree of fusion label nodes of graph neural network
CN110209860B (en) Template-guided interpretable garment matching method and device based on garment attributes
Zhang et al. End‐to‐end generation of structural topology for complex architectural layouts with graph neural networks
CN114254738A (en) Double-layer evolvable dynamic graph convolution neural network model construction method and application
Cao et al. QMEDNet: A quaternion-based multi-order differential encoder–decoder model for 3D human motion prediction
Xiao et al. Model transferability from ImageNet to lithography hotspot detection
Kim et al. Automated door placement in architectural plans through combined deep-learning networks of ResNet-50 and Pix2Pix-GAN

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant