CN113628059A - Associated user identification method and device based on multilayer graph attention network - Google Patents
Associated user identification method and device based on multilayer graph attention network Download PDFInfo
- Publication number
- CN113628059A CN113628059A CN202110794741.1A CN202110794741A CN113628059A CN 113628059 A CN113628059 A CN 113628059A CN 202110794741 A CN202110794741 A CN 202110794741A CN 113628059 A CN113628059 A CN 113628059A
- Authority
- CN
- China
- Prior art keywords
- network
- attention
- matrix
- embedded
- node
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 239000011159 matrix material Substances 0.000 claims abstract description 128
- 239000013598 vector Substances 0.000 claims abstract description 80
- 230000007246 mechanism Effects 0.000 claims abstract description 22
- 230000006870 function Effects 0.000 claims description 41
- 230000004927 fusion Effects 0.000 claims description 37
- 230000004913 activation Effects 0.000 claims description 25
- 238000000605 extraction Methods 0.000 claims description 25
- 238000004364 calculation method Methods 0.000 claims description 18
- 230000008569 process Effects 0.000 claims description 17
- 238000012549 training Methods 0.000 claims description 17
- 238000010586 diagram Methods 0.000 claims description 12
- 230000002787 reinforcement Effects 0.000 claims description 9
- 238000005516 engineering process Methods 0.000 claims description 7
- 230000003213 activating effect Effects 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- OAICVXFJPJFONN-UHFFFAOYSA-N Phosphorus Chemical compound [P] OAICVXFJPJFONN-UHFFFAOYSA-N 0.000 claims description 3
- 238000009472 formulation Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000000875 corresponding effect Effects 0.000 claims 4
- 230000002596 correlated effect Effects 0.000 claims 1
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000005295 random walk Methods 0.000 description 5
- 238000007418 data mining Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 235000010678 Paulownia tomentosa Nutrition 0.000 description 1
- 240000002834 Paulownia tomentosa Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- 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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
Abstract
The invention provides a method and a device for identifying associated users based on a multilayer graph attention network, wherein the method firstly simulates structure noise and attribute noise through a random permutation matrix so as to improve the adaptivity of a model; a correlation user identification model is established, and the importance of each neighbor node to the target node is calculated based on an attention mechanism; calculating to obtain an embedded vector of each node by using the normalized attention coefficient and the weight matrix, and fusing the embedded vectors of all target nodes into an embedded matrix; fusing multilayer embedded matrixes obtained by the multilayer graph attention network according to the weight, and obtaining a final embedded matrix based on a greedy strategy; and performing correlation identification on the final embedded matrix to obtain a final result. Because the structure noise and the attribute noise are simulated in advance, the self-adaptability of the model is greatly improved; the embedding result of the multi-layer graph attention network is effectively extracted, so that the identification accuracy of the associated user is improved, and the error is effectively reduced.
Description
Technical Field
The invention relates to the technical field of associated user identification, in particular to an associated user identification method and device based on a multilayer graph attention network.
Background
The associated user identification is a technology for detecting associated users in a plurality of social networks. Associated users refer to users in different social networks, but they belong to the same natural person in the real world. The associated user identification is often applied to the problems of recommendation systems, criminal behavior prediction, personalized services and 'cold start', so that the associated user identification becomes a hot spot of the current social network research; however, node information in the social network is various, and the network structure is large and complex, so that feature extraction of the social network becomes a difficult point. Most of the related user identification algorithms at present are based on a deep Random Walk-based Social network feature extraction method proposed by Perozzi, Al-Rfou and Skiona in the literature (deep Walk: one Learning of Social responses [ C ]. Proceedings of the 20th ACM SIGKDD interactive conference on Knowledge and data mining-KDD'14,2014:701 and 710.), wherein Perozzi represents a vertex sequence as a multi-dimensional vector by using a natural language processing tool (word2vec) on the basis of the vertex sequence extracted by the Random Walk algorithm (Random Walk), and finally obtains a better feature extraction effect. However, the feature extraction method based on random walk cannot directly process the non-european structure of the social network, but converts the non-european structure into a vertex sequence to perform feature extraction, which results in errors.
Just because the random walk algorithm causes great errors and noises in feature extraction, the method of graph neural network, which can directly extract features of non-european structures (e.g., social networks), is becoming a hot spot. How to combine the graph neural network with the associated user identification, the social network characteristics are efficiently extracted, and the associated user identification is accurately performed, which becomes a difficult problem. There are also related studies aimed at solving this problem, such as Chen et al in the literature (h.chen, h.yin, x.sun, t.chen, b.gabrys, and k.music, "Multi-level graph relational network for cross-platform and link prediction," in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,2020, pp.1503-Data 1.) propose to simultaneously use simple graph convolutional neural networks and hyper graph convolutional neural networks for feature extraction of social networks, and un et al in the literature (h.t.trunk, t.van video, n.t.tam, h.yin, m.m.idlich, n.q.v.and, "tung mapping network" propose to stabilize the recognition of neural networks by IEEE network-link recognition method, IEEE-based on the product of interest-Data Mining,2020, pp.1503-Data), and "the correlation of multiple-network recognition" 14-Data Mining "2020, and extension network recognition by IEEE-network-link recognition method, and the stability-node recognition method of IEEE, increase by IEEE-map recognition. In the process of implementing the present invention, the inventors of the present application find that the above algorithms still have some problems, and define the weights of all neighbor nodes as 1, neglecting the difference of the weights of different neighbor nodes to the target node, thereby causing the feature extraction process to be inaccurate.
By combining the above analysis, although various improved associated user identification methods achieve certain detection effects, the problem of low identification accuracy still needs to be improved.
Disclosure of Invention
The invention provides a method and a device for identifying associated users based on a multilayer graph attention network, which are used for solving or at least partially solving the technical problem of low identification precision in the prior art.
In order to solve the above technical problem, a first aspect of the present invention provides a method for identifying an associated user based on a multi-layer graph attention network, including:
s1: acquiring two social networks, taking one of the social networks as a source social network and the other social network as a target social network, and performing data reinforcement learning on the two social networks, wherein the social networks comprise three kinds of information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
s2: constructing a correlation user identification model based on a multilayer graph attention network, wherein the correlation user identification model based on the multilayer graph attention network comprises the multilayer graph attention network, an embedded fusion module and an output module, and the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
s3: training a correlation user identification model based on the multilayer diagram attention network by using the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedded matrix, and obtaining a model corresponding to the optimal embedded matrix as the trained correlation user identification model based on the multilayer diagram attention network;
s4: and performing relevant user identification on the input social network by using a trained relevant user identification model based on the multilayer graph attention network.
In one embodiment, the data-enhanced learning of the two social networks in step S1 includes: the data enhancement is realized by simulating structural noise and attribute noise through a random permutation matrix.
In one embodiment, the calculation process of the node relation extracting module in step S2 includes:
based on an attention mechanism, calculating the importance of each neighbor node to the target node, and taking the importance as an attention coefficient, wherein the calculation formula is as follows:
euv＝a(WFu,WFv)
wherein ,euvDenotes an attention coefficient calculated with a node u as a target node, W denotes a weight matrix, FuFeature vector, F, representing node uvA feature vector representing node v, a () representing the attention mechanism;
the attention coefficient is normalized by a softmax function:
wherein the expression represents that attention coefficients are normalized by a softmax function, and alpha'uvDenotes the normalized attention coefficient, MuThe neighborhood representing node u, i.e. all neighbor nodes, represents an exponential function with e as base, exp (e)uv) Is represented by e as base, euvIs a function of the index of the light,
means to perform the operation on all the neighbor nodes and sum;
activating the normalized attention coefficient by using a LeakyReLU activation function:
wherein the above formula represents activating the normalized attention coefficient by using LeakyReLU function as an activation function, alphauvIndicating the attention coefficient after activation, the LeakyReLU indicates the activation function,
weight vector representing attention mechanism parameterization [ | | ·]The concatenation of the vectors is represented and,
a formulation representing the attention mechanism a (·).
In one embodiment, the calculation process of the feature fusion module in step S2 includes:
calculating to obtain an embedded vector F 'of each node by using the activated attention coefficient and the weight matrix'u：
Where σ denotes the activation function, αuvDenotes the attention coefficient after activation, FvA feature vector representing node v;
fusing the embedded vectors of all target nodes into an embedded matrix:
wherein the multi-level graph attention network comprises a plurality of graph attention networks, each graph attention network obtaining a corresponding embedded matrix, H(l)Representing the l-th layer graph attention network derived embedded matrix,
the embedded vector of node 1 obtained by the l-th layer graph attention network is shown.
In one embodiment, the calculation process of embedding the fusion module in step S2 includes;
obtaining the association identification score of each graph attention layer through a hierarchical alignment matrix:
fusing the layered alignment matrix according to a preset weight, and obtaining a final associated user identification matrix by adopting a greedy strategy:
wherein ,
the layer l representing the source social network is aware of the network derived embedded matrix,
an embedding matrix, S, representing the Lth layer of the target social network, derived by the attention network(l)A hierarchical alignment matrix representing the graph attention network of the ith layer, an association recognition score, theta, representing the graph attention layer(l)Representing the weight of the l-th level hierarchical alignment matrix.
In one embodiment, the loss function in step S3 is:
wherein u and upRepresenting an original network G and a reinforcement learned network GpOf the same corresponding node, H(l)(u) node embedding, H, at level l of node u(l)(up) Representing a node upAnd embedding nodes of the l-th layer.
The optimal embedding matrix obtained by minimizing the objective function is:
wherein ,
representing a regularized Laplacian matrix, | | · |. non-woven phosphorFRepresents the Frobenius norm,
the layer I of the representation target social network is aware of the embedded matrix derived by the force network.
Based on the same inventive concept, a second aspect of the present invention provides an associated user identification device based on a multi-layer graph attention network, comprising:
the data enhancement module is used for acquiring two social networks, taking one of the two social networks as a source social network and the other as a target social network, and performing data enhancement learning on the two social networks, wherein the social networks comprise three information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
the model building module is used for building an associated user identification model based on the multilayer graph attention network, wherein the associated user identification model based on the multilayer graph attention network comprises the following steps: the system comprises a multilayer graph attention network, an embedded fusion module and an output module, wherein the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
the model training module is used for training the associated user recognition model based on the multilayer diagram attention network by taking the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedding matrix, obtaining a model corresponding to the optimal embedding matrix and taking the model as the trained associated user recognition model based on the multilayer diagram attention network;
and the associated user identification module is used for performing associated user identification on the input social network by utilizing the trained associated user identification model based on the multilayer graph attention network.
One or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:
the invention provides a multi-layer graph attention network-based associated user identification method, which constructs a multi-layer graph attention network-based associated user identification model, comprises a multi-layer graph attention network and an embedded fusion module, wherein the multi-layer graph attention network in the model is used for calculating the importance of each neighbor node to a target node based on an attention mechanism, calculating an embedded vector of each node by using an activated attention coefficient and a weight matrix, and fusing the embedded vectors of all the target nodes into an embedded matrix, so that the weights of different neighbor nodes to the target node can be considered in the process of extracting social network characteristics, and the accuracy of characteristic extraction is greatly improved; and further fusing the embedded matrixes of the source social network and the target social network according to preset weights through an embedded fusion module to obtain an associated user identification matrix representing the friend relationship between users, namely weighting and fusing the multi-layer graph embedding in the associated identification, and using the fused embedded matrix for associated identification, thereby greatly improving the identification accuracy.
Furthermore, the structural noise and the attribute noise are simulated by adopting data reinforcement learning, so that the adaptability of the associated user identification model based on the multilayer graph attention network is greatly improved.
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 introduced below, and it is obvious that the drawings in the following description are 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 an overall flowchart of a method for identifying associated users based on a multi-layer graph attention network in the implementation of the present invention.
Detailed Description
The invention aims to provide a method for identifying associated users based on a multilayer graph attention network, which solves the problems of low identification precision, high error rate and low self-adaptability of the existing similar algorithm
The main concept of the invention is as follows:
firstly, performing data reinforcement learning on two social networks, then constructing a multi-layer graph attention network-based associated user identification model, then training the identification model to obtain a trained model, and finally identifying associated users by using the identification model; calculating to obtain an embedded vector of each node by using the normalized and activated attention coefficient and the weight matrix, and fusing the embedded vectors of all target nodes into an embedded matrix; fusing the embedded matrixes obtained by the multilayer graph attention network according to the weight, and obtaining a final embedded matrix based on a greedy strategy; and performing correlation identification on the final embedded matrix to obtain a final result.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
The embodiment of the invention provides a method for identifying associated users based on a multilayer graph attention network, which comprises the following steps:
s1: acquiring two social networks, taking one of the social networks as a source social network and the other social network as a target social network, and performing data reinforcement learning on the two social networks, wherein the social networks comprise three kinds of information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
s2: constructing a correlation user identification model based on a multilayer graph attention network, wherein the correlation user identification model based on the multilayer graph attention network comprises the multilayer graph attention network, an embedded fusion module and an output module, and the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
s3: training a correlation user identification model based on the multilayer diagram attention network by using the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedded matrix, and obtaining a model corresponding to the optimal embedded matrix as the trained correlation user identification model based on the multilayer diagram attention network;
s4: and performing relevant user identification on the input social network by using a trained relevant user identification model based on the multilayer graph attention network.
In particular, since the associated user identification is to identify an associated user between two social networks, one of the social networks is typically defined as a source social network and the other is defined as a target social network. In the two social networks obtained in step S1, either one of the two social networks may be used as the source social network, and the other social network may be used as the target social network. The attribute information of the user includes the age, sex, place of residence, and the like of the user. The connecting edges between the nodes represent the friend relationship between the users, and if the two users are friends, the connecting edges exist. Structural noise refers to errors in social network structural information, and attribute noise refers to the ghosting that exists in social network attribute information. The purpose of this step is to simulate the structural noise and the attribute noise existing in the real social network, so as to improve the adaptability of the model to different noises.
In the step S2, when the multi-layer graph attention network performs feature extraction, the importance of each neighbor node to the target node is calculated based on the attention mechanism, where the target node refers to the node that is calculating the attention coefficient, and the neighbor nodes are all nodes that have a friend relationship with the target node. For example, for a source social network, if 100 nodes are included, node 1 may be first used as a target node, and the importance of the neighbor node of node 1 to the target node is calculated by using an attention mechanism, and then similar calculation is performed by using other nodes as target nodes in turn. For the target social network, the calculation process is similar, and the description is omitted here.
The weight matrix represents the functional relationship between the input and the output, and is continuously updated in the process of deep learning. The embedded vector represents a vector representation obtained by fusing the attention coefficient, the network structure and other embedded information on the basis of the feature vector, and the embedded vector can more accurately reflect the information of the node. And merging the embedded vectors of each node in the social network, and obtaining a matrix called an embedded matrix.
The multi-layer graph attention network comprises a plurality of attention networks, wherein each layer graph attention network can obtain an embedded matrix, the first layer graph attention network corresponds to the embedded matrix obtained through the first-order neighbor friend relationship, the second layer graph attention network corresponds to the embedded matrix obtained through the second-order neighbor friend relationship, and the like (wherein the output of the previous layer graph attention network is used as the input of the next layer graph attention network). The first-order neighbors are nodes which are directly friends with the target node, the second-order neighbors are nodes which are friends with the target node, and the like.
In one embodiment, the data-enhanced learning of the two social networks in step S1 includes: the data enhancement is realized by simulating structural noise and attribute noise through a random permutation matrix.
In a specific implementation process, data enhancement can be realized through the following steps:
step 1.1, simulating structural noise: randomly deleting connecting edges in the source social network by using the probability Ps, and simulating the structure noise;
step 1.2, simulating attribute noise: for the binary attribute, randomly moving the position of a non-zero item by a probability Pa; for the real-value attribute Fij, randomly sampling in the range of [0, Pa x Fij ] to replace the original attribute value;
wherein, the binary attribute only exists in two forms of 0 or 1, and the non-zero item is 1;
step 1.3, noise addition is carried out through the random permutation matrix, and an adjacent matrix A of the enhanced network is obtainedp：
Ap＝PAPT
Wherein, the random permutation matrix is P, Pij ═ 1 indicates that the node i of the enhanced network corresponds to the node j in the original network, and a represents the adjacency matrix of the original network.
The adaptability of the model can be improved by simulating the structure noise and the attribute noise through the random permutation matrix.
In one embodiment, the calculation process of the node relation extracting module in step S2 includes:
based on an attention mechanism, calculating the importance of each neighbor node to the target node, and taking the importance as an attention coefficient, wherein the calculation formula is as follows:
euv＝a(WFu,WFv)
wherein ,euvDenotes an attention coefficient calculated with a node u as a target node, W denotes a weight matrix, FuFeature vector, F, representing node uvA feature vector representing node v, a () representing the attention mechanism;
the attention coefficient is normalized by a softmax function:
wherein the expression represents that attention coefficients are normalized by a softmax function, and alpha'uvDenotes the normalized attention coefficient, MuThe neighborhood representing node u, i.e. all neighbor nodes, represents an exponential function with e as base, exp (e)uv) Is represented by e as base, euvIs a function of the index of the light,
means to perform the operation on all the neighbor nodes and sum;
activating the normalized attention coefficient by using a LeakyReLU activation function:
wherein the above formula represents activating the normalized attention coefficient by using LeakyReLU function as an activation function, alphauvIndicating the attention coefficient after activation, the LeakyReLU indicates the activation function,
weight vector representing attention mechanism parameterization [ | | ·]The concatenation of the vectors is represented and,
a formulation representing the attention mechanism a (·).
Specifically, the weight matrix is a parameter of the associated user identification model, and is continuously updated in the model training process. The eakyReLU activation function is used to improve the non-linearity of the model.
In one embodiment, the calculation process of the feature fusion module in step S2 includes:
calculating to obtain an embedded vector F 'of each node by using the activated attention coefficient and the weight matrix'u：
Where σ denotes the activation function, αuvDenotes the attention coefficient after activation, FvA feature vector representing node v;
fusing the embedded vectors of all target nodes into an embedded matrix:
wherein the multi-level graph attention network comprises a plurality of graph attention networks, each graph attention network obtaining a corresponding embedded matrix, H(l)Representing the l-th layer graph attention network derived embedded matrix,
the embedded vector of node 1 obtained by the l-th layer graph attention network is shown.
In one embodiment, the calculation process of embedding the fusion module in step S2 includes;
obtaining the association identification score of each graph attention layer through a hierarchical alignment matrix:
fusing the layered alignment matrix according to a preset weight, and obtaining a final associated user identification matrix by adopting a greedy strategy:
wherein ,
the layer l representing the source social network is aware of the network derived embedded matrix,
an embedding matrix, S, representing the Lth layer of the target social network, derived by the attention network(l)A hierarchical alignment matrix representing the attention network of the ith layer, storing the associated identification score, θ, of the attention layer of the graph(l)Representing the weight of the l-th level hierarchical alignment matrix.
Specifically, the hierarchical alignment matrix is an associated user identification matrix obtained by using only an embedded matrix of a certain layer of graph attention network, and the final associated user identification matrix is obtained by weighting and fusing the hierarchical alignment matrix. For example, if user u is associatedUser prediction, where v is the associated user of u. Through the model of the invention, the association scores between all other users and u can be obtained, and if the association score of u and v is ranked at the top q (which can be set according to practical situations, such as 1, 5, 10), v is considered as the associated user of u. The preset weight is a proportion set according to actual conditions, theta(l)May be 50%, 60%, 70%, etc.
In the method disclosed by the invention, a two-layer graph attention network is taken as an example, a source social network is input into a correlation user identification model of a multi-layer graph attention network, an embedded matrix of the source social network is obtained through a first-layer graph attention network, then the embedded matrix obtained through the first-layer graph attention network is taken as input to continue the calculation of an attention coefficient and an embedded vector, and the embedded matrix of a second-layer graph attention network is obtained by fusing the embedded vector. And then, in an embedded fusion module, setting a preset weight for an embedded matrix of each layer of graph attention network for fusion to obtain a relevant user identification matrix, wherein in the calculation process, the previous layer of graph attention network is used as the input of the next layer of graph attention network, and the influence of each layer of graph attention network is considered in the embedded fusion, so that the difference of the importance of different neighbor nodes on the target node is fully considered, the accuracy of feature extraction is improved, and the identification accuracy of the model is further improved.
In one embodiment, the loss function in step S3 is:
wherein u and upRepresenting an original network G and a reinforcement learned network GpOf the same corresponding node, H(l)(u) node embedding, H, at level l of node u(l)(up) Representing a node upAnd embedding nodes of the l-th layer.
The optimal embedding matrix obtained by minimizing the objective function is:
wherein ,
representing a regularized Laplacian matrix, | | · |. non-woven phosphorFRepresents the Frobenius norm,
the layer I of the representation target social network is aware of the embedded matrix derived by the force network.
Specifically, the self-adaptability of the model is improved by minimizing a loss function, the influence of structural noise and attribute noise on the model is reduced, and the optimal embedded matrix is the optimal parameter of the model.
Please refer to fig. 1, which is a flowchart of an associated user identification method based on a multi-layer graph attention network according to an embodiment. And respectively carrying out noise simulation, attention coefficient calculation, embedded matrix calculation and embedded matrix fusion operation on the source social network and the target social network through the associated user identification model. Then obtaining the correlation identification score of each graph attention layer through layering of the matrix, fusing the layered alignment matrixes according to the weight, and obtaining a final correlation user identification matrix by adopting a greedy strategy; and finally, performing associated user identification by using the associated user identification matrix, and outputting a detection result.
And calculating key indexes of the final associated user identification matrix to obtain a final result:
where Precision @ q represents the probability of the standard answer appearing in top-q candidates,
representing a pair of associated users in a real answer,
representing the matching scores of the pair of associated users in the model,
representing a user
The set of users that match the top q ranks in the model, # { truean formallinks } represents all the associated user pairs.
The applicant runs on a Computer of Intel (R) core (TM) i7-7700K CPU @4.20GHz, 2080Ti GPU, and uses the disclosed Data set Douban of flex-online and documents (D.Koutra, H.Tong, and D.lube, "Big-alignment: Fast binary alignment," in2013IEEE 13th International Conference Data mining.IEEE,2013, pp.389-398.) (F.Zhou, L.Liu, K.Zhang, G.Tracjjjkvski, J.Wu, and D.Zhang, "deep: A deep approximation approach for use alignment," in IEEE FOCOM 8-model correlation, Sa 2018, Sa.2018, and D.8. echo of the published correlation, "I.8. for use alignment," I, J.J.Wu, and D.8. for comparison, "I.8 J.M.8, J.8. for comparison," Big-alignment, "I.8. for use of the same, M.8, M.J.J., Crime behavior prediction, personalized service and the like.
Compared with the prior art, the invention has the beneficial effects that:
according to the method, the attention network of the multilayer graph is utilized, the weights of different neighbor nodes to the target node are effectively considered in the process of extracting the social network characteristics, and the accuracy rate of characteristic extraction is greatly improved; before feature extraction, the method adopts data reinforcement learning to simulate structure noise and attribute noise, thereby greatly improving the adaptivity of the model; in the correlation identification, the invention carries out weighting and fusion on the embedding of the multilayer graph, and uses the fused embedding matrix to carry out correlation identification, thereby greatly improving the identification accuracy.
Example two
Based on the same inventive concept, the present embodiment provides an associated user identification device based on a multi-layer graph attention network, including:
the data enhancement module is used for acquiring two social networks, taking one of the two social networks as a source social network and the other as a target social network, and performing data enhancement learning on the two social networks, wherein the social networks comprise three information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
the model building module is used for building an associated user identification model based on the multilayer graph attention network, wherein the associated user identification model based on the multilayer graph attention network comprises the following steps: the system comprises a multilayer graph attention network, an embedded fusion module and an output module, wherein the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
the model training module is used for training the associated user recognition model based on the multilayer diagram attention network by taking the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedding matrix, obtaining a model corresponding to the optimal embedding matrix and taking the model as the trained associated user recognition model based on the multilayer diagram attention network;
and the associated user identification module is used for performing associated user identification on the input social network by utilizing the trained associated user identification model based on the multilayer graph attention network.
Since the apparatus described in the second embodiment of the present invention is an apparatus used for implementing the method for identifying associated users based on a multi-layer graph attention network in the first embodiment of the present invention, a person skilled in the art can understand the specific structure and the deformation of the apparatus based on the method described in the first embodiment of the present invention, and thus the details are not described herein again. All the devices adopted in the method of the first embodiment of the present invention belong to the protection scope of the present invention.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
Claims (7)
1. A method for identifying associated users based on a multi-layer graph attention network is characterized by comprising the following steps:
s1: acquiring two social networks, taking one of the social networks as a source social network and the other social network as a target social network, and performing data reinforcement learning on the two social networks, wherein the social networks comprise three kinds of information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
s2: constructing a correlation user identification model based on a multilayer graph attention network, wherein the correlation user identification model based on the multilayer graph attention network comprises the multilayer graph attention network, an embedded fusion module and an output module, and the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
s3: training a correlation user identification model based on the multilayer diagram attention network by using the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedded matrix, and obtaining a model corresponding to the optimal embedded matrix as the trained correlation user identification model based on the multilayer diagram attention network;
s4: and performing relevant user identification on the input social network by using a trained relevant user identification model based on the multilayer graph attention network.
2. The method for identifying associated users according to claim 1, wherein the step of performing data-enhanced learning on two social networks in step S1 comprises: the data enhancement is realized by simulating structural noise and attribute noise through a random permutation matrix.
3. The method for identifying associated users according to claim 1, wherein the calculation process of the node relation extracting module in step S2 includes:
based on an attention mechanism, calculating the importance of each neighbor node to the target node, and taking the importance as an attention coefficient, wherein the calculation formula is as follows:
euv＝a(WFu,WFv)
wherein ,euvDenotes an attention coefficient calculated with a node u as a target node, W denotes a weight matrix, FuFeature vector, F, representing node uvA feature vector representing node v, a () representing the attention mechanism;
the attention coefficient is normalized by a softmax function:
wherein the expression represents that attention coefficients are normalized by a softmax function, and alpha'uvDenotes the normalized attention coefficient, MuThe neighborhood representing node u, i.e. all neighbor nodes, represents an exponential function with e as base, exp (e)uv) Is represented by e as base, euvIs a function of the index of the light,
means to perform the operation on all the neighbor nodes and sum;
activating the normalized attention coefficient by using a LeakyReLU activation function:
wherein the above formula represents activating the normalized attention coefficient by using LeakyReLU function as an activation function, alphauvIndicating the attention coefficient after activation, the LeakyReLU indicates the activation function,
weight vector representing attention mechanism parameterization [ | | ·]The concatenation of the vectors is represented and,
a formulation representing the attention mechanism a (·).
4. The method for identifying correlated users according to claim 1, wherein the calculation process of the feature fusion module in step S2 includes:
calculating to obtain an embedded vector F 'of each node by using the activated attention coefficient and the weight matrix'u：
Where σ denotes the activation function, αuvDenotes the attention coefficient after activation, FvA feature vector representing node v;
fusing the embedded vectors of all target nodes into an embedded matrix:
wherein the multi-level graph attention network comprises a plurality of graph attention networks, each graph attention network obtaining a corresponding embedded matrix, H(l)Representing the embedding matrix obtained by the l-th layer graph attention network, F1 lThe embedded vector of node 1 obtained by the l-th layer graph attention network is shown.
5. The method for identifying associated users according to claim 1, wherein the step S2 of embedding calculation of the fusion module includes;
obtaining the association identification score of each graph attention layer through a hierarchical alignment matrix:
fusing the layered alignment matrix according to a preset weight, and obtaining a final associated user identification matrix by adopting a greedy strategy:
wherein ,
the layer l representing the source social network is aware of the network derived embedded matrix,
an embedding matrix, S, representing the Lth layer of the target social network, derived by the attention network(l)A hierarchical alignment matrix representing the graph attention network of the ith layer, an association recognition score, theta, representing the graph attention layer(l)And representing the weight of the l-th layer hierarchical alignment matrix, and S represents the final associated user identification matrix.
6. The associated user identification method of claim 1, wherein the loss function in step S3 is:
wherein u and upRepresenting an original network G and a reinforcement learned network GpOf the same corresponding node, H(l)(u) node embedding, H, at level l of node u(l)(up) Representing a node upAnd embedding nodes of the l-th layer.
The optimal embedding matrix obtained by minimizing the objective function is:
wherein ,
representing a regularized Laplacian matrix, | | · |. non-woven phosphorFDenotes the Frobenius norm, Ht (l)The layer I of the representation target social network is aware of the embedded matrix derived by the force network.
7. An associated user identification device based on a multi-layer graph attention network, comprising:
the data enhancement module is used for acquiring two social networks, taking one of the two social networks as a source social network and the other as a target social network, and performing data enhancement learning on the two social networks, wherein the social networks comprise three information, namely nodes, connecting edges among the nodes and feature vectors of the nodes, the nodes represent users existing in the social networks, the connecting edges among the nodes represent friend relationships among the users, and the feature vectors of the nodes represent vector representations obtained by encoding attribute information of the users through a thermal encoding technology;
the model building module is used for building an associated user identification model based on the multilayer graph attention network, wherein the associated user identification model based on the multilayer graph attention network comprises the following steps: the system comprises a multilayer graph attention network, an embedded fusion module and an output module, wherein the multilayer graph attention network comprises a node relation extraction module and a feature fusion module; the node relation extraction module is used for calculating the importance of each neighbor node to the target node based on an attention mechanism, namely an attention coefficient, and then carrying out normalization and activation processing to obtain an activated attention coefficient; the feature fusion module is used for calculating to obtain an embedded vector of each node by using the activated attention coefficient and the weight matrix, fusing the embedded vectors of all target nodes into an embedded matrix, and expressing the embedded vector by fusing the attention coefficient and the network structure information on the basis of the feature vector; the embedded fusion module is used for fusing embedded matrixes of the source social network and the target social network according to preset weights and obtaining an associated user identification matrix representing the friend relationship between users based on a greedy strategy; the output module is used for obtaining an identification result according to the associated user identification matrix;
the model training module is used for training the associated user recognition model based on the multilayer diagram attention network by taking the network data subjected to data enhancement as training data, minimizing a loss function to obtain an optimal embedding matrix, obtaining a model corresponding to the optimal embedding matrix and taking the model as the trained associated user recognition model based on the multilayer diagram attention network;
and the associated user identification module is used for performing associated user identification on the input social network by utilizing the trained associated user identification model based on the multilayer graph attention network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110794741.1A CN113628059B (en) | 2021-07-14 | 2021-07-14 | Associated user identification method and device based on multi-layer diagram attention network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110794741.1A CN113628059B (en) | 2021-07-14 | 2021-07-14 | Associated user identification method and device based on multi-layer diagram attention network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113628059A true CN113628059A (en) | 2021-11-09 |
| CN113628059B CN113628059B (en) | 2023-09-15 |
Family
ID=78379739
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110794741.1A Active CN113628059B (en) | 2021-07-14 | 2021-07-14 | Associated user identification method and device based on multi-layer diagram attention network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113628059B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114067349A (en) * | 2022-01-12 | 2022-02-18 | 阿里巴巴达摩院(杭州)科技有限公司 | Target object processing method and device |
| CN114386524A (en) * | 2022-01-17 | 2022-04-22 | 深圳市城图科技有限公司 | Power equipment identification method for dynamic self-adaptive graph layering simulation learning |
| CN116248522A (en) * | 2023-05-11 | 2023-06-09 | 中国电信股份有限公司四川分公司 | Service willingness and deterministic network slice association method |
| CN116541273A (en) * | 2023-04-13 | 2023-08-04 | 中国兵器工业信息中心 | Binary code similarity detection method and system based on graph attention |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105095475A (en) * | 2015-08-12 | 2015-11-25 | 武汉大学 | Incomplete attribute tagged pedestrian re-identification method and system based on two-level fusion |
| US9317567B1 (en) * | 2011-02-16 | 2016-04-19 | Hrl Laboratories, Llc | System and method of computational social network development environment for human intelligence |
| CN107194438A (en) * | 2017-05-24 | 2017-09-22 | 武汉大学 | A kind of depth characteristic method for expressing based on multiple stack own coding |
| CN110837602A (en) * | 2019-11-05 | 2020-02-25 | 重庆邮电大学 | User recommendation method based on representation learning and multi-mode convolutional neural network |
| US20200356858A1 (en) * | 2019-05-10 | 2020-11-12 | Royal Bank Of Canada | System and method for machine learning architecture with privacy-preserving node embeddings |
| CN111931903A (en) * | 2020-07-09 | 2020-11-13 | 北京邮电大学 | Network alignment method based on double-layer graph attention neural network |
| CN112035745A (en) * | 2020-09-01 | 2020-12-04 | 重庆大学 | Recommendation algorithm based on counterstudy and bidirectional long-short term memory network |
| CN112084335A (en) * | 2020-09-09 | 2020-12-15 | 电子科技大学 | Social media user account classification method based on information fusion |
| CN112100514A (en) * | 2020-08-31 | 2020-12-18 | 浙江工业大学 | Social network service platform friend recommendation method based on global attention mechanism representation learning |
| JP2020205029A (en) * | 2019-06-17 | 2020-12-24 | 大連海事大学 | High-speed network representation learning method based on broad learning system |
-
2021
- 2021-07-14 CN CN202110794741.1A patent/CN113628059B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9317567B1 (en) * | 2011-02-16 | 2016-04-19 | Hrl Laboratories, Llc | System and method of computational social network development environment for human intelligence |
| CN105095475A (en) * | 2015-08-12 | 2015-11-25 | 武汉大学 | Incomplete attribute tagged pedestrian re-identification method and system based on two-level fusion |
| CN107194438A (en) * | 2017-05-24 | 2017-09-22 | 武汉大学 | A kind of depth characteristic method for expressing based on multiple stack own coding |
| US20200356858A1 (en) * | 2019-05-10 | 2020-11-12 | Royal Bank Of Canada | System and method for machine learning architecture with privacy-preserving node embeddings |
| JP2020205029A (en) * | 2019-06-17 | 2020-12-24 | 大連海事大学 | High-speed network representation learning method based on broad learning system |
| CN110837602A (en) * | 2019-11-05 | 2020-02-25 | 重庆邮电大学 | User recommendation method based on representation learning and multi-mode convolutional neural network |
| CN111931903A (en) * | 2020-07-09 | 2020-11-13 | 北京邮电大学 | Network alignment method based on double-layer graph attention neural network |
| CN112100514A (en) * | 2020-08-31 | 2020-12-18 | 浙江工业大学 | Social network service platform friend recommendation method based on global attention mechanism representation learning |
| CN112035745A (en) * | 2020-09-01 | 2020-12-04 | 重庆大学 | Recommendation algorithm based on counterstudy and bidirectional long-short term memory network |
| CN112084335A (en) * | 2020-09-09 | 2020-12-15 | 电子科技大学 | Social media user account classification method based on information fusion |
Non-Patent Citations (2)
| Title |
|---|
| 张志扬;张凤荔;陈学勤;王瑞锦;: "基于分层注意力的信息级联预测模型", 计算机科学, no. 06 * |
| 杨远奇;: "基于注意力机制的神经网络贝叶斯群组推荐算法", 数字技术与应用, no. 08 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114067349A (en) * | 2022-01-12 | 2022-02-18 | 阿里巴巴达摩院(杭州)科技有限公司 | Target object processing method and device |
| CN114386524A (en) * | 2022-01-17 | 2022-04-22 | 深圳市城图科技有限公司 | Power equipment identification method for dynamic self-adaptive graph layering simulation learning |
| CN116541273A (en) * | 2023-04-13 | 2023-08-04 | 中国兵器工业信息中心 | Binary code similarity detection method and system based on graph attention |
| CN116541273B (en) * | 2023-04-13 | 2024-01-26 | 中国兵器工业信息中心 | Binary code similarity detection method and system based on graph attention |
| CN116248522A (en) * | 2023-05-11 | 2023-06-09 | 中国电信股份有限公司四川分公司 | Service willingness and deterministic network slice association method |
| CN116248522B (en) * | 2023-05-11 | 2023-09-22 | 中国电信股份有限公司四川分公司 | Service willingness and deterministic network slice association method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113628059B (en) | 2023-09-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2023065545A1 (en) | Risk prediction method and apparatus, and device and storage medium | |
| CN111444340B (en) | Text classification method, device, equipment and storage medium | |
| CN111061856B (en) | Knowledge perception-based news recommendation method | |
| CN111695415B (en) | Image recognition method and related equipment | |
| CN113628059A (en) | Associated user identification method and device based on multilayer graph attention network | |
| CN110659723B (en) | Data processing method and device based on artificial intelligence, medium and electronic equipment | |
| CN111291190B (en) | Training method of encoder, information detection method and related device | |
| WO2022105117A1 (en) | Method and device for image quality assessment, computer device, and storage medium | |
| CN113627447B (en) | Label identification method, label identification device, computer equipment, storage medium and program product | |
| CN110705301A (en) | Entity relationship extraction method and device, storage medium and electronic equipment | |
| WO2023065859A1 (en) | Item recommendation method and apparatus, and storage medium | |
| WO2022105118A1 (en) | Image-based health status identification method and apparatus, device and storage medium | |
| CN114330966A (en) | Risk prediction method, device, equipment and readable storage medium | |
| CN111783903A (en) | Text processing method, text model processing method and device and computer equipment | |
| CN114282059A (en) | Video retrieval method, device, equipment and storage medium | |
| CN111523586A (en) | Noise-aware-based full-network supervision target detection method | |
| CN113723238B (en) | Face lightweight network model construction method and face recognition method | |
| CN117114063A (en) | Method for training a generative large language model and for processing image tasks | |
| Du et al. | Structure tuning method on deep convolutional generative adversarial network with nondominated sorting genetic algorithm II | |
| CN111079930A (en) | Method and device for determining quality parameters of data set and electronic equipment | |
| CN116975350A (en) | Image-text retrieval method, device, equipment and storage medium | |
| CN115620342A (en) | Cross-modal pedestrian re-identification method, system and computer | |
| CN112529637B (en) | Service demand dynamic prediction method and system based on context awareness | |
| CN114494809A (en) | Feature extraction model optimization method and device and electronic equipment | |
| CN114186039A (en) | Visual question answering method and device and electronic equipment |
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 |