CN110442758A - A kind of figure alignment schemes, device and storage medium - Google Patents
A kind of figure alignment schemes, device and storage medium Download PDFInfo
- Publication number
- CN110442758A CN110442758A CN201910667549.9A CN201910667549A CN110442758A CN 110442758 A CN110442758 A CN 110442758A CN 201910667549 A CN201910667549 A CN 201910667549A CN 110442758 A CN110442758 A CN 110442758A
- Authority
- CN
- China
- Prior art keywords
- node
- source domain
- aiming field
- generator
- diagnostic
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/90335—Query processing
-
- 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
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the invention discloses a kind of figure alignment schemes, device and storage mediums, wherein by obtaining the source domain figure and aiming field figure that need to be aligned;Unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and unsupervised learning is carried out to aiming field figure and obtains aiming field node diagnostic set;Building generates confrontation network, generates confrontation network, and be trained to confrontation network is generated;It is connect according to the anchor chain between the generator prediction source domain figure and aiming field figure in the generation confrontation network after training, and connects the generator after corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic, the generator after being optimized according to anchor chain;The source domain node in source domain figure is aligned with the target domain node in aiming field figure according to the generator after optimization.Hereby it is achieved that the figure under the conditions of unsupervised is aligned, the dependence in the prior art to labeled data is got rid of, not only reduces the human cost of realization figure alignment, also improves the efficiency of figure alignment.
Description
Technical field
The present invention relates to data mining technology fields, and in particular to a kind of figure alignment schemes, device and storage medium.
Background technique
Currently, figure is widely used in social networks, commodity as a kind of level of abstraction height, articulate data structure
In the expression of network and protein structure, closed by the definition to node and side to describe the association between entity and entity
System.
By taking social networks as an example, social networks can be expressed in the form of figure, when needs are in the social activity of two isomeries
When finding same user in network, the social networks of the social networks of two isomeries can be expressed respectively using figure, it will wherein one
A seal is source domain figure, is aiming field figure by another seal, as a result, by the mesh in source domain figure in source domain node and aiming field figure
Mark domain node is aligned, and matched source domain node and target domain node represent same user.
In the research and practice process to the prior art, it was found by the inventors of the present invention that the existing skill for scheming alignment
Art connects dependent on known anchor chain, and anchor chain, which connects, characterizes known matched node in two figures.However, this needs technical staff
It is connect to mark a large amount of anchor chain manually, the human cost for being aligned realization figure is higher, and the efficiency for scheming alignment is lower.
Summary of the invention
The embodiment of the present invention provides a kind of figure alignment schemes, device and storage medium, can reduce the people of realization figure alignment
Power cost, and improve the efficiency of figure alignment.
The embodiment of the present invention provides a kind of figure alignment schemes, comprising:
The source domain figure and aiming field figure for needing to be aligned are obtained, the source domain figure includes multiple source domain nodes, the aiming field
Figure includes multiple target domain nodes;
Unsupervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and the aiming field figure is carried out
Unsupervised learning obtains aiming field node diagnostic set;
Building generates confrontation network, and the generation confrontation network includes for source domain node diagnostic to be mapped to aiming field section
The generator of feature space where point feature, and for distinguishing the source domain node diagnostic after mapping and aiming field node diagnostic
Arbiter;
According to the source domain node diagnostic set and the aiming field node diagnostic set to the generator and described
Arbiter carries out alternately training, the generator after being trained;
Predict that the anchor chain between the source domain figure and the aiming field figure connects according to the generator after the training, and according to
The anchor chain connects corresponding source domain node diagnostic and aiming field node diagnostic and optimizes the generator after the training, is optimized
Generator afterwards;
According to the generator after the optimization by the source domain figure source domain node and the aiming field figure in target
Domain node alignment.
The embodiment of the present invention also provides a kind of figure alignment means, comprising:
Figure obtains module, and for obtaining the source domain figure and aiming field figure that need to be aligned, the source domain figure includes multiple source domain
Node, the aiming field figure include multiple target domain nodes;
Feature learning module obtains source domain node diagnostic set for carrying out unsupervised learning to the source domain figure, and
Unsupervised learning is carried out to the aiming field figure and obtains aiming field node diagnostic set;
Network struction module generates confrontation network for constructing, and it includes for by source domain node that generations, which fights network,
The generator of feature space where Feature Mapping to aiming field node diagnostic, and for distinguishing the source domain node diagnostic after mapping
With the arbiter of aiming field node diagnostic;
Network training module, for according to the source domain node diagnostic set and the aiming field node diagnostic set pair
The generator and the arbiter carry out alternately training, the generator after being trained;
Optimization module, for being predicted between the source domain figure and the aiming field figure according to the generator after the training
Anchor chain connects, and connects corresponding source domain node diagnostic and aiming field node diagnostic according to the anchor chain and optimize the life after the training
It grows up to be a useful person, the generator after being optimized;
Figure alignment module, for according to the generator after the optimization by the source domain figure source domain node and the mesh
Mark the target domain node alignment in the figure of domain.
In one embodiment, it is predicted between the source domain figure and the aiming field figure according to the generator after the training
Anchor chain when connecing, the optimization module is used for:
By the mesh in the source domain node diagnostic and the aiming field node diagnostic set in the source domain node diagnostic set
It marks domain node feature and carries out combination of two, obtain multiple node diagnostic combinations;
A node diagnostic is chosen to combine, and according to the generator after the training to the source in the node diagnostic combination chosen
Domain node feature is mapped, and mapping node feature is obtained;
Other aiming field node diagnostics except being combined according to the mapping node feature with the node diagnostic chosen
Cosine similarity and the node diagnostic combination chosen in aiming field node diagnostic and the other target domain nodes it is special
The cosine similarity of sign, the node diagnostic chosen described in acquisition combine corresponding cross-domain similarity;
Continue that other node diagnostics is chosen to combine, until getting the cross-domain similarity of all node diagnostic combinations, and will
The cross-domain highest node diagnostic combination of similarity corresponding source domain node and target domain node connect as the anchor chain.
In one embodiment, other except being combined according to the mapping node feature with the node diagnostic chosen
The cosine similarity of aiming field node diagnostic and it is described choose node diagnostic combination in aiming field node diagnostic and it is described its
The cosine similarity of its aiming field node diagnostic, when the node diagnostic chosen described in acquisition combines corresponding cross-domain similarity, institute
Optimization module is stated to be used for:
Determining other aiming field node diagnostics with the highest preceding preset quantity of the mapping node feature cosine similarity,
And obtain other aiming field node diagnostics of the corresponding preceding preset quantity of the mapping node feature with mean cosine phase
Like degree, the first mean cosine similarity is obtained;
It determines and is preset before the aiming field node diagnostic cosine similarity in combining with the node diagnostic chosen is highest
Other aiming field node diagnostics of quantity, and the aiming field node diagnostic in the node diagnostic combination chosen described in acquisition is right with it
The other aiming field node diagnostics for the preceding preset quantity answered and mean cosine similarity, obtain the second mean cosine similarity;
It is similar with the cosine of mapping node feature according to aiming field node diagnostic in the node diagnostic combination chosen
Degree, the first mean cosine similarity and the second mean cosine similarity, the node diagnostic group chosen described in acquisition
Close corresponding cross-domain similarity.
In one embodiment, according to the generator after the optimization by the source domain figure source domain node and the mesh
When marking the target domain node alignment in the figure of domain, the figure alignment module is used for:
For any source domain node in the source domain figure, any source domain is obtained according to the generator after the optimization
The cross-domain similarity of each target domain node in node and the aiming field figure;
Using the maximum target domain node of the cross-domain similarity of correspondence as the target domain node of any source domain node matching.
In one embodiment, corresponding source domain node diagnostic and the optimization of aiming field node diagnostic are being connect according to the anchor chain
Generator after the training, when generator after being optimized, the optimization module is used for:
So that the anchor chain connects the Euclidean distance between corresponding source domain node diagnostic and aiming field node diagnostic most
It is closely optimization aim, the generator after the training is optimized, the generator after obtaining the optimization.
In one embodiment, when obtaining the source domain figure and aiming field figure that need to be aligned, the figure obtains module and is used for:
The user information in the first social networks is obtained, and more according to the user information generation in first social networks
The side of a node and the different nodes of connection, obtains the source domain figure;
The user information in the second social networks is obtained, and multiple sections are generated according to the user information in the second social networks
Point and the side for connecting different nodes, obtain the aiming field figure, second social networks is with first social networks
Isomery social networks.
In one embodiment, according to the generator after the optimization by the source domain figure source domain node and the mesh
Before marking the target domain node alignment in the figure of domain, the optimization module is also used to:
Processing is orthogonalized to the generator after the optimization, obtains the generator of orthogonalization;
When the node in the source domain figure and the aiming field figure is aligned according to the generator after optimization, the figure pair
Neat module is used for:
The node in the source domain figure and the aiming field figure is aligned according to the generator of the orthogonalization.
In one embodiment, unsupervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and to institute
It states aiming field figure to carry out before unsupervised learning obtains aiming field node diagnostic set, the feature learning module is also used to:
The anchor chain judged whether there is between the known source domain figure and the aiming field figure connects;
If it does not exist, then unsupervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and to described
Aiming field figure carries out unsupervised learning and obtains aiming field node diagnostic set;
If it exists, then supervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and to the mesh
Mark domain figure carries out supervised learning and obtains aiming field node diagnostic set, and is transferred to and executes the building generation confrontation network.
In addition, the embodiment of the present invention also provides a kind of storage medium, the storage medium is stored with a plurality of instruction, the finger
It enables and being loaded suitable for processor, to execute the step in any figure alignment schemes provided by the embodiment of the present invention.
The embodiment of the present invention obtains the source domain figure and aiming field figure for needing to be aligned, and source domain figure includes multiple source domain nodes, mesh
Marking domain figure includes multiple target domain nodes;And unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and
Unsupervised learning is carried out to aiming field figure and obtains aiming field node diagnostic set;And building generates confrontation network, generates confrontation
Network includes the generator for feature space where source domain node diagnostic is mapped to aiming field node diagnostic, and is used for area
The arbiter of source domain node diagnostic and aiming field node diagnostic after dividing mapping;And according to source domain node diagnostic set and
Aiming field node diagnostic set carries out alternately training to generator and arbiter, the generator after being trained;And according to instruction
Generator prediction source domain figure after white silk and the anchor chain between aiming field figure connect, and corresponding source domain node diagnostic is connect according to anchor chain
Generator after being trained with the optimization of aiming field node diagnostic, the generator after being optimized;And according to the generator after optimization
Source domain node in source domain figure is aligned with the target domain node in aiming field figure.Hereby it is achieved that the figure under the conditions of unsupervised
Alignment, gets rid of the dependence in existing figure alignment techniques to labeled data, not only reduces the human cost of realization figure alignment, also
Improve the efficiency of figure alignment.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those skilled in the art, without creative efforts, it can also be obtained according to these attached drawings other attached
Figure.
Fig. 1 is the schematic diagram of a scenario of figure alignment schemes provided in an embodiment of the present invention;
Fig. 2 is a flow diagram of figure alignment schemes provided in an embodiment of the present invention;
Fig. 3 is showing the target domain node alignment in the source domain node and aiming field figure in source domain figure of the embodiment of the present invention
It is intended to;
Fig. 4 is another flow diagram of figure alignment schemes provided in an embodiment of the present invention;
Fig. 5 is a structural schematic diagram of figure alignment means provided in an embodiment of the present invention;
Fig. 6 is the structural schematic diagram of terminal provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those skilled in the art's every other implementation obtained without creative efforts
Example, shall fall within the protection scope of the present invention.
Machine learning (Machine Learning, ML) is a multi-field cross discipline, be related to probability theory, statistics,
The multiple subjects such as Approximation Theory, convextiry analysis, algorithm complexity theory.Specialize in the study that the mankind were simulated or realized to computer how
Behavior reorganizes the existing structure of knowledge and is allowed to constantly improve the performance of itself to obtain new knowledge or skills.Engineering
Habit is the core of artificial intelligence, is the fundamental way for making computer have intelligence, and application spreads the every field of artificial intelligence.
Machine learning and deep learning generally include artificial neural network, confidence network, intensified learning, transfer learning, inductive learning, formula
The technologies such as teaching habit.
Scheme provided by the embodiments of the present application is related to the machine learning techniques of artificial intelligence, especially by following examples into
Row explanation:
The embodiment of the present invention provides a kind of figure alignment schemes, device and storage medium.
Referring to Fig. 1, Fig. 1 is the schematic diagram of a scenario of figure alignment schemes provided by the embodiment of the present invention, the figure alignment side
Method can be applied to figure alignment means, which specifically can integrate in tablet computer, mobile phone and laptop etc.
Have reservoir and processor is installed and is had in the terminal of operational capability, for example, for example, the available needs pair of the terminal
Neat source domain figure and aiming field figure, the source domain figure include multiple source domain nodes, which includes multiple target domain nodes, than
Such as, the figure which can be A social networks indicates that source domain node is the user represented in A social networks, the aiming field figure
The figure that can be B social networks indicates that target domain node is the user represented in B social networks;Then nothing is carried out to source domain figure
Supervised learning obtains source domain node diagnostic set, and carries out unsupervised learning to aiming field figure and obtain aiming field node diagnostic collection
It closes, for example, unsupervised learning can be carried out to source domain figure and aiming field figure respectively using DeepWalk algorithm;And building
Confrontation network is generated, generation confrontation network includes for feature where source domain node diagnostic is mapped to aiming field node diagnostic
The generator in space, and the arbiter for distinguishing source domain node diagnostic and aiming field node diagnostic after mapping, for example, can
Confrontation network is generated to construct standard;And according to source domain node diagnostic set and aiming field node diagnostic set to generation
Device and arbiter carry out alternately training, the generator after being trained;Further according to the generator prediction source domain figure and mesh after training
Anchor chain between mark domain figure connects, and connects corresponding source domain node and aiming field node diagnostic to the life after training according to the anchor chain
It grows up to be a useful person and optimizes, the generator after being optimized;Finally, can utilize the generator after optimization by the source domain in source domain figure
The target domain node of node and aiming field figure alignment, to match the source domain node and target domain node of corresponding same entity.
It should be noted that the schematic diagram of a scenario of figure alignment schemes shown in FIG. 1 is only an example, the present invention is implemented
The scene of the figure alignment schemes of example description is the technical solution in order to more clearly illustrate the embodiment of the present invention, composition pair
In the restriction of technical solution provided in an embodiment of the present invention, those of ordinary skill in the art are it is found that drilling with figure alignment schemes
Become and the appearance of new business scene, technical solution provided in an embodiment of the present invention are equally applicable for similar technical problem.
It is described in detail separately below.
In the present embodiment, it will be described from the angle of figure alignment means, which specifically can integrate
Tablet computer, mobile phone and laptop etc. have reservoir and are equipped with processor and have in the terminal of operational capability.
A kind of figure alignment schemes, comprising: obtain the source domain figure and aiming field figure for needing to be aligned, source domain figure includes multiple source domain
Node, aiming field figure include multiple target domain nodes;Unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, with
And unsupervised learning is carried out to aiming field figure and obtains aiming field node diagnostic set;Building generates confrontation network, generates confrontation net
Network includes the generator for feature space where source domain node diagnostic is mapped to aiming field node diagnostic, and for distinguishing
The arbiter of source domain node diagnostic and aiming field node diagnostic after mapping;According to source domain node diagnostic set and aiming field section
Point feature set carries out alternately training to generator and arbiter, the generator after being trained;According to the generator after training
Anchor chain between prediction source domain figure and aiming field figure connects, and corresponding source domain node diagnostic and target domain node are connect according to anchor chain
Generator after characteristic optimization training, the generator after being optimized;According to the generator after optimization by the source domain in source domain figure
Node is aligned with the target domain node in aiming field figure.
Referring to Fig. 2, Fig. 2 is the flow diagram for the figure alignment schemes that one embodiment of the invention provides.The figure alignment side
Method may include:
In 201, the source domain figure and aiming field figure for needing to be aligned are obtained, source domain figure includes multiple source domain nodes, aiming field
Figure includes multiple target domain nodes.
In the embodiment of the present invention, the source domain figure and aiming field figure for needing to be aligned are got first, wherein include in source domain figure
Multiple nodes, and the side of these nodes is connected, the node in source domain figure is denoted as source domain node in the embodiment of the present invention, equally
, it include multiple nodes in aiming field, and connect the side of these nodes, by the node in aiming field figure in the embodiment of the present invention
It is denoted as target domain node.
It should be noted that source domain figure and aiming field figure are aligned, that is to say by source domain figure source domain node and
On target domain node corresponds in aiming field figure.
Currently, people are not limited solely to using a social networks, but can use simultaneously with the prosperity of social networks
Multiple social networks are to enjoy more applications, for example, same user can pay close attention to some stars, writer either on pushing away spy
It is main etc. to pay close attention to different videos on YouTube by interested people and thing etc..These social networks are often mutually indepedent, still
It can reflect the information of user jointly again, this is provided in conjunction with isomery social networks and to excavate the information that they are implied jointly
Chance and challenge.The first step that same user is many data mining tasks is found in different social networks, for example, not
With social networks in determine same user, can excavate user information in conjunction with more resources, such issues that referred to as society
Hand over network alignment.And social networks can actually be expressed as figure, therefore social networks alignment problem can also be expressed as figure pair
Neat problem.In one embodiment, " the source domain figure and aiming field figure for needing to be aligned are obtained ", comprising:
(1) user information in the first social networks is obtained, and more according to the user information generation in the first social networks
The side of a node and the different nodes of connection, obtains source domain figure;
(2) user information in the second social networks is obtained, and more according to the user information generation in the second social networks
The side of a node and the different nodes of connection, obtains aiming field figure, and the second social networks and the first social networks are that isomery is social
Network.
In the embodiment of the present invention, when needing to be aligned the social networks of two isomeries, it can be expressed first
For figure.Wherein, the first social networks and the second social networks are isomery social networks, it should be noted that first and second simultaneously
It is non-to refer in particular to a certain social networks, it is only used for distinguishing two isomery social networks, the first social networks and the second social networks can be with
For any two isomery social networks.For example, the first social networks can be to push away spy, nervus opticus network can be facebook etc..
When the social networks for needing to be aligned is expressed as figure, the available user information in the first social networks,
The user information describes the relationship between different user and different user, as a result, can be according to the user information by first
Social networks is expressed as multiple nodes, and connects the side of these nodes, obtains the social network diagram of the first social networks, the society
Handing over the node in network is the user represented in the first social networks, and the side between two nodes represents two users
Between relationship, in the embodiment of the present invention, the social network diagram that the first social networks is expressed is denoted as source domain figure, it is therein
Node is source domain node.
Similarly, the user information in the second social networks can also be got, and social by second according to the user information
Network is expressed as multiple nodes, and connects the side of these nodes, obtains the social network diagram of the second social networks, the social network
Node in network is the user represented in the second social networks, and the side between two nodes represents between two users
The social networks that second social networks is expressed in the embodiment of the present invention, is denoted as target and figure, node therein is by relationship
For target domain node.
It should be noted that source domain figure and the alignment of aiming field figure, understanding that can be popular are as follows: by the source domain section in source domain figure
Point is mapped in aiming field figure, finds matched target domain node.
In 202, unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and aiming field figure is carried out
Unsupervised learning obtains aiming field node diagnostic set.
In the embodiment of the present invention, after getting the source domain figure and aiming field figure that need to be aligned, further pass through no prison
Educational inspector practises learning the character representation of source domain figure and aiming field figure.Wherein, using identical unsupervised learning algorithm come to source domain
Figure and aiming field figure carry out unsupervised learning, and the character representation for obtaining source domain figure is source domain node diagnostic set, obtain aiming field
The character representation of figure is aiming field node diagnostic set.
It should be noted that for which kind of unsupervised learning method to learn to obtain source domain figure using in the embodiment of the present invention
It with the character representation of aiming field figure, can be chosen according to actual needs by those of ordinary skill in the art, the present invention is implemented
This is not specifically limited in example.
For example, unsupervised learning can be carried out to source domain figure using DeepWalk algorithm, source domain node diagnostic set is obtainedWherein, n indicates the number of source domain node in source domain figure,Indicate first source domain in source domain figure
The feature of node,Indicate the feature of n-th of source domain node in aiming field figure, and so on.
Likewise, unsupervised learning can be carried out to aiming field figure using DeepWalk algorithm, it is special to obtain target domain node
Collection is closedWherein, m indicates the number of target domain node in aiming field figure,Indicate aiming field figure
The feature of middle first aim domain node,Indicate the feature of n-th of target domain node in aiming field figure, and so on.
In 203, building generates confrontation network, and generating confrontation network includes for source domain node diagnostic to be mapped to target
The generator of feature space where domain node feature, and it is special for distinguishing the source domain node diagnostic after mapping and target domain node
The arbiter of sign.
It should be noted that due to the source domain figure and aiming field figure isomery that need to be aligned, respectively to source domain figure and target
After domain figure carries out unsupervised learning, the source domain node diagnostic learnt is concentrated conjunction and aiming field node diagnostic set and will be located at
Different feature spaces.
In order to realize that figure is aligned, expectation source domain node diagnostic can be with target after being mapped by certain mapping matrix in the present invention
Domain node feature is similar as far as possible.Therefore, dual training is introduced, in the embodiment of the present invention, obtains the source of source domain figure in study
After the aiming field node diagnostic set of domain node characteristic set and aiming field figure, further building generates confrontation network, should
It generates confrontation network to be made of generator and arbiter, generator can be a mapping matrix, be used for source domain node diagnostic
Source domain node diagnostic in set is mapped to the feature space where aiming field node diagnostic, after arbiter is for distinguishing mapping
Source domain node diagnostic and aiming field node diagnostic.
It should be noted that in the embodiment of the present invention for construct which kind of generation confrontation network framework with no restrictions, can
It is chosen according to actual needs by those of ordinary skill in the art, for example, the generation confrontation network of standard can be constructed, it can also be with structure
Improved generation confrontation network is built, such as the generation based on Wasserstein distance fights network.
In 204, according to source domain node diagnostic set and aiming field node diagnostic set to generator and arbiter into
Row is alternately trained, the generator after being trained.
In the embodiment of the present invention, after the building for completing to generate confrontation network, net further is fought to the generation of building
Network is trained.It should be noted that that can regard asking for zero-sum two-person game as on the training essential for generating confrontation network
Topic.
Illustratively, by taking the generation for constructing standard fights network as an example, the loss function of arbiter are as follows:
Wherein, W indicates generator, θDIndicate that arbiter, n indicate the quantity of source domain node, m indicates the number of target domain node
Amount,Indicate i-th of source domain node diagnostic after generator maps,Indicate j-th of aiming field node diagnostic,It is special to indicate that arbiter distinguishes i-th of source domain node after generator maps as far as possible
Sign, vice versa.
It is corresponding, the loss function of generator are as follows:
In training, generator W and arbiter θ is optimized by alternating iterationDTo minimize the loss function of generatorWith the loss function of arbiterPopular says, i.e., when arbiter can not accurately be distinguished through life
Grow up to be a useful person mapping after source domain node diagnostic when, i.e., training complete.
In 205, connect according to the generator prediction source domain figure after training and the anchor chain between aiming field figure, and according to anchor chain
Generator after connecing corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic, the generator after being optimized.
It should be noted that since, in training generator, there is no to the source domain in source domain figure in the embodiment of the present invention
Target domain node in node and aiming field figure carries out one-to-one constraint, needs to introduce additional supervision message to obtain to training
Generator optimize, to realize the one-to-one alignment of source domain node and target domain node, to avoid dual training process
In potential mode avalanche problem.
Therefore, further according to the generator prediction source domain figure after training and between aiming field figure in the embodiment of the present invention
Anchor chain connects.It should be noted that anchor chain is connected in the source domain node and target domain node of known corresponding same entity.Due to this hair
Physical presence anchor chain does not connect in bright implementation, but predicts that the anchor chain for obtaining, therefore can obtaining the prediction connects and be denoted as pseudo- anchor chain
It connects.
It obtains that the puppet anchor chain can be connect to the generator for obtaining training as supervision message after pseudo- anchor chain connects in prediction
It optimizes, namely connects the source domain node diagnostic of the source domain node of instruction and the aiming field section of target domain node according to anchor chain
Point feature optimizes come the generator obtained to training.
In one embodiment, it " is connect according to the generator prediction source domain figure after training and the anchor chain between aiming field figure ", packet
It includes:
(1) by the aiming field section in the source domain node diagnostic and aiming field node diagnostic set in source domain node diagnostic set
Point feature carries out combination of two, obtains multiple node diagnostic combinations;
(2) node diagnostic is chosen to combine, and according to the generator after training to the source in the node diagnostic combination chosen
Domain node feature is mapped, and mapping node feature is obtained;
(3) cosine of other aiming field node diagnostics except being combined according to mapping node feature with the node diagnostic chosen
Similarity, and choose node diagnostic combination in aiming field node diagnostic it is similar to the cosine of other aiming field node diagnostics
Degree obtains the node diagnostic chosen and combines corresponding cross-domain similarity;
(4) continue that other node diagnostics is chosen to combine, until the cross-domain similarity of all node diagnostic combinations is got, and
The highest node diagnostic combination of cross-domain similarity corresponding source domain node and target domain node are connect as anchor chain.
Popular says, predicting that the anchor chain between source domain figure and aiming field and figure connects that is to say that find may corresponding same entity
Source domain node and target domain node.
In the embodiment of the present invention, (anchor chain predicted connects) is connect in order to obtain pseudo- anchor chain, introduces cross-domain similarity, it should
Cross-domain similarity can be expressed by cosine similarity.
Wherein it is possible to by the source domain node diagnostic and aiming field node diagnostic set feature in source domain node diagnostic set
Aiming field node diagnostic carry out combination of two, obtain the combination of multiple node diagnostics;Later, a node diagnostic is chosen to combine, and
The source domain node diagnostic in the node diagnostic combination chosen is mapped according to the generator after training, maps that target
Feature space where domain node feature obtains mapping node feature;Then, according to the mapping node feature and the node chosen
Feature combination except other aiming field node diagnostics, and choose node diagnostic combination in aiming field node diagnostic with it is other
The cosine similarity of aiming field node diagnostic, come express mapping node feature and choose node diagnostic combination in target domain node
The cross-domain similarity of feature.
In this way, continuing that other node diagnostics is chosen to combine, until getting the cross-domain similarity of all node diagnostic combinations.
After the cross-domain similarity for getting all node diagnostic combinations, cross-domain similarity highest will be corresponded in the embodiment of the present invention
The corresponding source domain node and target domain node of node diagnostic combination as predicting that obtained anchor chain connects.
In one embodiment, (anchor chain predicted connects) is connect in order to obtain the pseudo- anchor chain of high quality, " according to reflecting
The cosine similarity of other aiming field node diagnostics except node diagnostic is combined with the node diagnostic chosen is penetrated, and choose
The cosine similarity of aiming field node diagnostic and other aiming field node diagnostics in node diagnostic combination, it is special to obtain the node chosen
Sign combines corresponding cross-domain similarity ", comprising:
(1) determining other aiming field node diagnostics with the highest preceding preset quantity of mapping node feature cosine similarity,
And obtain the corresponding preceding preset quantity of mapping node feature other aiming field node diagnostics and mean cosine similarity,
Obtain the first mean cosine similarity;
(2) the highest preceding present count of aiming field node diagnostic cosine similarity in combining with the node diagnostic chosen is determined
Amount other aiming field node diagnostics, and obtain choose node diagnostic combination in aiming field node diagnostic it is corresponding before
Other aiming field node diagnostics of preset quantity and mean cosine similarity, obtain the second mean cosine similarity;
(3) according to aiming field node diagnostic and the cosine similarity of mapping node feature in the node diagnostic combination chosen,
It is corresponding cross-domain similar to obtain the node diagnostic combination chosen for first mean cosine similarity and the second mean cosine similarity
Degree.
It in the embodiment of the present invention, is connect in order to obtain the pseudo- anchor chain of high quality, only retains source domain node and aiming field section
Mutual immediate neighbor node in point, but in view of immediate neighbor node is usually asymmetric, and this can shadow
The effect for ringing figure alignment, is further introduced into cross chart similitude to mitigate this central issue.
Wherein, other aiming field node diagnostics except being combined according to mapping node feature with the node diagnostic chosen
Cosine similarity, and choose node diagnostic combination in aiming field node diagnostic and other aiming field node diagnostics cosine phase
Like degree, when obtaining the node diagnostic corresponding cross-domain similarity of combination chosen, can determine first and mapping node feature cosine
Other aiming field node diagnostics of the highest preceding preset quantity of similarity, and obtain corresponding preceding default of mapping node feature
Other aiming field node diagnostics of quantity and mean cosine similarity, obtain the first mean cosine similarity.It is expressed as
Formula:
Wherein, WzsIndicate mapping node feature, K indicates that preset quantity (can be by those of ordinary skill in the art according to reality
Need value), rT(Wzs) indicate the first mean cosine similarity,It indicates with mapping node feature cosine similarity most
Cosine similarity is sought in target domain node corresponding to other aiming field node diagnostics of high preceding preset quantity, cos () expression.
Similarly, the aiming field node diagnostic cosine phase in combining with the node diagnostic chosen also is determined in the embodiment of the present invention
Like the other aiming field node diagnostics for spending highest preceding preset quantity, and obtain the aiming field section in the node diagnostic combination chosen
Other aiming field node diagnostics of the corresponding preceding preset quantity of point feature and mean cosine similarity, it is average to obtain second
Cosine similarity rS(zt)。
After acquiring the first mean cosine similarity and acquiring the second mean cosine similarity, i.e. basis
Aiming field node diagnostic is similar with the cosine similarity of mapping node feature, the first mean cosine in the node diagnostic combination chosen
Degree and the second mean cosine similarity obtain the node diagnostic chosen and combine corresponding cross-domain similarity, shown in following formula:
CGSS(Wzs, Zt)=2cos (Wzs,Zt)-rT(Wzs)-rs(Zt);
Wherein, CGSS (Wzs, Zt) aiming field node diagnostic and mapping node are special in the node diagnostic combination chosen that indicates
The cross-domain similarity of sign.
In one embodiment, " corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic are connect according to anchor chain
Generator afterwards, the generator after being optimized ", comprising:
So that it is European between corresponding source domain node diagnostic and aiming field node diagnostic to predict that obtained anchor chain connects
Distance is recently optimization aim, is optimized to the generator after training, the generator after being optimized.
In the embodiment of the present invention, when being optimized to the generator after training, shown in following formula:
Wherein, W* indicates that the generator of optimization, W indicate the generator before optimization, is the mapping matrix of d*d dimension, | |
Wzs-Zt||2Indicate the obtained anchor chain of prediction connect between corresponding source domain node diagnostic and aiming field node diagnostic it is European away from
From.
It should be noted that 205 can recycle execution, until meeting preset optimal conditions, wherein for the optimization item
It is not particularly limited, can be set according to actual needs by those of ordinary skill in the art in the setting embodiment of the present invention of part
It sets, for example, it is to optimize N times to the generator after training that optimal conditions, which can be set, N is positive integer.
In 206, according to the generator after optimization by source domain figure source domain node and aiming field figure in aiming field section
Point alignment.
In the embodiment of the present invention, optimization to the generator after training is completed, and after the generator after being optimized, i.e.,
The source domain node in source domain figure can be aligned with the target domain node in aiming field figure according to the generator after optimization.For example, asking
Referring to Fig. 3, the figure that source domain figure is social networks A indicates that the figure that aiming field figure is social networks B indicates, in source domain figure
Source domain node 1, the aiming field node diagnostic for being mapped to the source domain node diagnostic of the source domain node 1 according to the generator after optimization
The feature space at place, determine with the maximum aiming field node diagnostic of source domain node diagnostic cosine similarity after the mapping,
By the corresponding target domain node 1 of the aiming field node diagnostic as with the matched target domain node of source domain node 1, identify source domain
Same user " user A " in node 1 and corresponding two social networks of target domain node 1.
In one embodiment, in order to more accurately be aligned source domain node and target domain node, " according to the life after optimization
Grow up to be a useful person and be aligned the source domain node in source domain figure with the target domain node in aiming field figure ", comprising:
(1) for any source domain node in source domain figure, the source domain node and target are obtained according to the generator after optimization
The cross-domain similarity of each target domain node in the figure of domain;
(2) the cross-domain maximum target domain node of similarity will be corresponded to as the matched target domain node of foregoing source domain node.
In the embodiment of the present invention, source domain node and target domain node can be measured using the cross-domain similarity introduced before
Whether match.Wherein, for any source domain node in source domain figure, the source domain node and mesh are obtained according to the generator after optimization
The cross-domain similarity for marking each target domain node in the figure of domain that is to say the source domain node after the generator after acquisition is optimized maps
Source domain node diagnostic, the cross-domain similarity between the aiming field node diagnostic of each target domain node specifically can refer to above
Embodiment is accordingly implemented, and details are not described herein again.
It should be noted that in the embodiment of the present invention will the corresponding cross-domain maximum target domain node of similarity as with it is aforementioned
The target and node of source domain node matching, to realize the alignment of source domain node and target domain node.
In one embodiment, " according to the generator after optimization by source domain figure source domain node and aiming field figure in mesh
Mark domain node alignment " before, further includes:
(1) processing is orthogonalized to the generator after optimization, obtains the generator of orthogonalization;
" being aligned the node in source domain figure and aiming field figure according to the generator after optimization " includes:
(2) node in source domain figure and aiming field figure is aligned according to the generator of orthogonalization.
In order to further increase figure alignment accuracy, in the present invention is implemented, also to the generator after optimization into
Row orthogonalization process, so that generator is able to maintain close to orthogonal matrix.
Being orthogonalized processing to the generator after optimization can indicate are as follows:
W←(1+β)W-β(WWT)W;
Wherein, the W in left side indicates that the generator of orthogonalization, the W on right side indicate that the generator before orthogonalization (utilizes anchor chain
Connect as the generator after supervised learning optimization), β indicates preset step-length, and T indicates transposition.It can be by right according to orthogonality constraint(Zt indicates that the anchor chain that prediction obtains connects the source domain node diagnostic of corresponding source domain node,Indicate that the anchor chain of prediction connects
The transposition of the aiming field node diagnostic of corresponding target domain node) singular value decomposition obtain closed-form solution, it is as follows:
Wherein, SVD () indicates singular value decomposition.
By being orthogonalized processing to the generator after optimization, the generator of orthogonalization is obtained, so that generator is kept
Close to orthogonal matrix.The node in source domain figure and aiming field figure can be aligned according to the generator of orthogonalization, specifically may be used as a result,
Accordingly to implement referring to above embodiments, details are not described herein again.
In one embodiment, " unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field
Figure carries out unsupervised learning and obtains aiming field node diagnostic set " before, further includes:
(1) anchor chain judged whether there is between known source domain figure and aiming field figure connects;
(2) if it does not exist, then unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field
Figure carries out unsupervised learning and obtains aiming field node diagnostic set;
(3) if it exists, then supervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field figure
It carries out supervised learning and obtains aiming field node diagnostic set, and be transferred to execution 203.
In the embodiment of the present invention, after getting the source domain figure and aiming field figure that need to be aligned, not immediately to source domain
Figure carries out unsupervised learning and obtains source domain node diagnostic set, and carries out unsupervised learning to aiming field figure and obtain aiming field section
Point feature set, but it is first determined whether there are the anchor chains between known source domain figure and aiming field figure to connect, and if it exists, it is right again
Source domain figure carries out unsupervised learning and obtains source domain node diagnostic set, and carries out unsupervised learning to aiming field figure and obtain target
Domain node characteristic set, and the aiming field node diagnostic set that is obtained using unsupervised learning and source domain node diagnostic set are instructed
Practice and optimize for being aligned source domain figure and aiming field map generalization device, and be aligned, the correlation that specifically can refer to above embodiments is retouched
It states, details are not described herein again.
It, then can will be unsupervised in the present invention in addition, the anchor chain between known source domain figure and aiming field figure connects if it exists
Study replaces with supervised learning, for example, carrying out supervised learning to source domain figure using figure convolutional network obtains source domain node spy
Collection is closed, and is carried out supervised learning to aiming field figure using figure convolutional network and obtained aiming field node diagnostic set.
From the foregoing, it will be observed that obtaining the source domain figure and aiming field figure for needing to be aligned, source domain figure includes multiple in the embodiment of the present invention
Source domain node, aiming field figure include multiple target domain nodes;And unsupervised learning is carried out to source domain figure and obtains source domain node spy
Collection is closed, and is carried out unsupervised learning to aiming field figure and obtained aiming field node diagnostic set;And building generates confrontation net
Network, generating confrontation network includes the generation for feature space where source domain node diagnostic is mapped to aiming field node diagnostic
Device, and the arbiter for distinguishing source domain node diagnostic and aiming field node diagnostic after mapping;And according to source domain node
Characteristic set and aiming field node diagnostic set carry out alternately training to generator and arbiter, the generation after being trained
Device;And it is connect according to the generator prediction source domain figure after training and the anchor chain between aiming field figure, and connect according to anchor chain corresponding
Source domain node diagnostic and aiming field node diagnostic optimization training after generator, the generator after being optimized;And according to
Source domain node in source domain figure is aligned by the generator after optimization with the target domain node in aiming field figure.Hereby it is achieved that nothing
Figure alignment under surveillance requirements, gets rid of the dependence in existing figure alignment techniques to labeled data, not only reduces realization figure pair
Neat human cost also improves the efficiency of figure alignment.
Embodiment two,
Citing, is described in further detail by the method according to described in preceding embodiment below.
As shown in figure 4, the detailed process of the figure alignment schemes can be such that
401, terminal generates source domain figure according to the first social networks, and generates aiming field figure according to the second social networks,
Source domain figure includes the source domain node of user in corresponding first social networks, and aiming field figure includes user in corresponding second social networks
Target domain node.
Currently, people are not limited solely to using a social networks, but can use simultaneously with the prosperity of social networks
Multiple social networks are to enjoy more applications, for example, same user can pay close attention to some stars, writer either on pushing away spy
It is main etc. to pay close attention to different videos on YouTube by interested people and thing etc..These social networks are often mutually indepedent, still
It can reflect the information of user jointly again, this is provided in conjunction with isomery social networks and to excavate the information that they are implied jointly
Chance and challenge.The first step that same user is many data mining tasks is found in different social networks, for example, not
With social networks in determine same user, can excavate user information in conjunction with more resources, such issues that referred to as society
Hand over network alignment.And social networks can actually be expressed as figure, therefore social networks alignment problem can also be expressed as figure pair
Neat problem.
In the embodiment of the present invention, when needing to be aligned the social networks of two isomeries, it can be expressed first
For figure.Wherein, the first social networks and the second social networks are isomery social networks, it should be noted that first and second simultaneously
It is non-to refer in particular to a certain social networks, it is only used for distinguishing two isomery social networks, the first social networks and the second social networks can be with
For any two isomery social networks.For example, the first social networks can be to push away spy, nervus opticus network can be facebook etc..
When the social networks for needing to be aligned is expressed as figure, the available user information in the first social networks,
The user information describes the relationship between different user and different user, as a result, can be according to the user information by first
Social networks is expressed as multiple nodes, and connects the side of these nodes, obtains the social network diagram of the first social networks, the society
Handing over the node in network is the user represented in the first social networks, and the side between two nodes represents two users
Between relationship, in the embodiment of the present invention, the social network diagram that the first social networks is expressed is denoted as source domain figure, it is therein
Node is source domain node.
Similarly, the user information in the second social networks can also be got, and social by second according to the user information
Network is expressed as multiple nodes, and connects the side of these nodes, obtains the social network diagram of the second social networks, the social network
Node in network is the user represented in the second social networks, and the side between two nodes represents between two users
The social networks that second social networks is expressed in the embodiment of the present invention, is denoted as target and figure, node therein is by relationship
For target domain node.
It should be noted that source domain figure and the alignment of aiming field figure, understanding that can be popular are as follows: by the source domain section in source domain figure
Point is mapped in aiming field figure, finds matched target domain node.
402, terminal carries out unsupervised learning to source domain figure and obtains source domain node diagnostic set, and carries out to aiming field figure
Unsupervised learning obtains aiming field node diagnostic set.
Further learn the character representation of source domain figure and aiming field figure by unsupervised learning.Wherein, use is identical
Unsupervised learning algorithm to carry out unsupervised learning to source domain figure and aiming field figure, and the character representation for obtaining source domain figure is source domain section
Point feature set, the character representation for obtaining aiming field figure is aiming field node diagnostic set.
It should be noted that for which kind of unsupervised learning method to learn to obtain source domain figure using in the embodiment of the present invention
It with the character representation of aiming field figure, can be chosen according to actual needs by those of ordinary skill in the art, the present invention is implemented
This is not specifically limited in example.
For example, unsupervised learning can be carried out to source domain figure using DeepWalk algorithm, source domain node diagnostic set is obtainedWherein, n indicates the number of source domain node in source domain figure,Indicate first source domain in source domain figure
The feature of node,Indicate the feature of n-th of source domain node in aiming field figure, and so on.
Likewise, unsupervised learning can be carried out to aiming field figure using DeepWalk algorithm, it is special to obtain target domain node
Collection is closedWherein, m indicates the number of target domain node in aiming field figure,Indicate aiming field figure
The feature of middle first aim domain node,Indicate the feature of n-th of target domain node in aiming field figure, and so on.
403, terminal building generates confrontation network, and is based on source domain node diagnostic set and aiming field node diagnostic set pair
Confrontation network is generated to be trained.
It should be noted that due to the source domain figure and aiming field figure isomery that need to be aligned, respectively to source domain figure and target
After domain figure carries out unsupervised learning, the source domain node diagnostic learnt is concentrated conjunction and aiming field node diagnostic set and will be located at
Different feature spaces.
In order to realize that figure is aligned, expectation source domain node diagnostic can be with target after being mapped by certain mapping matrix in the present invention
Domain node feature is similar as far as possible.Therefore, dual training is introduced, in the embodiment of the present invention, obtains the source of source domain figure in study
After the aiming field node diagnostic set of domain node characteristic set and aiming field figure, further building generates confrontation network, should
It generates confrontation network to be made of generator and arbiter, generator can be a mapping matrix, be used for source domain node diagnostic
Source domain node diagnostic in set is mapped to the feature space where aiming field node diagnostic, after arbiter is for distinguishing mapping
Source domain node diagnostic and aiming field node diagnostic.
It should be noted that in the embodiment of the present invention for construct which kind of generation confrontation network framework with no restrictions, can
It is chosen according to actual needs by those of ordinary skill in the art, for example, the generation confrontation network of standard can be constructed, it can also be with structure
Improved generation confrontation network is built, such as the generation based on Wasserstein distance fights network.
In the embodiment of the present invention, after the building for completing to generate confrontation network, net further is fought to the generation of building
Network is trained.It should be noted that that can regard asking for zero-sum two-person game as on the training essential for generating confrontation network
Topic.
Illustratively, by taking the generation for constructing standard fights network as an example, the loss function of arbiter are as follows:
Wherein, W indicates generator, θDIndicate that arbiter, n indicate the quantity of source domain node, m indicates the number of target domain node
Amount,Indicate i-th of source domain node diagnostic after generator maps,Indicate j-th of aiming field node diagnostic,It is special to indicate that arbiter distinguishes i-th of source domain node after generator maps as far as possible
Sign, vice versa.
It is corresponding, the loss function of generator are as follows:
In training, generator W and arbiter θ is optimized by alternating iterationDTo minimize the loss function of generatorWith the loss function of arbiterPopular says, i.e., when arbiter can not accurately be distinguished through life
Grow up to be a useful person mapping after source domain node diagnostic when, i.e., training complete.
404, terminal is according between the generator prediction source domain figure and aiming field figure in the generation confrontation network after training
Anchor chain connects, and is connect according to the anchor chain of the prediction and optimized to the generator after training, until meeting default optimal conditions, obtains
Generator after optimization.
It should be noted that since, in training generator, there is no to the source domain in source domain figure in the embodiment of the present invention
Target domain node in node and aiming field figure carries out one-to-one constraint, needs to introduce additional supervision message to obtain to training
Generator optimize, to realize the one-to-one alignment of source domain node and target domain node, to avoid dual training process
In potential mode avalanche problem.
Therefore, further according to the generator prediction source domain figure after training and between aiming field figure in the embodiment of the present invention
Anchor chain connects.It should be noted that anchor chain is connected in the source domain node and target domain node of known corresponding same entity.Due to this hair
Physical presence anchor chain does not connect in bright implementation, but predicts that the anchor chain for obtaining, therefore can obtaining the prediction connects and be denoted as pseudo- anchor chain
It connects.
It obtains that the puppet anchor chain can be connect to the generator for obtaining training as supervision message after pseudo- anchor chain connects in prediction
It optimizes, namely connects the source domain node diagnostic of the source domain node of instruction and the aiming field section of target domain node according to anchor chain
Point feature optimizes come the generator obtained to training.
405, terminal is according to the generator of optimization by the target domain node in the source domain node and aiming field figure in source domain figure
Alignment, identifies the same subscriber in the first social networks and the second social networks.
In the embodiment of the present invention, optimization to the generator after training is completed, and after the generator after being optimized, i.e.,
The source domain node in source domain figure can be aligned with the target domain node in aiming field figure according to the generator after optimization, in this way,
The source domain node and the corresponding user of target domain node matched are same user.
Embodiment three,
In order to better implement above method, the embodiment of the present invention also provides a kind of figure alignment means, the figure alignment means
Specifically it can integrate in the terminals such as such as mobile phone, tablet computer, laptop.
For example, as shown in figure 5, the figure alignment means may include that figure obtains module 501, feature learning module 502, net
Network constructs module 503, network training module 504, optimization module 505 and figure alignment module 506, as follows:
Figure obtains module 501, and for obtaining the source domain figure and aiming field figure that need to be aligned, source domain figure includes multiple source domain sections
Point, aiming field figure include multiple target domain nodes;
Feature learning module 502 obtains source domain node diagnostic set and right for carrying out unsupervised learning to source domain figure
Aiming field figure carries out unsupervised learning and obtains aiming field node diagnostic set;
Network struction module 503 generates confrontation network for constructing, and generating confrontation network includes for source domain node is special
The generator of feature space where sign is mapped to aiming field node diagnostic, and for distinguishes the source domain node diagnostic after mapping with
The arbiter of aiming field node diagnostic;
Network training module 504 is used for according to source domain node diagnostic set and aiming field node diagnostic set to generation
Device and arbiter carry out alternately training, the generator after being trained;
Optimization module 505, for being connect according to the generator prediction source domain figure after training and the anchor chain between aiming field figure, and
Generator after connecing corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic according to anchor chain, after being optimized
Generator;
Figure alignment module 506, for according to the generator after optimization by source domain figure source domain node and aiming field figure in
Target domain node alignment.
In one embodiment, when being connect according to the generator prediction source domain figure after training and the anchor chain between aiming field figure,
Optimization module 505 is used for:
By the target domain node in the source domain node diagnostic and aiming field node diagnostic set in source domain node diagnostic set
Feature carries out combination of two, obtains multiple node diagnostic combinations;
A node diagnostic is chosen to combine, and according to the generator after training to the source domain section in the node diagnostic combination chosen
Point feature is mapped, and mapping node feature is obtained;
The cosine phase of other aiming field node diagnostics except being combined according to mapping node feature with the node diagnostic chosen
Like degree, and choose node diagnostic combination in aiming field node diagnostic and other aiming field node diagnostics cosine similarity,
It obtains the node diagnostic chosen and combines corresponding cross-domain similarity;
Continue that other node diagnostics is chosen to combine, until getting the cross-domain similarity of all node diagnostic combinations, and will
The cross-domain highest node diagnostic combination of similarity corresponding source domain node and target domain node connect as anchor chain.
In one embodiment, other aiming field sections except being combined according to mapping node feature with the node diagnostic chosen
Aiming field node diagnostic and other aiming field node diagnostics in the cosine similarity of point feature, and the node diagnostic combination chosen
Cosine similarity, when obtaining the node diagnostic chosen and combining corresponding cross-domain similarity, optimization module 505 is used for:
Determining other aiming field node diagnostics with the highest preceding preset quantity of mapping node feature cosine similarity, and obtain
The other aiming field node diagnostics for the preceding preset quantity for taking mapping node feature corresponding and mean cosine similarity, obtain
First mean cosine similarity;
Determine the highest preceding preset quantity of aiming field node diagnostic cosine similarity in combining with the node diagnostic chosen
Other aiming field node diagnostics, and obtain corresponding preceding pre- of aiming field node diagnostic in the node diagnostic combination chosen
If other aiming field node diagnostics of quantity and mean cosine similarity, obtain the second mean cosine similarity;
According to aiming field node diagnostic and the cosine similarity of mapping node feature in the node diagnostic combination chosen, the
It is corresponding cross-domain similar to obtain the node diagnostic combination chosen for one mean cosine similarity and the second mean cosine similarity
Degree.
In one embodiment, according to the generator after optimization by source domain figure source domain node and aiming field figure in mesh
When marking domain node alignment, figure alignment module 506 is used for:
For any source domain node in source domain figure, any source domain node and aiming field are obtained according to the generator after optimization
The cross-domain similarity of each target domain node in figure;
Using the maximum target domain node of the cross-domain similarity of correspondence as the target domain node of any source domain node matching.
In one embodiment, corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic are being connect according to anchor chain
Generator afterwards, when generator after being optimized, optimization module 505 is used for:
So that the Euclidean distance that anchor chain connects between corresponding source domain node diagnostic and aiming field node diagnostic is recently
Optimization aim, optimizes the generator after training, the generator after being optimized.
In one embodiment, when obtaining the source domain figure and aiming field figure that need to be aligned, figure obtains module 501 and is used for:
The user information in the first social networks is obtained, and multiple sections are generated according to the user information in the first social networks
Point and the side for connecting different nodes, obtain source domain figure;
The user information in the second social networks is obtained, and multiple sections are generated according to the user information in the second social networks
Point and the side for connecting different nodes, obtain aiming field figure, and the second social networks and the first social networks are isomery social networks.
In one embodiment, according to the generator after optimization by source domain figure source domain node and aiming field figure in mesh
Before marking domain node alignment, optimization module 505 is also used to:
Processing is orthogonalized to the generator after optimization, obtains the generator of orthogonalization;
When being aligned the node in source domain figure and aiming field figure according to the generator after optimization, figure alignment module 506 is used
In:
The node in source domain figure and aiming field figure is aligned according to the generator of orthogonalization.
In one embodiment, unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field
Before figure progress unsupervised learning obtains aiming field node diagnostic set, feature learning module 502 is also used to:
The anchor chain judged whether there is between known source domain figure and aiming field figure connects;
If it does not exist, then unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field figure
It carries out unsupervised learning and obtains aiming field node diagnostic set;
If it exists, then supervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and to aiming field figure into
Row supervised learning obtains aiming field node diagnostic set, and is transferred to execution building and generates confrontation network.
It should be noted that being suitable for server in figure alignment means provided in an embodiment of the present invention and foregoing embodiments
Figure alignment schemes belong to same design, and specific implementation process is detailed in above method embodiment, and details are not described herein again.
Example IV,
The embodiment of the present invention also provides a kind of terminal, which can set for mobile phone, tablet computer, laptop etc.
It is standby.As shown in fig. 6, it illustrates the structural schematic diagrams of terminal involved in the embodiment of the present invention, specifically:
The terminal may include one or processor 601, one or more calculating of more than one processing core
The components such as memory 602, power supply 603 and the input unit 604 of machine readable storage medium storing program for executing.It will be understood by those skilled in the art that
The restriction of terminal structure shown in Fig. 6 not structure paired terminal may include than illustrating more or fewer components or group
Close certain components or different component layouts.Wherein:
Processor 601 is the control centre of the terminal, using the various pieces of various interfaces and the entire terminal of connection,
By running or execute the software program and/or module that are stored in memory 602, and calls and be stored in memory 602
Data, execute terminal various functions and processing data.
Memory 602 can be used for storing software program and module, and processor 601 is stored in memory 602 by operation
Software program and module, thereby executing various function application and data processing.In addition, memory 602 may include height
Fast random access memory, can also include nonvolatile memory, a for example, at least disk memory, flush memory device,
Or other volatile solid-state parts.Correspondingly, memory 602 can also include Memory Controller, to provide processor
The access of 601 pairs of memories 602.
Terminal further includes the power supply 603 powered to all parts, it is preferred that power supply 603 can pass through power-supply management system
It is logically contiguous with processor 601, to realize the functions such as management charging, electric discharge and power managed by power-supply management system.
The terminal may also include input unit 604, which can be used for receiving the number or character letter of input
Breath, and generation keyboard related with user setting and function control, mouse, operating stick, optics or trackball signal are defeated
Enter.
Although being not shown, terminal can also be including display unit etc., and details are not described herein.Specifically in the present embodiment, eventually
Processor 601 in end can be corresponding executable by the process of one or more application program according to following instruction
File is loaded into memory 602, and the application program of storage in the memory 602 is run by processor 601, to realize
Various functions are as follows:
The source domain figure and aiming field figure for needing to be aligned are obtained, source domain figure includes multiple source domain nodes, and aiming field figure includes more
A target domain node;Unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and nothing is carried out to aiming field figure
Supervised learning obtains aiming field node diagnostic set;Building generates confrontation network, and generating confrontation network includes for by source domain section
The generator of feature space where point feature is mapped to aiming field node diagnostic, and it is special for distinguishing the source domain node after mapping
It seeks peace the arbiter of aiming field node diagnostic;According to source domain node diagnostic set and aiming field node diagnostic set to generator
Alternately training is carried out with arbiter, the generator after being trained;According to the generator prediction source domain figure and aiming field after training
Anchor chain between figure connects, and the life after corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic is connect according to anchor chain
It grows up to be a useful person, the generator after being optimized;It will be in the source domain node and aiming field figure in source domain figure according to the generator after optimization
The alignment of target domain node.
It should be noted that terminal provided in an embodiment of the present invention is aligned with the figure suitable for terminal in foregoing embodiments
Method belongs to same design, and specific implementation process is detailed in above method embodiment, and details are not described herein again.
Embodiment five,
It will appreciated by the skilled person that all or part of the steps in the various methods of above-described embodiment can be with
It is completed by instructing, or relevant hardware is controlled by instruction to complete, which can store computer-readable deposits in one
In storage media, and is loaded and executed by processor.
For this purpose, the embodiment of the present invention provides a kind of storage medium, wherein it is stored with computer program, the computer program packet
The a plurality of instruction included can be loaded by the processor of terminal, to execute provided by the embodiment of the present invention suitable for terminal
Figure alignment schemes, such as:
The source domain figure and aiming field figure for needing to be aligned are obtained, source domain figure includes multiple source domain nodes, and aiming field figure includes more
A target domain node;Unsupervised learning is carried out to source domain figure and obtains source domain node diagnostic set, and nothing is carried out to aiming field figure
Supervised learning obtains aiming field node diagnostic set;Building generates confrontation network, and generating confrontation network includes for by source domain section
The generator of feature space where point feature is mapped to aiming field node diagnostic, and it is special for distinguishing the source domain node after mapping
It seeks peace the arbiter of aiming field node diagnostic;According to source domain node diagnostic set and aiming field node diagnostic set to generator
Alternately training is carried out with arbiter, the generator after being trained;According to the generator prediction source domain figure and aiming field after training
Anchor chain between figure connects, and the life after corresponding source domain node diagnostic and the optimization training of aiming field node diagnostic is connect according to anchor chain
It grows up to be a useful person, the generator after being optimized;It will be in the source domain node and aiming field figure in source domain figure according to the generator after optimization
The alignment of target domain node.
Wherein, which may include: read-only memory (ROM, Read Only Memory), random access memory
Body (RAM, Random Access Memory), disk or CD etc..
Storage medium provided by the embodiment of the present invention can be realized corresponding diagram alignment side provided by the embodiment of the present invention
Beneficial effect achieved by method is detailed in the embodiment of front, and details are not described herein.
It is provided for the embodiments of the invention a kind of figure alignment schemes, device and storage medium above and has carried out detailed Jie
It continues, used herein a specific example illustrates the principle and implementation of the invention, and the explanation of above embodiments is only
It is to be used to help understand method and its core concept of the invention;Meanwhile for those skilled in the art, according to the present invention
Thought, there will be changes in the specific implementation manner and application range, in conclusion the content of the present specification should not be construed as
Limitation of the present invention.
Claims (10)
1. a kind of figure alignment schemes characterized by comprising
The source domain figure and aiming field figure for needing to be aligned are obtained, the source domain figure includes multiple source domain nodes, the aiming field figure packet
Include multiple target domain nodes;
Unsupervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and the aiming field figure is carried out without prison
Educational inspector's acquistion is to aiming field node diagnostic set;
Building generates confrontation network, and the generation confrontation network includes special for source domain node diagnostic to be mapped to target domain node
The generator of feature space where sign, and the differentiation for distinguishing source domain node diagnostic and aiming field node diagnostic after mapping
Device;
According to the source domain node diagnostic set and the aiming field node diagnostic set to the generator and the differentiation
Device carries out alternately training, the generator after being trained;
Predict that the anchor chain between the source domain figure and the aiming field figure connects according to the generator after the training, and according to described
Anchor chain connects corresponding source domain node diagnostic and aiming field node diagnostic and optimizes the generator after the training, after being optimized
Generator;
According to the generator after the optimization by the source domain figure source domain node and the aiming field figure in aiming field section
Point alignment.
2. figure alignment schemes according to claim 1, which is characterized in that the generator according to after the training is predicted
The step of anchor chain between the source domain figure and the aiming field figure connects, comprising:
By the aiming field in the source domain node diagnostic and the aiming field node diagnostic set in the source domain node diagnostic set
Node diagnostic carries out combination of two, obtains multiple node diagnostic combinations;
A node diagnostic is chosen to combine, and according to the generator after the training to the source domain section in the node diagnostic combination chosen
Point feature is mapped, and mapping node feature is obtained;
More than other aiming field node diagnostics except being combined according to the mapping node feature with the node diagnostic chosen
Aiming field node diagnostic and other aiming field node diagnostics in string similarity and the node diagnostic combination chosen
Cosine similarity, the node diagnostic chosen described in acquisition combine corresponding cross-domain similarity;
Continue that other node diagnostics is chosen to combine, until getting the cross-domain similarity of all node diagnostic combinations, and will be cross-domain
The highest node diagnostic combination of similarity corresponding source domain node and target domain node connect as the anchor chain.
3. figure alignment schemes according to claim 2, which is characterized in that it is described according to the mapping node feature with it is described
The cosine similarity and the node diagnostic chosen of other aiming field node diagnostics except the node diagnostic combination chosen
The cosine similarity of aiming field node diagnostic and other aiming field node diagnostics in combination, the node chosen described in acquisition are special
Sign combines the step of corresponding cross-domain similarity, comprising:
Determining other aiming field node diagnostics with the highest preceding preset quantity of the mapping node feature cosine similarity, and obtain
The other aiming field node diagnostics for the preceding preset quantity for taking the mapping node feature corresponding and mean cosine similarity,
Obtain the first mean cosine similarity;
Determine the highest preceding preset quantity of aiming field node diagnostic cosine similarity in combining with the node diagnostic chosen
Other aiming field node diagnostics, and obtain described in choose node diagnostic combination in aiming field node diagnostic it is corresponding
Other aiming field node diagnostics of preceding preset quantity and mean cosine similarity, obtain the second mean cosine similarity;
According to aiming field node diagnostic and the cosine similarity of mapping node feature, institute in the node diagnostic combination chosen
The first mean cosine similarity and the second mean cosine similarity are stated, the node diagnostic combination chosen described in acquisition corresponds to
Cross-domain similarity.
4. figure alignment schemes according to claim 2, which is characterized in that the generator according to after the optimization is by institute
Stating the step of source domain node in source domain figure is aligned with the target domain node in the aiming field figure includes:
For any source domain node in the source domain figure, any source domain node is obtained according to the generator after the optimization
With the cross-domain similarity of target domain node each in the aiming field figure;
Using the maximum target domain node of the cross-domain similarity of correspondence as the target domain node of any source domain node matching.
5. figure alignment schemes according to claim 1-4, which is characterized in that it is described according to the anchor chain connect pair
The source domain node diagnostic and aiming field node diagnostic answered optimize the generator after the training, the step of the generator after being optimized
Suddenly, comprising:
So that the Euclidean distance that the anchor chain connects between corresponding source domain node diagnostic and aiming field node diagnostic is recently
Optimization aim optimizes the generator after the training, the generator after obtaining the optimization.
6. figure alignment schemes according to claim 1-4, which is characterized in that described to obtain the source domain for needing to be aligned
Figure and the step of aiming field figure include:
The user information in the first social networks is obtained, and multiple sections are generated according to the user information in first social networks
Point and the side for connecting different nodes, obtain the source domain figure;
Obtain the user information in the second social networks, and according to the user information in the second social networks generate multiple nodes with
And the side of the different nodes of connection, the aiming field figure is obtained, second social networks and first social networks are isomery
Social networks.
7. figure alignment schemes according to claim 1-4, which is characterized in that the life according to after the optimization
It grows up to be a useful person before the step of being aligned the source domain node in the source domain figure with the target domain node in the aiming field figure, also wraps
It includes:
Processing is orthogonalized to the generator after the optimization, obtains the generator of orthogonalization;
The generator according to after optimization by the source domain figure and the aiming field figure node be aligned the step of include:
The node in the source domain figure and the aiming field figure is aligned according to the generator of the orthogonalization.
8. figure alignment schemes according to claim 1-4, which is characterized in that described to carry out nothing to the source domain figure
Supervised learning obtains source domain node diagnostic set, and carries out unsupervised learning to the aiming field figure and obtain target domain node spy
Before the step of collection is closed, further includes:
The anchor chain judged whether there is between the known source domain figure and the aiming field figure connects;
If it does not exist, then unsupervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and to the target
Domain figure carries out unsupervised learning and obtains aiming field node diagnostic set;
If it exists, then supervised learning is carried out to the source domain figure and obtains source domain node diagnostic set, and to the aiming field
Figure carries out supervised learning and obtains aiming field node diagnostic set, and is transferred to and executes the building generation confrontation network.
9. a kind of figure alignment means characterized by comprising
Figure obtains module, and for obtaining the source domain figure and aiming field figure that need to be aligned, the source domain figure includes multiple source domain nodes,
The aiming field figure includes multiple target domain nodes;
Feature learning module obtains source domain node diagnostic set for carrying out unsupervised learning to the source domain figure, and to institute
It states aiming field figure progress unsupervised learning and obtains aiming field node diagnostic set;
Network struction module generates confrontation network for constructing, and it includes for by source domain node diagnostic that generations, which fights network,
The generator of feature space where being mapped to aiming field node diagnostic, and for distinguishing source domain node diagnostic and mesh after mapping
Mark the arbiter of domain node feature;
Network training module is used for according to the source domain node diagnostic set and the aiming field node diagnostic set to described
Generator and the arbiter carry out alternately training, the generator after being trained;
Optimization module, for predicting the anchor chain between the source domain figure and the aiming field figure according to the generator after the training
It connects, and corresponding source domain node diagnostic and aiming field node diagnostic is connect according to the anchor chain and optimize the generation after the training
Device, the generator after being optimized;
Figure alignment module, for according to the generator after the optimization by the source domain figure source domain node and the aiming field
Target domain node alignment in figure.
10. a kind of storage medium, which is characterized in that the storage medium is stored with a plurality of instruction, and described instruction is suitable for processor
It is loaded, to execute figure alignment schemes as claimed in any one of claims 1 to 8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910667549.9A CN110442758B (en) | 2019-07-23 | 2019-07-23 | Graph alignment method, device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910667549.9A CN110442758B (en) | 2019-07-23 | 2019-07-23 | Graph alignment method, device and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110442758A true CN110442758A (en) | 2019-11-12 |
CN110442758B CN110442758B (en) | 2022-05-06 |
Family
ID=68431285
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910667549.9A Active CN110442758B (en) | 2019-07-23 | 2019-07-23 | Graph alignment method, device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110442758B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112507247A (en) * | 2020-12-15 | 2021-03-16 | 重庆邮电大学 | Cross-social network user alignment method fusing user state information |
CN112667864A (en) * | 2021-03-17 | 2021-04-16 | 腾讯科技(深圳)有限公司 | Graph alignment method and device, electronic equipment and storage medium |
CN113176917A (en) * | 2020-01-27 | 2021-07-27 | 富士胶片商业创新有限公司 | System and method for contrasting network analysis and visualization |
CN113284150A (en) * | 2021-07-26 | 2021-08-20 | 常州微亿智造科技有限公司 | Industrial quality inspection method and industrial quality inspection device based on unpaired industrial data |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107480714A (en) * | 2017-08-09 | 2017-12-15 | 东北大学 | Across social network user recognition methods based on full visual angle characteristic |
US20180196860A1 (en) * | 2017-01-06 | 2018-07-12 | International Business Machines Corporation | Efficient representation, access and modification of variable length objects |
CN109949174A (en) * | 2019-03-14 | 2019-06-28 | 哈尔滨工程大学 | A kind of isomery social network user entity anchor chain connects recognition methods |
-
2019
- 2019-07-23 CN CN201910667549.9A patent/CN110442758B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180196860A1 (en) * | 2017-01-06 | 2018-07-12 | International Business Machines Corporation | Efficient representation, access and modification of variable length objects |
CN107480714A (en) * | 2017-08-09 | 2017-12-15 | 东北大学 | Across social network user recognition methods based on full visual angle characteristic |
CN109949174A (en) * | 2019-03-14 | 2019-06-28 | 哈尔滨工程大学 | A kind of isomery social network user entity anchor chain connects recognition methods |
Non-Patent Citations (1)
Title |
---|
王莉等: "网络媒体大数据中的异构网络对齐关键技术和应用研究", 《太原理工大学学报》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113176917A (en) * | 2020-01-27 | 2021-07-27 | 富士胶片商业创新有限公司 | System and method for contrasting network analysis and visualization |
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 |
CN112667864A (en) * | 2021-03-17 | 2021-04-16 | 腾讯科技(深圳)有限公司 | Graph alignment method and device, electronic equipment and storage medium |
CN113284150A (en) * | 2021-07-26 | 2021-08-20 | 常州微亿智造科技有限公司 | Industrial quality inspection method and industrial quality inspection device based on unpaired industrial data |
Also Published As
Publication number | Publication date |
---|---|
CN110442758B (en) | 2022-05-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110442758A (en) | A kind of figure alignment schemes, device and storage medium | |
Yan et al. | HiGCIN: Hierarchical graph-based cross inference network for group activity recognition | |
Shu et al. | Host–parasite: Graph LSTM-in-LSTM for group activity recognition | |
Li et al. | Lotteryfl: Personalized and communication-efficient federated learning with lottery ticket hypothesis on non-iid datasets | |
Song et al. | Jkt: A joint graph convolutional network based deep knowledge tracing | |
Guo et al. | Iterative context-aware graph inference for visual dialog | |
Feng et al. | Poi2vec: Geographical latent representation for predicting future visitors | |
Xie et al. | Representation learning of knowledge graphs with entity descriptions | |
CN107784124B (en) | LBSN (location based service) hyper-network link prediction method based on space-time relationship | |
Shandilya et al. | AI-assisted Computer Network Operations testbed for Nature-Inspired Cyber Security based adaptive defense simulation and analysis | |
Wang et al. | Inductive learning on commonsense knowledge graph completion | |
CN104424507B (en) | Prediction method and prediction device of echo state network | |
CN109960755B (en) | User privacy protection method based on dynamic iteration fast gradient | |
CN110413704A (en) | Entity alignment schemes based on weighting neighbor information coding | |
Haihong et al. | Theme and sentiment analysis model of public opinion dissemination based on generative adversarial network | |
Deac et al. | Xlvin: executed latent value iteration nets | |
Portegys | A maze learning comparison of Elman, long short-term memory, and Mona neural networks | |
Mei | Simulation of Chinese online teaching invocational colleges based on complex evolution and improved neural network | |
Fang et al. | URPI-GRU: An approach of next POI recommendation based on user relationship and preference information | |
Tan et al. | Meta-learning enhanced neural ODE for citywide next POI recommendation | |
Sun et al. | HVLM: exploring human-like visual cognition and language-memory network for visual dialog | |
Dabaghi-Zarandi et al. | Community detection in complex network based on an improved random algorithm using local and global network information | |
Liu et al. | Understanding information diffusion with psychological field dynamic | |
Wang et al. | Research on prediction of physical fitness test results in colleges and universities based on deep learning | |
Zec et al. | Specialized federated learning using a mixture of experts |
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 |