CN108053454A - A kind of graph structure data creation method that confrontation network is generated based on depth convolution - Google Patents
A kind of graph structure data creation method that confrontation network is generated based on depth convolution Download PDFInfo
- Publication number
- CN108053454A CN108053454A CN201711261769.9A CN201711261769A CN108053454A CN 108053454 A CN108053454 A CN 108053454A CN 201711261769 A CN201711261769 A CN 201711261769A CN 108053454 A CN108053454 A CN 108053454A
- Authority
- CN
- China
- Prior art keywords
- mrow
- graph structure
- structure data
- degree
- picture
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Processing Or Creating Images (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of graph structure data creation methods that confrontation network is generated based on depth convolution, belong to big data technical field.Depth convolution is generated confrontation network application to the generation field of graph structure data by the method for the present invention, and the method for the present invention is first by the graph structure data conversion network picture with uniqueness of reality；Then confrontation network is generated by depth convolution to learn to generate model；Finally, the generation model obtained by study generates the analogous diagram structured data of specific scale according to input parameter on demand.The present invention is based on the graph structure data creation methods that depth convolution generates confrontation network, mainly carry out tectonic model from real graph structure data extraction feature, so that the attribute that the analogous diagram structured data of generation possesses is more in line with the graph structure data of reality.
Description
Technical field
The invention belongs to big data technical fields, and confrontation network is generated based on depth convolution more particularly, to a kind of
Graph structure data creation method.
Background technology
Graph structure model (Graph-Structured Model) is seen everywhere in actual life, such as social networks,
Transportation network, information network etc..In the big data epoch, data are exactly assets, and most data resource rests in only a few
Enterprise's hand in.Which results in the situations of most researcher's data resource scarcities.Meanwhile at ultra-large graph structure data
The exploitation of reason system, it is also desirable to which ultra-large graph structure data are tested.It is the needs of in order to more than meeting, polytype
Graph structure data generation model is suggested.The purpose of graph structure data generation model is that generation possesses real graph structure data attribute
Analogous diagram structured data, meet the needs of to graph structure data research.Although graph structure data Study on Generation Model Program obtains very
Big development, but its there are the problem of do not solve still.For example, the attribute that the analogous diagram structured data of generation possesses deviates in fact
The graph structure data on border or the analogous diagram structured data of generation only possess specific attribute and are unable to be provided simultaneously with a variety of categories
The problems such as property.How to generate in this context be more in line with reality analogous diagram structured data be big data generation area research one
A urgent problem to be solved.
Existing graph structure data generation model is roughly divided into five types：Random Graph generation model, preference depend on generation
Model generates model and specific properties generation model based on optimization generation model, based on tensor.Random Graph generation model be by
It is proposed earliest graph structure data generation model, when new data point is added to graph structure data, the point and graph structure data
In each point between there are the probability on side be impartial p.The model manipulation is simple, but there are the problem of it is extremely prominent, such as
The degree distribution of the graph structure data of model generation is Poisson distribution, and the degree distribution of real graph structure data is often power rate point
Cloth.Preference depends on generation model and is described as " richling is richer ", when adding new data point, the point to graph structure data
It is positively correlated with probability of the point in graph structure data there are side with the degree size put.Preference, which depends on generation model, can generate tool
The analogous diagram structured data for the attributes such as standby power rate and degree distribution are heavy-tailed, but the analogous diagram knot for possessing community structure attribute cannot be generated
Structure data.It is a kind of resources optimization model based on generation model is optimized, when adding new point to graph structure data, considers road
The reality factors such as journey distance, connection expense, the point always connect the minimum several points of expense.Specific properties generation model be in order to
The graph structure data studied specific attribute and generated, such as worldlet graph structure model, hierarchy chart structural model etc..
Four kinds of graph structure data generation model described above is manually to set the parameters to the figure that generation possesses particular community
Structured data, and most common graph structure data generation model is the generation model based on tensor.Generation mould based on tensor
Type can generate analogous diagram structured data from the graph structure data learning characteristic in reality, and by tensor product.It should
Model has been incorporated into Graph500 benchmark.Generation model based on tensor can be drawn in two stages：In the feature learning stage, eigenmatrix is provided firstBy asking for maximum likelihood function argmaxθP (G | θ) asks for eigenmatrix θ.Second stage is to feature square
Battle array θ carries out tensor product generating probability matrix, and the element value of matrix is the probability value there are side between 2 points.Life based on tensor
Into model compared to more preceding four kinds of models, the analogous diagram structured data of generation is more in line with the graph structure data in reality.
The to sum up generation model of five types, the model of either artificial arrange parameter is still from the graph structure data of reality
Learning parameter model, there are deviation between the analogous diagram structured data attribute of generation and real graph structure data attribute, and
And the analogous diagram structured data of generation can not be provided simultaneously with a variety of attributes well.
The content of the invention
For the disadvantages described above or Improvement requirement of the prior art, the present invention provides one kind based on the generation confrontation of depth convolution
The graph structure data creation method of network, its object is to first by real graph structure data conversion grid knot with uniqueness
Composition piece；Then confrontation network is generated by depth convolution to learn to generate model；Finally, the generation mould by learning to obtain
Type generates the analogous diagram structured data of specific scale according to input parameter on demand, and the analogous diagram structured data thus generated is more preferable
The many attributes for possessing real graph structure data.
To achieve the above object, the present invention provides a kind of graph structure data lifes that confrontation network is generated based on depth convolution
Into method, the described method includes：
(1) by the graph structure data conversion of reality into network picture with uniqueness；
(2) generate confrontation network with depth convolution and model training is carried out to network picture, obtain generation model；
(3) utilize and generate model generation analogous diagram structured data.
Further, the step (1) specifically includes：
(11) ordering rule is set, unique orderings are carried out to all the points in the graph structure data of reality；
(12) according to the adjacency matrix of the sequence structural map structured data at graph structure data midpoint；
(13) it will abut against matrix and be converted into network picture.
Further, the step (11) specifically includes：
(111) judge the directionality of the graph structure data of reality, if non-directed graph, then enter step (112)；Otherwise enter
Step (113)；
(112) degree of all the points in graph structure data is calculated, and descending sort is carried out to all the points according to degree size, successively
It is inserted into queue；When identical there are the degree of multiple points, then the element by the adjoining list for spending identical point arranges in descending order,
It spends afterwards between identical point and mutually compares the degree of adjacent list element successively, the degree identical point that the degree of element is big is initially inserted into team
Row；
(113) degree, in-degree and the out-degree of all the points in graph structure data are calculated, and all the points are dropped according to degree size
Sequence sorts, and is sequentially inserted into queue；When identical there are the degree of multiple points, then the out-degree and in-degree of degree of comparison identical point, if going out
Degree then selects greatly the adjoining list of the then out-degree, otherwise selects the adjoining list of the in-degree；Again to the neighbour of all degree identical points
The element for connecing list carries out descending arrangement, spends afterwards between identical point and mutually compares the degree of adjacent list element successively, element
It spends big degree identical point and is initially inserted into queue.
Further, the sequence at graph structure data midpoint constructs adjacency matrix element specific rules in the step (12)
For：When then the corresponding element value of adjacency matrix is set to 1 there are a line between in graph structure data 2 points, 0 is otherwise set to.
5th, a kind of graph structure data creation method according to claim 2, which is characterized in that in the step (13)
The specific rules that adjacency matrix is converted into network picture are：Network picture is two-value picture, pixel in two-value picture
Gray value and adjacency matrix element value correspond.
Further, the step (2) is specially：
(21) noise vector Z is generated；
(22) maker G () prototype network structure, initiation parameter are set；Resolving device D () prototype network structure is set,
Initiation parameter；
(23) maker G () changes into picture P by deconvolution operation handlebar noise vector ZG(z)；
It (24) will be by network picture P that step (1) obtainsdataWith picture PG(z)Resolving device D () is input to be instructed
Practice, the parameter of update resolving device D () and maker G (), until the output valve D ()=0.5 of resolving device, by noise vector Z with
Maker model G () after undated parameter is preserved as graph structure data generation model.
Further, the method specifically declined in the step (24) with gradient updates resolving device D () and maker G ()
Parameter：
Update the parameter θ of resolving device D ()dFormula be：
Wherein, m is sample size；γ is Learning Step；It is parameter θdGrad；
Update the parameter θ of maker G ()gFormula be：
Wherein,It is parameter θgGrad.
Further, the step (3) is specially：Training is input to maker G () with noise vector Z to be emulated
Picture, then picture binaryzation will be emulated, afterwards by binaryzation emulation picture construction adjacency matrix, finally adjacency matrix is converted into
High-dimensional graph structure data.
In general, by the above technical scheme conceived by the present invention compared with prior art, there is following technology spy
Sign and advantageous effect：
(1) the method for the present invention generates confrontation network from the graph structure data learning feature of reality to generate with depth convolution
Analogous diagram structured data can make the analogous diagram structured data of generation preferably possess many attributes of real graph structure data；
(2) conversion of real graph structure data is had uniqueness network picture by the method for the present invention before model learning,
Make graph structure data characteristics more obvious, and be easier to feature extraction.
Description of the drawings
Fig. 1 is the overall framework schematic diagram of the method for the present invention；
Fig. 2 is the overview flow chart of the method for the present invention；
Fig. 3 is the flow chart that the method for the present invention sorts to graph structure data point；
Fig. 4 is the schematic diagram of the method for the present invention point sequence identical to non-directed graph moderate size；
Fig. 5 is the schematic diagram of the method for the present invention point sequence identical to digraph moderate size；
Fig. 6 is the structure chart of depth convolution generation confrontation network in the method for the present invention；
Fig. 7 is the network structure of depth convolution generation confrontation network generator in the method for the present invention；
Fig. 8 is the network structure of depth convolution generation confrontation network resolving device in the method for the present invention.
Specific embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, it is right
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below that
Conflict is not formed between this to can be combined with each other.
Depth convolution generation confrontation network, depth convolution generation confrontation network (Deep Convolutional are introduced first
Generative Adversarial Networks, DCGAN) it is the 2015 generation confrontation networks being suggested.The network mainly wraps
Include maker G () and resolving device D () two parts.The noise vector Z of input is carried out deconvolution operation generation by maker G ()
Picture PG(Z).Resolving device D () is to the picture P of realitydataWith the picture P of generationG(Z)Carry out convolution operation output scalar D (PG(Z))
With D (Pdata).Resolving device D () output valve is scalar, and the picture for judging input is real picture or the picture of generation.
The network is updated the parameter of maker G () and resolving device D () with the method that gradient declines, until parameter restrains.
Fig. 1 show a kind of totality for the graph structure data creation method that confrontation network is generated based on depth convolution of the present invention
Frame diagram mainly includes graph structure data and picture transform portion and depth convolution confrontation generation network portion.Graph structure data
It is responsible for the conversion between high-dimensional graph structure data and network picture with picture transform portion.The generation confrontation of depth convolution
Network portion is responsible for learning the feature of picture, obtains generation model.
Fig. 2 show a kind of implementation for the graph structure data creation method that confrontation network is generated based on depth convolution of the present invention
The entire flow figure of example, specifically includes following steps：
(1) customized rules, immobilization graph structure data make graph structure data and the picture of conversion have man-to-man pass
System.
The point of the graph structure data of reality is ranked up first, the detailed sequence flow of ordering rule such as Fig. 3 displayings
Figure, specifically includes following sub-step：
(11) it is non-directed graph or digraph to judge graph structure data, if non-directed graph goes to step (12), digraph turns
To step (14)；
(12) the size Node of the degree of all the points in non-directed graph is calculateddeg；
(13) according to NodedegSize carries out descending sort to all the points, is sequentially inserted into queue
QueueNode_Permutation.When there are the Node of multiple pointsdegPoint is inserted into interim set Set by identical situationdeg_equal,
And the adjacent column Table A dj_List to each being put in setnodeElement carries out the descending sort by degree size.Then compare successively
The Node for the adjoining list element each putdegValue, NodedegIt is more big that the point is just initially inserted into queue
QueueNode_Permutation；Concrete operations example as shown in Figure 4, it is equal for the degree of multiple points such as V1, V2 and V3, first this
The adjoining list element put a bit carries out descending sort according to degree size, V1 adjoining listing sequences { v3, v5, v7, v10 ... }, and V2 is adjacent
Connect listing sequence { v4, v9, v11, v8 ... }, V3 adjoinings listing sequence { v1, v6, v11, v8 ... }.First compare v1, v2 and v3
One element { v3, v4, v1 }, the degree of v4 is maximum for 7, so v2 is first inserted into QueueNodePermutationIn.V1 and v3
One element degree is equal, so comparing second element { v5, v6 } of v1 and v3 again, v5 degree is bigger, so v1 is inserted into
QueueNodePermutationIn, remaining point is ranked up with identical method；
(14) the degree Node of all the points of digraph structure data is calculateddeg, in-degree Nodein_degWith out-degree Nodeout_deg。
(15) according to NodedegSize carries out descending sort to all the points, is sequentially inserted into queue
QueueNode_Permutation.When there are the Node of multiple pointsdegIn the case of identical, point is inserted into interim set
Setdeg_equal.Max (Node are chosen to each putting rank of advanced units table in setin_deg, Nodeout_deg), if the in-degree of the point
It is larger, just choose in-degree adjacent column Table A dj_Listnode_inIt is compared, otherwise chooses out-degree adjacent column Table A dj_
Listnode_out.Descending sort to adjoining list element degree of the progress size that selection is each put in set, it is then more adjacent successively
Meet the Node of list elementdegValue, NodedegIt is more big that the point is just initially inserted into queue QueueNode_Permutation.Concrete operations
Example as shown in Figure 5, it is equal for the degree of multiple points such as V1, V2 and V3, first select in-degree or the larger list of out-degree, V1 choosings
Select in-degree list, V2 selection out-degree lists, V3 selection in-degree lists.The adjoining list element of reselection is dropped according to degree size
Sequence sorts, V1 adjoining listing sequences { v12, v5, v7, v10 ... }, and V2 adjoinings listing sequence { v4, v9, v11, v8 ... }, V3 is adjacent
Connect listing sequence { v12, v6, v11, v8 ... }.First compare first element { v12, v4, v12 } of v1, v2 and v3, the degree of v4 is 7
Maximum, so v2 is first inserted into QueueNodePermutationIn.V1 and first element degree of v3 are equal, thus compare again v1 and
Second element { v5, v6 } of v3, v5 degree is bigger, so v1 is inserted into QueueNodePermutationIn, with identical method pair
Remaining point is ranked up；
(2) according to QueueNode_PermutationPut in order the adjacency matrix Adj_Matrix of structural map structured data.Cause
It is mainly generated for this method and haves no right graph structure data, therefore the rule of the element value of adjacency matrix is set：Exist between 2 points
Then its value is set to 1 to a line, is otherwise set to 0；
(3) adjacency matrix Adj_Matrix is changed into the picture P of networkdata；
(4) confrontation network is generated to carry out model learning with depth convolution, specific learning process is as shown in Figure 6.It is main
Comprising maker G () and resolving device D () two parts, maker G () model is mainly by deconvolution operation handlebar noise vector
Z is converted into picture, and picture is carried out constant output by resolving device D () model using convolution operation.Maker G () and resolving device D
As shown in Figure 7 and Figure 8, the sub-step specifically included is as follows for () specific network model：
(41) noise vector Z is generated；
(42) maker G () prototype network structure, initiation parameter are set；
(43) resolving device D () prototype network structure, initiation parameter are set；
(44) maker G () changes into picture P by deconvolution operation handlebar noise vector ZG(z)；
(45) the picture P of generationG(z)With picture PdataIt is input to resolving device D () to be trained, exports the scalar of resolution
Value D (Pdata) and D (PG(Z)).As resolving device output valve D (Pdata)=1 judges that the picture is real picture.As output valve D
(PG(Z))=0 judges that the picture is the picture of generation.The target of maker is to make D (PG(Z)) value becomes larger, for confusing resolving device.
The target of resolving device is to make D (PG(Z)) become smaller, it is easier to it is the picture generated to identify.As resolving device output valve D ()=0.5
Situation is optimal value, can not distinguish that picture is real or generates at this time；
(46) method declined with gradient updates the parameter of resolving device D () and maker G (), updates the parameter of resolving device
Formula (γ is Learning Step)：
Wherein, m is sample size；D is resolving device；θdIt is the parameter of resolving device；γ is Learning Step；It is parameter θdLadder
Angle value；
Update the parameter equation of maker：
Wherein, m is sample size；D is resolving device；G is maker；θgIt is the parameter of maker；γ is Learning Step；
It is for parameter θgGrad
(47) after the parameter for updating resolving device D () and maker G (), noise vector Z is re-entered into maker G
In (), step (44)-(47) are repeated, until parameter restrains, i.e. the output valve D ()=0.5 of resolving device, this moment resolving device D ()
The picture that input can not be told is real picture or the picture of generation；
(48) noise vector Z and Maker model G () is preserved.
Step 5 generates the emulation picture of particular size according to input parameter on demand, and the picture is converted into high-dimensional
Graph structure data；Specially when needing analogous diagram structured data, noise vector Z is input to maker G () and is emulated
Picture, then will emulation picture binaryzation (gray value of pixel is set to 1 more than 0 in picture, 0) gray value is just set to equal to 0, then
Adjacency matrix is generated according to the gray value of picture, adjacency matrix is finally converted into high-dimensional graph structure data, and (1 represents at 2 points
Between there are side, 0 side is not present).
More than content as it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention,
It is not intended to limit the invention, all any modification, equivalent and improvement made within the spirit and principles of the invention etc.,
It should all be included in the protection scope of the present invention.
Claims (8)
- A kind of 1. graph structure data creation method that confrontation network is generated based on depth convolution, which is characterized in that the method bag It includes：(1) by the graph structure data conversion of reality into network picture with uniqueness；(2) generate confrontation network with depth convolution and model training is carried out to the network picture, obtain generation model；(3) utilize and generate model generation analogous diagram structured data.
- 2. a kind of graph structure data creation method according to claim 1, which is characterized in that the step (1) is specifically wrapped It includes：(11) ordering rule is set, unique orderings are carried out to all the points in the graph structure data of reality；(12) according to the adjacency matrix of the sequence structural map structured data at graph structure data midpoint；(13) it will abut against matrix and be converted into network picture.
- 3. a kind of graph structure data creation method according to claim 2, which is characterized in that the step (11) is specifically wrapped It includes：(111) judge the directionality of the graph structure data of reality, if non-directed graph, then enter step (112)；Otherwise enter step (113)；(112) degree of all the points in graph structure data is calculated, and descending sort is carried out to all the points according to degree size, is sequentially inserted into To queue；When identical there are the degree of multiple points, then the element by the adjoining list for spending identical point arranges in descending order, afterwards It spends between identical point and mutually compares the degree of adjacent list element successively, the degree identical point that the degree of element is big is initially inserted into queue；(113) degree, in-degree and the out-degree of all the points in graph structure data are calculated, and descending row is carried out to all the points according to degree size Sequence is sequentially inserted into queue；When identical there are the degree of multiple points, then the out-degree and in-degree of degree of comparison identical point, if out-degree is big The adjoining list of the then out-degree is then selected, otherwise selects the adjoining list of the in-degree；Again to the adjacent column of all degree identical points The element of table carries out descending arrangement, spends afterwards between identical point and mutually compares the degree of adjacent list element successively, and the degree of element is big Degree identical point be initially inserted into queue.
- A kind of 4. graph structure data creation method according to claim 2, which is characterized in that figure knot in the step (12) The sequence at structure data midpoint constructs adjacency matrix element specific rules：When between in graph structure data 2 points there are a line then The corresponding element value of adjacency matrix is set to 1, is otherwise set to 0.
- 5. a kind of graph structure data creation method according to claim 2, which is characterized in that adjacent in the step (13) The specific rules that matrix is converted into network picture are：Network picture is two-value picture, the ash of pixel in two-value picture Angle value and the value of adjacency matrix element correspond.
- 6. a kind of graph structure data creation method according to claim 1, which is characterized in that the step (2) is specially：(21) noise vector Z is generated；(22) maker G () prototype network structure, initiation parameter are set；Resolving device D () prototype network structure is set, initially Change parameter；(23) maker G () changes into picture P by deconvolution operation handlebar noise vector ZG(z)；It (24) will be by network picture P that step (1) obtainsdataWith picture PG(z)It is input to resolving device D () to be trained, more The parameter of new resolving device D () and maker G (), until the output valve D ()=0.5 of resolving device, by noise vector Z and update Maker model G () after parameter is preserved as graph structure data generation model.
- 7. a kind of graph structure data creation method according to claim 6, which is characterized in that specific in the step (24) The method update resolving device D () and the parameter of maker G () declined with gradient：Update the parameter θ of resolving device D ()dFormula be：<mrow> <mover> <mi>V</mi> <mo>^</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mi>m</mi> </mfrac> <msubsup> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>m</mi> </msubsup> <mo>&lsqb;</mo> <mi>log</mi> <mi> </mi> <mi>D</mi> <mrow> <mo>(</mo> <msup> <msub> <mi>P</mi> <mrow> <mi>d</mi> <mi>a</mi> <mi>t</mi> <mi>a</mi> </mrow> </msub> <mi>i</mi> </msup> <mo>)</mo> </mrow> <mo>+</mo> <mi>log</mi> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <mi>D</mi> <mo>(</mo> <mrow> <mi>G</mi> <mrow> <mo>(</mo> <msup> <mi>Z</mi> <mi>i</mi> </msup> <mo>)</mo> </mrow> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>&rsqb;</mo> <mo>,</mo> </mrow><mrow> <msub> <mi>&theta;</mi> <mi>d</mi> </msub> <mo>=</mo> <msub> <mi>&theta;</mi> <mi>d</mi> </msub> <mo>+</mo> <mi>&gamma;</mi> <mo>&dtri;</mo> <mover> <mi>V</mi> <mo>^</mo> </mover> <mrow> <mo>(</mo> <msub> <mi>&theta;</mi> <mi>d</mi> </msub> <mo>)</mo> </mrow> <mo>,</mo> </mrow>Wherein, m is sample size；γ is Learning Step；It is parameter θdGrad；Update the parameter θ of maker G ()gFormula be：<mrow> <mover> <mi>T</mi> <mo>^</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mi>m</mi> </mfrac> <msubsup> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>m</mi> </msubsup> <mo>&lsqb;</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <mi>D</mi> <mo>(</mo> <mrow> <mi>G</mi> <mrow> <mo>(</mo> <msup> <mi>Z</mi> <mi>i</mi> </msup> <mo>)</mo> </mrow> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>&rsqb;</mo> <mo>,</mo> </mrow><mrow> <msub> <mi>&theta;</mi> <mi>g</mi> </msub> <mo>=</mo> <msub> <mi>&theta;</mi> <mi>g</mi> </msub> <mo>-</mo> <mi>&gamma;</mi> <mo>&dtri;</mo> <mover> <mi>T</mi> <mo>^</mo> </mover> <mrow> <mo>(</mo> <msub> <mi>&theta;</mi> <mi>g</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow>Wherein,It is parameter θgGrad.
- 8. a kind of graph structure data creation method according to claim 1 or 7, which is characterized in that the step (3) is specific For：Training is input to maker G () with noise vector Z and obtains emulation picture, then picture binaryzation will be emulated, afterwards by two Value emulation picture construction adjacency matrix, is finally converted into adjacency matrix high-dimensional graph structure data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711261769.9A CN108053454B (en) | 2017-12-04 | 2017-12-04 | Method for generating graph structure data of countermeasure network based on deep convolution generation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711261769.9A CN108053454B (en) | 2017-12-04 | 2017-12-04 | Method for generating graph structure data of countermeasure network based on deep convolution generation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108053454A true CN108053454A (en) | 2018-05-18 |
| CN108053454B CN108053454B (en) | 2020-05-19 |
Family
ID=62122259
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711261769.9A Active CN108053454B (en) | 2017-12-04 | 2017-12-04 | Method for generating graph structure data of countermeasure network based on deep convolution generation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108053454B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109146868A (en) * | 2018-08-27 | 2019-01-04 | 北京青燕祥云科技有限公司 | 3D Lung neoplasm generation method, device and electronic equipment |
| CN109214422A (en) * | 2018-08-02 | 2019-01-15 | 深圳先进技术研究院 | Parking data method for repairing and mending, device, equipment and storage medium based on DCGAN |
| CN109447081A (en) * | 2018-11-14 | 2019-03-08 | 广东小天才科技有限公司 | A kind of acquisition methods and system of optical character identification data set |
| CN109886936A (en) * | 2019-01-28 | 2019-06-14 | 深圳大学 | A kind of low contrast defect inspection method and device |
| WO2020024206A1 (en) * | 2018-08-02 | 2020-02-06 | 深圳先进技术研究院 | Dcgan-based parking data repairing method and apparatus, and device and storage medium |
| CN111080727A (en) * | 2019-12-17 | 2020-04-28 | 华中科技大学鄂州工业技术研究院 | Color image reconstruction method and device and image classification method and device |
| US11507803B2 (en) | 2020-01-31 | 2022-11-22 | Rohde & Schwarz Gmbh & Co. Kg | System for generating synthetic digital data for data multiplication |
| US11551061B2 (en) | 2020-01-31 | 2023-01-10 | Rohde & Schwarz Gmbh & Co. Kg | System for generating synthetic digital data of multiple sources |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106997380A (en) * | 2017-03-21 | 2017-08-01 | 北京工业大学 | Imaging spectrum safe retrieving method based on DCGAN depth networks |
| CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
| CN107293289A (en) * | 2017-06-13 | 2017-10-24 | 南京医科大学 | A kind of speech production method that confrontation network is generated based on depth convolution |
| CN107358203A (en) * | 2017-07-13 | 2017-11-17 | 西安电子科技大学 | A kind of High Resolution SAR image classification method based on depth convolution ladder network |
| US20170344879A1 (en) * | 2016-05-31 | 2017-11-30 | Linkedln Corporation | Training a neural network using another neural network |
-
2017
- 2017-12-04 CN CN201711261769.9A patent/CN108053454B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170344879A1 (en) * | 2016-05-31 | 2017-11-30 | Linkedln Corporation | Training a neural network using another neural network |
| CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
| CN106997380A (en) * | 2017-03-21 | 2017-08-01 | 北京工业大学 | Imaging spectrum safe retrieving method based on DCGAN depth networks |
| CN107293289A (en) * | 2017-06-13 | 2017-10-24 | 南京医科大学 | A kind of speech production method that confrontation network is generated based on depth convolution |
| CN107358203A (en) * | 2017-07-13 | 2017-11-17 | 西安电子科技大学 | A kind of High Resolution SAR image classification method based on depth convolution ladder network |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109214422A (en) * | 2018-08-02 | 2019-01-15 | 深圳先进技术研究院 | Parking data method for repairing and mending, device, equipment and storage medium based on DCGAN |
| WO2020024206A1 (en) * | 2018-08-02 | 2020-02-06 | 深圳先进技术研究院 | Dcgan-based parking data repairing method and apparatus, and device and storage medium |
| CN109214422B (en) * | 2018-08-02 | 2020-05-22 | 深圳先进技术研究院 | Parking data repairing method, device, equipment and storage medium based on DCGAN |
| CN109146868A (en) * | 2018-08-27 | 2019-01-04 | 北京青燕祥云科技有限公司 | 3D Lung neoplasm generation method, device and electronic equipment |
| CN109447081A (en) * | 2018-11-14 | 2019-03-08 | 广东小天才科技有限公司 | A kind of acquisition methods and system of optical character identification data set |
| CN109447081B (en) * | 2018-11-14 | 2021-07-23 | 广东小天才科技有限公司 | Method and system for acquiring optical character recognition data set |
| CN109886936A (en) * | 2019-01-28 | 2019-06-14 | 深圳大学 | A kind of low contrast defect inspection method and device |
| CN111080727A (en) * | 2019-12-17 | 2020-04-28 | 华中科技大学鄂州工业技术研究院 | Color image reconstruction method and device and image classification method and device |
| CN111080727B (en) * | 2019-12-17 | 2023-03-21 | 华中科技大学鄂州工业技术研究院 | Color image reconstruction method and device and image classification method and device |
| US11507803B2 (en) | 2020-01-31 | 2022-11-22 | Rohde & Schwarz Gmbh & Co. Kg | System for generating synthetic digital data for data multiplication |
| US11551061B2 (en) | 2020-01-31 | 2023-01-10 | Rohde & Schwarz Gmbh & Co. Kg | System for generating synthetic digital data of multiple sources |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108053454B (en) | 2020-05-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108053454A (en) | A kind of graph structure data creation method that confrontation network is generated based on depth convolution | |
| CN104866578B (en) | A kind of imperfect Internet of Things data mixing fill method | |
| Courtney et al. | Dense power-law networks and simplicial complexes | |
| CN109543043A (en) | A kind of automatic monitoring method of the power communication big data mistake of knowledge based map reasoning | |
| CN109191345A (en) | A kind of cognitive diagnosis method of Student oriented cognitive process | |
| CN106874955A (en) | A kind of 3D shape sorting technique based on depth convolutional neural networks | |
| CN111309927A (en) | Knowledge graph mining-based personalized learning path recommendation method and system | |
| CN110288878A (en) | Adaptive learning method and device | |
| CN115659807A (en) | Method for predicting talent performance based on Bayesian optimization model fusion algorithm | |
| CN113449878B (en) | Data distributed incremental learning method, system, equipment and storage medium | |
| CN112784899B (en) | Method, device and equipment for mining frequent pattern of power transformation operation and maintenance knowledge and storage medium | |
| CN106886510A (en) | The method and device that a kind of chart shows | |
| CN111415265A (en) | Social relationship data generation method of generative confrontation network | |
| CN111539517A (en) | Graph convolution neural network generation method based on graph structure matrix characteristic vector | |
| CN114741591B (en) | Method for recommending learning path to learner and electronic equipment | |
| CN111582595B (en) | Emergency decision method and device for natural disasters and electronic equipment | |
| Soto et al. | Top‐k Based Adaptive Enumeration in Constraint Programming | |
| CN109345047B (en) | Group teaching content quantitative optimization method based on knowledge topological relation | |
| CN113761289A (en) | Method, frame, computer system and readable storage medium for drawing learning | |
| Capuano et al. | Multi-criteria fuzzy ordinal peer assessment for MOOCs | |
| Lovelace | On the complexity of scheduling university courses | |
| Ettarres | Evaluation of online assignments and quizzes using Bayesian Networks | |
| Škraba et al. | Hybridization of stochastic local search and genetic algorithm for human resource planning management | |
| CN116401309B (en) | Student personalized learning recommendation method and device based on local influence and deep preference propagation | |
| Popescu et al. | Learning Testing Model using Test Generators and Mobile Applications. |
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 |