CN109960755B - User privacy protection method based on dynamic iteration fast gradient - Google Patents
User privacy protection method based on dynamic iteration fast gradient Download PDFInfo
- Publication number
- CN109960755B CN109960755B CN201910125911.XA CN201910125911A CN109960755B CN 109960755 B CN109960755 B CN 109960755B CN 201910125911 A CN201910125911 A CN 201910125911A CN 109960755 B CN109960755 B CN 109960755B
- Authority
- CN
- China
- Prior art keywords
- network
- gradient
- matrix
- neural network
- node
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 22
- 239000011159 matrix material Substances 0.000 claims abstract description 52
- 238000013528 artificial neural network Methods 0.000 claims abstract description 16
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 27
- 230000004913 activation Effects 0.000 claims description 6
- 238000009795 derivation Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 2
- 239000000126 substance Substances 0.000 claims 1
- 238000004422 calculation algorithm Methods 0.000 description 28
- 230000008859 change Effects 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 238000003012 network analysis Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 239000013598 vector Substances 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 241001622623 Coeliadinae Species 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000005295 random walk Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a user privacy protection method based on dynamic iteration fast gradient, which comprises the following steps: (1) for an original network, adopting the original network and a corresponding true class standard training graph convolutional neural network, and determining graph convolutional neural network parameters; (2) according to the parameters of the graph convolution neural network, taking the cross entropy before and after the target node reconnection edge in the original network as a hidden target function, calculating the gradient information of the hidden target function to the adjacent matrix, then calculating the momentum information by integrating the last gradient information, and updating the adjacent matrix corresponding to the original network according to the obtained momentum information. The user privacy protection method can quickly and effectively hide the information of the user, and achieves protection of the user privacy.
Description
Technical Field
The invention belongs to the field of network privacy protection, and particularly relates to a user privacy protection method based on dynamic iteration fast gradient.
Background
In real life, there are many network structures, and a social network is one of them. Each node in the social network represents a person, and the connecting edges represent communication information or a friend relationship between the persons. The third party can extract the characteristics of each node according to the connection relationship among the nodes, and then perform various analyses on the nodes, such as dividing the nodes into different clusters or predicting people or things which may be interested by the nodes. Many graph embedding models are proposed at present, and new methods are provided for analyzing the structure and characteristics of social networks.
The purpose of the graph embedding algorithm is to map the network structure into a low-dimensional space, thereby converting the traditional network analysis problem into a mathematical problem to be solved. Inspired by word2vec, the skip-gram model is widely applied in the field of graph embedding, resulting in a large number of graph embedding algorithms, such as deep walk, LINE and node2vec algorithms. They typically apply random walks to sequences of nodes and view these sequences of nodes as sentences in the word2vec model. Unlike graph embedding algorithms based on skip-gram models, Graph Convolution Networks (GCNs) are a graph convolution algorithm based on deep learning. It learns the local graph structure and node features and maps them into hidden layers. The graph-convolution network only needs a small number of node labels to effectively map all nodes in the network into a low-dimensional space, which is incomparable with other algorithms.
Although graph embedding algorithms have enjoyed great success in the field of network analysis, they also pose many information leakage problems and privacy protection is becoming an increasing concern. Unlike third parties desiring more information, many users do not want their own information to be used in commercial activities such as political communications of politicians, investigation activities of bedclothing policemen, and so on. None of these people want their tag attributes to be discovered and utilized by third parties. For such a situation, a network privacy protection algorithm is needed, and the network structure is slightly changed to realize larger change of the classification result of the target node, so that the results of downstream network analysis algorithms are influenced.
For a large network, the influence of different parts on the properties of the target node is different. Therefore, the suitable model is needed to find the continuous edge combinations with the largest influence on the target node attributes.
Disclosure of Invention
The invention aims to provide a user privacy protection method based on dynamic iteration fast gradient, which can quickly and effectively hide user information and realize protection of user privacy.
The technical scheme of the invention is as follows:
a user privacy protection method based on dynamic iteration fast gradient comprises the following steps:
(1) for an original Network, adopting the original Network and a corresponding true class standard training Graph Convolutional neural Network (GCN), and determining Graph Convolutional neural Network parameters;
(2) according to the parameters of the graph convolution neural network, taking the cross entropy before and after the target node reconnection edge in the original network as a hidden target function, calculating the gradient information of the hidden target function to the adjacent matrix, then calculating the momentum information by integrating the last gradient information, and updating the adjacent matrix corresponding to the original network according to the obtained momentum information.
In the user privacy protection method provided by the invention, the optimal solution of the reconnected edge is searched by means of GCN, so that the attribute change of the measurement target node is converted into the change problem of the characteristic vector of the measurement target node. Considering the situation that simple iteration based on gradient is easy to fall into local optimum, the last gradient information is also considered when the adjacency matrix is updated, and the two gradient information is combined into momentum information to determine the search direction, so that the algorithm is helped to jump out of a local optimum point and the mobility of the algorithm is increased.
Preferably, when training the graph convolution neural network, the forward transfer expression of the graph convolution neural network is as follows:
wherein X is a characteristic matrix of nodes in the original network G, is composed ofThe degree matrix of (c) is,for the original network G with a self-connected adjacency matrix, I is the identity matrix, A is the adjacency matrix of the original network G, W0∈RC×HAnd W1∈RH×|F|Respectively weighting an output layer to a hidden layer and a hidden layer to an output layer, wherein F is a class mark set, and F (-) and sigma (-) are a softmax activation function and a Relu activation function respectively;
and updating the parameters of the graph convolution neural network by taking the target function L as a target,
wherein, VLIs a node set containing class labels, F is the class label set, Y is the real class label matrix, if the l-th node vlClass (d) is denoted h, then Y lh1, otherwise Ylh=0,Y′lh(A) Is the output result of the graph convolution neural network.
When the convolutional neural network of the graph is trained, the fast gradient descent method is adopted to update the ith layer weight W in the convolutional neural network of the graphi:
Where η is the learning rate, WiIncluding the weight W of the ith output layer to the hidden layeri,0And the weight W of the ith output layer to the hidden layeri,1。
After training is finished, the graph convolution neural network is determined, and parameters of the graph convolution neural network determined when the confrontation network is generated subsequently are unchanged. When the target node in the original network is hidden, namely the adjacency matrix of the original network is regarded as a variable, and the variable is iteratively updated. Specifically, a hidden objective function L constructed for the target node ttComprises the following steps:
wherein, F is a class mark set, Y is a real class mark matrix, if the l-th node vlClass (d) is denoted h, then Y th1, otherwise Yth=0,Y′th(A) Output results of the graph convolution neural network;
hiding the objective function LtIndicating the difference size between the predicted class label and the real class label for the target node t. For privacy protection, the goal is to modify neighborsAnd connecting the matrix A to increase the value of the community hiding objective function as much as possible. Thus, the objective function L will be hiddentDerivation is carried out on the adjacent matrix A to obtain a gradient matrix
Where i and j are node indices.
Due to the gradient matrixAsymmetric, before computing momentum information, a gradient network g is also constructed:
in the gradient network g, any node pair has a value, and if the value is positive, it means that a connecting edge is added between the node pair to hide the objective function LtIncreasing, otherwise, indicating to delete the continuous edge between the node pair to hide the target function LtAnd is increased.
Wherein, the step of calculating the momentum information by synthesizing the last gradient information comprises the following steps:
wherein mu is an attenuation factor,||gk-1||1for the previous generation of gradient networks gk-1Is a norm of.
The invention does not use the gradient network g to directly update the adjacency matrix A, but uses the gradient network g of the (k-1) th iterationk-1And momentumCalculating new momentumUpdating the last adjacency matrix A by means of new momentumk-1. The method has the advantages that the new searching direction is not the strict current gradient direction, but a new direction obtained by integrating the previous gradient information is helpful for the algorithm to jump out of the local optimal solution.
The updating the adjacency matrix corresponding to the original network according to the obtained momentum information comprises:
from momentumIf the element with the maximum absolute value is positive, adding a connecting edge between the node pairs, otherwise, deleting the connecting edge between the node pairs to update the adjacency matrixTo obtainAttention is paid to
The invention has the following beneficial effects:
the adjacent matrix is updated based on the gradient momentum, so that the global optimal solution can be effectively searched, the local optimal solution is prevented from being trapped, and the mobility of the algorithm is enhanced. And (3) searching a reconnected edge strategy by using a depth model GCN, and converting the attribute change of the target node into the change of the distance between the characteristic vectors before and after the reconnected edge, so that the algorithm is more intuitive and is convenient to find the connected edge which has the largest influence on the target node more quickly and accurately. Finally, experimental results on a real data set show that the algorithm has good applicability and expandability, can quickly and effectively hide user information, and achieves protection of user privacy.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a flow chart of a dynamic iterative fast gradient-based user privacy protection method of the present invention;
FIG. 2 is a structural block diagram of the user privacy protection method based on dynamic iterative fast gradient according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
As shown in fig. 1 and fig. 2, the method for protecting user privacy based on dynamic iterative fast gradient provided by the present invention includes the following steps:
firstly, the embodiment trains the GCN by using the original network and the real label, so as to obtain the weights of the matrixes of each layer of the GCN, and the weights are not changed when the confrontation network is generated later. For a undirected network G, whose adjacency matrix is A, thenI is an identity matrix for the network G with self-connected adjacency matrices. The forward process of GCN can be simply expressed as:
wherein X is a characteristic matrix of the nodes, is composed ofDegree matrix of (W)0∈RC×HAnd W1∈RH×|F|Respectively weighting an output layer to a hidden layer and a hidden layer to an output layer, wherein F is a class mark set, and F (-) and sigma (-) are a softmax activation function and a Relu activation function respectively;
for node classification, the present invention uses cross entropy as an objective function, defined as shown in equation (2).
Wherein VLIs a node set containing class labels, F is the class label set, Y is the real class label matrix, wherein if the node vlClass (d) is denoted h, then Y lh1, otherwise Ylh=0,Y′lh(A) For the output result of GCN, A is the adjacency matrix of the original network.
In the mth iteration, the weights W of each layer in the neural network are updated by using a fast gradient descent methodi:
Where η is the learning rate.
After the GCN is trained, the adjacency matrix A is regarded as a variable and is iteratively updated. Firstly, defining a community hidden objective function L for a target node ttComprises the following steps:
the hidden objective function LtIndicating the difference size between the predicted class label and the real class label for the target node t. For privacy protection, the goal herein is to increase the value of the community hiding objective function as much as possible by modifying the adjacency matrix a. For this purpose, L istDerivation of the adjacency matrix A to obtain a gradient matrix
Because of the gradient matrixIs asymmetric, so a gradient network g is constructed as shown in equation (6).
In the gradient network g, any node has a value between it. If the value is positive, it indicates that adding a connecting edge between the node pair will cause the objective function LtIncreasing, otherwise indicating deletion of a connecting edge between the pair of nodes will cause LtAnd is increased. And the larger the absolute value of the value is, the more and more the edge pairs L are added and deletedtThe greater the effect of (c). Note that if the value is positive (negative), and there are (no) edges between this node pair, then such node pair is ignored.
The invention does not use the gradient network g to directly update the adjacency matrix A, but uses the gradient network g of the (k-1) th iterationk-1And momentumCalculating new momentumUpdating the last adjacency matrix A by means of new momentumk-1. The advantage of such calculation is that the new search direction is not the exact current gradient direction, but a new direction obtained by integrating the previous gradient information, so that it helps the algorithm to jump out of the local optimal solution. Defining the momentum produced by the k-th iterationAs follows.
Final driven loadSelecting the node pair with the maximum absolute value of the element, and updating the adjacency network adjacency matrix according to the positive and negative of the element with the maximum absolute valueTo obtainAttention is paid to
The steps of generating the countermeasure network are as follows:
a-1, training a GCN model by using an original network G;
a-5 fromThe node pair with the maximum absolute value is selected, and a new countermeasure network adjacency matrix is obtained according to the updating rule
and a-6, if K is less than K, repeating a-3 to a-5.
Simulation of experiment
Experiments the effectiveness of the MIFGS algorithm was tested using political blog data sets. The data set reflects political trends of blogs in the network, where nodes represent blogs and edges are automatically crawled from the top pages of blogs. There are 1490 nodes, 19090 edges, and 2 classes in the network.
And selecting privacy protection success rate (ASR) and the number of reconnected edges (AML) required by the successful average protection as evaluation indexes to measure the capability of each algorithm for hiding a certain node.
ASR: average protection success rate, i.e. the rate at which a certain graph embedding algorithm misclassifies the target nodes. Here, the disturbance size K is 1 to 20, and 20 are provided.
AML: and averaging the number of the reconnection edges required by successfully hiding the target node, wherein the upper limit of the reconnection edges is limited to 20, and if a certain node cannot be successfully hidden by modifying 20 reconnection edges, taking 20 as the number of the successfully hidden reconnection edges.
In order to verify that momentum in the user privacy protection method (MIFGS for short) based on dynamic iteration fast gradient provided by the invention really helps the algorithm jump out of a local optimum point, the MIFGS is directly used for updating an adjacent matrix by using a gradient networkIs compared with DICE. Briefly two comparison algorithms are as follows:
FGS: updating adjacency matrix of countermeasure network directly using gradient network gSpecifically, the node pair with the largest absolute value in the gradient network g is directly selected and updated according to the update rule mentioned above
The DICE: a simple heuristic network privacy protection algorithm randomly deletes the connecting edges of b (b is less than K) entry marking nodes if K is the disturbance size (the maximum iteration times in MIFGS and FGS), and randomly connects a target node with K-b nodes of other categories.
In the experiment, the final result of each algorithm is shown in table 1, taking μ ═ 0.5.
TABLE 1 politics blog data set Algorithm Final results
As can be seen from table 1, the indexes of the MIFGS are better than those of the FGS, which fully indicates that updating the countermeasure network with momentum information will help the algorithm jump out of the local optimum point faster, making the result more accurate. In addition, both the MIFGS algorithm and the FGS algorithm are obviously superior to the DICE algorithm, which shows that the conventional network analysis problem is feasible to be converted into a mathematical problem through the GCN, and more accurate results can be obtained.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (7)
1. A user privacy protection method based on dynamic iteration fast gradient comprises the following steps:
(1) for a blog network, the blog network is constructed by political blog data, wherein nodes represent blogs, edges represent blog articles automatically crawled from a first page of the blogs, the blog network and a corresponding real class mark training graph convolutional neural network are adopted, and graph convolutional neural network parameters are determined;
(2) and according to the parameters of the graph convolution neural network, taking the cross entropy before and after the target node reconnection edge in the blog network as a hidden target function, calculating the gradient information of the hidden target function to the adjacent matrix, then calculating momentum information by integrating the last gradient information, and updating the adjacent matrix corresponding to the blog network according to the obtained momentum information.
2. The dynamic iterative fast gradient-based user privacy protection method according to claim 1, wherein, when training the graph convolutional neural network, the forward transfer expression of the graph convolutional neural network is:
wherein X is a characteristic matrix of nodes in the blog network G, is composed ofThe degree matrix of (c) is,with self-connecting adjacency matrix for blog network G, I being identity matrix, A being adjacency matrix for blog network G, W0∈RC×HAnd W1∈RH×|F|Respectively weighting an output layer to a hidden layer and a hidden layer to an output layer, wherein F is a class mark set, and F (-) and sigma (-) are a softmax activation function and a Relu activation function respectively;
and updating the parameters of the graph convolution neural network by taking the target function L as a target,
wherein, VLIs a node set containing class labels, F is the class label set, Y is the real class label matrix, if the l-th node vlClass (d) is denoted h, then Ylh1, otherwise Ylh=0,Yl ′ h(A) Is the output result of the graph convolution neural network.
3. The dynamic iterative fast gradient-based user privacy protection method of claim 2, wherein a fast gradient descent method is used to update the ith layer weight W in the graph convolution neural network when training the graph convolution neural networki:
Where η is the learning rate, WiIncluding the weight W of the ith output layer to the hidden layeri,0And the weight W of the ith output layer to the hidden layeri,1。
4. The dynamic iterative fast gradient-based user privacy protection method according to claim 1, wherein a hidden objective function L is constructed for the target node ttComprises the following steps:
wherein the content of the first and second substances,f is a class mark set, Y is a real class mark matrix, if the l-th node vlClass (d) is denoted h, then Yth1, otherwise Yth=0,Y′th(A) Output results of the graph convolution neural network;
will hide the objective function LtDerivation is carried out on the adjacent matrix A to obtain a gradient matrix
Where i and j are node indices.
5. The dynamic iterative fast gradient-based user privacy protection method according to claim 1, wherein before computing the momentum information, a gradient network g is further constructed:
in the gradient network g, any node pair has a value, and if the value is positive, it means that a connecting edge is added between the node pair to hide the objective function LtIncreasing, otherwise, indicating to delete the continuous edge between the node pair to hide the target function LtAnd is increased.
6. The dynamic iterative fast gradient-based user privacy protection method of claim 5, wherein the integrating previous gradient information to compute momentum information comprises:
7. The dynamic iterative fast-gradient-based user privacy protection method as claimed in claim 6, wherein the updating the adjacency matrix corresponding to the blog network according to the obtained momentum information includes:
from momentumSelecting the node pair with the maximum element absolute value, if the element with the maximum element absolute value is positive, adding a connecting edge between the node pair, otherwise, deleting the connecting edge between the node pair to update the adjacency matrixTo obtainAttention is paid to
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910125911.XA CN109960755B (en) | 2019-02-20 | 2019-02-20 | User privacy protection method based on dynamic iteration fast gradient |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910125911.XA CN109960755B (en) | 2019-02-20 | 2019-02-20 | User privacy protection method based on dynamic iteration fast gradient |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109960755A CN109960755A (en) | 2019-07-02 |
CN109960755B true CN109960755B (en) | 2021-03-05 |
Family
ID=67023594
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910125911.XA Active CN109960755B (en) | 2019-02-20 | 2019-02-20 | User privacy protection method based on dynamic iteration fast gradient |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109960755B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110674869B (en) * | 2019-09-23 | 2023-04-18 | 腾讯科技(深圳)有限公司 | Classification processing and graph convolution neural network model training method and device |
CN111275189A (en) * | 2020-02-25 | 2020-06-12 | 支付宝(杭州)信息技术有限公司 | Network representation learning method, system and device |
CN111882054B (en) * | 2020-05-27 | 2024-04-12 | 杭州中奥科技有限公司 | Method for cross training of encryption relationship network data of two parties and related equipment |
CN111859454B (en) * | 2020-07-28 | 2024-03-29 | 桂林慧谷人工智能产业技术研究院 | Privacy protection method for defending link prediction based on graph neural network |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2010144645A2 (en) * | 2009-06-12 | 2010-12-16 | Brian Burk | Healthcare provider resources online |
CN103488637A (en) * | 2012-06-11 | 2014-01-01 | 北京大学 | Method for carrying out expert search based on dynamic community mining |
CN104834747A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Short text classification method based on convolution neutral network |
CN106097355A (en) * | 2016-06-14 | 2016-11-09 | 山东大学 | The micro-Hyperspectral imagery processing method of gastroenteric tumor based on convolutional neural networks |
CN106847302A (en) * | 2017-02-17 | 2017-06-13 | 大连理工大学 | Single channel mixing voice time-domain seperation method based on convolutional neural networks |
CN108196251A (en) * | 2017-12-25 | 2018-06-22 | 电子科技大学 | Accelerated iteration regularization super-resolution imaging method based on vector extrapolation |
CN108664687A (en) * | 2018-03-22 | 2018-10-16 | 浙江工业大学 | A kind of industrial control system space-time data prediction technique based on deep learning |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10614912B2 (en) * | 2014-08-17 | 2020-04-07 | Hyperfine, Llc | Systems and methods for comparing networks, determining underlying forces between the networks, and forming new metaclusters when saturation is met |
-
2019
- 2019-02-20 CN CN201910125911.XA patent/CN109960755B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2010144645A2 (en) * | 2009-06-12 | 2010-12-16 | Brian Burk | Healthcare provider resources online |
CN103488637A (en) * | 2012-06-11 | 2014-01-01 | 北京大学 | Method for carrying out expert search based on dynamic community mining |
CN104834747A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Short text classification method based on convolution neutral network |
CN106097355A (en) * | 2016-06-14 | 2016-11-09 | 山东大学 | The micro-Hyperspectral imagery processing method of gastroenteric tumor based on convolutional neural networks |
CN106847302A (en) * | 2017-02-17 | 2017-06-13 | 大连理工大学 | Single channel mixing voice time-domain seperation method based on convolutional neural networks |
CN108196251A (en) * | 2017-12-25 | 2018-06-22 | 电子科技大学 | Accelerated iteration regularization super-resolution imaging method based on vector extrapolation |
CN108664687A (en) * | 2018-03-22 | 2018-10-16 | 浙江工业大学 | A kind of industrial control system space-time data prediction technique based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN109960755A (en) | 2019-07-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109960755B (en) | User privacy protection method based on dynamic iteration fast gradient | |
CN111428147B (en) | Social recommendation method of heterogeneous graph volume network combining social and interest information | |
CN110555112B (en) | Interest point recommendation method based on user positive and negative preference learning | |
CN107145977B (en) | Method for carrying out structured attribute inference on online social network user | |
CN110097125B (en) | Cross-network account association method based on embedded representation | |
WO2023000574A1 (en) | Model training method, apparatus and device, and readable storage medium | |
CN108492200A (en) | A kind of user property estimating method and device based on convolutional neural networks | |
CN109190030B (en) | Implicit feedback recommendation method fusing node2vec and deep neural network | |
US9146941B2 (en) | Image tag pair graph for image annotation | |
CN112988917B (en) | Entity alignment method based on multiple entity contexts | |
CN109743196B (en) | Network characterization method based on cross-double-layer network random walk | |
CN110263236B (en) | Social network user multi-label classification method based on dynamic multi-view learning model | |
EP4350572A1 (en) | Method, apparatus and system for generating neural network model, devices, medium and program product | |
CN110502743A (en) | Social networks based on confrontation study and semantic similarity is across media search method | |
CN111274485A (en) | Personalized recommendation method based on community discovery | |
CN112580728B (en) | Dynamic link prediction model robustness enhancement method based on reinforcement learning | |
CN109271546A (en) | The foundation of image retrieval Feature Selection Model, Database and search method | |
CN109948242A (en) | Network representation learning method based on feature Hash | |
CN109933720B (en) | Dynamic recommendation method based on user interest adaptive evolution | |
CN111104975A (en) | Credit assessment model based on breadth learning | |
CN112800111B (en) | Position prediction method based on training data mining | |
CN113627550A (en) | Image-text emotion analysis method based on multi-mode fusion | |
CN110008411A (en) | It is a kind of to be registered the deep learning point of interest recommended method of sparse matrix based on user | |
CN105205075B (en) | From the name entity sets extended method of extension and recommended method is inquired based on collaboration | |
CN114281976A (en) | Model training method and device, electronic equipment and storage medium |
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 |