CN108509651A - The distributed approximation searching method with secret protection based on semantic consistency - Google Patents
The distributed approximation searching method with secret protection based on semantic consistency Download PDFInfo
- Publication number
- CN108509651A CN108509651A CN201810342218.3A CN201810342218A CN108509651A CN 108509651 A CN108509651 A CN 108509651A CN 201810342218 A CN201810342218 A CN 201810342218A CN 108509651 A CN108509651 A CN 108509651A
- Authority
- CN
- China
- Prior art keywords
- node
- transition matrix
- matrix
- sample
- approximation
- 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
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
- G06F21/6263—Protecting personal data, e.g. for financial or medical purposes during internet communication, e.g. revealing personal data from cookies
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Bioethics (AREA)
- General Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- Databases & Information Systems (AREA)
- Computer Security & Cryptography (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses the distributed approximation searching methods with secret protection based on semantic consistency.Classification marker is carried out to image, video, file first in the database of each node, initializes transition matrix and Lagrange multiplier, semantic consistency is then introduced and builds object function, solve above-mentioned object function, update transition matrix.Neighboring node is communicated, and judges whether the transition matrix of each node reaches unanimity, and updates Lagrange multiplier, finally carries out approximation search process.The problem of present invention solves large-scale data and is storing, and required scale is excessive when calculating, and centralization ground training algorithm model has been no longer appropriate for.And the communication between node is carried out by using transition matrix, the communication between node does not exchange raw information, can effectively solve the problems, such as that transmission communication is excessive, while can carry out effective protection to the data privacy on node.
Description
Technical field
The invention belongs to machine learning field, relate generally to ensure sample using learning distance metric under distributed environment
Consistency, and in particular to the distributed approximation searching method with secret protection based on semantic consistency.
Background technology
With the continuous development of social networks, e-commerce, mobile Internet etc., data need the scale of storage, processing
Increasing, one-of-a-kind system cannot be satisfied growing demand.The Internet companies such as Google, Alibaba are successfully expedited the emergence of
Cloud computing and this two big hot topics field of big data, cloud computing and big data are all the applications built on distributed storage.
The core of cloud storage is the large-scale distributed storage system of rear end, and big data not only needs to store the data of magnanimity, also to lead to
It crosses suitable frame and tool to analyze these data, obtains wherein useful part, if without distributed storage
It is just far from being and big data is analyzed.Although many years have been carried out in the research of distributed system, until in recent years, mutually
The rise for big data of networking is just so that distributed system is applied in engineering practice on a large scale.Distributed system is to utilize more
Collaborative computer solves calculating, the storage problem that single computer cann't be solved, and distributed system is maximum with one-of-a-kind system
Difference is that the scale of problem.The system that it is made of multiple nodes, often will be on a server or server
One process is known as a node, these nodes are generally not isolated, but are communicated by network, transmits information.
In addition, due to the fast development of the mobile terminals such as smart mobile phone, smart mobile phone stores the letters such as a large amount of picture, text and video
Breath, smart mobile phone can also regard an independent node as, between smart mobile phone by base station or between each other by point
Cloth cooperates to improve data-handling capacity.
Secret protection is an important research direction in data mining to be made just because of the extensive use of data mining
Secret protection is obtained greatly to be paid close attention to.In Distributed Calculation, mutual communication must be carried out between each node, in communication
In the process, it is possible to the case where there are node privacy leakages.How while protecting privacy distributed deposit is effectively performed
Storage and calculating are the hot spots studied now.Currently, the method for secret protection mainly has disturbance of data, adds under distributed environment
Close storage, multi-party computations etc..Disturbance of data refers to taking closing or hiding mode to initial data, generates new data
Source, but this mode can substantially reduce data accuracy, to impact analysis result.Encryption storage is a kind of generally acknowledged data guarantor
Shield technology mainly protects private data by way of encrypting and decrypting.Multi-party computations technology belongs to cryptography research
Scope goes out to protect the scheme of privacy agreement by the protocol construction on some bases.But algorithm above is required to increase additional meter
Calculation and storage overhead.
In addition, Euclidean distance widely used in machine learning cannot well between reflected sample semantic information, than
Such as " Ha Shiqi " and " chihuahua " belongs to classification " dog ", but " Ha Shiqi " may seem closer with " wolf ", their Europe
Family name's distance than the Euclidean distance of " Ha Shiqi " and " chihuahua " closer to, therefore pass through training study to one optimize conversion square
Battle array, new space is mapped to by transition matrix, and distance of the sample of classification of the same race between them is closer to this in new space
Sample is just more likely to find semantic neighbour, improves the precision of search.And mahalanobis distance can continuing to optimize by transition matrix, will be former
The feature space of beginning is mapped to new feature space so that distance of the sample with same tag in new feature space is most
It is possible small, it is constantly widened with not the distance between sample of isolabeling, is more in line with language in new feature space in this way
Adopted consistency.
In conclusion having the sample of same tag new for how to utilize under mahalanobis distance environment in the prior art
Transmission initial data be easy to cause information leakage between node when the smaller advantage of distance in feature space solves Distributed Calculation
The problem of still without disclosed disclosure.
Invention content
The purpose of the present invention is to provide a kind of, and the distributed approximation with secret protection based on semantic consistency is searched
Suo Fangfa is mainly used for solving image, and video, text equal samples number is big, can not accurately find semantic neighbour, be distributed simultaneously
Transmission initial data be easy to cause the leakage of information, the excessive problem of transmission quantity, the main mesh of this method between node when formula calculates
Be by distribution training, train to obtain the transition matrix of global optimization with lower computing cost, while protecting distribution
The data-privacy of each node in training, and realize the semantic consistency neighbor search of query sample.
In order to reach the above object, the present invention provides the distribution with secret protection based on semantic consistency is close
Like property searching method, this approach includes the following steps：
Step 1：Classification marker is carried out to image, video, file in the database of each node；
Step 2：Initialize transition matrix and Lagrange multiplier；
Step 3：It introduces semantic consistency and builds object function；
Step 4：Above-mentioned object function is solved, transition matrix is updated；
Step 5：Neighboring node is communicated, and judges whether the transition matrix of each node reaches unanimity, update Lagrange
Multiplier；
Step 6：Carry out approximation search process.
Further, in above-mentioned steps 1, it is assumed that share N number of node, each node corresponds to a database Xi, XiIndicate i-th
The database of a node, the database in different nodes are independent from each other, and shared letter is not intended between different nodes
It ceases, has L kind category labels in each database, different labels is stamped to different samples.
Further, in above-mentioned steps 2, initialization all is carried out to transition matrix and Lagrange multiplier on each node and is set
It sets, the transition matrix that initialization is arranged is the unit matrix of d × d dimensions, and the Lagrange multiplier of initialization is d × d dimensions
Full 0 matrix, d indicate the dimension of sample original feature space.
Further, in above-mentioned steps 3, semantic consistency is introduced in object function, by transition matrix by original spy
It levies in space reflection to new feature space so that the distance at sample to such center with same tag is close as far as possible,
The distance at different classes of center is remote as far as possible.
Preferably, considering the information of neighboring node simultaneously, punishment parameter is added can be with Accelerated iteration.
Further, in above-mentioned steps 4, due to transition matrix AiIt is positive semidefinite symmetrical matrix, can is A with feature decompositioni=
WiWi T, AiIndicate the transition matrix of i-th of node, WiIt is the matrix that a d × r is tieed up in i-th of node, r ＜ d, Wi TIndicate Wi's
Transposition.By the A in object functioniIt is converted to Ai=WiWi T, and solve Wi, then obtain transition matrix Ai。
Further, in above-mentioned steps 5, neighboring node there are two each nodes is set, passes through between neighboring node and converts square
Battle array exchanges information, and determines whether to update from new iteration according to transition matrix.All nodes may be constructed connected graph, i.e., arbitrarily
When two nodes can indirectly be connected by other nodes, if transition matrix reaches unanimity, deconditioning process.It is no
Then, Lagrange multiplier is updated, and repeats step 3.
Further, in above-mentioned steps 6, for a new query sample, it is input to some node, by converting square
It after battle array mapping, calculates in new feature space at a distance between query sample and the node other samples, f most before taking wherein
The result that sample corresponding to small distance is searched for as our approximations.
Compared with prior art, the present invention has following advantageous effects：
It is excessively unilateral when 1, solving many traditional methods for neighbor search, do not account for sample when finding neighbour
The semantic information that may have so that these algorithms bad problem of performance in the practical application of proximity search.
2, it solves large-scale data storing, required scale is excessive when calculating, and centralization ground training algorithm model is not
Suitable problem again.
3, the communication between node is carried out using transition matrix, the communication between node does not exchange raw information, can effectively solve
Transmission communicates excessive problem, while can carry out effective protection to the data privacy on node.
Description of the drawings
Fig. 1 is the system framework figure of this method；
Fig. 2 is the distribution training flow chart of this method；
Fig. 3 is the neighbor search flow chart of this method.
Specific implementation mode
Invention is described in further detail below in conjunction with Figure of description.
The system framework figure of this method as shown in Figure 1, entire procedure can be divided into distributed training process with it is approximate
Property search process.Detailed process difference it is as shown in Figures 2 and 3, wherein the first step to the 5th step by the way of as shown in Figure 2,
6th step is by the way of as shown in Figure 3.
The first step carries out classification marker in the database of each node to image, video, file etc..
Assuming that sharing N number of node, each node corresponds to a database Xi, XiThe database for indicating i-th of node, not
It is independent from each other with the database in node, and is not intended to shared information between different nodes, there is M in each database
Sample, while having L kind category labels in each database, different labels is stamped to different samples.
Second step initializes transition matrix and Lagrange multiplier.
Transition matrix and Lagrange multiplier are initialized in each node, and square is converted in the initialization that i-th of node is arranged
Battle array AiIt is the unit matrix of d × d dimensions, the initialization Lagrange multiplier of i-th of node is the full 0 matrix of d × d dimensions, and d indicates sample
The dimension of this original feature space.The transition matrix of all nodes and the initialization value of Lagrange multiplier are all.
Third walks, and introduces semantic consistency and builds object function.
It should be noted that the present invention focus on obtain the utilization of transition matrix and transition matrix, so no longer
Emphasis lists original object function, and to the specific optimization process of object function, after only giving objective function optimization
As a result.Object function after the optimization of i-th of node structure is as follows：
X in above formulamIndicate m-th of sample, zp、zqThe approximate center for indicating pth and q-th of classification respectively.With zpFor
Example, it is all mean values with p-th of marker samples.Value be 0 or 1, if sample xmThere is p-th of label, then
Value be 1, otherwise be 0.M is the total sample number of i-th of node, and L is the total species number of the node.λ is one of weight
Parameter can choose following values：(5,0.5,0.05,0.005).Yi (k)Indicate the kth time of the Lagrange multiplier of i-th of node repeatedly
Generation, T are a transposition symbol, (Yi (k))TIndicate Yi (k)Transposition.The mark of tr () representing matrix.i' represent the object of i-th of node
Neighbour is managed, we generally choose two neighboring nodes,Indicate the conversion square of the neighboring node of i-th of node when kth time iteration
Battle array,It indicatesTransposition.AiIt is the transition matrix of i-th of node, ρ is a punishment parameter, can be chosen：(10,
1,0.1).
Calculate two samples apart from when, it is contemplated that semantic consistency by original Euclidean distance replace with geneva away from
From.(x in above-mentioned formulam-zp)TAi(xm-zp) it is exactly to pass through transition matrix A in i-th of nodeiOriginal feature space is reflected
It is mapped in new feature space, meets semantic consistency in new feature space：If sample xmThere is p-th of label, just makes
Obtain sample xmIt is close as far as possible at a distance from the center marked with p-th in new feature space.Negative (zp-zq)TAi(zp-zq)
It is meant that in new feature space, the center of inhomogeneity label is remote as far as possible.The effect being finally reached in this way is：New
Feature space in, the distance between sample with phase cohort labelling is close as far as possible, has between the not sample of isolabeling
Distance it is remote as far as possible.
4th step solves object function, updates transition matrix.
Due to AiIt is positive semidefinite symmetrical matrix, by the property of positive semidefinite symmetrical matrix it is found that AiCan be with feature decompositionWherein WiIt is the matrix of d × r dimensions, r ＜ d, i indicate i-th of node, and WiEach row be all mutually just
It hands over, i.e.,IrIt is the unit matrix of r × r dimensions,Indicate WiTransposition.
AiAfter feature decomposition, object function (1) is expressed as to the form of tr (), it is as follows
S.t. indicates constraints in above formula, for the ease of writing and describing, defines the expression shape of new variable a U, U
Formula is as follows：
Then, optimization aim (2) can be converted into following form：
The solution of the final problem is equivalent to solve Wi, that is, the feature vector corresponding to the preceding r minimal eigenvalue of U is sought, profit
WithIt can obtain current AiSolution.Assuming that current AiSolution be+1 iteration of kth it is newer as a result, withIt indicates.
5th step, neighboring node are communicated, and judge whether the transition matrix of each node reaches unanimity, update Lagrange
Multiplier.
It should be noted that neighboring node refers to the physical close proximity of some node, each node, which is arranged, in we generally two
A physical close proximity node, all nodes have neighboring node, and these nodes may be constructed connected graph.It is communicated between node
When, the transition matrix of neighboring node is only transmitted, original data are not transmitted, to protect privacy, simultaneously because the rule of transition matrix
Mould is much smaller than initial data, it is possible to reduce the time overhead of communication and the scale for reducing storage.
First, it is communicated between neighboring node.If each node and the transition matrix of its neighboring node reach unanimity,
And these nodes may be constructed connected graph, then the transition matrix of these all nodes all reaches unanimity, at this time distributed training
Terminate.
If the transition matrix of all nodes tends not to unanimously, iteration updates Lagrange multiplier, and repeats third step.
The iteration of+1 Lagrange multiplier of kth is as follows：
Yi (k+1)Punishment parameter before indicating the update of kth+1 time of the Lagrange multiplier of i-th of node, ρ and being,
It is the result of+1 newer transition matrix of kth of the neighboring node of i-th of node.It is to return to update Lagrange multiplier
Third uses when walking iteration.
6th step, approximation search process.
For a new query sample xc, the node that is input to where the sample, it is assumed that enter i-th of section
Point, the transition matrix A obtained using distributed training processi, calculate sample xcTo other samples in i-th of node new
The distance of feature space.Such as：Query sample xcWith sample xmIt is expressed as in the distance of new feature space：
(xm-zp)TAi(xm-zp) (6)
T is transposition symbol, (xm-zp)TIndicate (xm-zp) transposition.
Since new feature space has more semantic congruence characteristic, sample more can be accurately measured when approximation is searched for
Similitude between this, similitude is higher, and reaction is exactly that the distance between them is smaller in feature space, uses semantic information
It can more effectively realize that approximation is searched for.
Obtain query sample xcAfter in the node at a distance from other samples, these distances are ranked up, f before obtaining
Minimum distance, the sample corresponding to the distance of preceding f minimum are the result that our approximations are searched for.
It should be noted that the invention is not limited in the above embodiment, all use equivalent replacement or equivalence replacement
The technical solution of formation belongs to the scope of protection of present invention.
Claims (3)
1. the distributed approximation searching method with secret protection based on semantic consistency, it is characterised in that including following step
Suddenly：
Step 1：Classification marker is carried out to image, video, file in the database of each node；
Step 2：Transition matrix and Lagrange multiplier are initialized, on each node all to transition matrix and Lagrange multiplier
Initialize installation is carried out, the transition matrix that initialization is arranged is the unit matrix of d × d dimensions, and the Lagrange of initialization multiplies
Son is the full 0 matrix of d × d dimensions, and d indicates the dimension of sample original feature space；
Step 3：It introduces semantic consistency and builds object function, semantic consistency is introduced in object function, passes through transition matrix
Original feature space is mapped in new feature space so that the distance of the sample with same tag to such center is most
Possible close, the distance at different classes of center is remote as far as possible；
Step 4：Above-mentioned object function is solved, transition matrix is updated；
Due to AiIt is positive semidefinite symmetrical matrix, by the property of positive semidefinite symmetrical matrix it is found that AiFeature decomposition is Ai=WiWi T,
Middle WiIt is the matrix of d × r dimensions, r ＜ d, i indicate i-th of node, and WiEach row be all mutually orthogonal, i.e. Wi TWi
=Ir, IrIt is the unit matrix of r × r dimensions, Wi TIndicate WiTransposition；
AiAfter feature decomposition, by object function：
It is expressed as the form of tr (), it is as follows：
s.t.Wi TWi=Ir
S.t. indicates constraints in above formula, and the representation for defining new variable a U, U is as follows：
Then, optimization aim：
It is converted into as follows
s.t.Wi TWi=Ir
Form：
s.t.Wi TWi=Ir
The solution of the final problem is equivalent to solve Wi, that is, the feature vector corresponding to the preceding r minimal eigenvalue of U is sought, A is utilizedi=
WiWi TObtain current AiSolution, it is assumed that current AiSolution be+1 iteration of kth it is newer as a result, withIt indicates, wherein xmIndicate the
M sample, zp、zqThe approximate center for indicating pth and q-th of classification respectively；
Step 5：Neighboring node is communicated, and judges whether the transition matrix of each node reaches unanimity, and updates Lagrange multiplier,
Each node is set there are two neighboring node, information is exchanged by transition matrix between neighboring node, and true according to transition matrix
Whether fixed to be updated from new iteration, all nodes constitute connected graph, i.e. any two node is indirectly connected by other nodes
When, if transition matrix reaches unanimity, deconditioning process；Otherwise, Lagrange multiplier is updated, and repeats step 3；
Step 6：It carries out approximation search process and some node is entered into for a new query sample, by conversion
After matrix mapping, calculate in new feature space at a distance between query sample and the node other samples, f before taking wherein
The result that sample corresponding to minimum distance is searched for as approximation.
2. the distributed approximation searching method with secret protection according to claim 1 based on semantic consistency,
It is characterized in that in step 1, it is assumed that share N number of node, each node corresponds to a database Xi, XiIndicate the number of i-th of node
According to library, the database in different nodes is independent from each other, and shared information is not intended between different nodes, each data
There are L kind category labels in library, different labels is stamped to different samples.
3. the distributed approximation searching method with secret protection according to claim 1 based on semantic consistency,
It is characterized in that, the method considers the information of neighboring node, punishment parameter Accelerated iteration is added.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810342218.3A CN108509651B (en) | 2018-04-17 | 2018-04-17 | The distributed approximation searching method with secret protection based on semantic consistency |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810342218.3A CN108509651B (en) | 2018-04-17 | 2018-04-17 | The distributed approximation searching method with secret protection based on semantic consistency |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108509651A true CN108509651A (en) | 2018-09-07 |
| CN108509651B CN108509651B (en) | 2019-03-12 |
Family
ID=63382152
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810342218.3A Expired - Fee Related CN108509651B (en) | 2018-04-17 | 2018-04-17 | The distributed approximation searching method with secret protection based on semantic consistency |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108509651B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110647804A (en) * | 2019-08-09 | 2020-01-03 | 中国传媒大学 | Violent video identification method, computer system and storage medium |
| CN111130913A (en) * | 2019-12-31 | 2020-05-08 | 联想(北京)有限公司 | Information processing method and device |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140019776A1 (en) * | 2012-07-01 | 2014-01-16 | Jerzy Lewak | Methods of providing fast search, analysis, and data retrieval of encrypted data without decryption |
| CN105808689A (en) * | 2016-03-03 | 2016-07-27 | 中国地质大学（武汉） | Drainage system entity semantic similarity measurement method based on artificial neural network |
| CN105843555A (en) * | 2016-03-18 | 2016-08-10 | 南京邮电大学 | Stochastic gradient descent based spectral hashing method in distributed storage |
| CN107133348A (en) * | 2017-05-23 | 2017-09-05 | 胡海峰 | Extensive picture concentrates the proximity search method based on semantic consistency |
| US20170300556A1 (en) * | 2014-02-24 | 2017-10-19 | Entefy Inc. | System and method of dynamic, encrypted searching |
| CN107291813A (en) * | 2017-05-19 | 2017-10-24 | 南京邮电大学 | Exemplary search method based on semantic segmentation scene |
| CN107315765A (en) * | 2017-05-12 | 2017-11-03 | 南京邮电大学 | A kind of method of the concentrated-distributed proximity search of extensive picture |
-
2018
- 2018-04-17 CN CN201810342218.3A patent/CN108509651B/en not_active Expired - Fee Related
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140019776A1 (en) * | 2012-07-01 | 2014-01-16 | Jerzy Lewak | Methods of providing fast search, analysis, and data retrieval of encrypted data without decryption |
| US20170300556A1 (en) * | 2014-02-24 | 2017-10-19 | Entefy Inc. | System and method of dynamic, encrypted searching |
| CN105808689A (en) * | 2016-03-03 | 2016-07-27 | 中国地质大学（武汉） | Drainage system entity semantic similarity measurement method based on artificial neural network |
| CN105843555A (en) * | 2016-03-18 | 2016-08-10 | 南京邮电大学 | Stochastic gradient descent based spectral hashing method in distributed storage |
| CN107315765A (en) * | 2017-05-12 | 2017-11-03 | 南京邮电大学 | A kind of method of the concentrated-distributed proximity search of extensive picture |
| CN107291813A (en) * | 2017-05-19 | 2017-10-24 | 南京邮电大学 | Exemplary search method based on semantic segmentation scene |
| CN107133348A (en) * | 2017-05-23 | 2017-09-05 | 胡海峰 | Extensive picture concentrates the proximity search method based on semantic consistency |
Non-Patent Citations (1)
| Title |
|---|
| HU H 等: "Anchor graph hashing with semantically consistent graph", 《2016 8TH INTERNATIONAL CONFERENCE ON WIRELESS COMMUNICATIONS & SIGNAL PROCESSING (WCSP)》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110647804A (en) * | 2019-08-09 | 2020-01-03 | 中国传媒大学 | Violent video identification method, computer system and storage medium |
| CN111130913A (en) * | 2019-12-31 | 2020-05-08 | 联想(北京)有限公司 | Information processing method and device |
| CN111130913B (en) * | 2019-12-31 | 2021-11-16 | 联想(北京)有限公司 | Information processing method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108509651B (en) | 2019-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhao et al. | User preference learning for online social recommendation | |
| Cai et al. | Batch mode active learning for regression with expected model change | |
| CN109241405A (en) | A kind of associated education resource collaborative filtering recommending method of knowledge based and system | |
| CN105138601B (en) | A kind of graphic mode matching method for supporting fuzzy constraint relationship | |
| CN107077487A (en) | Personal photo is tagged using depth network | |
| CN110489585A (en) | Distributed image searching method based on supervised learning | |
| CN104700190A (en) | Method and device for matching item and professionals | |
| CN108509651B (en) | The distributed approximation searching method with secret protection based on semantic consistency | |
| Iyer et al. | VeeAlign: multifaceted context representation using dual attention for ontology alignment | |
| Wang et al. | A community detection algorithm based on topology potential and spectral clustering | |
| Zhang et al. | A parallel compact sine cosine algorithm for TDOA localization of wireless sensor network | |
| Ying et al. | Transfer learning on high variety domains for activity recognition | |
| Wang et al. | An Improved Hybrid Algorithm Based on Biogeography/Complex and Metropolis for Many‐Objective Optimization | |
| CN110399564B (en) | Account classification method and device, storage medium and electronic device | |
| Zhang et al. | Scenario-adaptive and self-supervised model for multi-scenario personalized recommendation | |
| CN115618097A (en) | Entity alignment method for prior data insufficient multi-social media platform knowledge graph | |
| Bi et al. | MM-GNN: Mix-moment graph neural network towards modeling neighborhood feature distribution | |
| Yuan et al. | Adaptive incentivize for cross-silo federated learning in IIoT: A multi-agent reinforcement learning approach | |
| Su et al. | A new random-walk based label propagation community detection algorithm | |
| CN109857892A (en) | Semi-supervised cross-module state Hash search method based on category transmitting | |
| Zhao et al. | A novel higher-order neural network framework based on motifs attention for identifying critical nodes | |
| Zhang et al. | A framework of data fusion through spatio-temporal knowledge graph | |
| Qiao et al. | Improving stochastic block models by incorporating power-law degree characteristic | |
| Yu et al. | Community detection in the textile-related trade network using a biased estimation of distribution algorithm | |
| CN114297381A (en) | Text processing method, device, 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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190312 Termination date: 20200417 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |