CN108228782A - A kind of implication relation based on deep learning finds method - Google Patents
A kind of implication relation based on deep learning finds method Download PDFInfo
- Publication number
- CN108228782A CN108228782A CN201711469074.XA CN201711469074A CN108228782A CN 108228782 A CN108228782 A CN 108228782A CN 201711469074 A CN201711469074 A CN 201711469074A CN 108228782 A CN108228782 A CN 108228782A
- Authority
- CN
- China
- Prior art keywords
- tutor
- matrix
- author
- paper
- probability
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000013135 deep learning Methods 0.000 title claims abstract description 15
- 239000011159 matrix material Substances 0.000 claims abstract description 86
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 230000001351 cycling effect Effects 0.000 claims description 9
- 239000004576 sand Substances 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000036962 time dependent Effects 0.000 claims description 6
- 238000005065 mining Methods 0.000 claims description 5
- 239000008186 active pharmaceutical agent Substances 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 4
- 238000009412 basement excavation Methods 0.000 description 10
- 238000004422 calculation algorithm Methods 0.000 description 10
- 238000011160 research Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 9
- 238000011478 gradient descent method Methods 0.000 description 6
- 230000003993 interaction Effects 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 230000000306 recurrent effect Effects 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000015654 memory Effects 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 244000046052 Phaseolus vulgaris Species 0.000 description 1
- 235000010627 Phaseolus vulgaris Nutrition 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000009193 crawling Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013079 data visualisation Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- BULVZWIRKLYCBC-UHFFFAOYSA-N phorate Chemical compound CCOP(=S)(OCC)SCSCC BULVZWIRKLYCBC-UHFFFAOYSA-N 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000005295 random walk Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000011282 treatment Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Economics (AREA)
- Probability & Statistics with Applications (AREA)
- Software Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Fuzzy Systems (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of implication relations based on deep learning to find method, belongs to information technology field, specifically includes from scholar and delivers generation paper coauthorship network G ' in network G;Calculate the collaboration matrix X of paper publishing situation Matrix C, D, S and paperS, XD, XT;Propose RGRU models;It is designed on the basis of RGRU and builds tARMM models to predict " tutor student " relationship.Prediction accuracy of the tARMM models proposed by the present invention on data set is higher than other methods, can reach 95% or so, other social relationships with time dependence are excavated with certain reference and reference value.
Description
Technical field
The invention belongs to information technology fields, and in particular to a kind of implication relation based on deep learning finds method.
Background technology
It with the universal of the social medias such as Facebook, Twitter, wechat and promotes, social media has become people
Between interaction and communication Important Platform.Different types of social relationships have opposite impacts on people, people’s lives, study and
Be operated in these relationships it is subtle under change occurs, such as in social networks, the hobby of people can be by friend
Influence, the research direction of student can be influenced by tutor.Meanwhile also imply a large amount of additional letter in these relationships
Breath, for example by studying " tutor-student " relationship, learned society can be excavated, scientific research community network is established, further appreciate that phase
The development course of research topic is closed, finds the developing direction of next step.
There are many explicit relations, such as friends, concern relation, comment relationship, reply relationship in network, however,
Also it is to imply in a network, such as to have many relationships:" tutor-student " relationship is lain in paper coauthorship network.Paper is collaborateed
Network is the cooperative relationship network that scientific research personnel gradually forms in literature procedure is published cooperatively, such as DBLP.At present, have several
Project to safeguard relationship as the target of oneself, such as LinkedIn and AI family trees.The former requires user special to each
Object be labeled, such as colleague, tutor, student etc., the latter equally leads research field using the method that marks by hand
Teacher's information is labeled.Obviously, these methods are largely dependent on artificial mark, and not only efficiency is low, and accuracy is not also high, this is big
Its Generalization Ability is limited greatly.It is a kind of method of design for this ideal solution of phenomenon one, automatically from network
Excavate or predict wherein implicit relationship.
In paper coauthorship network, it is desirable to only judge who be tutor is relatively difficult in list from publishing.Sometimes according to
Intuition it is assumed that relationship type can be distinguished in certain social networks using heuristic rule.But research finds to use allusion quotation
The heuristic rule of type can only achieve precision as 70-80%, even with the multiple rules trained based on multiple and different features
With reference to supervised learning model, precision is average still to only have 80%, moreover, it is difficult often to collect supervision letter in practical training
Breath.
" tutor-student " relationship in paper coauthorship network has following several characteristics:
1. implicity." tutor-student " relationship is hidden in paper coauthorship network, in paper coauthorship network, only
Have the partner of paper, the topic of paper, paper deliver the time, paper publishing the information such as publication/meeting, can not be explicitly
Know " tutor-student " relationship between partner.
2. time dependence.Tutor-student's relationship has the time dependence of height, for any one author,
In its numerous partner, the partner of early stage is more likely its tutor than the partner in later stage.In addition, a people is permissible
It is tutor role from the Role composition of student, and this diversification in role may be without any apparent sign.
It is 3. difficult predictive.Since paper coauthorship network only has the relevant information for publishing paper cooperatively, with other social matchmakers
Body is compared to being very simple, simultaneously as " tutor-student " relationship is hidden in paper coauthorship network, this is resulted in paper
It artificially goes to infer that " tutor-student " relationship is relatively difficult in coauthorship network.
In recent years, social networks research causes the extensive concern of academia.It currently can to the research work of social networks
To be divided into three aspects:The interaction prediction of social networks prediction, social networks type identification and relationship.
Social networks prediction, also known as link prediction, refer to the feature according to nodes or already existing side, in advance
There is side between two nodes of survey.Liben-Nowell etc. is directed to specific social networks, the similarity measurements based on figure
Similitude between amount method calculate node recycles the link possibility between similitude prediction node.Lee etc. proposes one
Kind calculates the smaller model based on social vector clock feature of cost to solve link forecasting problem.The it is proposeds such as Cunchao Tu
CANE models carry out internet startup disk so as to reach the target of link prediction by text data information relevant to user.
Backstrom etc. proposes the Random Walk Algorithm based on supervised learning for the strength problem of social networks.The it is proposeds such as Zhao
One kind is based on the Forecasting Methodology of " trusted path ", this is one of a small number of Forecasting Methodologies for being suitable for weighted network.
Relationship type identifies, refers to, for one or more social networks, automatically identify and excavate and wherein contained
Relationship type.Coppola etc. proposes semantic-based automatic relation excavation frame.Leskovec etc. utilizes logistic regression models
Identify the positive relationship in social networks or negative relationship, i.e. friends or non-friends.Diehl etc. uses study ranking functions
" manager-subordinate " relationship of identification.Pentland etc. proposes several obile data mining models, for speculating friends.By
" tutor-student " relation excavation problem relation belonging to type identification problem of literary coauthorship network, in the problem, the propositions such as Tang Jie
TPFG models are used to excavate " director-tutee " relationship from paper coauthorship network, in addition, they are towards heterogeneous network
(such as mail network, scientific collaboration network) proposes a kind of Unified frame based on factor graph, it is intended to solve social networks type
Identification problem.Li Yongjun etc. speculates " tutor-student " relationship in paper coauthorship network using maximum entropy model.
Relationship interaction prediction, mainly studies how unidirectional social networks develop into two-way social networks and its hair
The reason of changing.Most common unidirectional relationship is the relationship between star and their beans vermicelli, and bidirectional relationship is friends.
Hopcroft etc. explores the research social networks such as relationship interaction prediction problem, Lou are how to develop into ternary closure.They
A kind of learning framework that relationship interaction prediction problem is abstracted as to figure is proposed jointly.
Invention content
For the above-mentioned technical problems in the prior art, the present invention proposes a kind of implicit pass based on deep learning
It is discovery method, reasonable design overcomes the deficiencies in the prior art, has good effect.
To achieve these goals, the present invention adopts the following technical scheme that:
A kind of implication relation based on deep learning finds method, makes formalization to implication relation Mining Problems and determines
Justice:
It defines 1 scholar and delivers network G
Time dependent scholar is delivered into network formalism and is expressed as a bigraph (bipartite graph), enables G=(A, P, E), wherein Represent that scholar delivers the set of all authors in network;It is the collection of all papers
It closes;E={ eik|1<=i<=na,1<=k<=np,aiIt is pkAuthor, represent that scholar delivers author in network and paper
Works relationship;
Define 2 paper coauthorship network G '
It is generated from GWherein,It is that author gathers, a0It is one
Virtual author, for author ai, it is assumed that its tutor isIfIt is considered thatE '={ eij|1<=i
<=na,1<=j<=na,aiAnd ajWith cooperative relationship and ai≠aj};Wherein, pnijIt is and eijA relevant vector, pnij
∈R1×40Represent aiAnd ajThe Quantity of Papers collaborateed in some time-domain;For single author, pn is usediIt can be with table
Show author aiPaper publishing situation;
It defines 3 papers and collaborates Matrix C
For author x arbitrary in A, it is assumed that it has collaboration relationship with m authors, and partner's collection shares AxIt represents, Ax=
{b0,b1,b2,···,bm, wherein b0=a0;If in a certain year t, x and bjThe paper number of collaboration isThen for author
X has collaboration matrix:
Wherein, T is the overall time domain that author cooperates, herein with 1 year for a time span, if author collaborates the time
For [1970,2010], totally 40 years, then the T=39 in above-mentioned matrix, collaborates Matrix C ∈ R(m+1)×40;
Define 4 tutor student's relationship R
Enable R={ yij|0<=i<=na,0<=j<=na, represent author between whether be " tutor-student " relationship,
Specific value is as follows:
The implication relation based on deep learning finds method, specifically comprises the following steps:
Input:Scholar delivers network G;
Output:The prediction result of " tutor-student " relationship;
Step 1:The link delivered scholar in network G is analyzed, and is delivered from scholar and paper collaboration net is generated in network G
Network G ';
Step 2:According to paper coauthorship network G ', paper publishing situation Matrix C, D, S are calculated, and then calculate the collaboration of paper
Matrix XS, XD, XT;
Step 3:Establish tARMM (time-aware Advisor-advisee Relationship Mining Model,
Tutor student's relation excavation model of Time Perception) model;
Step 4:It is handled by tARMM models collaborateing matrix;
Step 4.1:Probability P is calculated using RGRUT;
Step 4.2:Probability P is calculated using DNNF;
Step 4.3:Calculate final tutor's probability P;
Step 5:The candidate tutor of maximum probability is the prediction tutor of x in P, so as to obtain " tutor-student " relationship
Prediction result.
Preferably, in step 2, it for the collaboration situation of paper, is analyzed in terms of following two:
In a first aspect, being analyzed from the details of collaboration, for author x, represent that x is waited with it by collaborateing Matrix C
The co-authored papers between tutor is selected to deliver situation;
The paper publishing situation of candidate tutor is represented with D:
The paper publishing situation pn of author xxIt is represented with S:
S=(S0 … ST-1) (2.3);
The paper publishing situation for being utilized respectively author and candidate tutor is normalized to collaborateing Matrix C:
XS=CS (2.5);
XD=DS (2.6);
Wherein, XSFor the collaboration submatrix based on student, XSij∈XS, represent author x and its candidate tutor b in jth yeari
Co-authored papers number accounts for the ratio of author's x jth years total paper number;XDFor the collaboration submatrix based on tutor, XDij∈XD, represent
Author x and its candidate tutor b in jiCo-authored papers number accounts for candidate tutor biJth year total paper number ratio;
Second aspect, from the time angle of collaboration, according to collaboration Matrix C by the time structure for the situation of collaborateing with matrix
Form be indicated, be defined as follows:
XTFor the collaboration submatrix based on time structure, it is meant that and represents that author x is led with its candidate with the form of matrix
Teacher biBetween co-authored papers time structure.
Preferably, in step 4.1, in tARMM models, to RNN (Recursive Neural Network, cycle
Neural network) it is transformed, (Refresh Gate Recurrent Unit, update door follow generation update door cycling element RGRU
Ring element), by updating door cycling element RGRU, to XTIt is handled, obtains tutor's probability PT;
For moment t, have:
rt=σ (wr[ht+1,xt]+br) (2.9);
ht=wh[(1-rt)ht+1,rtxt] (2.10);
Wherein, rtIt is to update door in the state of time t, wrBe update door weight matrix, brIt is the offset for updating door,
ht+1Be update gate cell moment t+1 state, xtIt is the input matrix of moment t, htIt is the state for updating gate cell in time t,
wtAnd btIt is generation state h respectivelytWeight matrix and offset;
Tutor's probability P based on RGRUT:
PT=hT(2.11);
Wherein, hTIt is the state for updating gate cell in time T;Its formula and htIt is identical;
It is as follows:
Input:Paper collaborates matrix XT;
Output:Tutor's probability P based on RGRUT;
Step 4.1.1:Initialize PTFor null matrix;
Step 4.1.2:The state r of the update door of t is calculated by formula (2.9)t;
Step 4.1.3:The state h of the update gate cell of t is calculated by formula (2.10)t;
Step 4.1.4:Tutor's probability P of x is calculated by formula (2.11)T。
Preferably, in step 4.2, by tARMM models, using deep neural network, to XS、XDIt is handled, is obtained
Tutor's probability P based on class figure matrixF;
By XSAnd XDIt is combined, forms the bitmap of a double Color Channels, referred to as class figure matrix X;Target is to find class
The line number where special pattern in figure matrix X;Since this is the target orientation problem of a Pixel-level, so structure one
DNN is identified, and according to the calculation formula of perceptron, for each node in DNN, output is:
Wherein, wi, weights and offset parameter of the b for model, piThe probability value predicted for each node;
Tutor's probability P based on class figure matrix that then DNN is finally generatedFOutput for last layers of DNN:
PF=Relu (f (XS,XD)) (2.13);
It is as follows:
Input:Paper collaborates matrix XSAnd XD;
Output:Tutor's probability P based on class figure matrixF;
Step 4.2.1:Initialize PFFor null matrix;
Step 4.2.2:The output of each node in DNN is calculated by formula (2.12);
Step 4.2.3:Probability P is calculated by formula (2.13)F。
Preferably, in step 4.3, by PTAnd PFFinal tutor's probability matrix is generated by full articulamentum, is therefrom chosen
Highest probability value P, corresponding candidate tutor are the prediction tutor of x;
P=σ (PF·PT) (2.14)。
Advantageous effects caused by the present invention:
The present invention has used for reference long memory models (LSTM) and logic gate cycling element (GRU) etc. in short-term and has become body circulation nerve net
The theory of network (RNN) model, RNN is transformed, and proposes update door cycling element (RGRU), is collaborateed in matrix for handling
Time structure;Since RGRU only has, there are one gate cells, simpler than LSTM and GRU in structure, but are closed " tutor-student "
There is higher accuracy on the Mining Problems of system;
The present invention handles " tutor-student " relation excavation problem in paper coauthorship network using the thought of deep learning,
It is proposed time dependent " tutor-student " relation excavation neural network (tARMM), prediction accuracy of the model on data set
Higher than other methods, 95% or so can be reached, other social relationships with time dependence are excavated and are borrowed with certain
Meaning of reflecting and reference value.
Description of the drawings
Fig. 1 is " tutor-student " relation excavation schematic diagram.
Fig. 2 is tARMM schematic diagrames.
Fig. 3 is RGRU schematic diagrames.
Fig. 4 (a) is XSClass figure matrix schematic diagram.
Fig. 4 (b) is XDClass figure matrix schematic diagram.
Fig. 5 is DNN schematic diagrames.
Fig. 6 is full articulamentum schematic diagram.
Specific embodiment
Below in conjunction with the accompanying drawings and specific embodiment is described in further detail the present invention:
1st, the formal definitions of problem
In this section, the basic symbol of some herein and definition are provided.
1.1 main symbol of table and its meaning
It defines 1 scholar and delivers network G
Time dependent scholar is delivered into network formalism and is expressed as a bigraph (bipartite graph), enables G=(A, P, E), wherein Represent that scholar delivers the set of all authors in network;It is the collection of all papers
It closes;E={ eik|1<=i<=na,1<=k<=np,aiIt is pkAuthor, represent that scholar delivers author in network and paper
Works relationship.
Define 2 paper coauthorship network G '
Generated from G G '=(A ', E ', { pnij}eij∈E’), wherein,It is that author gathers, a0It is
One virtual author, for author ai, it is assumed that its tutor isIfIt is considered thatE '={ eij|1<
=i<=na,1<=j<=na,aiAnd ajWith cooperative relationship and ai≠aj}。pnijIt is and eijA relevant vector, pnij∈
R1×40Represent aiAnd ajThe Quantity of Papers collaborateed in some time-domain.For single author, pn is usediIt can represent
Author aiPaper publishing situation.
It defines 3 papers and collaborates Matrix C
For author x any one in A, it is assumed that it has collaboration relationship with m authors, and partner's collection shares AxIt represents, Ax=
{b0,b1,b2,···,bm, wherein b0=a0.If in a certain year t, x and bjThe paper number of collaboration isThen for author
X has collaboration matrix:
Wherein, T is the overall time domain that author cooperates, herein with 1 year for a time span, if author collaborates the time
For [1970,2010], totally 40 years, then the T=39 in above-mentioned matrix, collaborates Matrix C ∈ R(m+1)×40。
Define 4 tutor student's relationship R
Enable R={ yij|0<=i<=na,0<=j<=na, represent author between whether be " tutor-student " relationship,
Specific value is as follows:
The goal in research of this paper is exactly that the tutor y of x is predicted from Cx, it is the tutor of x, with great that this, which needs that whom is solved,
Probability is tutor's both of these problems of x.
2 model constructions
2.1 collaborate the structure of matrix
In order to excavate " tutor-student " relationship, generation paper coauthorship network in network G is delivered from original scholar first
G ', then therefrom matrix ((such as Fig. 1 shows)) is collaborateed in extraction.
For the collaboration situation of paper, can be analyzed in terms of following two:
First aspect is analyzed from the details of collaboration, for author x, has collaboration Matrix C to represent that x is led with its candidate
Co-authored papers between teacher deliver situation.
The paper publishing situation of candidate tutor is represented with D:
The paper publishing situation pn of author xxIt is represented with S:
S=(S0 … ST-1) (2.3);
Then, the paper publishing situation for being utilized respectively author and candidate tutor is normalized to collaborateing Matrix C:
XS=CS (2.5);
XD=DS (2.6);
XSAnd XDSubmatrix and the collaboration submatrix based on tutor are collaborateed respectively based on student.For XSij∈XS, table
Show author x and its candidate tutor b in jth yeariCo-authored papers number accounts for the ratio of author's x jth years total paper number.XDij∈XDIt represents
Author x and its candidate tutor b in jth yeariCo-authored papers number accounts for candidate tutor biJth year total paper number ratio.
Second aspect, from the time angle of collaboration, according to collaboration Matrix C by the time structure for the situation of collaborateing with matrix
Form be indicated, be defined as follows:
XTAs the collaboration submatrix based on time structure, it is meant that and represents that author x is candidate with it with the form of matrix
The time structure of co-authored papers between tutor.
2.2 time dependent relation excavation model constructions
This section is it is proposed that a kind of time dependent relation excavation neural network model tARMM (as shown in Figure 2), the model
By respectively to XTAnd XS、XDIt is handled, obtains tutor's probability matrix based on time structure and based on class figure matrix, then
Final tutor's probability matrix is generated by full articulamentum.To XTWhen being handled, the update door cycle designed between the inverse time is single
Member, for XSAnd XDWhen being handled, using deep neural network.
2.2.1 the method for calculating probability based on RGRU
It adds a update gate cell on the basis of standard RNN herein, is formed only there are one the Recognition with Recurrent Neural Network of update door,
Referred to as update door cycling element RGRU (as shown in Figure 3).For based on the collaboration matrix X for delivering the timeT, pass through formula
(2.7) understand that non-zero element column is more forward, then it is general to have higher tutor by the candidate tutor of be expert at characterization in a matrix
Rate.So by matrix XTWith row for unit reversely through RGRU processing, obtain tutor's probability matrix based on RGRU.
For moment t, have:
rt=σ (wr[ht+1,xt]+br) (2.9);
ht=wh[(1-rt)ht+1,rtxt] (2.10);
Wherein, rtIt is to update door in the state of time t, wrBe update door weight matrix, brIt is the offset for updating door,
ht+1Be update gate cell moment t+1 state, xtIt is the input matrix of moment t, htIt is the state for updating gate cell in time t,
wtAnd btIt is generation state h respectivelytWeight matrix and offset;
Tutor's probability P based on RGRUT:
PT=hT(2.11);
Wherein, hTIt is the state for updating gate cell in time T;hTFormula and htIt is identical.
In conclusion tutor's probability calculation based on improved update door cycling element (RGRU) is as shown in algorithm 1.
2.2.2 the method for calculating probability based on class figure matrix
XSAnd XDRespectively collaboration situation is characterized in terms of student and candidate tutor two.With XSFor, by the collaboration based on student
Matrix regards 66 × 40 gray-scale map as, and bitmap mode is taken to be shown, it can be found that ought time that wherein certain a line is characterized
When selecting tutor as practical tutor, the particular image similar to " one " can be formed there are continuous one section of pixel value in the row, but
It is to have different characteristics in different bitmaps, so being handled herein by deep neural network bitmap, extracts it
Eigenmatrix excavates the position where particular image.Fig. 4 (a) is XSClass figure matrix;Fig. 4 (b) is XDClass figure matrix.
By XSAnd XDIt is combined, forms the bitmap of a double Color Channels, referred to as class figure matrix X.So next step
Target is the line number where the special pattern found in class figure matrix X.Since this is the target orientation problem of a Pixel-level, institute
It is identified with building a DNN (as shown in Figure 5).According to the calculation formula of perceptron, for each node in DNN,
Its output is:
Enable the output that y ' is last layer in DNN.Tutor's probability P based on class figure matrix that then DNN is finally generatedFFor:
PF=Relu (f (XS,XD))=y′(2.13);
In conclusion the realization process of DNN can use following algorithm description.
Finally, by PTAnd PFBy tutor's probability matrix that the generation (as shown in Figure 6) of full articulamentum is final, therefrom choose most
High probability value, corresponding candidate tutor is the tutor predicted.
P=σ (PF·PT) (2.14);
2.3 the learning algorithm of model
The learning algorithm that model will be introduced in this part, the update method including loss function and parameter.The mould carried herein
Type uses cross entropy as loss function, specific as follows:
In terms of the update of parameter, it is after all parameters are initialised herein, goes to optimize using Adam methods
Parameter.Adam methods are a kind of learning methods of autoadapted learning rate, can be that each parameter calculates the learning rate of oneself.It is public
Formula is as follows:
mt=β1mt-1+(1-β1)gt(2.16);
Wherein, mtIt is the single order moments estimation to gradient, can be regarded as to it is expected E | gt| estimation, vtIt is two to matrix
Rank moments estimation can be regarded as to it is expectedEstimation,WithIt is to mtAnd vtCorrection, be approximately to desired nothing
Estimation partially.It is a dynamic constrained of learning rate.
2.4 algorithm description
The complete algorithm of tARMM models proposed in this paper is described as follows:
3 design and analysis of experiment
3.1 experimental setup
Data set.Using the DBLP computer science bibliographic data base that Michael Ley are developed as the data set of experiment
It goes to speculate " tutor-student " relationship therein.The part wherein from 1970 to 2010 year is chosen, it includes 654628 authors
With 1076946 publications.As label data, using MAN, the union of tri- data sets of MathGP, AIGP is as verification number
According to collection, wherein MAN is by crawling acquisition on the personal homepage of tutor, and MathGP is from Mathematics
Acquisition is crawled in Genealogy projects, AIGP is to crawl to obtain from AI Genealogy projects.
Done it is a series of experiment go correctness of the search model in " tutor-student " relation excavation problem and effectively
Property.Selected section data are trained model from data set at random, then randomly selected from data set again data set into
Row test.
In order to intuitively compare estimation result, herein using the most common evaluation index of sorting algorithm:Accuracy rate ACC,
Calculation formula is as follows:
Wherein, TP is real example number, and FP is false positive example number.
Experimental situation is:Intel Core i5-2520M double-cores (2.5GHz), windows10 64,8G memories,
NVIDA GeForce GT635M video cards.Programming language is:Matlab and Python uses TensorFlow frames.
3.2 programming technique
Data preprocessing phase writes code using Matlab, and tARMM model realizations part is write using python,
The part has used TensorFlow machine learning frames.As a result exposition realizes that page end is mainly adopted using JavaWeb
The ECharts components increased income with Baidu.
(1)TensorFlow
TensorFlow is the software library that the second generation that Google increases income is used for numerical calculation, it is one very flexible
Machine learning frame can be operated on the single or multiple CPU and GPU of the even mobile equipment of server or PC.
TensorFlow is the processing frame based on data flow diagram, and the node table in data flow diagram shows mathematical operation, side table
Show the data interaction between operation node.Tensor represents the data transmitted between node in TensorFlow, and Flow represents number
It is exactly each node that Tensor enters data operation figure according to the form of stream according to stream.
It needs to represent calculating task using figure (graph) in programming, then in the context of referred to as session (Session)
(Context) figure is performed in, meanwhile, it represents data using tensor, state is safeguarded by variable (Variable), uses feed
Either fetch is arbitrary operation assignment or data therefrom.
(2)Echarts
Echarts is the icon library of a pure javascript, is embedded it in html webpage, can be in computer
With operation smooth in mobile equipment, compatible current most of browser, bottom layer realization depends on the Canvas classes of lightweight
Library Zrender, provide it is lively, intuitive, can interact, can the personalized data visualization icon of height.Text is used using Echarts
In the displaying of experimental result.
3.3 experimental result
For deep learning, different optimization methods will have different influences to trained efficiency and validity.
In general, it is common to use training method of the gradient descent method as model.Gradient declines again there are many classifying, wherein batch ladder
Degree descent method BGD is the form of most original in gradient descent method, and concrete thought is to update each ginseng using all samples
Number.Since batch gradient descent method is when updating each parameter, all training samples are required for, so training process can be with
It the increase of sample size and becomes more and more slowly.Stochastic gradient descent method SGD is precisely in order to solve batch gradient descent method
This drawback and propose, employ stochastic gradient descent method herein and carry out training pattern.
3.3.1 the validity of RGRU
In order to prove the validity of RGRU, RNN, LSTM is used alone and update door cycling element RGRU is tested, obtains
The result arrived is as follows:
The performance of the different neural networks of table 3.2
It can be seen that RGRU than simple Recognition with Recurrent Neural Network and long memory models in short-term " tutor-student " by experiment
There is higher accuracy on the Mining Problems of relationship.Proof is correct effective to the improvement of Recognition with Recurrent Neural Network.
3.3.2 the comparison of tARMM and other algorithms
In the part, the TPFG models proposed for " tutor-student " relations problems and the SVM models for classification are chosen
It is compared with tARMM models proposed in this paper.Test of many times is carried out, is averaged, it is as a result as follows:
The results contrast of 3.3 algorithms of different of table
By experiment as can be seen that the accuracy rate of tARMM models is above SVM and TPFG, tARMM moulds are further proved
The correctness of type.
4 summarize and look forward to
The identification problem of " tutor-student " relationship is had studied in paper coauthorship network herein.For the problem, pass through first
Matrix is collaborateed to the pretreatment generation of data, tARMM model treatments is then established and collaborates matrix excavation " tutor-student " relationship.
Generation RGRU is transformed to RNN in tARMM models, which can excavate the relationship with time dependence.Utilize DBLP
In data tested, it was demonstrated that the correctness and validity of tARMM models.
In this study, since the data set of tape label can not cover entire DBLP databases, so there are certain mistakes
Difference.In this regard, the later stage will be improved by the data the set pair analysis model for expanding tape label, the accuracy of model is improved.Meanwhile this mould
To the social relationships with time dependence with certain expansion, the later stage will make model for different social medias type
It is further to improve, improve the versatility of model.
Certainly, above description is not limitation of the present invention, and the present invention is also not limited to the example above, this technology neck
The variations, modifications, additions or substitutions that the technical staff in domain is made in the essential scope of the present invention should also belong to the present invention's
Protection domain.
Claims (5)
1. a kind of implication relation based on deep learning finds method, it is characterised in that:Shape is made to implication relation Mining Problems
The definition of formula:
It defines 1 scholar and delivers network G
Time dependent scholar is delivered into network formalism and is expressed as a bigraph (bipartite graph), enables G=(A, P, E), wherein Represent that scholar delivers the set of all authors in network;It is the collection of all papers
It closes;E={ eik|1<=i<=na,1<=k<=np,aiIt is pkAuthor, represent that scholar delivers author in network and paper
Works relationship;
Define 2 paper coauthorship network G '
It is generated from GWherein,It is that author gathers, a0It is one virtual
Author, for author ai, it is assumed that its tutor isIfIt is considered thatE '={ eij|1<=i<=
na,1<=j<=na,aiAnd ajWith cooperative relationship and ai≠aj};Wherein, pnijIt is and eijA relevant vector, pnij∈R1 ×40Represent aiAnd ajThe Quantity of Papers collaborateed in some time-domain;For single author, pn is usediIt can represent to make
Person aiPaper publishing situation;
It defines 3 papers and collaborates Matrix C
For author x any one in A, it is assumed that it has collaboration relationship with m authors, and partner's collection shares AxIt represents, Ax={ b0,
b1,b2,···,bm, wherein b0=a0;If in a certain year t, x and bjThe paper number of collaboration isThen for author x, have
Collaborate matrix:
Wherein, T is the overall time domain that author cooperates, and herein with 1 year for a time span, is if author collaborates the time
[1970,2010], totally 40 years, then the T=39 in above-mentioned matrix, collaborates Matrix C ∈ R(m+1)×40;
Define 4 tutor student's relationship R
Enable R={ yij|0<=i<=na,0<=j<=na, represent author between whether be " tutor-student " relationship, specifically take
Value is as follows:
The implication relation based on deep learning finds method, specifically comprises the following steps:
Input:Scholar delivers network G;
Output:The prediction result of " tutor-student " relationship;
Step 1:The link delivered scholar in network G is analyzed, and generation paper coauthorship network in network G is delivered from scholar
G’;
Step 2:According to paper coauthorship network G ', paper publishing situation Matrix C, D, S are calculated, and then calculate the collaboration matrix of paper
XS, XD, XT;
Step 3:Establish tARMM models;
Step 4:It is handled by tARMM models collaborateing matrix;
Step 4.1:Probability P is calculated using RGRUT;
Step 4.2:Probability P is calculated using DNNF;
Step 4.3:Calculate final tutor's probability P;
Step 5:The candidate tutor of maximum probability is the prediction tutor of x in P, so as to obtain the prediction of " tutor-student " relationship
As a result.
2. the implication relation according to claim 1 based on deep learning finds method, it is characterised in that:In step 2,
For the collaboration situation of paper, analyzed in terms of following two:
In a first aspect, being analyzed from the details of collaboration, for author x, represent that x is led with its candidate by collaborateing Matrix C
Co-authored papers between teacher deliver situation;
The paper publishing situation of candidate tutor is represented with D:
The paper publishing situation pn of author xxIt is represented with S:
S=(S0 … ST-1) (2.3);
The paper publishing situation for being utilized respectively author and candidate tutor is normalized to collaborateing Matrix C:
XS=CS (2.5);
XD=DS (2.6);
Wherein, XSFor the collaboration submatrix based on student, XSij∈XS, represent author x and its candidate tutor b in jth yeariIt collaborates
Paper number accounts for the ratio of author's x jth years total paper number;XDFor the collaboration submatrix based on tutor, XDij∈XD, represent in jth year
Middle author x and its candidate tutor biCo-authored papers number accounts for candidate tutor biJth year total paper number ratio;
Second aspect, from the time angle of collaboration, according to collaboration Matrix C by the time structure for the situation of collaborateing with the shape of matrix
Formula is indicated, and is defined as follows:
XTFor the collaboration submatrix based on time structure, it is meant that and represents author x and its candidate tutor b with the form of matrixiIt
Between co-authored papers time structure.
3. the implication relation according to claim 1 based on deep learning finds method, it is characterised in that:In step 4.1
In, in tARMM models, RNN is transformed, generation update door cycling element RGRU, by updating door cycling element RGRU,
To XTIt is handled, obtains tutor's probability PT;
For moment t, have:
rt=σ (wr[ht+1,xt]+br) (2.9);
ht=wh[(1-rt)ht+1,rtxt] (2.10);
Wherein, rtIt is to update door in the state of time t, wrBe update door weight matrix, brBe update door offset, ht+1It is
Update the state of gate cell moment t+1, xtIt is the input matrix of moment t, htIt is to update gate cell in the state of time t, wtAnd bt
It is generation state h respectivelytWeight matrix and offset;
Tutor's probability P based on RGRUT:
PT=hT(2.11);
Wherein, hTIt is the state for updating gate cell in time T;Its formula and htIt is identical;
It is as follows:
Input:Paper collaborates matrix XT;
Output:Tutor's probability P based on RGRUT;
Step 4.1.1:Initialize PTFor null matrix;
Step 4.1.2:The state r of the update door of t is calculated by formula (2.9)t;
Step 4.1.3:The state h of the update gate cell of t is calculated by formula (2.10)t;
Step 4.1.4:Tutor's probability P of x is calculated by formula (2.11)T。
4. the implication relation according to claim 1 based on deep learning finds method, it is characterised in that:In step 4.2
In, by tARMM models, using deep neural network, to XS、XDIt is handled, obtains tutor's probability based on class figure matrix
PF;
By XSAnd XDIt is combined, forms the bitmap of a double Color Channels, referred to as class figure matrix X;Target is to find class figure square
The line number where special pattern in battle array X;Since this is the target orientation problem of a Pixel-level, thus structure one DNN into
Row identification, according to the calculation formula of perceptron, for each node in DNN, output is:
Wherein, wi, weights and offset parameter of the b for model, piThe probability value predicted for each node;
Tutor's probability P based on class figure matrix that then DNN is finally generatedFOutput for last layers of DNN:
PF=Relu (f (XS,XD)) (2.13);
It is as follows:
Input:Paper collaborates matrix XSAnd XD;
Output:Tutor's probability P based on class figure matrixF;
Step 4.2.1:Initialize PFFor null matrix;
Step 4.2.2:The output of each node in DNN is calculated by formula (2.12);
Step 4.2.3:Probability P is calculated by formula (2.13)F。
5. the implication relation according to claim 1 based on deep learning finds method, it is characterised in that:In step 4.3
In, by PTAnd PFFinal tutor's probability matrix is generated by full articulamentum, therefrom chooses highest probability value P, corresponding time
Select the prediction tutor that tutor is x;
P=σ (PF·PT) (2.14)。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711469074.XA CN108228782B (en) | 2017-12-29 | 2017-12-29 | Implicit relation discovery method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711469074.XA CN108228782B (en) | 2017-12-29 | 2017-12-29 | Implicit relation discovery method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108228782A true CN108228782A (en) | 2018-06-29 |
CN108228782B CN108228782B (en) | 2020-04-21 |
Family
ID=62646745
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711469074.XA Active CN108228782B (en) | 2017-12-29 | 2017-12-29 | Implicit relation discovery method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108228782B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110008279A (en) * | 2019-03-27 | 2019-07-12 | 北京工商大学 | One kind is based on " relationship-weight " the orderly visual analysis method of matrix and application |
CN111126758A (en) * | 2019-11-15 | 2020-05-08 | 中南大学 | Academic team influence propagation prediction method, device and storage medium |
CN112148776A (en) * | 2020-09-29 | 2020-12-29 | 清华大学 | Academic relation prediction method and device based on neural network introducing semantic information |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160099010A1 (en) * | 2014-10-03 | 2016-04-07 | Google Inc. | Convolutional, long short-term memory, fully connected deep neural networks |
CN106170800A (en) * | 2014-09-12 | 2016-11-30 | 微软技术许可有限责任公司 | Student DNN is learnt via output distribution |
CN106557809A (en) * | 2015-09-30 | 2017-04-05 | 富士通株式会社 | Nerve network system and the method is trained by the nerve network system |
-
2017
- 2017-12-29 CN CN201711469074.XA patent/CN108228782B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106170800A (en) * | 2014-09-12 | 2016-11-30 | 微软技术许可有限责任公司 | Student DNN is learnt via output distribution |
US20160099010A1 (en) * | 2014-10-03 | 2016-04-07 | Google Inc. | Convolutional, long short-term memory, fully connected deep neural networks |
CN106557809A (en) * | 2015-09-30 | 2017-04-05 | 富士通株式会社 | Nerve network system and the method is trained by the nerve network system |
Non-Patent Citations (1)
Title |
---|
ZHONGYING ZHAO 等: "Identifying advisor-advisee relationships from co-author networks via a novel deep model", 《INFORMATION SCIENCES》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110008279A (en) * | 2019-03-27 | 2019-07-12 | 北京工商大学 | One kind is based on " relationship-weight " the orderly visual analysis method of matrix and application |
CN110008279B (en) * | 2019-03-27 | 2021-03-23 | 北京工商大学 | Visual analysis method based on 'relation-weight' ordered matrix and application |
CN111126758A (en) * | 2019-11-15 | 2020-05-08 | 中南大学 | Academic team influence propagation prediction method, device and storage medium |
CN112148776A (en) * | 2020-09-29 | 2020-12-29 | 清华大学 | Academic relation prediction method and device based on neural network introducing semantic information |
CN112148776B (en) * | 2020-09-29 | 2024-05-03 | 清华大学 | Academic relationship prediction method and device based on neural network introducing semantic information |
Also Published As
Publication number | Publication date |
---|---|
CN108228782B (en) | 2020-04-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Guo et al. | Deep graph neural network-based spammer detection under the perspective of heterogeneous cyberspace | |
Balaji et al. | Machine learning algorithms for social media analysis: A survey | |
Liu et al. | Anomaly detection in dynamic graphs via transformer | |
Wang et al. | An improved density peaks-based clustering method for social circle discovery in social networks | |
Hopcroft et al. | Who will follow you back? reciprocal relationship prediction | |
Xiao et al. | Rumor diffusion model based on representation learning and anti-rumor | |
CN110462612A (en) | The method and apparatus for carrying out machine learning using the network at network node with ageng and ranking then being carried out to network node | |
CN106295186A (en) | A kind of method and system of aided disease diagnosis based on intelligent inference | |
Huang et al. | Multi-modal Bayesian embedding for point-of-interest recommendation on location-based cyber-physical–social networks | |
CN110347932A (en) | A kind of across a network user's alignment schemes based on deep learning | |
Zhang et al. | Attribute-aware graph recurrent networks for scholarly friend recommendation based on internet of scholars in scholarly big data | |
Ran et al. | MGAT-ESM: Multi-channel graph attention neural network with event-sharing module for rumor detection | |
CN108228782A (en) | A kind of implication relation based on deep learning finds method | |
Zhang et al. | Link prediction across heterogeneous social networks: A survey | |
Yin et al. | A real-time dynamic concept adaptive learning algorithm for exploitability prediction | |
Liu et al. | Pair-wise ranking based preference learning for points-of-interest recommendation | |
González-Vidal et al. | A transfer learning framework for predictive energy-related scenarios in smart buildings | |
Xu et al. | Ssser: Spatiotemporal sequential and social embedding rank for successive point-of-interest recommendation | |
You et al. | QoS evaluation for web service recommendation | |
CN107807919A (en) | A kind of method for carrying out microblog emotional classification prediction using random walk network is circulated | |
Yang et al. | Rumor detection driven by graph attention capsule network on dynamic propagation structures | |
Wang et al. | Enhancing rumor detection in social media using dynamic propagation structures | |
Zhang et al. | Dynamic network embedding via structural attention | |
Liaghat et al. | Application of data mining methods for link prediction in social networks | |
Li et al. | Expertise network discovery via topic and link analysis in online communities |
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 |