CN105931122A - Ranking method of magic research communities in academic social network - Google Patents

Abstract

The invention discloses a ranking method of magic research communities in an academic social network and belongs to the data mining technical field. According to the ranking method, a time window is given, the magic research communities are found in the academic social network, wherein the magic research communities are attractive research communities which gradually become popular. The method specifically includes the steps of research community internal feature calculation, external feature FE calculation and ranking. The method of the invention can assist researchers to better understand and grasp current research trends and hotspots. According to the ranking method, hypotheses in existing researches can be broken through, namely, one person is only located at one community at one time point; and the characteristics of the magic research communities are extracted, and a unified algorithm is designed to rank potential popularity degrees of the research communities.

Description

The sort method of magic power research community in a kind of academic social networks

Technical field

The present invention is applied to find magic power research community in academic social networks, belongs to data mining technology field.

Background technology

In recent years, social network-i i-platform (such as: Facebook and Twitter) quickly grew.Meanwhile, social network analysis Also the extensive concern of academia is received.Academic network, as ingredient very important in social networks, also becomes research The research emphasis of personnel, sees list of references [1] J.Tang, J.Zhang, L.Yao, J.Li, L.Zhang, and Z.Su, “Arnetminer:extraction and mining of academic social networks,”in Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining.ACM,2008,pp. 990 998. lists of references [2] J.Tang, R.Jin, and J.Zhang, " A topic modeling approach and its integration into the random walk framework for academic search,”in Data Mining Eighth IEEE International Conference on.IEEE,2008,pp.1055–1060。

List of references [3] (G.Wang, Y.Zhao, X.Shi, and P.S.Yu, " Magnet community identification on social networks,”in Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining.ACM, 2012, pp.588 596.) propose the concept of magic power community and it It is applied to IT company and the sequence of financial company's captivation.This list of references [3] is intended to find the community that certain is popular, assumes simultaneously Being independent of each other between community, a people at a time can be only in a community.But in a lot of reality scenes (as Academic social networks), if the research worker of research identical content is regarded as a community, then find that those are the newest, The popular community of following meeting rather than those the most popular communities are more valuable.

Summary of the invention

It is an object of the invention to help research worker, particular without the research worker of experience, grind existing from the angle of macroscopic view The development studying carefully community has recognized, and helps research worker preferably to select the research topic of oneself.The present invention provides a kind of academic society Handing over the sort method of magic power research community in network, in described sort method, a given time window, at academic social networks Middle discovery magic power research community.Described magic power research community does not refer to those the most popular communities, but those are gradually The attractive research community come into vogue.The present invention is suitable for the application such as information retrieval and community's recommendation.

The sort method of magic power research community in the academic social networks that the present invention provides, specifically includes following steps:

The first step, research community internal feature calculates；

Select novelty degree as research community internal feature F_C, jth research community C_jNovel degree N_jRefer to the list of community's theme Word frequency change from time window s to s+1, uses N_j∈F_CRepresent；Research community C_jNovel degree N_jBe equivalent under theme every The sum of individual word novelty degree；

Research community C_jNovel degree N_jCalculate in the following manner:

$N_j=\underset{v=1}{\overset{\left|W\right|}{\mathrm{\Σ}}}{\mathrm{CW}}_{jv}\·log\frac{\left|C\right|}{{\mathrm{\δ}}_{C_{W_v}}}\·\frac{{\mathrm{\δ}}_{W_v}^{(s+1)}}{{\mathrm{\δ}}_{W_v}^{\left(s\right)}+{\mathrm{\δ}}_{W_v}^{(s+1)}}---\left(1\right)$

Wherein,Represent and comprise word W_vResearch communityNumber,WithRepresent word W respectively_vAt time window s With the number of times occurred in s+1；| W | is word W_vQuantity, v=1,2 ..., | W |；| C | is to study community in research community network Quantity；

Second step, research community surface F_ECalculate；

Given research worker R_iIn the community of time window s and s+1, distribution is respectivelyWithCalculation and Study personnel R_i Total transfer amount, Calculation and Study personnel R_iFrom research community C_j'To research community C_jTransfer amountWith transfer amount as matrix Element obtains transfer matrixTo each research worker R_iTransfer matrixIt is added, obtains final transfer matrix T；

3rd step, sequence；

Weighted Directed Graph G=(C, E, the F of given research community network_C,F_E), to arbitrary C_j∈ C, definition research community C_jInhale Draw other research community C_j'Ability u of research worker attention_jj′:

$u_{{\mathrm{jj}}^{\′}}=\mathrm{\α}\·\frac{T_{{\mathrm{jj}}^{\′}}^T}{O_j}+(1-\mathrm{\α})\·N_j---\left(2\right)$

Wherein, α is weight parameter,It is research community C_jTo research community C_j'Transfer matrix T_jj'Transposition, O_jIt is to grind Study carefully community C_jSize；

For arbitrary research community C_j∈ C, it is propagated attention and studies community C to other_j'Ability be defined as:

$v_{{\mathrm{jj}}^{\′}}=\frac{T_{{\mathrm{jj}}^{\′}}}{O_j}.---\left(3\right)$

Based on formula (2) and formula (3), for two scores of each research community definition: front score PS and negative score NS； Front score weighs the captivation of research community from the angle attracting attention, and negative score is weighed from the angle propagating attention and ground Study carefully the captivation of community, for research community C_jFront score PS_jWith negative score NS_jIt is defined as follows:

$\left\{\begin{array}{c}{\mathrm{PS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\mathrm{\Σ}}}\frac{1}{\left|Z\right|}{u}_{{\mathrm{jj}}^{\′}}{\mathrm{NS}}_{j^{\′}}\\ {\mathrm{NS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\mathrm{\Σ}}}\frac{1}{\left|Z^{\′}\right|}{v}_{{\mathrm{jj}}^{\′}}{\mathrm{PS}}_{j^{\′}}\end{array}\right.---\left(4\right)$

Wherein,WithIt it is normalization factor；

Based on the two score, study community C_jCaptivation grade A_jCalculated by equation below:

A_j=PS_j-NS_j (5)。

It is an advantage of the current invention that:

(1) present invention proposes for the first time, how to find potential attractive research community, the most just in academic social networks Being the problem of magic power research community, the present invention can help research worker to be best understood from and hold current research tendency and focus；

(2) it is during the present invention breaches existing research it is assumed that i.e. one people is only in a community a moment；

(3) present invention has extracted the feature of magic power research community, and devises the unified algorithm potential popular journey to research community Degree is ranked up.

Accompanying drawing explanation

Figure 1A and Figure 1B is the inventive method with existing algorithm Indegree, PageRank and MIM respectively at two different numbers According to the comparison schematic diagram on collection.

Fig. 2 A and Fig. 2 B is that related algorithm is illustrated with HotRank, TrandRank contrast on two different pieces of information collection respectively Figure.

Fig. 3 is community network G=(C, E, a F given in embodiment_C,F_E) schematic diagram.

Detailed description of the invention

Below in conjunction with drawings and Examples, the present invention is described in further detail.

The present invention provides the sort method of magic power research community in a kind of academic social networks, is that one is found to have captivation research society The method (Attractive Research Community Ranking is called for short ARTRank) in district.Described attractive research society District is also referred to as magic power research community, present invention A_iThe captivation grade of the magic power research community described in expression, A_i∈ A, A are institute Charm and study the captivation class set of community.Given research community network G=(C, E, FC, FE), then magic power research society Captivation class set A in district is defined as A=f (F_C,F_E), A is the internal feature F of research community network_CWith surface F_E Copula.Definition research community C_jWith research community C_j′Captivation grade be respectively A_jAnd A_j', then research community C_jRatio Research community C_j′More attractive and if only if A_j＞ A_j'.So finding the problem of magic power research community is a sequencing problem.

Research community network is defined as Weighted Directed Graph G=(C, E, F by the present invention_C,F_E), wherein C represents research community C_j's Set, each research community C_jBeing the research worker of one group of polymerization, E represents the set of link between research community, represents research people Member's transfer between research community.F_CRepresent the internal feature of each research community, F_EOutside for link between expression research community Portion's feature.

Present invention R_iRepresent i-th research worker, use Latent Dirichlet Allocation (LDA) (list of references [4]: D. M.Blei,A.Y.Ng,and M.I.Jordan,“Latent dirichlet allocation,”the Journal of machine Learning Research, vol.3, pp.993 1022,2003. list of references [5]: T.L.Griffiths and M.Steyvers, " Finding scientific topics,”Proceedings of the National Academy ofSciences,vol.101,no.suppl 1,pp. 5228 5235,2004.) research worker is clustered by model.By cluster, each research worker can be expressed as one and lead Probability distribution in topic (research community).The result of LDA model can be expressed as two matrixes, respectively matrix RC and square Battle array CW:

Matrix RC is the matrix of | R | × | C |, and | R | is the quantity of the research worker in research community network, and | C | is research community Network is studied the quantity of community, is also the research community total in the set of research community.RC_ijIt it is research worker R_iBelong to research Community C_jProbability.I=1,2 ..., | R |, j=1,2 ..., | C |.

Matrix C W is the matrix of | C | × | W |, and | W | is word W_vQuantity, v=1,2 ..., | W |.CW_jvIt is word W_vAssign Give research community C_jProbability.

Based on above-mentioned definition, the invention provides the sort method of magic power research community in a kind of academic social networks, including grinding Study carefully community's internal feature calculating, surface calculates and the sequence of grade captivation, specifically comprises the following steps that

The first step, research community internal feature calculates.

The present invention selects novelty degree as research community internal feature F_C.Jth research community C_jNovel degree N_jRefer to community master The frequency change from time window s to s+1 of the word of topic, uses N_j∈F_CRepresent.The concept source of the novel degree of research community is in society Hand over the concept of the topic detection that happens suddenly in network event detection.In event detection, when the word the most more frequency of certain event topic When occurring, then this event is defined as accident and (sees list of references [6]: Q.Diao, J.Jiang, F.Zhu, and numerously E.-P.Lim,“Finding bursty topics from microblogs,”in Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics:Long Papers-Volume 1.Association for Computational Linguistics,2012,pp.536–544.).One research community having novel subject matter can attract to grind more Study carefully the attention of personnel.The definition of the novel degree according to research community, studies community C_jNovel degree N_jBe equivalent under theme every The sum of individual word novelty degree.

First, word W_vResearch community C should be able to be represented well_jTheme.Use word W_vIt is assigned to study community C_j's Probability, namely CW_jv, represent word W_vTo research community C_jImportance.Wherein CW_jvObtained by LDA model. But this single standard is inadequate, such as word " network " occurs in multiple research communities, and word " social " only exists Some research community therein occurs, it is clear that than " network ", " social " can preferably represent that this studies community.So, this Invention uses IDF (inverse document frequency) value to weigh word W as supplementing_vTo research community C_jImportant Degree.According further to the definition of the novel degree of research community, word W_vFrequency change can image study community C_jNovelty Degree.Analyzed by above, study community C_jNovel degree N_jCan calculate in the following manner:

$N_j=\underset{v=1}{\overset{\left|W\right|}{\mathrm{\Σ}}}{\mathrm{CW}}_{jv}\·log\frac{\left|C\right|}{{\mathrm{\δ}}_{C_{W_v}}}\·\frac{{\mathrm{\δ}}_{W_v}^{(s+1)}}{{\mathrm{\δ}}_{W_v}^{\left(s\right)}+{\mathrm{\δ}}_{W_v}^{(s+1)}}---\left(1\right)$

Wherein,Represent and comprise word W_vResearch communityNumber,WithRepresent word W respectively_vAt time window s With the number of times occurred in s+1.

Second step, research community surface F_ECalculate.

Research worker transfer between research community reflects the development trend of research community, and the present invention divides the community of research worker From research community C in cloth_j'To research community C_jChange be defined as research worker from research community C_j'To research community C_jTurn Move.Given research worker R_iIn the community of time window s and s+1, distribution is respectivelyWithGenerally research worker R_i Attention can from some research community transfer to other research community.Assume for research worker R_i, research community obtains Transfer amount come from those research communities losing concern in proportion.After the transfer matrix calculating each research worker, The transfer matrix of all research worker is added, and just obtains the transfer matrix between final research community, is research community outside special Levy F_E。

The generation method of described transfer matrix includes: (1) Calculation and Study personnel R_iTotal transfer amount, (2) Calculation and Study personnel R_iFrom Research community C_j'To research community C_jTransfer amountTransfer matrix is obtained for matrix element with transfer amountAnd (3) To each research worker R_iTransfer matrixIt is added, obtains final transfer matrix T.Concrete methods of realizing is as follows:

3rd step, sequence.

Weighted Directed Graph G=(C, E, the F of given research community network_C,F_E), to arbitrary C_j∈ C, definition research community C_jInhale Draw other research community C_j'Ability u of research worker attention_jj:

$u_{{\mathrm{jj}}^{\′}}=\mathrm{\α}\·\frac{T_{{\mathrm{jj}}^{\′}}^T}{O_j}+(1-\mathrm{\α})\·N_j---\left(2\right)$

Wherein, α is weight parameter,It is research community C_jTo research community C_j'Transfer matrix T_jj'Transposition, O_jIt is to grind Study carefully community C_jSize, say, that the sort method of the present invention is to research community magnitude.Such a popular grinding Study carefully community and will obtain little contribution from other research communities, and the tribute that communities obtain is studied, from other, by attractive research community Offering a lot, because popular research community is the biggest, and attractive research community is the most relatively small.With Sample, when the ability of Calculation and Study Community Communication attention, it is also considered that research community size, to avoid finding those the least Research community.It is to say, for arbitrary research community C_j∈ C, it is propagated attention and studies community C to other_j'Energy Power can be defined as:

$v_{{\mathrm{jj}}^{\′}}=\frac{T_{{\mathrm{jj}}^{\′}}}{O_j}.---\left(3\right)$

Based on formula (2) and formula (3), for two scores of each research community definition: front score PS and negative score NS. Front score weighs the captivation of research community from the angle attracting attention, and negative score is weighed from the angle propagating attention and ground Study carefully the captivation of community.For research community C_jFront score PS_jWith negative score NS_jIt is defined as follows:

$\left\{\begin{array}{c}{\mathrm{PS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\mathrm{\Σ}}}\frac{1}{\left|Z\right|}{u}_{{\mathrm{jj}}^{\′}}{\mathrm{NS}}_{j^{\′}}\\ {\mathrm{NS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\mathrm{\Σ}}}\frac{1}{\left|Z^{\′}\right|}{v}_{{\mathrm{jj}}^{\′}}{\mathrm{PS}}_{j^{\′}}\end{array}\right.---\left(4\right)$

Wherein,WithIt it is normalization factor.One research community has high front score PS and means that it studies community from other Obtain a lot of contributions, had high negative score NS on the contrary and mean that it is that a lot of contributions has been done by other research communities.One Can there be high front score PS and high negative score NS in individual popular research community simultaneously, and attractive research community has High front score PS and low negative score NS.Based on the two score, study community C_jCaptivation grade A_jPermissible Calculated by equation below:

A_j=PS_j-NS_j (5)

According to formula (4) and formula (5), the detailed step of described research community captivation grade sort method is as follows:

Embodiment:

The data set used in the present embodiment is from ArnetMiner (list of references [1], [2]).

Computer science data set: this data set comprises 2 relevant with computer science, 084,055 paper, and every paper comprises Exercise question, author, deliver time and summary etc..Author, exercise question and summary letter is extracted from the data of 2005 to 2010 years Breath, then obtains and 2005-2006, the data that two time windows of 2007-2008 are relevant, respectively contains 47565 works The relevant information of person.

Data fields conferencing data collection: this data set is the subset of computer science data set, these data take from nine top-level meetings (SIGMOD, KDD, VLDB, SIGIR, ICDE, CIKM, WWW, ICDM and WSDM), by pretreatment, Two time windows of 2005-2006 and 2007-2008 remain the information of 2399 authors respectively.

The present embodiment uses LDA model to carry out community's detection.On computer science data set, if number of topics | C | is 300, That is 300 research communities are found.On data fields conferencing data collection, if number of topics | C | is 50, namely want Find 50 research communities.Two Dirichlet hyper parameter in LDA model are set toWith 0.01.

The sort method that the present invention proposes compares with following four method.

Control methods 1: in-degree (Indegree): Indegree weighs the captivation of research community only by the amount of proceeding to of research worker.

Control methods 2:PageRank (sees list of references [7]: L.Page, S.Brin, R.Motwani, and T.Winograd, " The Pagerank citation ranking:Bringing order to the web. " 1999.): PageRank research worker between community A kind of ballot is regarded in transfer as, when weighing community's captivation, it not only allows for the quantity of transfer, it is also contemplated that the quality of transfer.

Control methods 3: magic power community discovery model (MIM) (sees list of references [3]): this model uses one based on PageRank Optimization Framework weigh the captivation of community, in an experiment, in this model that community's novelty degree factor is joined.

Control methods 4: trend sequence (TrendRank): TrendRank uses linear regression method, utilizes topic keyword every year Shared ratio generates the Trendline (list of references [8]: A.Hoonlor, B.K.Szymanski, and M. of theme popularity change J.Zaki,“Trends in computer science research,”Communications of the ACM,vol.56,no.10,pp. 74–83,2013.).With with time window 2007-2008,2009,2010 relevant data carry out trend sequence.Instinctively, more Attractive community more should have a good development trend in the near future.In the present embodiment using TrendRank as The standard of experiment.

The present embodiment uses recommends intensity (recommendation intensity) (to see list of references [9]: R.Hampel and M. Hauck,“Towards an effective use of audio conferencing in distance language courses,”Language Learning&Technology, vol.8, no.1, pp.66 82,2004.) as estimating standard, it is defined as follows:

$RI\left(L_i\right)@k=\left\{\begin{array}{cc}2-\frac{\left|O_r-O_g\right|}{k}& L_i\∈TrendRank\\ 0& L_i\∉TrendRank\end{array}\right.---\left(6\right)$

In formula (6), L is the sorted lists of front k the research community that certain sort method generates.O_rIt is L_iIn L Sorting position, O_gIt is L_iSorting position in TrendRank.This also implies that, if L_iOccur in TrendRank's In the list of front k, and its sorting position is the most accurate, then L_iRecommendation the highest will be obtained Intensity value.So the recommendation intensity of front k the sequence of L can be defined as:

$RI\left(L\right)@k=\underset{L_i\∈L}{\mathrm{\Σ}}RI\left(L_i\right)@k---\left(7\right)$

Table 1 illustrates by front 15 research communities of PageRank, MIM, ARTRank and Trendrank sequence.According to Word keyword in theme is that these themes label, and the community adding bold Italic occurs from Trendrank sequence front 15 Community.Owing to the result of Indegree and PageRank is closely similar, so table 1 omits the result of Indegree.

It will be seen that the ARTRank method that the present invention provides has the biggest advantage than relevant method from table 1, ARTRank Front the 15 of sorted lists have 9 to occur in Trendrank, and MIM and PageRank only 5 and 3 respectively.Ratio It is the most attractive community in TrendRank as studied community " cloud ", only occurs in the sequence of 15 before ARTRank In result, and before PageRank and MIM, the ranking results of 15 does not occur.On data fields conferencing data collection, Have also been obtained similar result, owing to space limits, omit herein.

Except accuracy, the present invention also compares these methods quantized result under recommendation intensity, as Shown in Figure 1A and Figure 1B.ARTRank is substantially better than additive method, because InDegree only considered the quantity of transfer, this Sample it can tend to find those popular communities.Although PageRank considers the architectural feature of network, it can be looked for equally To the community that those are popular.Some is similar, this is because the present invention all ensures that often for the result of MIM and the result of the method for the present invention The amount of proceeding to of individual community wants the unnecessary amount of producing.But, the effect of ARTRank is still good than MIM because MIM be based on PageRank's, its purpose is the community finding a certain class popular, this target obviously and the present invention to find new gradually The target of popular research community is different.

15 research community sequence before table 1 computer science data set

It is an object of the present invention to find those attractive research communities rather than community that those are popular.Then, this Bright related algorithm and community temperature sort algorithm (HotRank) compare, and HotRank is based on the time of community's theme What community was ranked up by intensity (sees list of references [10]: X.Wang, C.Zhai, and D.Roth, " Understanding evolution of research themes:a probabilistic generative model for citations,”in Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining.ACM, 2013, pp.1115 1123. lists of references [11]: D.Zhou, X.Ji, H.Zha, and C.L.Giles, " Topic evolution and social interactions:how authors effect research,”in Proceedings of the 15th ACM international conference on Information and knowledge management.ACM,2006,pp.248–257.)。

Fig. 2 A and Fig. 2 B gives result of the comparison, and wherein the data of HotRank take from 2009.It will be seen that Indegree With the ranking results of PageRank and HotRank closer to, and and the ranking results difference of TrendRank far, say, that this Two algorithms are more likely to find popular community.But the sequence to community's temperature, the trend of community development can not be reflected, But because awfully hot community starting to go down hill still can in temperature sorts ranking the highest.The ranking results of ARTRank Then the most consistent with TrendRank, say, that the ranking results of ARTRank more can reflect the trend of community development.

Present invention could apply in academic social networks, research on utilization personnel are at the intercommunal transfer of difference research, Yi Jiyan Study carefully the subject novel degree of community, the captivation grade of research community is ranked up.ARTRank algorithm is with research worker At the article of different year as input, with study community captivation grade ranking for output.Such as, Fig. 3 gives one Embodiment.Community network in Fig. 3 is at the article content of different year by research worker, shifts square according to novelty degree and community Battle array algorithm generates, and the size of circle represents the size of community, and the value of square frame represents the novel degree of community, and the value on limit represents research The transfer amount of personnel.By the sort method of the present invention, obtain studying the captivation ranking of community: 4,1,2,3.Obviously grind Study carefully community 4 and have the theme of novelty, while attracting a lot of research worker, have again little people to leave, simultaneously the scale of community The least, it is clear that to be new and the most popular research community to be looked in the present embodiment.

Claims (3)

1. the sort method of magic power research community in an academic social networks, it is characterised in that comprise the steps:

The first step, research community internal feature calculates；

Select novelty degree as research community internal feature F_C, jth research community C_jNovel degree N_jRefer to the list of community's theme Word frequency change from time window s to s+1, uses N_j∈F_CRepresent；Research community C_jNovel degree N_jBe equivalent under theme every The sum of individual word novelty degree；

Research community C_jNovel degree N_jCalculate in the following manner:

$N_j=\underset{v=1}{\overset{\left|W\right|}{\Σ}}{\mathrm{CW}}_{jv}\·\log \frac{\left|C\right|}{{\mathrm{\δ}}_{C_{W_v}}}\·\frac{{\mathrm{\δ}}_{W_v}^{(s+1)}}{{\mathrm{\δ}}_{W_v}^{\left(s\right)}+{\mathrm{\δ}}_{W_v}^{(s+1)}}---\left(1\right)$

Wherein,Represent and comprise word W_vResearch communityNumber,WithRepresent word W respectively_vAt time window s With the number of times occurred in s+1；| W | is word W_vQuantity, v=1,2 ..., | W |；| C | is to study community in research community network Quantity；

Second step, research community surface F_ECalculate；

Given research worker R_iIn the community of time window s and s+1, distribution is respectivelyWithCalculation and Study personnel R_i Total transfer amount, Calculation and Study personnel R_iFrom research community C_j'To research community C_jTransfer amountWith transfer amount as matrix Element obtains transfer matrixTo each research worker R_iTransfer matrixIt is added, obtains final transfer matrix T；

3rd step, sequence；

Weighted Directed Graph G=(C, E, the F of given research community network_C,F_E), to arbitrary C_j∈ C, definition research community C_jInhale Draw other research community C_j'Ability u of research worker attention_jj':

$u_{{\mathrm{jj}}^{\′}}=\mathrm{\α}\·\frac{T_{{\mathrm{jj}}^{\′}}^T}{O_j}+(1-\mathrm{\α})\·N_j---\left(2\right)$

Wherein, α is weight parameter,It is research community C_jTo research community C_j'Transfer matrix T_jj'Transposition, O_jIt is to grind Study carefully community C_jSize；

For arbitrary research community C_j∈ C, it is propagated attention and studies community C to other_j'Ability be defined as:

$v_{{\mathrm{jj}}^{\′}}=\frac{T_{{\mathrm{jj}}^{\′}}}{O_j}.---\left(3\right)$

Based on formula (2) and formula (3), for two scores of each research community definition: front score PS and negative score NS； Front score weighs the captivation of research community from the angle attracting attention, and negative score is weighed from the angle propagating attention and ground Study carefully the captivation of community, for research community C_jFront score PS_jWith negative score NS_jIt is defined as follows:

$\{\begin{array}{c}{\mathrm{PS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\Σ}}\frac{1}{\left|Z\right|}{u}_{{\mathrm{jj}}^{\′}}{\mathrm{NS}}_{j^{\′}}\\ {\mathrm{NS}}_j=\underset{j^{\′}=1}{\overset{\left|C\right|}{\Σ}}\frac{1}{\left|Z^{\′}\right|}{v}_{{\mathrm{jj}}^{\′}}{\mathrm{PS}}_{j^{\′}}\end{array}---\left(4\right)$

Wherein,WithIt it is normalization factor；

Based on the two score, study community C_jCaptivation grade A_jCalculated by equation below:

A_j=PS_j-NS_j (5)。

The sort method of magic power research community in a kind of academic social networks the most according to claim 1, it is characterised in that second The concrete calculation procedure of transfer matrix T final in step is as follows:

Step 1, each research worker R_i, i=1,2 ..., | R |, perform step 2-11；

Step 2, make research worker R_iTotal transfer amount Q=0,

Step 3, for each research community C_j∈ C, performs step 4；

Step 4, given research worker R_iIn the community of time window s and s+1, distribution is respectivelyWithIfPerform step 5；

Step 5, update total transfer amount

Step 6, for each research community C_j∈ C, performs step 7-10；

If step 7Perform step 8；

Step 8, for each research community C_j'∈ C, performs step 9；

If step 9Perform step 10；

Step 10, research worker R_iFrom research community C_j' to C_jTransfer amount:

$T_{j^{\′}j}^{\left(R_i\right)}=\frac{{\mathrm{RC}}_{ij}^{(s+1)}-{\mathrm{RC}}_{ij}^{\left(s\right)}}{Q}\·\left|{\mathrm{RC}}_{{\mathrm{ij}}^{\′}}^{(s+1)}-{\mathrm{RC}}_{{\mathrm{ij}}^{\′}}^{\left(s\right)}\right|;$

Step 11, to the transfer amount of each research worker Ri as transfer matrixElement, and by all of transfer matrix Add and, obtain final transfer matrix T.

The sort method of magic power research community in a kind of academic social networks the most according to claim 1, it is characterised in that the 3rd The detailed step studying community's captivation grade sort method in step is as follows:

Step 1, a given positive number ξ ＞ 0, front score and negative score for each research community compose initial value PS respectively₀With NS₀；

Step 2, (1,1,1,1) ∈ R^|C|It is assigned to PS₀；R^|C|Represent that numerical value dimension is the vector space of | C |；

Step 3, (1,1,1,1) ∈ R^|C|It is assigned to NS₀；

Step 4, repeated execution of steps 5 to step 9；

Step 5, use NS_k-1Update PS_k；Formula (4)；

Step 6, use PS_k-1Update NS_k；Formula (4)；

Step 7, standardization PS_k；

Step 8, standardization NS_k；

Step 9, k increase by 1；

Step 10, until meet condition | PS_k-PS_k-1| ＜ ξ, and | NS_k-NS_k-1| ＜ ξ.