CN103136331A - Micro blog network opinion leader identification method - Google Patents

Abstract

The invention discloses a micro blog network opinion leader identification method and is used for solving the technical problem of poor recall rate of the existing opinion leader identification method. The technical scheme includes storing network topology information collected from the internet into a database by using a web crawler tool; constructing a directed network diagram G=(V, E); calculating an effective fan collection Ef (u); calculating a node weight IRL (ui) generated by a link relationship; calculating a node weight IRTR (ui) generated by a node interactive relationship; calculating a node comprehensive weight IR (ui); and calculating comprehensive weights of all nodes in the network diagram, performing sequencing in a descending order according to the comprehensive weights, and selecting n nodes with larger comprehensive weights as candidates of opinion leaders. The fan number possessed by the nodes, factors such as the node link relationship and the interactive relationship are considered when the node weights are calculated, so that the recall rate and the accuracy are improved. Through detection, the return rate is improved from 81.7-88.5% of the background technology to above 89.3%, and the accuracy is improved from 84.7-90.4% of the background technology to above 91.7%.

Description

Microblogging network leader of opinion recognition methods

Technical field

The present invention relates to a kind of recognition methods, be specifically related to a kind of microblogging network leader of opinion recognition methods.

Background technology

Along with the development of Web2.0 technology, some novel network applications have appearred in the internet, as social networks, microblogging network etc., aspect Information Communication and interpersonal relation carrying, show increasing value and influence power.

Social networks is intended to help people more effectively to set up and keep human relation network as the expansion on the internet of real social networks.Different from the website take aggregation information as characteristics, social networks is take the polymerization crowd as characteristics, people can set up and keep the friend's circle of oneself by social networks, become a kind of novel individual social mode and information interchange platform, by means of the Information Communication pattern that friend's public praise passes on from one to another, accelerated the propagation of information.

Microblogging network (Micro-B1ogging Network) is also a kind of social networks, the user can pass through multiple channel issue 140 words such as browser, mobile phone, instant communication software with interior information, the Information Communication characteristic of this instantaneity, fragmentation, polymerism is subject to users' welcome, and the registered user of domestic Sina microblogging has surpassed 300,000,000 people.

In the spreading network information process, the leader of opinion has brought into play important effect.Local suggestion develops under leader of opinion's guiding and impact and is network public opinion.The leader of opinion claims again opinion leader, refer in interpersonal communication's network " activist " that often information be provided and exert one's influence for other people, they play important intermediary or filtration in the forming process of mass media effect, by them with diffusion of information to the audient, the two-step flow of communication flow that formation information is transmitted.Along with continuing to increase of network public opinion influence power, people are also constantly going deep into network leader of opinion's research.

The statistics demonstration, the most of user in network does not often participate in manufacturing and the propagation of information, and the leader of opinion is often followed in the decision that they make.Recognition network leader of opinion effectively, delivering guided bone information by the leader of opinion affects the place network user but not directly persuades them, can effectively trigger the influence power of whole network or society, for promoting Information Communication, improve demonstration effect and have important practical significance.

The microblogging network is a kind of complex network, usually adopts Complex Networks Theory to carry out modeling analysis to this class complex network both at home and abroad, to disclose intrinsic characteristic and the scientific law of complex network.According to Complex Networks Theory, the microblogging network can be abstracted into a kind of directed networks figure, each user consists of the node in network, limit between user's Relations Among configuration node, the friend who has due to each user is different with bean vermicelli quantity, so each node has different weights, and the node weights are larger, the influence power that this node is described is larger, and the possibility that becomes the leader of opinion is also just larger.Therefore how leader of opinion's identification problem can be summed up as the computing node weighted problem.By setting up the directed networks graph model, structural relation between analysis node is calculated the weights of each node, and the node weights are larger, and the possibility that becomes the leader of opinion is just larger.

Document 1 " Discovering Important Bloggers based on Analyzing Blog Threads[WWW; Chiba; Japan; May 10-14; 2005] " has proposed a kind of blog responsible consumer analytical approach ThreadRank based on the model content analysis, the method judges its user's importance by analyzing a large amount of Blog contents, need to expend a large amount of time for the content cleaning and analyze, and efficient is lower.

Document 2 " Identifying Opinion Leaders in the Blogosphere[CIKM; pp.971-974; 2007] " has proposed a kind of leader of opinion's recognition methods InfluenceRank, the method is according to comparing to judge user's importance with other blogs, and user's weights are calculated in the contribution that these users do whole network, and adopt the cosine law to calculate the similarity of different blog entities, and complicacy is higher, and expense is large.

Document 3 " TwitterRank:Finding topic-sensitive Inuential Twitterers[WSDM; 2010] " has proposed a kind of Twitter network node computing method TwitterRank, the method is carried out weight calculation according to the distribution between the customer relationship in Twitter, bean vermicelli and follower and various user groups play in the process of Information Communication effect, this algorithm is mainly analyzed based on topic, and recall rate is not high.

Summary of the invention

Recall the deficiency of rate variance in order to overcome existing leader of opinion's recognition methods, the invention provides a kind of microblogging network leader of opinion recognition methods.The method is come identification nodes influence power and importance by node weights, and the node weights are larger, and the possibility that becomes the leader of opinion is just larger.When the computing node weight, consider bean vermicelli quantity and the many factors such as node link relation and interactive relation that node has, can improve recall rate, improve simultaneously accuracy rate.

The technical solution adopted for the present invention to solve the technical problems is: a kind of microblogging network leader of opinion recognition methods is characterized in comprising the following steps:

Step 1, utilize the web crawlers instrument, gather actual microblogging network data from the internet, extracting wherein network topological information, to deposit database in pending.

Step 2, structure microblogging directed networks figure

G＝（E，V）

In formula, E represents the node relationships set, and V represents node set.

Step 3, calculating effective bean vermicelli set Ef (u)

Ef(u)={v|v∈Follower(u)∧Response(u)>δ}

In formula, δ is non-negative constant threshold, the feedback degree thresholding of the bean vermicelli node v of expression node u to node u, and the bean vermicelli that surpasses this threshold value and belong to node u just can be can be regarded as effective bean vermicelli.

The node weights IRL (u that step 4, calculating are produced by linking relationship _i)

$\left(u_i\right)=\frac{\mathrm{\σ}}{N}+(1-\mathrm{\σ})\underset{u_j\∈\mathrm{Follower}\left(u_i\right)}{\mathrm{\Σ}}\frac{\mathrm{IRL}\left(u_j\right)}{L\left(u_j\right)}$

In formula, Follower (u _i) be node u _iAll bean vermicelli set, L (u _j) be node u _jThe bean vermicelli number, σ is the ratio of damping between 0 and 1, N is the total nodes in network chart.

The node weights IRTR (u that step 5, calculating are produced by the node interactive relation _i)

$\left(u_i\right)=\underset{t_j\∈\mathrm{Tweet}\left(u_i\right)}{\mathrm{\Σ}}\frac{\underset{u_j\∈\mathrm{Re\; sponse}\left(t_j\right)}{\mathrm{\Σ}}|N_s\left(u_j\right)-N_{\mathrm{\μ}}\left(u_j\right)|}{\left|A\right|}$

In formula, Tweet (u _i) be node u _iThe model set, A represents that all have the model collection of mutual situation | A| is the set of A, N _s(u _j) be node u _jFor model t _j, response times, N _μ(u _j) being response mean value, Response comprises that the user turns note, money order receipt to be signed and returned to the sender, comment and collection.

Step 6, the comprehensive weights IR of computing node (u _i)

IR(u _i)=(1-β)×IRL(u _i)+β×IRTR(u _i)

In formula, parameter beta (β ∈ [0,1]) decision linking relationship and two factors of node interactive relation are residing status in the calculating of node weights; When β hour, the node weights determine by linking relationship, calculate weights by linking relationship fully especially when β=0.

The comprehensive weights of all nodes in step 7, computational grid figure, and by the descending sequence of comprehensive weights, choose the larger n of a comprehensive weights node, as leader of opinion's candidate target.

The invention has the beneficial effects as follows: owing to coming identification nodes influence power and importance by node weights, the node weights are larger, and the possibility that becomes the leader of opinion is just larger.When the computing node weight, consider bean vermicelli quantity and the many factors such as node link relation and interactive relation that node has, improved recall rate, improved simultaneously accuracy rate.After testing, recall rate is brought up to more than 89.3% by 81.7～88.5% of background technology, and accuracy rate is brought up to more than 91.7% by 84.7～90.4% of background technology.

Below in conjunction with drawings and Examples, the present invention is elaborated.

Description of drawings

Fig. 1 is the process flow diagram of microblogging network leader of opinion of the present invention recognition methods.

Embodiment

With reference to Fig. 1.Microblogging network leader of opinion recognition methods concrete steps of the present invention are as follows:

1. obtain the microblogging network data: utilize the web crawlers instrument, gather actual microblogging network data from the internet, extracting the network topological informations such as wherein node, connection, to deposit database in pending.

2. build microblogging directed networks figure:

G＝（E，V）

In formula, E represents the node relationships set, and V represents node set.

3. calculate effective bean vermicelli set Ef (u):

Ef(u)={v|v∈Follower(u)∧Response(u)>δ} (1)

In formula, δ is non-negative constant threshold, the feedback degree thresholding of the bean vermicelli node v of expression node u to node u, and the bean vermicelli that surpasses this threshold value and belong to node u just can be can be regarded as effective bean vermicelli.

4. calculate the node weights IRL (u that is produced by linking relationship _i):

$\left(u_i\right)=\frac{\mathrm{\σ}}{N}+(1-\mathrm{\σ})\underset{u_j\∈\mathrm{Follower}\left(u_i\right)}{\mathrm{\Σ}}\frac{\mathrm{IRL}\left(u_j\right)}{L\left(u_j\right)}---\left(2\right)$

In formula, Follower (u _i) be node u _iAll bean vermicelli set, L (u _j) be node u _jThe bean vermicelli number, σ is the ratio of damping between 0 and 1, N is the total nodes in network chart.

5. calculate the node weights IRTR (u that is produced by the node interactive relation _i):

$\mathrm{IRTR}\left(u_i\right)=\underset{t_j\∈\mathrm{Tweet}\left(u_i\right)}{\mathrm{\Σ}}\frac{\underset{u_j\∈\mathrm{Re\; sponse}\left(t_j\right)}{\mathrm{\Σ}}|N_s\left(u_j\right)-N_{\mathrm{\μ}}\left(u_j\right)|}{\left|A\right|}---\left(3\right)$

In formula, Tweet (u _i) be node u _iThe model set, A represents that all have the model collection of mutual situation | A| is the set of A, N _s(u _j) be node u _jFor model t _j, response times, N _μ(u _j) being response mean value, Response comprises that the user turns note, money order receipt to be signed and returned to the sender, comment and collection.

6. the comprehensive weights IR of computing node (u _i):

IR(u _j)=(1-β)×IRL(u _i)+β×IRTR(u _i) (4)

In formula, parameter beta (β ∈ [0,1]) decision linking relationship and two factors of node interactive relation are residing status in the calculating of node weights; When β hour, the node weights mainly determine by linking relationship, calculate weights by linking relationship fully especially when β=0.

7. comprehensive weights of all nodes in computational grid figure, and by the descending sequence of comprehensive weights, choose the larger n of a comprehensive weights node, as leader of opinion's candidate target.

The present invention has improved existing methodical deficiency from counting yield and degree of accuracy two aspects.At first, by defining effective bean vermicelli set, will be not or the node that has a small amount of bean vermicelli exclude, it is minimum that they become leader of opinion's possibility, because leader of opinion or high weight node must have a large amount of beans vermicelli, so just can significantly reduce the network chart scale, be conducive to improve counting yield.Secondly, when the computing node weights, not only consider the linking relationship that produced by bean vermicelli, also considered the node interactive relation that issue, forwarding, reply and the collection etc. of model produce, therefore improved counting accuracy.

The present invention and existing method contrast and experiment are as shown in table 1.

The recall rate of the various node weight value calculating methods of table 1, accuracy rate and the contrast of average nodal processing time

Computing method

Out-degree

In-degree/out-degree combination

Document 1

Document 2

Document 3

The present invention

Recall rate

57.3％

65.4％

82.2％

81.7％

88.5％

89.3％

Accuracy rate

62.2％

67.3％

86.1％

84.7％

90.4％

91.7％

Time/node

0.14min

0.23min

3.37min

2.81min

2.76min

0.31min

This experiment is to process 100,000 microblogging network nodes as benchmark test.As can be seen from Table 1, although higher based on out-degree, in-degree/method counting yielies such as out-degree combination, accuracy rate and recall rate are very low; Although the method that document 1, document 2 and document 3 propose has higher accuracy rate and recall rate, counting yield is lower; And the present invention not only has higher counting yield, and has higher accuracy rate and recall rate.

Claims (1)

1. microblogging network leader of opinion recognition methods is characterized in that comprising the following steps:

Step 1, utilize the web crawlers instrument, gather actual microblogging network data from the internet, extracting wherein network topological information, to deposit database in pending;

Step 2, structure microblogging directed networks figure

G＝（E，V）

In formula, E represents the node relationships set, and V represents node set;

Step 3, calculating effective bean vermicelli set Ef (u)

Ef(u)={v|v∈Follower(u)∧Response(u)>δ}

In formula, δ is non-negative constant threshold, the feedback degree thresholding of the bean vermicelli node v of expression node u to node u, and the bean vermicelli that surpasses this threshold value and belong to node u just can be can be regarded as effective bean vermicelli;

The node weights IRL (u that step 4, calculating are produced by linking relationship _i)

$\left(u_i\right)=\frac{\mathrm{\σ}}{N}+(1-\mathrm{\σ})\underset{u_j\∈\mathrm{Follower}\left(u_i\right)}{\mathrm{\Σ}}\frac{\mathrm{IRL}\left(u_j\right)}{L\left(u_j\right)}$

In formula, Follower (u _i) be node u _iAll bean vermicelli set, L (u _j) be node u _jThe bean vermicelli number, σ is the ratio of damping between 0 and 1, N is the total nodes in network chart;

The node weights IRTR (u that step 5, calculating are produced by the node interactive relation _i)

$\mathrm{IRTR}\left(u_i\right)=\underset{t_j\∈\mathrm{Tweet}\left(u_i\right)}{\mathrm{\Σ}}\frac{\underset{u_j\∈\mathrm{Re\; sponse}\left(t_j\right)}{\mathrm{\Σ}}|N_s\left(u_j\right)-N_{\mathrm{\μ}}\left(u_j\right)|}{\left|A\right|}$ In formula, Tweet (u _i) be node u _iThe model set, A represents that all have the model collection of mutual situation | A| is the set of A, N _s(u _j) be node u _jFor model t _j, response times, N _μ(u _j) being response mean value, Response comprises that the user turns note, money order receipt to be signed and returned to the sender, comment and collection;

Step 6, the comprehensive weights IR of computing node (u _i)

IR(u _i)=(1-β)×IRL(u _i)+β×IRTR(u _i)

In formula, parameter beta (β ∈ [0,1]) decision linking relationship and two factors of node interactive relation are residing status in the calculating of node weights; When β hour, the node weights determine by linking relationship, calculate weights by linking relationship fully especially when β=0;

The comprehensive weights of all nodes in step 7, computational grid figure, and by the descending sequence of comprehensive weights, choose the larger n of a comprehensive weights node, as leader of opinion's candidate target.

Images