CN103902690B - Method for improving accuracy of influence of user generate content (UGC) information of social network - Google Patents

Abstract

The invention discloses a method for improving the influence accuracy of content information generated by social network users. The user-generated content UGC includes M keywords, and a total of N users participate in the UGC. The method includes: establishing social network UGC members Participation mechanism: Construct fan network unauthorized directed graph according to user-fan relationship of UGC and carry out community division; build interest network authority undirected graph and carry out community division according to user reply relationship of UGC; Calculate the social influence U _X of user X based on the degree of correlation among the various influencing factors of the member participation mechanism; Calculate the social influence of user X publishing keyword K, m is the number of times keyword K spreads on user X, if m=0, then S _KX =0; according to the formula Calculating the comprehensive social influence of the keyword K in the UGC; calculating the sum of the comprehensive social influences of the M keywords in the UGC to obtain the information influence INF of the UGC.

Description

A method for improving the influence accuracy of content information generated by social network users

technical field

The invention relates to information monitoring technology, in particular to a method for improving the influence accuracy of content information generated by social network users.

Background technique

The Internet has entered the era of web2.0, where every user is free to express their opinions. Many important content or news are first generated through User Generated Content (UGC, User Generate Content), and then widely disseminated through social networks, and finally in a certain within a specific social circle or even society as a whole. Therefore, research on UGC influence plays a very important role in information collection, monitoring, and prediction. However, because the number of UGCs is too large and the speed of increase is very fast, it is difficult to deal with all UGCs. It is necessary to screen out UGCs with good quality and high influence for research and utilization. Therefore, more and more attention has been paid to the research on the quality and information influence evaluation of UGC.

The current research on information influence mainly uses the Influence Diffusion Model (IDM, Influence Diffusion Model) and its improved models (such as the Influence Diffusion Probability Model IDPM, Influence Diffusion Probability Model, etc.) for analysis. The influence diffusion model based on text conversation IDM uses the reply chain structure in the conversation to calculate the similarity between texts based on word frequency to calculate the influence diffusion ability of the source. The sum of the influence of each reply is the influence of the text. Diffusion ability. After the model was put forward, it became an important cornerstone of information influence research. Most of the research on information influence in later generations is to improve the model; the influence diffusion probability model IDPM spreads probability influence by defining a single keyword in the entire interest space. To solve the structural fault problem of influence transmission in the IDM model and the false influence propagation problem caused by irrigation, and to solve the content fault problem of influence transmission in the IDM model by considering the effective keywords in the sentence.

However, these models have some very obvious defects, such as the weight of each comment or reply is the same, and the relationship between users is not considered. Take a post on BBS as UGC as an example, as shown in Figure 1:

User 1 is the information publisher, users 2 to 5 are the replies of user 1, A, B, C, D, E, and F are the keywords included in the post, and the thick solid line indicates the influence of the post among users relationship, the direction of the thick solid line is the influence propagation direction of the post, the dotted line indicates the fan relationship between users, the dotted line indicates the relationship that the user belongs to the same community in the interest network, and the thin solid line indicates that the user belongs to the same community in the fan network. group relationship.

In Figure 1, users 2 to 5 all replied to user 1’s post, but user 2 is a fan of user 1, user 3 and user 1 belong to the same interest network community, and user 4 and user 1 belong to the same fan network community ( but not a fan of user 1), user 5 is new and probably had little to no relationship with user 1 before.

It can be seen from this that, in the IDPM model, the keywords of UGC are not separately weighted, which will lead to deviations in the calculation of the information influence of UGC.

Contents of the invention

In view of this, the present invention proposes a method for improving the accuracy of the influence of content information generated by social network users, which effectively solves the problem in the prior art that there is a deviation in the calculation of UGC information influence due to the fact that the keywords of UGC are not separately weighted. defect. The technical scheme that the present invention proposes is:

A method for improving the influence accuracy of content information generated by social network users, the method comprising:

A. Establish a social network UGC member participation mechanism, determine the path coefficient between the various influencing factors of the member participation mechanism, and the path coefficient is the degree of correlation between the various influencing factors of the member participation mechanism;

B. Construct a fan network unauthorized directed graph based on the user-fan relationship of the UGC, and perform community division on the fan network unauthorized directed graph; construct an interest network authorized undirected graph based on the user reply relationship of the UGC , performing community division on the right undirected graph of the interest network;

C. Calculate the social influence U _X of user X according to the degree of correlation between the various influencing factors of the member participation mechanism;

D. According to Calculate the social influence of user X publishing keyword K, m is the number of times keyword K spreads on user X, if m=0, then S _KX =0;

E. According to the formula Calculating the comprehensive social influence of keyword K in the UGC;

F. Calculate the sum of the comprehensive social influences of the M keywords in the UGC to obtain the information influence INF of the UGC.

In the above scheme, the member participation mechanism includes four influencing factors: information quality, group identity, value perception and participation. The path coefficient of information quality and group identity is a ₁ , and the path coefficient of information quality and value perception is a ₂ , the path coefficient of value perception and group identity is a ₃ , and the path coefficient of participation and group identity is a ₄ .

In the above scheme, the step C further includes:

According to the formula Calculate the social influence of user X in the UGC,

Wherein, b is the number of times that user X is directly replied in the UGC, if the user does not have a direct reply person, then U _X =0;

If user X and his direct responder belong to the same interest network community, C ₁ =a ₁ , otherwise, C ₁ =1;

If user X and his direct reply belong to the same fan network community, C ₂ =a ₂ ×a ₃ , otherwise, C ₂ =1;

If user X is a fan of the UGC information publisher, f=a ₂ , otherwise f=1.

In the above scheme, the step F further includes:

According to the formula The information influence of the UGC is INF.

In the above scheme, the path coefficient a ₁ = 0.333 for information quality and group identity, a ₂ = 0.824 for information quality and value perception, a 3 = 0.624 for value perception and group identity, and a ₃ = 0.624 for participation and group identity The path coefficient a ₄ =0.437.

In summary, the present invention proposes a method for improving the accuracy of the influence of content information generated by social network users, and uses the Sociability-based Influence Diffusion Probability Model (S-IDPM, Sociability-based Influence Diffusion Probability Model) to calculate the influence of UGC information. Power, mainly using user social network (including fan network and interest network) and reply chain structure to weight the replies of different users, so as to weight the keywords of UGC respectively, and improve the calculation of the influence of content information generated by social network users accuracy.

Description of drawings

Figure 1 is a diagram of the relationship between users and posts.

Figure 2 is a fan network map.

Figure 3 is a diagram of an interest network.

Figure 4 shows the user member participation mechanism.

FIG. 5 is a UGC structure diagram of the first method embodiment.

FIG. 6 is a flow chart of the first method embodiment.

Fig. 7 is a comparison chart of cumulative content ratios of manually marked high-quality posts in the second method embodiment.

Fig. 8 is a class-eigenvalue comparison chart of the second method embodiment.

Fig. 9 is a comparison chart of accumulative content ratios of machine-labeled high-quality posts in the second method embodiment.

detailed description

In order to make the object, technical solution and advantages of the present invention more clearly, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

The technical scheme of an embodiment of the present invention is:

A. Establish a social network UGC member participation mechanism, determine the path coefficient between the various influencing factors of the member participation mechanism, and the path coefficient is the degree of correlation between the various influencing factors of the member participation mechanism;

B. Construct a fan network unauthorized directed graph based on the user-fan relationship of the UGC, and perform community division on the fan network unauthorized directed graph; construct an interest network authorized undirected graph based on the user reply relationship of the UGC , performing community division on the right undirected graph of the interest network;

C. Calculate the social influence U _X of user X according to the degree of correlation between the various influencing factors of the member participation mechanism;

D. According to Calculate the social influence of user X publishing keyword K, m is the number of times keyword K spreads on user X, if m=0, then S _KX =0;

E. According to the formula Calculating the comprehensive social influence of keyword K in the UGC;

F. Calculate the sum of the comprehensive social influences of the M keywords in the UGC to obtain the information influence INF of the UGC.

The technical solution of the present invention adds user factors to the calculation of the influence of user-generated content information, and divides all users of social networks such as BBS, Weibo, and Renren into information publishers and information replyers. According to the fan relationship of users who participate in UGC, construct the unrighted directed graph of fan network, as shown in Figure 2, user 1 is a fan of user 2, then there is an edge from user 1 to user 2 between user 1 and user 2; according to User responses to UGC construct an interest network rights undirected graph, as shown in Figure 3, user 1 and user 2 jointly participated in the information discussion released by 7 information publishers, then there is a right relationship between user 1 and user 2. An undirected edge with a value of 7.

According to the existing technology, the above-mentioned fan network and interest network are divided into communities, and the divided communities are numbered. The community numbers of users belonging to the same community are the same, and the same community number (the same fan network community number or the same interest network Community ID) users have similar values to a certain extent. The user community division technology is an existing technology, and will not be described in detail here. Table 1 shows the fan network community after the community division in Figure 2, and Table 2 shows an example of the interest network community after the community division. It can be seen from Table 1 , user 1 and user 3 belong to the same fan network community, and it can be seen from Table 2 that user 1 and user 2 belong to the same interest network community.

Table 1

user Fan network community number f 1 1 2 2 3 1 4 3 5 4 6 5

Table 2

user Interest network community number r 1 3 2 3 3 1 4 2 5 4 6 5

The technical solution of the present invention builds a user member participation mechanism, as shown in FIG. 4 .

Information quality refers to the information influence of UGC in the social network, which represents the stability of the UGC, the correctness of information, timeliness, novelty and the quality of service.

Value perception is represented by fan relationships among users in social networks. For an information publisher U1, if U1’s fan user U2 replies to the information publisher, it is considered that the factors that drive the fan user U2 to reply to U1 are not only the recognition of U1, but also the factors of maintaining interpersonal relationships. In this case, it is not only the influence of the information released by U1 that drives the participation of fan user U2, but also the factors of interpersonal relationships between users. In the calculation of the influence of UGC information on social networks, the content of replies to fans should be reduced. small corresponding weights.

Group identity indicates the influence of community division and reply evaluation among users in social networks on the influence of UGC information. The community division is the community division of the aforementioned fan network and interest network. If the respondent and the respondent belong to the same community (the same fan network community and/or the same interest network community), the corresponding weight will be reduced; otherwise, the corresponding weight will be increased. Large corresponding weight; if a UGC content is of high quality or has the potential to generate a certain influence, most users are willing to participate in such a UGC, so reply evaluation becomes an important factor in the calculation of UGC influence. The more the keywords in the reply evaluation are spread in the UGC, the greater the influence of the UGC.

Figure 4 The path coefficient in the user member participation mechanism measures the degree of correlation between two variables, represented by a ₁ , a ₂ , a ₃ , and a ₄ respectively, and 0<a ₁ <1, 0<a ₂ <1, 0<a ₃ <1, 0<a ₄ <1.

Calculate the social influence U _X of the participating user X of the social network UGC according to the degree of correlation among the various influencing factors of the member participation mechanism:

Among them, b is the number of times that user X is directly replied in a UGC. U _X =0 if user X is not directly replied to.

C ₁ indicates whether the respondent of user X belongs to the same interest network community as user X, if yes, since information quality and group identity have a correlation degree of a ₁ , C ₁ =a ₁ ; otherwise, C ₁ =1. C ₁ =1 indicates that if users of different interest network communities can be attracted to reply, the influence will be stronger.

C ₂ indicates whether the respondent of user X belongs to the same fan network community as user X, if yes, because information quality and value perception have a ₂ correlation degree, and value perception and group identity have a ₃ correlation degree, so C ₂ =a ₂ ×a ₃ ; otherwise, C ₂ =1. C ₂ =1 indicates that if users from different fan network communities can be attracted to reply, the influence will be stronger.

f indicates whether user X is a fan of the information publisher of the UGC, if yes, f=a ₂ due to the correlation degree of a ₂ between information quality and value perception; otherwise f=1. This shows that if user X is a fan of the information publisher, user X's reply is not only an affirmation of the content, but also has reasons for maintaining social relations, so the weight should be reduced.

After the social influence U _X of user X in a UGC in the social network is determined, the social influence of the keyword K published by user X can be further determined:

Among them, m is the number of propagation times of keyword K on user X, that is, the number of direct replies of user X who also reply to keyword K, if m=0, then S _KX =0.

The comprehensive social influence of keyword K in the entire UGC is the sum of the social influence of all users of the UGC (including information publishers and respondents) publishing keyword K, namely:

Among them, N is the number of users participating in the UGC (that is, the sum of the number of information publishers and respondents).

UGC information influence is the sum of the comprehensive social influence of all keywords in the UGC, namely:

Among them, M is the number of all keywords contained in the UGC.

The technical solution of the present invention will be further described below with examples.

Method embodiment one

FIG. 5 is a structural diagram of a UGC in this embodiment. As shown in FIG. 5 , the UGC includes four users: User 1, User 2, User 3, and User 4. User 1 is the information publisher, and the keywords released are A, B, and C; users 2 and 3 respectively reply to user 1 directly, and the keywords posted by user 2 are A, C, D, and the keywords posted by user 3 are B, F; User 4 directly replied to User 2 with the keywords C and F. The interest network community number is represented by r, r ₁ =1, r ₂ =1, r ₃ =2, r ₄ =3; the fan network community number is represented by f, f ₁ =1, f ₂ =2, f ₃ =1, f ₄ =3. User 2 and User 4 are fans of the information publisher User 1. In this embodiment, the path coefficients among various factors in the member participation mechanism are assigned as follows: a ₁ =0.333, a ₂ =0.824, a ₃ =0.624, a ₄ =0.437. Fig. 6 is the flowchart of this embodiment, as shown in Fig. 6, comprises the following steps:

Step 601: Calculate the social influence of each user in the UGC.

According to the formula Calculate the social influence of each user separately. Taking the calculation of the social influence of user 1 as an example for illustration, the calculation methods of users 2 to 4 are the same and will not be repeated here.

The number of direct responders of user 1 is 2, that is, b=2; user 2 and user 1 are in different interest network communities and the same fan network community, so C ₁ =a ₁ =0.333, C ₂ =1; user 3 In the same interest network community as user 1, but in different fan network communities, therefore C ₁ =1, C ₂ =a ₂ ×a ₃ =0.514; user 1 is not a fan of himself, f=1, therefore

U ₁ ={1+ln[1+a ₁ ×1+1×a ₂ ×a ₃ ]}×1=1.614

Similarly, U ₂ =1.395, U ₃ =0, U ₄ =0.

Step 602: Calculate the comprehensive social influence of each keyword in the UGC.

Taking the calculation of the comprehensive social influence of keyword C as an example for illustration, the calculation methods of the comprehensive social influence of keywords A, B, D, and F are the same and will not be repeated here.

The users who released keyword C include user 1, user 2, and user 4. For user 1, keyword C was only propagated once (user 2, the direct responder of user 1, published keyword C), then S _C1 = U ₁ ; for user 2, keyword C has only been propagated once (user 4, the direct responder of user 2, released keyword C), then S _C2 = U ₂ ; for user 3 and user 4, keyword C There is no dissemination (there is no direct responder to publish keyword C in user 3 and user 4), S _C3 = 0, S _C4 = 0, therefore, the social influence of keyword C is:

In the same way,

Step 603: Calculate the information influence of UGC.

According to the formula Calculate the information influence of the UGC.

Method embodiment two

Method Embodiment 1 Taking UGC with less user participation as an example, how to calculate the influence of social network UGC information by the technical solution of the present invention is explained. The program is further explained.

User information includes a total of 181,841 user IDs, IDs of their fans, IDs of posts published in this forum, and IDs of replies to posts in this forum; post information includes a total of 43,609 post IDs, the serial number of each floor in the post, and the publisher ID and its content. By judging whether the post contains the boutique symbol of the forum administrator, 827 posts are selected from the post information as the manual tagged boutique post set, and the other posts are regarded as the non-fine post collection. Due to the huge amount of data, 9173 posts were randomly selected from the collection of non-excellent posts and 827 artificially marked high-quality posts were mixed into a sample of 10,000 posts, and the samples were used to compare, analyze and analyze S-IDPM, IDM, and IDPM respectively. Evaluate. In addition, the clustering method of statistical information is used to machine-label the posts, and the machine-labeled high-quality post collection is obtained, and S-IDPM, IDM, and IDPM are also compared and analyzed.

Table 3 shows the comparative experimental results of the above three methods. Due to the large number of posts, only the top 5 posts are given here.

table 3

It can be observed from Table 3 that the main difference between IDM and S-IDPM is in posts 2894103 and 2366245. By observing the corresponding corpus, it is found that post 2894103 is an advertisement solicitation post. He released the corresponding advertisement template, and all users must reply in a fixed format. Therefore, according to the IDM model, co-occurrence words are mainly used to calculate influence. Therefore, This post has high influence under IDM. However, it can be seen from the title that this post attracts a group of users who like cars, and from the calculated user interest network, it can also be seen that many users have responded to certain posts together, so they are in the same interest network , have the same interest network ID, which means that the post was only circulated in a small circle. Hence, it is not ranked very high in S-IDPM, not in the top 5. However, post 2366245 attracted widespread attention and replies, with a total of 1476757 replies. The users in the post did not have obvious large-scale fans and interest networks, and the user circle was relatively scattered, indicating that this post attracted widespread attention from various user groups in the community. Therefore, it ranks higher in post influence in S-IDPM.

And the main difference between IDPM and S-IDPM is in post 2510082 and post 2713599. By observing the corresponding corpus, it is found that the user name of the poster of post 2713599 is "I am a Japanese car owner", which is very similar to the title of the post. By observing her user page, it is also found that the user does not have any fans, attention, or reply Any post and only this one post, which fully demonstrates that the username is a vest name and has no social or interest network relationship with anyone. This post caused a total of 14,944 replies, and although post 2510082 attracted 28,211 replies, which is more than the number of replies to post 2713599, but the fans and interest networks of post 2713599 reply users are more scattered. Therefore, in S-IDPM influence ranks higher.

It can be seen from the above qualitative analysis that S-IDPM can solve some problems not considered by IDM and IDPM to a certain extent.

Next, quantitatively analyze the effect of the three methods in post influence calculation.

First of all, this embodiment compares the quality posts manually marked (marked by forum administrators) as the standard of quality posts, and compares the cumulative content ratio of IDM, IDPM and S-IDPM quality posts, as shown in Figure 7. It can be seen from Figure 7 that the accumulative content rate of S-IDPM high-quality posts reaches about 70% the fastest among the top 3,000. It shows that in the influence ranking of S-IDPM, the top 30% covers 70% of high-quality posts, and the top 10% and 20% are higher than the IDM and IDPM models. Therefore, it shows that the calculation result of S-IDPM's influence on posts is better, which is more in line with the manual labeling results.

Next, the cluster-based opinion leader discovery algorithm in the prior art is used in the post influence analysis of the present invention, and an algorithm for discovering high-quality posts by using the statistical information clustering method is realized.

Choose the floor number F of the post, the duration T, the number of replies P, the number of replies per hour is expressed as F/T, the average number of words per floor is W/F, and the difference between the number of replies from non-hosts and the number of replies from hosts is D As a feature value, N represents the number of members of the class. Eight classes are obtained by using the method for selecting the number of subclasses and the clustering algorithm (both existing technologies), as shown in Figure 8.

The filter conditions in the clustering-based opinion leader discovery algorithm are adjusted as follows: the members in the class with fewer members and larger average feature value of class members are regarded as high-quality posts marked by machines. Therefore, the members in categories 5 and 7 are the high-quality posts for the next experiment. There are 1,001 members in categories 5 and 7, and only 291 of them are the same as the 827 high-quality posts marked by the forum administrator. Therefore, , which is different from the experiment shown in Fig. 7. Next, use these 1001 high-quality posts to compare the cumulative content ratio of IDM, IDPM and S-IDPM high-quality posts, as shown in Figure 9:

It can be seen from Figure 9 that the cumulative rate curve of S-IDPM high-quality posts has always been at the middle level, and the cumulative rate of high-quality posts has reached more than 85% in the first 2000 articles, which shows that the use of statistical information machines to mark high-quality posts Under this condition, the calculation effect of S-IDPM on the influence of posts is still very good.

Finally, compare the accuracy of the high-quality posts of the three algorithms in the case of manual labeling and machine labeling, as shown in Table 4, S-IDPM has higher accuracy than other high-quality posts in the two cases of manual labeling and machine labeling. Two models.

Table 4

Pt ₀ Pt ₁ IDM 28.1% 68.1% IDPM 30.2% 67.3% S-IDPM 32.4% 68.4%

Through the comparison experiment of the accumulative rate of high-quality posts and the comparison experiment of the accuracy of high-quality posts using manual and machine marking, it can be seen from the above experimental results that S-IDPM is more accurate in calculating the influence of posts, which is better than IDM and IDPM methods.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (5)

1. it is a kind of to improve the method that social network user produces content information power of influence accuracy, it is applied to social network user product Raw content UGC, the UGC include M key word, have N number of user and participate in the UGC, it is characterised in that the method includes：

A. social networkies UGC member's participation mechanism is set up, the path coefficient between each influence factor of member's participation mechanism is determined, The path coefficient is the degree of correlation between each influence factor of member's participation mechanism；

B. vermicelli network is built according to user's vermicelli relation of the UGC and haves no right directed graph, the vermicelli network is had no right oriented Figure carries out community division；Relation is replied according to the user of the UGC build correlation network and have the right non-directed graph, to the correlation network Non-directed graph of having the right carries out community division；

C. social influence power U of user X is calculated according to the degree of correlation between each influence factor of member's participation mechanism_X；

D. basisThe social influence power that user X issues key word K is calculated, m is biographies of the key word K on user X Number of times is broadcast, if m=0, S_KX=0；

E. according to formulaCalculate comprehensive social influence power of the key word K in the UGC；

F. comprehensive social influence power sum of the M key word in the UGC is calculated, the informational influence of the UGC is obtained Power INF.

2. method according to claim 1, it is characterised in that member's participation mechanism includes that information quality, colony are recognized It is a that the same feeling, value perceive and participate in the path coefficient of four influence factors, information quality and group identification sense₁, information quality and The path coefficient that value is perceived is a₂, value is perceived and the path coefficient of group identification sense is a₃, participate in the road with group identification sense Footpath coefficient is a₄。

3. method according to claim 2, it is characterised in that step C is further included：

According to formulaThe social influence power of user X in the UGC is calculated,

Wherein, b is the number of times directly replied in the UGC by user X, if user does not have direct reply person, U_X=0；

If directly reply person belongs to same interest network social association, C to user X₁=a₁, otherwise, C₁=1；

If directly reply person belongs to identical vermicelli network social association, C to user X₂=a₂×a₃, otherwise, C₂=1；

If user X is the vermicelli of the UGC information publishers, f=a₂, otherwise f=1.

4. method according to claim 1, it is characterised in that step F is further included：

According to formulaCalculate informational influence power INF of the UGC.

5. method according to claim 2, it is characterised in that the path coefficient a of information quality and group identification sense₁= 0.333, the path coefficient a that information quality and value are perceived₂=0.824, value perceives the path coefficient a with group identification sense₃= 0.624, participate in the path coefficient a with group identification sense₄=0.437.