A method for measuring social user influence based on multi-angle analysis

Abstract

The present invention belongs to the technical field of measuring social influence, and particularly relates to a social user influence measurement method based on multi-angle analysis. First, in order to fully consider the contribution of different relationships to user influence, the user's relationship network is constructed from multiple angles of attention relationship, forwarding relationship, comment relationship and mention relationship; secondly, in different relationship networks, user activity, interaction intensity and interest similarity are used to quantify user influence; finally, the user's influence from different angles is integrated to obtain the user's comprehensive influence in the global scope. The present invention has better performance than existing methods in terms of accuracy, relevance, coverage, etc., with higher accuracy, more consistent with the actual true value, and wider coverage of user influence.

Description

Technical Field

[0001] The present invention belongs to the technical field of measuring social influence, and in particular relates to a social user influence measurement method based on multi-angle analysis. Background Art

[0002] With the development of the Internet, social networks have become an important tool for maintaining human social relations and disseminating information. They are no longer limited to information exchange, but have developed into a comprehensive information platform that integrates communication, information, entertainment, search, e-commerce, office collaboration and corporate customer service. With the rapid expansion of the scale of social network users, social content data has also increased dramatically. These massive social data have great value. Based on rich social data, the study of influence in social networks has attracted widespread attention from scholars. Social influence is reflected in the phenomenon of being able to change the emotions, opinions or behaviors of other users. Analyzing the formation and dissemination process of user influence in the network and then identifying influential communicators is of great significance to understanding the behavior of users in the network, the diffusion of information in the network and the evolution of social networks. At the same time, it provides technical support for better public opinion control and has broad application prospects.

[0003] The current methods for measuring social influence characterize user influence from different dimensions. In some aspects, they can effectively discover influential users in social networks, but there are still some shortcomings. First, the relationship between users in social networks often has multiple links with different semantics. For example, in Weibo, there are not only direct link relationships such as following, but also multiple interactive relationships such as forwarding, commenting, mentioning and liking. If the user influence is analyzed only from the dimension of one link relationship, the influence of other relationships in the information dissemination of social networks will be ignored. Secondly, the follow-up relationship between users is a relatively static social relationship. Only the network topology constructed by the follow-up relationship is considered, and the dynamic interactive behavior between users is not used as an evaluation indicator of user influence. It is difficult to reflect the difference between "zombie users" and normal users. Finally, when analyzing the closeness between users, most of the existing methods only consider the explicit feature of the link strength between users, but do not consider the implicit feature of interest preferences between users. Users with similar interest preferences may have stronger influence. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention proposes a social user influence measurement method based on multi-angle analysis, which quantifies user influence by using factors such as user activity, interaction intensity between users and interest preference. It has higher accuracy, is more consistent with the actual true value, and has a wider coverage of the dissemination of user influence.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A social user influence measurement method based on multi-angle analysis of the present invention comprises the following steps:

[0007] Build user relationship networks from multiple perspectives;

[0008] In different relationship networks, user activity, interaction intensity, and interest similarity are used to quantify user influence;

[0009] By integrating the user's influence from different angles, we can get the user's comprehensive influence on a global scale.

[0010] Furthermore, constructing the user's relationship network from multiple perspectives includes:

[0011] The user relationship network is constructed from multiple perspectives, including the following relationship, forwarding relationship, comment relationship, and mention relationship. The relationship network is defined as G i = {V i ,E i}, where V i is the set of users, E i is the link relationship, i∈{f,r,c,m} corresponds to the follow, forward, comment and mention relationships respectively.

[0012] Furthermore, the user activity indicates the frequency of various activities of the user in the network, and the user activity is defined as formula (1):

[0013] Act i (v)=Post(v)+Interact i (v),i∈{f,r,c,m} (1)

[0014] Among them, Act i (v) represents the activity of user v in different networks, Post(v) represents the number of microblogs posted by user v, Interact i (v) indicates the number of times user v follows / reposts / comments / mentions other users.

[0015] Furthermore, the interaction intensity is used to describe and quantify the interaction behavior between users. The interaction intensity is defined as follows:

[0016]

[0017] Among them, RS i (v,u) represents the interaction intensity between user v and user u in different networks. i (v,u) represents the number of times user u forwards / comments / mentions user v, Interact i(u) indicates the number of times user u forwards / comments / mentions other users, Interacted i (v) represents the number of times other users forwarded / commented / mentioned user v.

[0018] Further, the interest similarity is regarded as an indicator of the possibility of information propagation from one user to another user, and the interest similarity between users is measured by comparing the similarity of attribute keyword vectors between users;

[0019] For any user v, extract the keywords of the user's attributes, expressed as KW(v) = (Keyword1: Weight1; Keyword2: Weight2; ...), where Keywordi represents the keyword in the user v's attributes, and Weighti represents the weight of Keywordi in the user v's attributes;

[0020] For two users v and u, the interest similarity value between the users is obtained by calculating the inner product sum of the same subvectors in the two keyword vectors, that is, by calculating the intersection of the keywords in KW(v) and KW(u), the two subvectors of v and u are obtained respectively, expressed as SK(v)=(ckw1:w v1 ;ckw2:w v2 ; ......), SK(u)=(ckw1:w u1 ;ckw2:w u2 ; ......), where ckw* represents the same keyword as KW(v) and KW(u), w * represents the weight, and the interest similarity between users v and u is defined as formula (3):

[0021]

[0022] Among them, IS(v,u) represents the interest similarity between user v and user u, w vk and w uk They represent the weights corresponding to user v and user u respectively, and s represents the number of the same keywords.

[0023] Furthermore, in the forwarding, commenting and mentioning subnetworks, the interaction influence between users is quantified by the interaction intensity and interest similarity, and the interaction influence between users is defined as follows:

[0024]

[0025] Among them, Inf i (v,u) represents the interaction between user v and user u in different networks, RS i(v,u) represents the interaction intensity between user v and user u in different networks, and IS(v,u) represents the interest similarity between user v and user u.

[0026] Furthermore, the influence of the user in the forwarding subnet is calculated:

[0027] The transition probability of random walk from user v to user u in the forwarding subnet is defined as formula (5):

[0028]

[0029] Where Ree(u) represents the set of users who forward user u, Inf r (v,u) represents the interaction between user v and user u in the forwarding subnet;

[0030] The ratio of a single user to the sum of the activities of all users in the forwarding subnet is defined as the jump probability, as shown in formula (6):

[0031]

[0032] Among them, Act r (v) represents the activity of user v in the forwarding subnet, and the value of θ is a non-zero decimal;

[0033] The influence of user v in the forwarding subnet is calculated as follows:

[0034]

[0035] Among them, α is set to 0.85, and Ret(v) represents the set of users forwarded by user v.

[0036] Furthermore, the influence of the user in the comment subnetwork is calculated:

[0037] The transition probability of random walk from user v to user u in the comment subnetwork is defined as formula (8):

[0038]

[0039] Among them, Coe(u) represents the set of users who comment on user u, Inf c (v,u) represents the interaction influence between user v and user u in the comment subnetwork;

[0040] The ratio of a single user to the sum of the activities of all users in the comment subnet is defined as the jump probability, as shown in formula (9):

[0041]

[0042] Among them, Act c(v) represents the activity of user v in the comment subnetwork, and the value of θ is a non-zero decimal;

[0043] The influence of user v in the comment subnetwork is calculated as follows:

[0044]

[0045] Among them, α is 0.85, Com(v) represents the set of users commented by user v;

[0046] Similarly, calculate the influence of the user in the mention subnet:

[0047] In the mentioned subnet, the transition probability is as follows:

[0048]

[0049] Among them, Mee(u) represents the set of users who mention user u, Inf m (v,u) represents the interaction influence between user v and user u in the mention subnet;

[0050] The jump probability is as shown in formula (12):

[0051]

[0052] Among them, Act m (v) represents the activity of user v in the mention subnet, and the value of θ is a non-zero decimal;

[0053] The influence of user v in the mention subnetwork is calculated as follows:

[0054]

[0055] Among them, α is set to 0.85, and Men(v) represents the set of users mentioned by user v.

[0056] Furthermore, the influence of the user in the attention subnet is calculated:

[0057] The transition probability from user v to user u in the attention subnet is defined as formula (14):

[0058]

[0059] Among them, Foe(u) represents the set of users followed by user u, and Post(v) represents the number of microblogs posted by user v;

[0060] The jump probability is as shown in formula (15):

[0061]

[0062] Among them, Actf (v) represents the activity of user v in the attention subnet, and the value of θ is a non-zero decimal;

[0063] The influence of user v in the attention subnet is calculated as follows:

[0064]

[0065] Among them, α is set to 0.85, and Fow(v) represents the set of fans of user v.

[0066] Furthermore, through linear fusion, the comprehensive influence of the user in the global scope is calculated as shown in formula (17):

[0067] ASInf(v)=β1ASInf f (v)+β2ASInf r (v)+β3ASInf c (v)+β4ASInf m (v) (17)

[0068] Among them, β1, β2, β3, and β4 are the influence weights of the user in the attention network, forwarding network, comment network, and mention network, respectively, and β1+β2+β3+β4=1.

[0069] Compared with the prior art, the present invention has the following advantages:

[0070] The present invention improves on the PageRank algorithm, and proposes a user influence measurement method combining user activity, user interaction intensity and interest preference under different relationship network dimensions-ASIRank (Actively Strength Interest based Rank) to measure the user's influence in social networks; first, in order to fully consider the contribution of different relationships to user influence, the user's relationship network is constructed from multiple perspectives such as attention relationship, forwarding relationship, comment relationship and mention relationship; secondly, in order to eliminate the influence of "zombie users", in different relationship networks, the quality of user nodes is evaluated by analyzing the user's activity, as the PageRank jump probability; thirdly, based on factors such as user interaction intensity and interest similarity, the closeness between users is characterized as the basis for PageRank weight allocation to reflect the influence between users; finally, the user influence ranking is obtained by weighted summing the user influence in different relationship networks. The present invention has better performance than existing methods in terms of accuracy, relevance, coverage, etc., with higher accuracy, more consistent with the actual true value, and wider coverage of user influence. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0072] Figure 1 is a flow chart of a method for measuring the influence of social users based on multi-angle analysis according to an embodiment of the present invention;

[0073] Figure 2 It is a heterogeneous information microblog network link diagram of an embodiment of the present invention;

[0074] Figure 3 This is the accuracy comparison result of the embodiment of the present invention for N=2 using the cross-validation method;

[0075] Figure 4 This is the accuracy comparison result of the embodiment of the present invention for N=3 using the cross-validation method;

[0076] Figure 5 This is the accuracy comparison result of the embodiment of the present invention for N=4 using the cross-validation method;

[0077] Figure 6 is the accuracy comparison result of the Sina Weibo official influence evaluation method in the embodiment of the present invention;

[0078] Figure 7 This is a comparison result of the coverage of the five algorithms in the embodiment of the present invention. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0080] First, let's briefly introduce the PageRank algorithm in social networks. The details are as follows:

[0081] PageRank is an algorithm used to analyze the importance of web pages. Its basic idea is to calculate the importance of web pages by analyzing the topological characteristics of the network. Drawing on the method of judging academic papers, the importance of a node is not only determined by the number of links, but also by the quality of the link node itself, similar to "analyzing a person by looking at his friends". The PageRank algorithm calculation formula is defined as:

[0082]

[0083] Where PR(v) is the PageRank value of web page v, N is the total number of web pages in the network, and B v is the set of web pages linking to page v, L(u) is the number of web pages linking to page u, and d is the damping factor, which represents the probability of a web page being randomly accessed, usually set to 0.85.

[0084] In social networks, the relationships between users are also point-to-point directed edges, similar to the links between web pages. The PageRank algorithm can be used to analyze the importance of nodes in social networks. By linking PageRank with the social influence model, it is believed that the authority of the node is equivalent to the influence of the user. Therefore, the PageRank algorithm helps to identify influential nodes or opinion leaders in the network.

[0085] In a social network, if the opinions expressed by a certain user are always recognized by other users, or if a certain user can always give his or her own unique evaluation of current hot events and arouse direct discussion and attention from other users, then such users have a certain degree of influence. In Weibo, the influence between users can not only be measured by the possibility of a user reading other users’ Weibo, but also the interactive behaviors such as forwarding, commenting and mentioning between users are also important ways to spread influence, because they more accurately represent the user’s response. The definition of user influence in the Weibo network is based on the following assumptions: if a user is followed, forwarded, commented or mentioned by more users, then the user is considered to have a higher influence. Inspired by the PageRank algorithm, if a user is followed, forwarded, commented or mentioned by another user with high influence, the user’s influence will also increase accordingly. To this end, the present invention comprehensively evaluates the influence of users by analyzing the generation and transmission process of user influence under various link relationship networks such as attention, forwarding, commenting and mentioning; based on this, if Figure 1 As shown, this embodiment proposes a social user influence measurement method based on multi-angle analysis (abbreviated as ASIRank), which includes the following steps:

[0086] Step S11, in order to fully consider the contribution of different relationships to the user's influence, the user's relationship network is constructed from multiple perspectives such as follow relationship, forwarding relationship, comment relationship and mention relationship;

[0087] Step S12, in different relationship networks, user activity, interaction intensity and interest similarity are used to quantify user influence;

[0088] Step S13, synthesizing the influence of the user from different perspectives to obtain the user's comprehensive influence in the global scope.

[0089] The step S11 of constructing the user's relationship network from multiple perspectives specifically includes:

[0090] The connections between social users are often formed through friendships. By using friendship networks to establish information dissemination and interaction mechanisms, users can easily obtain and publish information and influence other users. Traditional evaluations of user influence are achieved by analyzing the topological characteristics of friendship networks. However, most social networks are directed heterogeneous networks. For example, on Weibo, the links between users are not just follow-up relationships, but also include multiple interactive relationships such as forwarding, commenting, and mentioning. Users in the network can comment on other users at will and forward content they are interested in without following each other. The channels for information dissemination are complex and diverse, creating a heterogeneous information network link graph composed of different behaviors between users, such as Figure 2 shown.

[0091] The original heterogeneous social network is split according to the link relationship, and the homogeneous subnetworks based on different link relationships are split. This not only does not lose the original network structure, but can also mine more user relationships. The subnetwork is defined as G i = {V i ,E i}, where V i is the set of users, E i is the link relationship, i∈{f,r,c,m} corresponds to the follow, forward, comment and mention relationships respectively. For a specific edge (v,u), it means that user u follows / forwards / comments / mentions user v. Each user is attached with personal information, such as keywords, the number of microblogs posted by the user, etc. At the same time, the number of interactions between each pair of users is recorded.

[0092] The three influencing factors of user activity, interaction intensity and interest similarity in step S12 are analyzed in detail as follows.

[0093] a. User activity

[0094] A social network is a network structure that reflects social activities, in which the activity level of each user is different, which determines the difference in the quality of its nodes. That is, the more frequently a user publishes, reads, forwards, likes and comments on Weibo, the more active the user's status is, indicating that the user participates in more network activities and plays a greater role.

[0095] Activity directly indicates the frequency of various activities of users in the network. It is a behavioral attribute of users and an important indicator for measuring user influence. In Weibo, it is very important to improve user activity. Users must continue to participate in hot topics and create topics in order to attract fans' attention and increase their popularity. Since there are a large number of "zombie users" in real social networks, they follow a large number of users but rarely publish original content or interact with each other. Therefore, by analyzing the frequency of each user's activities, we can find active nodes in the network, effectively reduce the interference of "zombie users" on influence analysis, and improve the accuracy of calculation results. Define user activity as formula (1):

[0096] Act i (v)=Post(v)+Interact i (v),i∈{f,r,c,m} (1)

[0097] Among them, Act i (v) represents the activity of user v in different networks, Post(v) represents the number of microblogs posted by user v, Interact i (v) indicates the number of times user v follows / forwards / comments / mentions other users. The manifestation of user activity in different relationship networks is also different. For example, in the following subnet, interactive behaviors such as forwarding and commenting are not considered, so only the number of microblogs posted is counted, while in the forwarding, commenting and mentioning subnets, the number of microblogs posted and the number of interactive behaviors must be considered at the same time.

[0098] b. Interaction strength

[0099] Interactions such as forwarding, commenting, and mentioning between social network users can be represented as directed weighted edges. The interaction between users can be described and quantified by defining the interaction intensity. The more interactions between users, the deeper the influence between users. For example, in a forwarding network, the forwarding intensity consists of two parts. The first part is the number of times user u forwards user v divided by the total number of times user u forwards other users. The second part can be expressed as the number of times user u forwards user v divided by the total number of times other users forward user v. Both parts conform to the fact that the larger the value, the greater the influence of user v on user u. Similarly, the two interactive relationships of commenting and mentioning also have this phenomenon, so in the forwarding, commenting, and mentioning relationship networks, the interaction intensity is defined as formula (2):

[0100]

[0101] Among them, RS i (v,u) represents the interaction intensity between user v and user u in different networks. i (v,u) represents the number of times user u forwards / comments / mentions user v, Interact i (u) indicates the number of times user u forwards / comments / mentions other users, Interacted i (v) represents the number of times other users forwarded / commented / mentioned user v.

[0102] c. Interest similarity

[0103] In social networks, the interest similarity between users can be seen as an indicator of the possibility of information spreading from one user to another, that is, potential interaction. Generally, users with greater interest similarity are more likely to be interested in the same object, which makes it easier for users to influence each other. At the same time, behaviors such as forwarding, commenting, and mentioning between users are also largely based on the same interests and hobbies. Therefore, considering interest similarity in the forwarding, commenting, and mentioning relationship network can better characterize the influence of a user.

[0104] The interest similarity between users can be measured by comparing the similarity of the attribute keyword vectors between users; for any user v, the keywords of the user's attributes can be extracted, expressed as KW(v)=(Keyword1:Weight1;Keyword2:Weight2;...), where Keywordi represents the keyword in the user v's attributes, and Weighti represents the weight of Keywordi in the user v's attributes. For two users v and u, the interest similarity value between users can be obtained by calculating the inner product sum of the same subvectors in the two keyword vectors, that is, by calculating the intersection of the keywords in KW(v) and KW(u), two subvectors of v and u are obtained respectively, expressed as SK(v)=(ckw1:w v1 ;ckw2:w v2 ; ......), SK(u)=(ckw1:w u1 ;ckw2:w u2 ; ......), where ckw* represents the same keyword as KW(v) and KW(u), w * represents the weight. Then the interest similarity between users v and u is defined as formula (3):

[0105]

[0106] Among them, IS(v,u) represents the interest similarity between user v and user u, w vk and w uk They represent the weights corresponding to user v and user u respectively, and s represents the number of the same keywords.

[0107] In summary, the interaction influence between users represents the degree of influence of user v on user u. The larger the value, the greater the influence of user v on user u, and the more likely the information will spread from user v to user u. The two different interaction metrics mentioned above can be used to measure the degree of influence between users. Therefore, for simplicity, in the forwarding, commenting, and mentioning subnetworks, the interaction influence between users is quantified by the interaction intensity and interest similarity, and the interaction influence between users is defined as formula (4):

[0108]

[0109] Among them, Inf i (v,u) represents the interaction between user v and user u in different networks, RS i (v,u) represents the interaction intensity between user v and user u in different networks, and IS(v,u) represents the interest similarity between user v and user u. Here, the above two indicators are considered to play the same role, and before calculating the interaction impact, each indicator is normalized, the maximum value is set to 1.0, and other values ​​are changed accordingly in proportion.

[0110] In step S13, the user's influence from different perspectives includes the user's influence in the forwarding subnet, commenting subnet, mentioning subnet, and following subnet.

[0111] a. Calculate the user’s influence in the forwarding subnet

[0112] The influence of a user in the network is evaluated by using user activity and the interaction between users. The present invention adopts the idea of ​​PageRank algorithm and further considers user activity and interaction characteristics between users to analyze user influence through random walk. There are two key probabilities that need to be processed in the PageRank algorithm. One is the transfer probability between users and their linked users; the other is the jump probability of users reading non-linked information, which is an important factor for random walk restart. These two key points are also where the present invention applies user activity and interaction between users to the PageRank algorithm.

[0113] The transition probability of random walk from user v to user u in the forwarding subnet is defined as formula (5):

[0114]

[0115] Where Ree(u) represents the set of users who forward user u, Inf r (v,u) represents the interaction between user v and user u in the forwarding subnet.

[0116] For the jump probability, the traditional PageRank algorithm sets it to 1 / n, which means that the probability of each user being visited by other users is the same. However, in social networks, it can be considered that the more active a user is in the network, the more likely other users are to be attracted to this user and then visit this user. Therefore, the present invention calculates the jump probability by user activity, which is defined as formula (6):

[0117]

[0118] That is, the ratio of the activity of a single user to the sum of the activity of all users in the forwarding subnet is used as the jump probability. r (v) represents the activity of user v in the forwarding subnet. The value of θ is a non-zero decimal, such as 0.01 or 0.1. Here, the value of θ is to avoid To ensure that every user in the network can be accessed.

[0119] In this way, the influence of user v in the forwarding subnet can be calculated as shown in formula (7):

[0120]

[0121] Among them, α is set to 0.85, and Ret(v) represents the set of users forwarded by user v.

[0122] b. Calculate the user’s influence in the comment subnetwork

[0123] In the comment subnet, just like calculating the user influence in the forwarding subnet, the jump probability can be calculated through the user activity, and the transfer probability can be obtained by using the interaction influence between users. Finally, the influence of each user in the network can be obtained. The calculation process is as follows:

[0124] The transition probability of random walk from user v to user u in the comment subnetwork is defined as formula (8):

[0125]

[0126] Among them, Coe(u) represents the set of users who comment on user u, Inf c (v,u) represents the interaction influence between user v and user u in the comment subnetwork.

[0127] The ratio of a single user to the sum of the activities of all users in the comment subnet is defined as the jump probability, as shown in formula (9):

[0128]

[0129] Among them, Act c (v) represents the activity of user v in the comment subnetwork, and the value of θ is a non-zero decimal.

[0130] The influence of user v in the comment subnetwork is calculated as follows:

[0131]

[0132] Among them, α is set to 0.85, and Com(v) represents the set of users commented by user v.

[0133] c. Calculate the user’s influence in the mention subnet

[0134] Similarly, in the mentioned subnet, the transition probability is as follows:

[0135]

[0136] Among them, Mee(u) represents the set of users who mention user u, Inf m (v,u) represents the interaction influence between user v and user u in the mention subnet.

[0137] The jump probability is as shown in formula (12):

[0138]

[0139] Among them, Act m (v) represents the activity of user v in the mention subnet, and the value of θ is a non-zero decimal.

[0140] The influence of user v in the mention subnetwork is calculated as follows:

[0141]

[0142] Among them, α is set to 0.85, and Men(v) represents the set of users mentioned by user v.

[0143] d. Calculate the user’s influence in the following subnetwork

[0144] In the attention subnetwork, since the attention relationship between users is an unweighted directed edge, the interaction influence between users cannot be quantified by the interaction intensity. Here, the transfer probability between users is defined by analyzing the behavior of pushing microblogs between users. Assume that user u follows many users, and these users have posted different numbers of microblogs, all of which will be pushed directly to user u. In this way, if the number of microblogs posted by user v in the followed user set is more, the more information received by user u comes from user v, which leads to a greater influence of user v on user u, and the corresponding transfer probability from user v to user u is higher.

[0145] The transition probability from user v to user u in the attention subnet is defined as formula (14):

[0146]

[0147] Among them, Foe(u) represents the set of users followed by user u, and Post(v) represents the number of microblogs posted by user v.

[0148] The jump probability is as shown in formula (15):

[0149]

[0150] Among them, Act f (v) represents the activity of user v in the attention subnet, and the value of θ is a non-zero decimal.

[0151] The influence of user v in the attention subnet is calculated as follows:

[0152]

[0153] Among them, α is set to 0.85, and Fow(v) represents the set of fans of user v.

[0154] In step S13, the user's comprehensive influence in the global scope is calculated

[0155] Based on the above considerations of the user's influence in different network dimensions such as following, forwarding, commenting and mentioning, the user's comprehensive influence is calculated by linear fusion, as shown in formula (17):

[0156] ASInf(v)=β1ASInf f (v)+β2ASInf r (v)+β3ASInf c (v)+β4ASInf m (v) (17)

[0157] Among them, β1, β2, β3, and β4 are the influence weights of users in the attention network, forwarding network, comment network, and mention network, respectively, showing the importance weights of the four link relationships, and β1+β2+β3+β4=1.

[0158] The following experiments are conducted on the real Tencent Weibo dataset, and the accuracy, relevance, and coverage are compared with algorithms such as MPPRank, TunkRank, PageRank, and DegreeRank.

[0159] A. Experimental Setup

[0160] Tencent Weibo is a typical social network, which contains the personal attribute information of users, the content information of users' Weibo, and the interactive relationship information such as attention, forwarding, and commenting between users. The present invention uses the Tencent Weibo dataset released by KDD Cup 2012, which contains all the data required by the present invention. In order to ensure accuracy, the dataset is first preprocessed, and a sub-dataset containing 100,000 nodes and 1,136,750 edges is extracted, and relevant experimental parameters are set, where β1=0.15, β2=0.35, β3=0.3, β4=0.2, and α=0.85. This experiment was carried out in the python3.5 environment and tested on a server. The configuration of the server is Intel Core E5-2620 CPU, 64GB RAM, and Windows10professional x64.

[0161] B. Comparison method

[0162] In order to verify the effectiveness of the ASIRank method proposed in this invention, it is compared with the following methods for measuring user influence in microblog datasets.

[0163] PageRank: Use the original PageRank algorithm to measure the influence of users in the attention network. Calculation formula Where P(v,u) = 1 / Followees(u), the α value is 0.85, and this method only considers the topological structure under the attention network.

[0164] TunkRank: TunkRank algorithm is a classic method for ranking the influence of Weibo users. The calculation formula is: Where P(v , u) = 1 / Followees(u), p is the restart factor, and the ideal value is 0.287.

[0165] MPPRank: This method considers both the user’s attention relationship and the content-based interaction relationship to analyze the user’s influence.

[0166] DegreeRank: Since the relationship network of Weibo can be represented as a directed graph, this method uses the user's in-degree (number of followers) to rank the influence.

[0167] C. Accuracy Verification

[0168] Since there are many criteria for measuring the influence of Weibo users, it is difficult to give a unified standard. This paper uses the cross-validation method and the Sina Weibo official influence evaluation method to obtain a reference ranking, while examining indicators such as accuracy and relevance. The cross-validation method is to take the correct results considered by multiple algorithms as the final correct results. For example, given 5 algorithms, the ranking sets obtained by each algorithm are f A ,f B ,f C ,f D ,f E When N=2 is set, the correct results considered by the two algorithms are taken as the final reference results, which is called the reference ranking RefRank2, and is defined as formula (18):

[0169] RefRank2=(f A ∩f B )∪(f A ∩f C )∪(f A ∩f D )∪(f A ∩f E )∪(f B ∩f C )∪(f B ∩f D )

[0170] ∪(f B ∩f E )∪(f C ∩f D )∪(f C ∩f E )∪(f D ∩f E ) (18)

[0171] The accuracy calculation of algorithm A is as shown in formula (19), and the accuracy calculation of other algorithms is similar.

[0172] accuracy A =|f A ∩RefRank2| / f A (19)

[0173] The official influence calculation formula of Sina Weibo is: Influence = a×Activity+b×Interaction+c×Coverage, with a full score of 100 points. It is composed of three parts: Activity, Interaction, and Coverage, accounting for a=25%, b=55%, and c=20% respectively. Among them, the calculation indicator of activity is the number of Weibo posts; the calculation indicator of interaction is the number of times a user is forwarded, commented on, and mentioned; the calculation indicator of coverage is the number of fans of the user. According to the above evaluation criteria, the score of each user is calculated and sorted, so as to obtain a reference ranking RefRank. w Similarly, formula (19) is used to evaluate the accuracy of different influence ranking methods. This paper verifies the accuracy of five algorithms under top10, top20, top30, top50, top80, and top100. For N = 2, 3, and 4, the results of the cross-validation method are as follows: Figure 3 , Figure 4 and Figure 5 As shown in Figure 2, the results of using the Sina Weibo official influence evaluation method are as follows: Figure 6 shown.

[0174] The results show that the ASIRank method has the highest accuracy in the four reference sets, followed by MPPRank and TunkRank. PageRank and DegreeRank have the lowest accuracy. For the cross-validation method, the trend of decreasing accuracy of all methods increases with the increase of N. This is because as the value of N increases, the number of nodes in the reference set decreases, resulting in a decrease in the number of nodes that the top-k of each method intersects with the reference set. At the same time, as the top-k increases, the accuracy of all methods shows an increasing trend. When N=2, ASIRank performs best, with an accuracy of almost 90%. The low accuracy of DegreeRank indicates that a large number of fans does not bring greater influence. It is difficult to reveal the influence of users based on topological measurements (such as the number of fans). The poor accuracy of PageRank also shows that considering only one relationship cannot accurately quantify the influence of users. It is necessary to comprehensively consider multiple link relationships as well as factors such as activity, interaction intensity, and interest preferences.

[0175] exist Figure 6In the results, compared with the official influence evaluation method of Sina Weibo, it can be clearly observed that the accuracy of the DegreeRank method is much lower than that of the other methods. No matter how many rankings are in the top, the ASIRank method performs best. When comparing the top 100 users, the method proposed in the present invention is 4%, 7%, 10% and 20% higher in calculation accuracy than the MPPRank, TunkRank, PageRank and DegreeRank algorithms, respectively. In summary, it can be proved from two aspects that the method proposed in the present invention is more accurate in analyzing user influence.

[0176] D. Correlation comparison

[0177] In order to further compare these five algorithms and highlight which algorithm is more consistent with the actual true value, the present invention uses the Spearman rank correlation coefficient to calculate the performance of the overall ranking and uses the Spearman rank correlation coefficient ρ to analyze the correlation between different rankings:

[0178]

[0179] in, and are the positions of the same user in the reference ranking and the ranking to be evaluated, respectively, where the superscript t represents each user variable, It indicates the position difference of the two variables after they are sorted. The value range of the factor ρ is [-1,1]. If the two ranks are exactly the same, ρ = 1, otherwise, ρ = -1. The higher the Spearman rank correlation value, the stronger the correlation between the reference ranking and the ranking to be evaluated, and the better the effect of the method to be evaluated.

[0180] Table 1 Correlation comparison results

[0181]

[0182] Table 1 shows the Spearman rank correlation coefficient of each ranking method under different evaluation criteria. Using RefRank as the reference ranking, it can be seen that the ASIRank method has a higher ρ value under both evaluation criteria, that is, it has a higher rank correlation. This result shows that the ranking obtained by the ASIRank method is generally credible compared with the other four ranking algorithms.

[0183] D. Coverage Comparison

[0184] The present invention uses an independent cascade model to test the propagation coverage of user influence. When a user's microblog information is browsed by other users and corresponding interactive behaviors are made, it means that the influence is transmitted between users. This propagation behavior conforms to the activation behavior in the independent cascade model. Therefore, the IC cascade model is selected for coverage experiment. The top k% users in terms of influence are selected as seed nodes, and the remaining users in the network are selected as users to be activated. Comparing the coverage of the above five algorithms, the experimental results are as follows: Figure 7 .

[0185] Figure 7 In the figure, the horizontal axis represents the proportion of seed nodes in the total number of users, and the vertical axis represents the proportion of affected users in the total number of users. The larger the ratio, the more users are affected, that is, the wider the scope of influence. From the results in the figure, it can be seen that the coverage of the ASIRank algorithm is better than that of the MPPRank, TunkRank, PageRank and DegreeRank algorithms. When the user ratio is 0.6, the number of activated users accounts for 90% of the total, and is 5% more than the MPPRank algorithm (about 5,000 users), which shows that the algorithm proposed in the present invention obtains users with great influence, which has a wider influence range.

[0186] This paper measures the influence of users in social networks by improving the traditional PageRank algorithm. Under different relationship network dimensions, it selects multiple targeted features such as user activity, interaction intensity and interest similarity, proposes the ASIRank algorithm, and compares it with four existing methods. The effectiveness of the algorithm is verified by indicators such as accuracy, coverage and relevance. Experimental results show that the performance of the ASIRank method is better than other ranking methods.

[0187] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0188] Finally, it should be noted that the above is only a preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, and is not used to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A social user influence measurement method based on multi-angle analysis, characterized in that: The following steps are involved: Step 1: Build the user's relationship network from multiple perspectives; The user relationship network is constructed from multiple perspectives, including the following relationship, forwarding relationship, comment relationship, and mention relationship. The relationship network is defined as G i = {V i ,E i }, where V i is the set of users, E i is the link relationship, i∈{f,r,c,m} corresponds to the follow, forward, comment and mention relationships respectively; Step 2: In different relationship networks, use user activity, interaction intensity, and interest similarity to quantify user influence; The user activity refers to the frequency of various activities of the user in the network, and the user activity is defined as follows: Act i (v)= Post(v)+Interact i (v),i∈{f,r,c,m} (1) Among them, Act i (v) represents the activity of user v in different networks, Post(v) represents the number of microblogs posted by user v, Interact i (v) indicates the number of times user v follows, reposts, comments on, or mentions other users; The interaction intensity is used to describe and quantify the interaction behavior between users. The interaction intensity is defined as follows: Among them, RS i (v,u) represents the interaction intensity between user v and user u in different networks. i (v,u) represents the number of times user u forwards, comments, or mentions user v. Interact i (u) indicates the number of times user u forwards, comments, or mentions other users. i (v) indicates the number of times other users forwarded, commented on, or mentioned user v; The interest similarity is regarded as an indicator of the possibility of information propagation from one user to another user, and the interest similarity between users is measured by comparing the similarity of attribute keyword vectors between users; For any user v, extract the keywords of the user's attributes, expressed as KW(v) = (Keyword1: Weight1; Keyword2: Weight2; ...), where Keywordi represents the keyword in the user v's attributes, and Weighti represents the weight of Keywordi in the user v's attributes; For two users v and u, the interest similarity value between the users is obtained by calculating the inner product sum of the same subvectors in the two keyword vectors, that is, by calculating the intersection of the keywords in KW(v) and KW(u), the two subvectors of v and u are obtained respectively, expressed as SK(v)=(ckw1:w v1 ;ckw2:w v2 ; ......), SK(u)=(ckw1:w u1 ;ckw2:w u2 ; ......), where ckw* represents the same keyword as KW(v) and KW(u), w * represents the weight, and the interest similarity between users v and u is defined as formula (3): Among them, IS(v,u) represents the interest similarity between user v and user u, w vk and w uk They represent the weights corresponding to user v and user u respectively, and s represents the number of the same keywords; In the forwarding, commenting and mentioning subnetworks, the interaction influence between users is quantified by the interaction intensity and interest similarity, and the interaction influence between users is defined as follows: Among them, Inf i (v,u) represents the interaction between user v and user u in different networks, RS i (v,u) represents the interaction intensity between user v and user u in different networks, and IS(v,u) represents the interest similarity between user v and user u; Step 3: Integrate the user's influence from different perspectives to obtain the user's comprehensive influence in the global scope; Calculate the influence of a user in the forwarding subnet: The transition probability of random walk from user v to user u in the forwarding subnet is defined as formula (5): Where Ree(u) represents the set of users who forward user u, Inf r (v,u) represents the interaction between user v and user u in the forwarding subnet; The ratio of a single user to the sum of the activities of all users in the forwarding subnet is defined as the jump probability, as shown in formula (6): Among them, Act r (v) represents the activity of user v in the forwarding subnet, and the value of θ is a non-zero decimal; The influence of user v in the forwarding subnet is calculated as follows: Among them, α is 0.85, Ret(v) represents the set of users forwarded by user v; Calculate the influence of a user in the comment subnetwork: The transition probability of random walk from user v to user u in the comment subnetwork is defined as formula (8): Among them, Coe(u) represents the set of users who comment on user u, Inf c (v,u) represents the interaction influence between user v and user u in the comment subnetwork; The ratio of a single user to the sum of the activities of all users in the comment subnet is defined as the jump probability, as shown in formula (9): Among them, Act c (v) represents the activity of user v in the comment subnetwork, and the value of θ is a non-zero decimal; The influence of user v in the comment subnetwork is calculated as follows: Among them, α is 0.85, Com(v) represents the set of users commented by user v; Similarly, calculate the influence of the user in the mention subnet: In the mentioned subnet, the transition probability is as follows: Among them, Mee(u) represents the set of users who mention user u, Inf m (v,u) represents the interaction influence between user v and user u in the mention subnet; The jump probability is as shown in formula (12): Among them, Act m (v) represents the activity of user v in the mention subnet, and the value of θ is a non-zero decimal; The influence of user v in the mention subnetwork is calculated as follows: Among them, α is 0.85, Men(v) represents the set of users mentioned by user v; Calculate the influence of a user in the following subnetwork: The transition probability from user v to user u in the attention subnet is defined as formula (14): Among them, Foe(u) represents the set of users followed by user u, and Post(v) represents the number of microblogs posted by user v; The jump probability is as shown in formula (15): Among them, Act f (v) represents the activity of user v in the attention subnet, and the value of θ is a non-zero decimal; The influence of user v in the attention subnet is calculated as follows: Among them, α is 0.85, Fow(v) represents the set of fans of user v; Through linear fusion, the comprehensive influence of users in the global scope is calculated as shown in formula (17): ASInf(v)=β1ASInf f (v)+β2ASInf r (v)+β3ASInf c (v)+β4ASInf m (v) (17) Among them, β1, β2, β3, and β4 are the influence weights of the user in the attention network, forwarding network, comment network, and mention network, respectively, and β1+β2+β3+β4=1.

Images