Initial user selection method for social network cross-group competition influence income maximization
By introducing the competition impact dissemination model CIPRM and evolutionary algorithm framework in social networks, the problem of initial user selection in cross-group competition impact environment is solved, and the impact in multiple groups is maximized, which is better than the initial user set of other algorithms. choose.
Patent Information
- Application Number
- CN202411164546.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2024-11-22
AI Technical Summary
In the competitive influence environment of social networks, how to select initial users across multiple groups to maximize the communication benefits of your own influence is an urgent problem, especially how to obtain the largest number among groups divided by different religions, communities, etc. Public opinion may affect communication income and have important applications.
A new competition impact propagation model CIPRM is proposed, based on the classic linear threshold model expansion, combined with evolutionary algorithm strategy, by simulating the competition impact propagation process, calculating the impact weight and propagation range between users, and using the evolutionary algorithm framework to select the initial user to maximize Competitive across groups affects returns.
It provides a competitive impact communication model that is more in line with the real situation, and can obtain the maximum impact communication income from the group's perspective under limited resources. Compared with other algorithms, it can obtain a better initial user set, which maximizes the impact in different groups. .
Smart Images

Figure CN119006198A_ABST
Abstract
Description
Technical field:
[0001] The present invention belongs to the technical field of social network influence maximization, and more specifically, relates to an initial user selection method for maximizing competitive influence benefits across groups in a social network. Background technology:
[0002] In daily life, people exchange information and opinions through social networks to obtain information and influence each other's thoughts and decisions. Using relationships between people to conduct viral marketing is a classic application of social network influence communication. Due to its huge application prospects in the fields of public opinion prevention and control, political propaganda, etc., research on influence communication has become a hot topic, and influence maximization is one of the most important research issues in influence communication. In the study, users are regarded as nodes and the relationships between users are regarded as edges, forming a topological graph of the social network. Selecting k nodes from the network as the initial users and maximizing the final influence benefits through the cascade propagation of influence between users is the main goal of influence maximization research. By modeling influence communication, it is important to dig out the initial users who can maximize the benefits of influence communication, which has important social and commercial value.
[0003] In social networks, there are competing influences spreading in the network, such as positive and negative public opinions or competitive promotion of similar products. How to make the spread of one's own influence accepted by more individuals in the competitive influence spread has attracted the attention of scholars. However, in the real influence competition environment, it is more important to study the popularity of one's own influence in different groups. For example, when two competing public opinions spread in a social network, the popularity of one's own information in groups divided by different religions, communities, etc. is of great significance for the control of public opinion or the recommendation of products. When the spread of one's own influence wins in a certain group, the influence benefit of the group is obtained. How to select k nodes as the initial users in the competition for influence in the social network, and ultimately maximize the sum of the benefits of one's own influence across groups, is an urgent problem to be solved. Summary of the invention:
[0004] In order to solve the problem of maximizing the cross-group influence benefits of social networks under competitive influence environments, the present invention provides for the first time an initial user selection method for maximizing the competitive influence benefits across groups in social networks. The method first proposes a new competitive influence propagation model to simulate the competitive influence propagation process. The model is based on the extension of the classic linear threshold model and is named the competitive influence propagation model for reconsideration, abbreviated as CIPRM. The model can reflect the realistic scenario that when the same user faces two competing influences, if the advantage of the dominant influence is not obvious, the user will enter the reconsideration stage and will not make a choice directly. On the basis of the given CIPRM, since the initial user selection for maximizing the competitive influence benefits across groups is an NP-Hard problem, an evolutionary algorithm strategy is adopted to provide an initial user selection method for maximizing the competitive influence benefits across groups in social networks.
[0005] For the sake of convenience, the two competitive influences involved in the present invention are represented by D and R, respectively, where R is the influence of the party itself and D is the influence of the competitor. The social network is divided into m disjoint groups, and the i-th group is represented by C i In group C i If the number of users affected by R is greater than the number of users affected by D, then R is said to be in C i The profit obtained by R in the group is recorded as pf(C i ). When the initial user set S of R R For competitor D’s initial user set S D When R obtains the total benefit in m groups, it is recorded as PF(S R |S D ). To achieve the above-mentioned invention object, the method for selecting initial users for maximizing the competitive influence benefits across groups in a social network of the present invention comprises the following steps:
[0006] S100: Obtaining the social network topology structure graph G(V,E) and the groups C to which the social network users belong from the social network i . Calculate the D influence weight between users in the social network through the topological relationship between user u and his friend user v in the social network and R influence weight Here are the steps.
[0007] S101: Take users in the social network as nodes and friendships between users as edges, construct a social network topology graph G(V,E), and annotate the nodes with group labels to which the users belong.
[0008] S102: Further, the closeness of the relationship between neighbor nodes u and v is calculated: Among them, N(v) is the set of neighbor nodes of node v, η u,v=|N(v)∩N(u)|+1. Obviously, ζ u,v ∈[0,1].
[0009] S103: Further, calculate the influence weight D between neighbor nodes u and v and R influence weight Among them U D and U R Represents the quality of R influence and D influence, U D ∈[0,1],U R ∈[0,1].
[0010] S200: Construct a competitive influence propagation model CIPRM, based on which the initial user S of a given D influence is calculated. D and the initial user S affected by R R When S is given, the propagation range of D influence and R influence is calculated. D and S R When R affects the inter-group competition, the profit PF(S R |S D ). The steps are as follows.
[0011] S201: Set the initial user affected by a given R to an R-active state, set the initial user affected by a given D to a D-active state, and set other users to an inactive state.
[0012] S202: For each node v∈V, select the activation threshold θ of the node from [0,1] with equal probability v , and record it.
[0013] S203: Nodes in a social network can only be in one of the three states: R-active, D-active, or inactive. Once a node changes from inactive to R-active or D-active, it will remain in R-active or D-active state and will not change. The propagation rules of D-influence and R-influence are as follows.
[0014] At time t (t is a non-negative integer), a node v in the social network that is in the inactive state will be affected by its neighboring nodes that are in the R-active or D-active state, and will change its state according to the following rules:
[0015] (1) If and Then node v will be activated by D and become D-active. Here, N D (v) and N R(v) represent the neighbor nodes of node v in R-active and D-active states respectively.
[0016] (2) If and Then node v will be activated by R and become R-active.
[0017] (3) If and Then the node v is not activated and remains in the inactive state.
[0018] (4) If and The node will enter the rethinking state. When the node v The probability of choosing to be activated by D, becoming D-active, The probability of choosing is activated by R and becomes R-active. When the node v The probability of choosing to be activated by R, becoming R-active, The probability of choosing is activated by D and becomes D-active state.
[0019] At time t+1, the node activated at time t can try to activate its neighboring nodes. The nodes that are still in the inactive state will detect whether they are activated and change their activation state according to the above rules. This continues until the node state in the network stops changing or reaches the preset upper limit T. At the end of the statistical propagation, the number of nodes in the R-active state and the D-active state in each group is obtained, and the propagation range of the D influence and the R influence is obtained and recorded.
[0020] S204: Set the number of simulations to MC times, and execute MC times of steps S202 and S203. Record the number of times each group C i The average number of nodes in the R-active state inf(S R |C i ), and the average number of nodes in the D-active state inf(S D |C i ).
[0021] S205: Further, traverse each group C i , when inf(S R |C i )>inf(S D |C i ),group C iThe return pf(C i ) accumulated into PF(S R |S D ). Get the given S D and S R When R affects the cross-group competition, the impact of the profit PF (S R |S D ) and record it.
[0022] S300: Based on the competitive influence propagation model CIPRM, an evolutionary algorithm framework is constructed to solve the initial user selection problem of maximizing the benefits of cross-group competitive influence. That is, the number of initial users k influenced by the party R and the number of initial users S influenced by the competitor D are known. D , R in group C i The winner gets the group's profit pf(C i ), select the initial user who maximizes the total cross-group benefits of R.
[0023] S301: Number the nodes in the social network, first select the top 1% nodes with the largest node median, and then select the top 1% nodes with the largest node betweenness centrality without duplication. If there is a node that is repeated with the previous one, select it backwards in sequence. According to THA(v,λ)=∑ u∈N(v) W v,u ·∑ w∈N(u) W u,w +λ·CN, calculate the local importance index of node v in the social network, select the top 1% nodes of the index without duplication, if there are nodes that are repeated with the previous ones, select them backwards in sequence, where CN represents the number of different groups of the neighbor nodes of node v, and λ is the weight factor.
[0024] S302: From the candidate nodes selected in S301, k nodes are randomly selected each time to construct individual X i {x i1 ,x i2 ,…,x ik}, where i∈[1,…,num]. The num individuals randomly constructed form the initial population. Each individual X i {x i1 ,x i2 ,…,x ik} is a candidate initial user set.
[0025] S303: Randomly divide the individuals in the population into groups of 2, according to the probability pr c The probability of crossover is to generate new individuals while retaining the original individuals. Choose two-point crossover or one-point crossover with equal probability. i {x i1 ,xi2 ,…,x ik}, use step S200 to calculate each gene node x in the individual ih The marginal revenue PF(X i |S D )-PF(X i \x ih |S D ), and sort the gene nodes in each individual in descending order according to their marginal benefits.
[0026] Two-point crossover: Randomly select an integer o between 1 and k, and exchange the gene nodes before the two parent individuals o without any duplication, forming two child individuals.
[0027] Single-point crossover: Randomly select an integer o between 1 and k, and exchange the gene nodes at the o-th position of the two parent individuals to form two child individuals if there is no duplication of gene nodes within the individual after exchanging them.
[0028] S304: Further, all individuals are sorted according to the probability pr m The probability of mutation produces new individuals while retaining the original individuals. i {x i1 ,x i2 ,…,x ik}, randomly select a gene node x ih , and use step S200 to calculate, if node x ih ∈C y , and X i {x i1 ,x i2 ,…,x ik} As the initial user set of R, we can get group C y The income of x ih The candidate nodes constructed by random mutation for S301 do not belong to group C y , otherwise, x ih Random mutation is any node in the candidate node set constructed by S301.
[0029] S305: Calculate the X of each individual in the population using step S200 i The total cross-group benefit PF(X i |S D ), and select the individual with the largest total benefit and record it. i |S D ), starting from the individual with the smallest benefit value, until the population recovers to num individuals.
[0030] S306: Repeat steps S302, S303, S304, and S305 until the set number of iterations is reached. The largest individual with the largest cross-group total benefit recorded in each round is selected, and the individual represented by the largest individual is the selected initial user set.
[0031] The present invention provides a method for selecting initial users for maximizing the benefits of competitive influence across groups in a social network, and provides a solution framework for the problem of maximizing the benefits of initial users in a competitive influence environment from a group perspective. First, the topological structure diagram and the group to which the user belongs are obtained from the social network, and the weight of the user's influence on his friends is calculated through the topological relationship between the user and his friends. Secondly, based on the extension of the classic linear threshold model, a CIPRM competitive influence propagation model is constructed to simulate the propagation mechanism of competitive influence in social networks. The influence propagation model takes into account the realistic scenario that when the same user faces two competing influences, if the advantage of the dominant influence is not obvious, the user will enter the reconsideration stage and will not make a direct choice. Finally, since the initial user selection for maximizing the benefits of competitive influence across groups is an NP-Hard problem, an initial user selection method based on evolutionary strategies is provided.
[0032] The present invention has the following beneficial effects:
[0033] (1) The present invention provides a method for selecting the initial user that maximizes the influence benefit from the group in a competitive influence environment. This method is the first solution proposed for this type of problem. Solving this type of problem has important applications in how to obtain the maximum public opinion or influence dissemination benefit from the group perspective in groups divided from different religions, communities, classes, etc. under limited resources.
[0034] (2) The CIPRM competitive influence propagation model proposed by the present invention takes into account the realistic scenario that when a user faces two competing influences and the advantage of the dominant influence is not obvious, the user will enter the reconsideration stage and will not make a choice directly. This provides a competitive influence propagation model that is more in line with realistic scenarios.
[0035] (3) The present invention aims at the problem of initial user selection for maximizing the benefits of cross-group competition in social networks. Compared with the CIMMIC seed set selection algorithm, degree maximization algorithm, and PageRank algorithm that consider community homogeneity factors, the evolutionary algorithm framework proposed in the present invention can obtain a better initial user set. Description of the drawings:
[0036] Figure 1 It is a flowchart of a specific implementation method of the method for maximizing the initial user selection of cross-group competition influence benefits in social networks of the present invention;
[0037] Figure 2 It is a schematic diagram of the two-point crossover operation of the initial user selection algorithm framework;
[0038] Figure 3 It is a schematic diagram of the single-point crossover operation of the initial user selection algorithm framework;
[0039] Figure 4 It is the impact of the value of the weight factor λ in the initial user selection algorithm framework on the algorithm;
[0040] Figure 5 is the crossover probability pr in the initial user selection algorithm framework c The impact of the value of on the algorithm;
[0041] Figure 6 is the mutation probability pr in the initial user selection algorithm framework m The impact of the value of on the algorithm;
[0042] Figure 7 This is a comparison chart of the total revenue obtained by the initial users selected by the present invention under the HtpTh dataset and the other three comparison methods;
[0043] Figure 8 This is a comparison chart of the total benefits obtained by the initial users selected by the present invention under the EmailEnron data set and the other three comparison methods;
[0044] Fig. 9 This is a comparison chart of the total revenue obtained by the initial users selected by the present invention under the WebStan data set and the other three comparison methods; Specific implementation method:
[0045] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.
[0046] Example
[0047] Figure 1 This is a flowchart of a specific implementation method of the method for maximizing initial user selection of cross-group competition influence benefits in social networks of the present invention. Figure 1 As shown, the initial user selection method for maximizing the benefits of cross-group competition in social networks of the present invention specifically comprises the following steps:
[0048] S100: Obtaining the social network topology structure graph G(V,E) and the groups C to which the social network users belong from the social network i . Calculate the D influence weight between users in the social network through the topological relationship between user u and his friend user v in the social network and R influence weight Here are the steps.
[0049] S101: Take users in the social network as nodes and friendships between users as edges, construct a social network topology graph G(V,E), and annotate the nodes with group labels to which the users belong.
[0050] S102: Further, the closeness of the relationship between neighbor nodes u and v is calculated: Among them, N(v) is the set of neighbor nodes of node v, η u,v =|N(v)∩N(u)|+1. Obviously, ζ u,v ∈[0,1].
[0051] S103: Further, calculate the influence weight D between neighbor nodes u and v and R influence weight Among them U D and U R Represents the quality of R influence and D influence, U D ∈[0,1],U R ∈[0,1].
[0052] S200: Construct a competitive influence propagation model CIPRM, based on which the initial user S of a given D influence is calculated. D and the initial user S affected by R R When S is given, the propagation range of D influence and R influence is calculated. D and S R When R affects the inter-group competition, the profit PF(S R |S D ). The steps are as follows.
[0053] S201: Set the initial user affected by a given R to an R-active state, set the initial user affected by a given D to a D-active state, and set other users to an inactive state.
[0054] S202: For each node v∈V, select the activation threshold θ of the node from [0,1] with equal probability v , and record it.
[0055] S203: Nodes in a social network can only be in one of the three states: R-active, D-active, or inactive. Once a node changes from inactive to R-active or D-active, it will remain in R-active or D-active state and will not change. The propagation rules of D-influence and R-influence are as follows.
[0056] At time t (t is a non-negative integer), a node v in the social network that is in the inactive state will be affected by its neighboring nodes that are in the R-active or D-active state, and will change its state according to the following rules:
[0057] (1) If and Then node v will be activated by D and become D-active. Here, N D (v) and N R (v) represent the neighbor nodes of node v in R-active and D-active states respectively.
[0058] (2) If and Then node v will be activated by R and become R-active.
[0059] (3) If and Then the node v is not activated and remains in the inactive state.
[0060] (4) If and The node will enter the rethinking state. When the node v The probability of choosing to be activated by D, becoming D-active, The probability of choosing is activated by R and becomes R-active. When the node v The probability of choosing to be activated by R, becoming R-active, The probability of choosing to be activated by D is to become D-active. It can be found that when ε→0, the rethinking state will not work, and the node will be directly activated by the party with greater influence; when ε→+∞, the node entering the rethinking state will be activated by R or D with equal probability.
[0061] At time t+1, the node activated at time t can try to activate its neighboring nodes. The nodes that are still in the inactive state will detect whether they are activated and change their activation state according to the above rules. This continues until the node state in the network stops changing or reaches the preset upper limit T. At the end of the statistical propagation, the number of nodes in the R-active state and the D-active state in each group is obtained, and the propagation range of the D influence and the R influence is obtained and recorded.
[0062] It can be found that when the initial user of competitor D is And when ε→0, the propagation model CIPRM proposed in the present invention will degenerate into a classic linear threshold model.
[0063] S204: Set the number of simulations to MC times, and execute S202 and S203 MC times. Record the number of times each group C i The average number of nodes in the R-active state inf(S R |C i ), and the average number of nodes in the D-active state inf(S D |C i ).
[0064] S205: Further, traverse each group C i , when inf(S R |C i )>inf(S D |C i ),group C i The return pf(C i ) accumulated into PF(S R |S D ). Get the given S D and S R When R affects the cross-group competition, the impact of the profit PF (S R |S D ) and record it.
[0065] S300: Based on the competitive influence propagation model CIPRM, an evolutionary algorithm framework is proposed to solve the problem of initial user selection to maximize the benefits of cross-group competitive influence. That is, given the number of initial users k influenced by one party R and the number of initial users S influenced by the competitor D, D , R in group C i The winner gets the group's profit pf(C i ), select the initial user who maximizes the total cross-group benefits of R.
[0066] When the initial user of competitor D is ε→0, the entire social network has only one group, and the benefit of influencing R's victory in the group is the number of nodes activated by R, the problem degenerates into the classic initial user mining problem of maximizing influence under the LT model, which is an NP-Hard problem. Therefore, the initial user selection problem of maximizing the competitive influence benefit across groups in the social network to be solved by the present invention is at least an NP-Hard problem, and the framework for solving the problem using an evolutionary algorithm is as follows.
[0067] S301: Number the nodes in the social network, first select the top 1% nodes with the largest node median, and then select the top 1% nodes with the largest node betweenness centrality without duplication. If there is a node that is repeated with the previous one, select it backwards in sequence. According to THA(v,λ)=∑ u∈N(v) W v,u ·∑ w∈N(u) W u,w +λ·CN, calculate the local important row index of node v in the social network, select the top 1% nodes of the index without duplication, if there are nodes that are repeated with the previous ones, select them backwards in sequence, where CN represents the number of different groups of the neighbor nodes of node v, and λ is the weight factor.
[0068] S302: From the candidate nodes selected in S301, k nodes are randomly selected each time to construct individual X i {x i1 ,x i2 ,…,x ik}, where i∈[1,…,num]. The num individuals randomly constructed form the initial population. Each individual X i {x i1 ,x i2 ,…,x ik} is a candidate initial user set.
[0069] S303: Randomly divide the individuals in the population into groups of 2, according to the probability pr c The probability of crossover is to generate new individuals while retaining the original individuals. Choose two-point crossover or one-point crossover with equal probability. i {x i1 ,x i2 ,…,x ik}, use step S200 to calculate each gene node x in the individual ih The marginal revenue PF(X i |S D )-PF(X i \x ih |S D ), and sort the gene nodes in each individual in descending order according to their marginal benefits.
[0070] Two-point crossover: Randomly select an integer o between 1 and k, and exchange all the gene nodes before the two parent individuals o without any duplication, forming two child individuals. Figure 2 As shown, a pair of parent individuals X that need to be crossed i {x i1 ,x i2 ,…,x ik} and X j {xj1 ,x j2 ,…,x jk After two-point crossover, a new individual X is generated a {x a1 ,x a2 ,…,x ak} and X b {x b1 ,x b2 ,…,x bk}.
[0071] Single-point crossover: Randomly select an integer o between 1 and k, and exchange the gene nodes at the oth position of the two parent individuals to form two child individuals if there is no duplication of gene nodes within the individual after exchanging them. Figure 2 As shown, a pair of parent individuals X that need to be crossed i {x i1 ,x i2 ,…,x ik} and X j {x j1 ,x j2 ,…,x jk}After single-point crossover, a new individual X is generated a {x a1 ,x a2 ,…,x ak} and X b {x b1 ,x b2 ,…,x bk}.
[0072] S304: Further, all individuals are sorted according to the probability pr m The probability of mutation produces new individuals while retaining the original individuals. i {x i1 ,x i2 ,…,x ik}, randomly select a gene node x ih , and use step S200 to calculate, if node x ih ∈C y , and X i {x i1 ,x i2 ,…,x ik} As the initial user set of R, we can get group C y The income of x ih The candidate nodes constructed by random mutation for S301 do not belong to group C y , otherwise, x ih Random mutation is any node in the candidate node set constructed by S301.
[0073] S305: Calculate the X of each individual in the population using step S200 i The total cross-group benefit PF(X i |S D ), and select the individual with the largest total benefit and record it. i |S D ), starting from the individual with the smallest benefit value, until the population recovers to num individuals.
[0074] S306: Repeat steps S302, S303, S304, and S305 until the set number of iterations is reached. The largest individual with the largest cross-group total benefit recorded in each round is selected, and the individual represented by the largest individual is the selected initial user set.
[0075] In order to better illustrate the technical effect of the present invention, the present invention is experimentally verified using specific examples. This experimental verification uses a real social network data set. In the HepTh data set, nodes represent papers in the field of high-energy physics theory, and edges represent citation relationships between papers. In the EmailEnron data set, nodes represent email addresses. If there is an email exchange between two addresses, there is an edge. The WebStan data set is a hyperlink network of the official website of Stanford University. In this data set, nodes represent web pages, and edges represent hyperlinks between web pages. The LPA algorithm is used in advance to divide the three network data sets into communities and mark them. Different communities represent different groups. The detailed statistical information of the data set is shown in the table below.
[0076] Table 1 Statistical information of experimental datasets
[0077]
[0078] Since the present invention is the first to propose a solution to the problem of maximizing the initial user selection of competitive influence benefits across groups in social networks, we choose other solutions that can be transplanted to this problem to compare the present invention. First, we verify the influence of hyperparameters on the performance of the present invention on three data sets. In the initialization stage, the parameter λ is used as a weighting factor in the THA calculation. In the crossover stage of step S303, the crossover probability pr c And the mutation probability pr in the mutation stage m In this experiment, the HighDegree algorithm was used to select 25 nodes as the initial users affected by competitor D. From the remaining nodes, the method of the present invention was used to select 30 nodes as the initial users affected by R. Since other comparison algorithms in subsequent experiments did not consider group benefits, R was selected in group C. i The profit from winning pf(Ci ), set to group C i The number of nodes in the simulation is set to ensure fairness. The number of simulations MC in step S200 is set to 3000. The experimental results are shown in Figure 4-6 As shown. The hyperparameter λ affects the quality of the initial candidate nodes selected by the evolutionary algorithm. When λ = 0, THA(v,λ) in step S301 only calculates the influence of node v in its local neighborhood. As the value of λ increases, bridge nodes connecting different communities are more likely to be selected. The experiment tested the cases where λ values were 0, 0.5, 1.0, 1.5 and 2.0. The results show that for the HepTh, EmailEnron and WebStan datasets, better solutions were obtained when λ values were 1.0, 1.5 and 1.5 respectively. The mutation probability pr m The selection range is [0.05, 0.25], and the step size is 0.05. The results show that for the HepTh, EmailEnron and WebStan datasets, the optimal pr m The values are 0.20, 0.15 and 0.20 respectively. The hyperparameter crossover probability pr c The selection range is [0.75, 0.95], and the step size is 0.05. The results show that for the HepTh, EmailEnron and WebStan datasets, the optimal pr c The values are 0.85, 0.9 and 0.9 respectively.
[0079] Secondly, in order to facilitate comparison between methods, the initial user selection method of the present invention is referred to as EA. Comparative method 1 adopts the method in the document "Xie, Xiaoqin, et al." Competitive influence maximization considering inactive nodes and community homophily." Knowledge-Based Systems 233 (2021): 107497." This document defines the community boundary influence, abbreviated as CBI, which is used to measure the ability of a node to influence other nodes within or outside the community, and proposes the CIMMIC algorithm, which uses the CBI value as a metric for initial user mining to identify highly influential nodes, but the algorithm does not consider the benefits of the community. Comparative method 2, the HighDegree algorithm selects the node with the highest degree in the network as the initial user. Comparative method 3, the PageRank algorithm selects the node with the highest PageRank value as the initial user. This method is very effective for identifying important web pages. In the method EA of the present invention, the hyperparameters are set as follows in the experiment: for the HepTh dataset, λ=1.0, pr m =0.20,pr c=0.85; for the EmailEnron dataset, λ=1.5, pr m =0.15,pr c =0.9; for the WebStan dataset, λ=1.5, pr m =0.20,pr c =0.9. In the experiment, the HighDegree algorithm was used to select 25 nodes as the initial user set of competitor D. The EA, CIMMIC, HighDegree and PageRank algorithms were used to select the seed set of R, and the cross-group influence benefits of the initial users selected by these four algorithms were compared. The results are shown in Figure 7-9 As shown. Since the original CIMMIC algorithm does not consider community benefits, the following adjustments are made to improve its performance in this problem setting. CIMMIC is used to select initial users in descending order of community benefits in communities with high group benefits and R is likely to win. The experimental results show that in all three data sets, the cross-group influence benefits achieved by the initial users selected by EA are higher than those of the other three methods. The HighDegree algorithm selects nodes with higher degrees as the initial user set, while the PageRank algorithm selects nodes with higher PageRank values as the initial user set. These two algorithms may select initial users scattered in different groups, so there is no guarantee that R influence will win in any group. The seed set selected by the improved CIMMIC is of higher quality than the HighDegree and PageRank algorithms. However, compared with EA, CIMMIC cannot consider cross-group influence benefits globally.
[0080] Finally, the above comparative experiments show the effectiveness of the method of the present invention in maximizing the initial user selection of the benefits of cross-group competition in social networks.
[0081] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for maximizing initial user selection based on competitive influence benefits across social networks, characterized in that: The following steps are involved: S100: Obtaining the social network topology structure graph G(V,E) and the groups C to which the social network users belong from the social network i . Calculate the D influence weight between users in the social network through the topological relationship between user u and his friend user v in the social network and R influence weight S200: Construct a competitive influence propagation model CIPRM, based on which the initial user S of a given D influence is calculated. D and the initial user S affected by R R When S is given, the propagation range of D influence and R influence is calculated. D and S R When R affects the inter-group competition, the profit PF(S R |S D ). S300: Based on the competitive influence propagation model CIPRM, an evolutionary algorithm framework is constructed to solve the initial user selection problem of maximizing the benefits of cross-group competitive influence. That is, the number of initial users k influenced by the party R and the number of initial users S influenced by the competitor D are known. D , R in group C i The winner gets the group's profit pf(C i ), choose to make R obtain the total benefit PF(S R |S D )The largest initial user.
2. The method for selecting initial users for maximizing competitive influence benefits across groups in a social network according to claim 1, characterized in that: The influence weight D between users in step S100 and R influence weight The calculation steps include: S101: Take users in the social network as nodes and friendships between users as edges, construct a social network topology graph G(V,E), and annotate the nodes with group labels to which the users belong. S102: Calculate the closeness of the relationship between neighbor nodes u and v: Among them, N(v) is the set of neighbor nodes of node v, η u,v =|N(v)∩N(u)|+1. Obviously, ζ u,v ∈[0,1]. S103: Further, calculate the influence weight D between neighbor nodes u and v and R influence weight Among them U D and U R Represents the quality of R influence and D influence, U D ∈[0,1],U R ∈[0,1].
3. The method for selecting initial users for maximizing competitive influence benefits across groups in a social network according to claim 1, characterized in that: In step S200, S is given D and S R When D influence and R influence are spread, and R influence's cross-group competition influence benefits PF(S R |S D ) calculation steps include: S201: Set the initial user affected by a given R to an R-active state, set the initial user affected by a given D to a D-active state, and set other users to an inactive state. S202: For each node v∈V, select the activation threshold θ of the node from [0,1] with equal probability v , and record it. S203: Nodes in a social network can only be in one of the three states: R-active, D-active, or inactive. Once a node changes from inactive to R-active or D-active, it will remain in R-active or D-active state and will not change. The propagation rules of D-influence and R-influence are as follows. At time t (t is a non-negative integer), a node v in the social network that is in the inactive state will be affected by its neighboring nodes that are in the R-active or D-active state, and will change its state according to the following rules: (1) If and Then node v will be activated by D and become D-active. Here, N D (v) and N R (v) represent the neighbor nodes of node v in R-active and D-active states respectively. (2) If and Then node v will be activated by R and become R-active. (3) If and Then the node v is not activated and remains in the inactive state. (4) If and The node will enter the rethinking state. When the node v The probability of choosing to be activated by D, becoming D-active, The probability of choosing is activated by R and becomes R-active. Similarly, when When the node v The probability of choosing to be activated by R, becoming R-active, The probability of choosing is activated by D and becomes D-active state. At time t+1, the node activated at time t can try to activate its neighboring nodes. The nodes that are still in the inactive state will detect whether they are activated and change their activation state according to the above rules. This continues until the node state in the network stops changing or reaches the preset upper limit T. At the end of the statistical propagation, the number of nodes in the R-active state and the D-active state in each group is obtained, and the propagation range of the D influence and the R influence is obtained and recorded. S204: Set the number of simulations to MC times, and execute MC times of steps S202 and S203. Record the number of times each group C i The average number of nodes in the R-active state inf(S R |C i ), and the average number of nodes in the D-active state inf(S D |C i ). S205: Traverse each group C i , when inf(S R |C i )>inf(S D |C i ),group C i The return pf(C i ) accumulated into PF(S R |S D ). Get the given S D and S R When R affects the cross-group competition, the profit PF(S R |S D ) and record it.
4. The method for selecting initial users for maximizing competitive influence benefits across groups in a social network according to claim 1, characterized in that: In step S300, an evolutionary algorithm framework is constructed to select a method that enables R to obtain the total benefit PF (S R |S D ) The biggest initial user steps include: S301: Number the nodes in the social network, first select the top 1% nodes with the largest node median, and then select the top 1% nodes with the largest node betweenness centrality without duplication. If there is a node that is repeated with the previous one, select it backwards in sequence. According to THA(v,λ)=∑ u∈N(v) W v,u ·∑ w∈N(u) W u,w +λ·CN, calculate the local importance index of node v in the social network, select the top 1% nodes of the index without duplication, if there are nodes that are repeated with the previous ones, select them backwards in sequence, where CN represents the number of different groups of the neighbor nodes of node v, and λ is the weight factor. S302: From the candidate nodes selected in S301, k nodes are randomly selected each time to construct individual X i {x i1 ,x i2 ,…,x ik }, where i∈[1,…,num]. The num individuals randomly constructed form the initial population. Each individual X i {x i1 ,x i2 ,…,x ik } is a candidate initial user set. S303: Randomly divide the individuals in the population into two groups according to pr c The probability of crossover is to generate new individuals while retaining the original individuals. Choose two-point crossover or one-point crossover with equal probability. i {x i1 ,x i2 ,…,x ik }, use step S200 to calculate each gene node x in the individual ih The marginal revenue PF(X i |S D )-PF(X i \x ih |S D ), and sort the gene nodes in each individual in descending order according to their marginal benefits. Two-point crossover: Randomly select an integer o between 1 and k, and exchange the gene nodes before the two parent individuals o without any duplication, forming two child individuals. Single-point crossover: Randomly select an integer o between 1 and k, and exchange the gene nodes at the o-th position of the two parent individuals to form two child individuals if there is no duplication of gene nodes within the individual after exchanging them. S304: All individuals are sorted according to pr m The probability of mutation produces new individuals while retaining the original individuals. i {x i1 ,x i2 ,…,x ik }, randomly select a gene node x ih , and use step S200 to calculate, if node x ih ∈C y , and X i {x i1 ,x i2 ,…,x ik } As the initial user set of R, we can get group C y The income of x ih The candidate nodes constructed by random mutation for S301 do not belong to group C y , otherwise, x ih Random mutation is any node in the candidate node set constructed by S301. S305: Calculate the X of each individual in the population using step S200 i The total cross-group benefit PF(X i |S D ), and select the individual with the largest total benefit and record it. i |S D ), starting from the individual with the smallest benefit value, until the population recovers to num individuals. S306: Repeat steps S302, S303, S304, and S305 until the set number of iterations is reached. The largest individual with the largest total cross-group benefit recorded in each round is selected, and the individual represented by this individual is the selected initial user set.
