A dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model

Abstract

The application belongs to the technical field of social network influence maximization, and specifically discloses a dynamic duplicate-removal seed selection algorithm based on a multilevel composite evaluation model, which comprises the following steps: given a marketing budget L1, obtaining a social network data set, pre-processing the data set, and obtaining the connection relationship between nodes in the social network; calculating the influence Inf(V i ) of a single node of the processed data set; according to the node influence Inf(V i ) in the data set, taking all nodes as a seed set and performing hierarchical evaluation, and sorting according to the evaluation value of the decision function; and selecting the seed set seeds as the initial active node set. The dynamic duplicate-removal seed selection algorithm based on the multilevel composite evaluation model can accurately reflect the behavior of putting advertisements in a social network in real life, and can make the influence range more accurate and extensive.

Description

Technical Field

[0001] This invention belongs to the field of network information dissemination technology, and specifically relates to a dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model. Background Technology

[0002] With the rise of social networks and the surge in users, social networks have become a crucial platform for information acquisition and dissemination, leading more and more businesses to shift their marketing focus there. Viral marketing on social networks, utilizing word-of-mouth marketing to spread information widely, has ushered in an era of self-media promotion. This has created new challenges: how to conduct effective advertising on social networks while controlling marketing costs and anticipating returns. This has led to the introduction of algorithms for maximizing influence within budget constraints. Existing algorithms for social network influence propagation have limited research on the seed selection phase, and studies on influence coverage issues arising from seed selection only extend to static topology deduplication. Therefore, research on seed selection algorithms in social network propagation and the dynamic deduplication problem during seed selection is of significant practical importance.

[0003] The problem of maximizing influence is one that many researchers are exploring in depth. The problem of maximizing influence is to find the optimal set of nodes in a large network topology so that these nodes can radiate throughout the entire network to the greatest extent possible, ultimately affecting as many nodes as possible. In other words, in the process of information propagation, it is to find an initial set of seed nodes with the greatest influence, so that the scope of influence is as large as possible with the least cost.

[0004] As can be seen from the current state of research on the influence maximization problem, there are already many relatively mature algorithms for maximizing influence in social networks. However, these works mainly focus on the preliminary influence estimation of all nodes in the network, such as the Greedy algorithm, which is very time-consuming. At the same time, the existing research adopts the idea of ​​prioritizing the node with the greatest influence as the seed. This algorithm, which focuses too much on its own influence, can lead to the phenomenon of "rich clubs," where nodes with greater influence are neighbors of each other, causing the influence range of the seeds to overlap. If these nodes are selected as seed nodes, a large number of common neighbors will appear during the propagation process, causing the influence range of the seed nodes to overlap. The current solution to the overlap is based on static deduplication of the topology, which is no longer effective for real dynamic networks.

[0005] The goal of maximizing influence is to maximize the combined influence of multiple seed nodes. However, in reality, the combined influence of seed nodes selected based on their influence ranking in communication studies is often not the greatest. A good set of seed nodes should have a wide distribution across the network. At the same time, influence models with budget constraints often have fixed marketing content. Ignoring the alignment between the nodes themselves and the marketing direction, as well as the cost-effectiveness of selecting nodes when choosing a seed set, can lead to the exclusion of highly cost-effective nodes with moderate cost and moderate influence within certain budget ranges. This results in a narrow scope for the final selected seed set, with centrally located, high-cost-effective nodes being overlooked, and the invested budget failing to achieve the expected range of influence. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies that fail to fully consider real-world influencing factors, thus preventing their effective application, and to provide a dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model.

[0007] This invention provides a dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model, which includes two parts: a multi-level composite evaluation model and a dynamic coverage deduplication algorithm.

[0008] The constant L1 represents the total budget for selecting seeds. A directed weighted graph G = (V, E) is used to represent the relationships between users in an online social network, where V represents the set of nodes, each node represents a user in the social network, E represents the set of edges between nodes, each edge represents the influence relationship between connected vertices, and the weight of the edge represents the strength of the relationship between two users in the social network. This is used to find the seed set with the greatest influence in the social network, and the total cost required to select seeds does not exceed the budget L1.

[0009] in:

[0010] Multi-level composite evaluation modeling is used to achieve:

[0011] For users who are on the periphery of the social network, with few connections to other users, or whose historical posts indicate a lack of interest in the current marketing content and immunity to it, this group of users is initially removed from the network to reduce computational complexity. Specifically, based on the user profile of the social network, the individual node influence of all nodes in node set V is calculated. Node set V is then filtered based on this individual node influence, discarding nodes whose individual node influence ranks in the bottom third of the overall network node set influence. The individual node influence is expressed as Inf(V). i The calculation basis is:

[0012] 1. The eigenvector centrality P(V) of a node itself i The eigenvector centrality of a node can be used as a selection criterion, with a weight of β1.

[0013] 2. The overlap R between the domain of the marketing content and the domain of the historical posts at this node. The overlap can be used as a selection criterion and its weight is β2. For example, if the marketing direction is a product in the beauty field, then the overlap between beauty bloggers and users who like beauty is high, while the overlap between beauty bloggers and parenting bloggers is low.

[0014] The following algorithms are mainly used for node influence estimation and node filtering in the multi-level composite evaluation modeling:

[0015] The relation matrix A is derived based on the node relationships in the dataset. ij Calculate the eigenvector centrality P(V) of each node. i The calculation formula is as follows:

[0016]

[0017] in It is the characteristic constant, X j This represents the relation matrix A. ij The eigenvectors corresponding to the eigenvalues ​​λ.

[0018] In this experiment, the overlap R between the domain of the marketing content and the domain of the node's historical posts was set to a randomly generated fixed value. β1 and β2 represent the weight coefficients of the corresponding parameters. The weight of the node's eigenvector centrality as a selection criterion is β1, and the weight of the overlap as a selection criterion is β2, where β1 + β2 = 1. Therefore, the influence Inf(V) of a single node in the social network can be derived. i The calculation function for ) uses the following formula:

[0019] Inf(V i )=β1·P(V i )+β2·R,

[0020] R represents the overlap between the domain of the node's historical posts and the current marketing campaign, and β1 and β2 represent the weight coefficients of the corresponding parameters, with β1 + β2 = 1. In this invention, β1 and β2 are both taken as 0.5, indicating the eigenvector centrality P(V i The overlap ratio R and the relevance R between the node and the current marketing content each account for half of the influence, and the relevance R between the node and the current marketing content is a determined random value.

[0021] Calculate Inf(V) of nodes in the dataset iThe highest and lowest values ​​of ) are used as the boundary between the highest and lowest values, and the low-order nodes (i.e., values ​​less than γ) are filtered out, and the remaining nodes are used as the candidate seed set.

[0022] Based on the specific influence of users on social networks, Inf(V) i All nodes in the candidate seed set are evaluated hierarchically, based on Inf(V) i Based on the size of the x, y, and z components, nodes with large components are classified into class S, medium components into class P, and small components into class D. Class S nodes represent users with significant influence and greater say in the marketing product's domain. The cost of selecting an S node as a seed is C. s High; P-class nodes indicate that the user has moderate influence and moderate say in the field of the marketing product. The cost of choosing a P-node as a seed is C. P Centered, it can supplement the coverage of S-type nodes; D-type nodes indicate that the user's influence is at a medium to low level and has a certain relevance to the field of this marketing product. The cost of choosing D-type nodes as seeds is C. d It is relatively small and is mainly used to cover isolated small and medium-sized groups, as a supplement to the influence range of S-class and P-class nodes;

[0023] The dynamic overlay deduplication algorithm is used to implement:

[0024] During the seed selection process in a social network, the candidate seed set is dynamically updated according to the following rules: When seed selection begins, the user's Inf(V) is... i The cost C required for its corresponding grading s C P C d The evaluation is performed by selecting a decision function, and the seeds are selected according to the order of the evaluation values ​​of the decision function.

[0025] In the process of selecting one by one, assuming the selected seed set is seeds, the budget C has been used, and the candidate seed set sequence after sorting according to the decision function in the social network is V1, V2, V3..., then we determine whether the remaining budget L is greater than the cost of selecting node V1 (i.e., if the node belongs to class S, then the cost is C). s If the node belongs to class P, then the cost is C. P If the node belongs to class D, then the cost is C. D If L is greater than the choice cost of V1, that is... Then, add node V1 to the seed set, and perform K MC simulations on node V1. Find the intersection of the influence sets of the K simulations to obtain the high-frequency influence set of node V1. Will Remove nodes from the candidate seed set and update the selected seed set and the candidate seed set; if L1-C is less than the selection cost of V1, discard V1 from the candidate seed set and re-enter the selection process; the seed selection process ends when L is 0 or the candidate seed set is empty.

[0026] Based on this, this application discloses a dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model, including the following steps:

[0027] Step 1: Given a marketing budget L1, obtain a social network dataset, preprocess the dataset, and obtain the connection relationships between nodes in the social network;

[0028] Step 2: Calculate the influence of a single node (Inf(V)) on the processed dataset. i ) calculation;

[0029] Step 3: Based on the node influence Inf(V) in the dataset i All nodes are used as a seed set and evaluated hierarchically, and then sorted according to the evaluation values ​​of the decision function.

[0030] Step 4: Select the seed set as the initial set of active nodes.

[0031] A further approach is to use a directed weighted graph G = (V, E) to represent the relationships between nodes in the social network, where V represents the set of nodes, each node represents a user in the social network, E represents the set of edges between nodes, each edge represents the influence relationship between connected vertices, and the weight of the edge represents the strength of the relationship between two users in the social network. This is used to find a seed set with the greatest influence in the social network, and the total cost required to select the seed set does not exceed the budget L1.

[0032] A further approach is to derive the relation matrix A based on the directed weighted graph G = (V, E). ij Calculate the eigenvector centrality P(V) of each node. i The calculation formula is as follows:

[0033]

[0034] in, It is the characteristic constant, X j This represents the relation matrix A. ij The eigenvectors corresponding to the eigenvalues ​​λ.

[0035] A further solution is to increase the influence of a single node Inf(V) as described in step 2. iThe calculation function for ) uses the following formula:

[0036] Inf(V i )=β1·P(V i )+β2·R

[0037] Where R is the overlap between the domain of the marketing content and the domain of the node's historical posts, β1 and β2 represent the weight coefficients, and the probability that the node's feature vector centrality can be used as a selection criterion is β1, where β1+β2=1.

[0038] A further approach is that, in step 3, the hierarchical evaluation of the seed set includes:

[0039] Step 3.1: Taking γ∈(0,1) as the boundary, let k=P(V i )*γ, x=R*γ, when P(V i When ) < k or R < x, filter the nodes and calculate the Inf(V) of the nodes in the dataset. i The highest and lowest values ​​of the () are used to filter the low-order nodes between the highest and lowest values, and the remaining nodes are used as the candidate seed set.

[0040] Step 3.2, sort the candidate seed set according to Inf(V) i The values ​​are divided into three equal parts from high to low, with large component nodes classified as S-class, medium component nodes as P-class, and small component nodes as D-class; the selection costs for S-class, P-class, and D-class nodes are set as C0, respectively. s C P C d ;

[0041] Step 3.3, based on the Inf(V) of each node in the dataset i The decision function is as follows: (This includes the cost of selecting this node, and the subsequent sorting of the candidate seed set using a decision function.)

[0042]

[0043] In the formula, f represents the ranking decision value of the candidate seed set. Indicates the selection of node V i The required costs.

[0044] A further approach involves selecting the seed set (seeds) using the following steps:

[0045] Step 4.1: Determine if the current candidate seed set is empty and if the budget C has been used. Determine if the remaining budget L is greater than the cost of selecting node V1 (i.e., if the node belongs to class S, the cost is C). s If the node belongs to class P, then the cost is C. PIf the node belongs to class D, then the cost is C. D If L is greater than the choice cost of V1, that is... Then add node V1 to the seed set and go to step 4.2; if L1-C is less than the selection cost of V1, go to step 4.3; if L is 0 or the candidate seed set is empty, the seed selection process ends.

[0046] Step 4.2: Perform K MC simulations on node V1, and find the intersection of the influence sets of the K simulations to obtain the high-frequency influence set of node V1. Will Remove the nodes from the candidate seed set, update the selected seed set and the candidate seed set, and proceed to step 4.1;

[0047] Step 4.3: Discard V1 from the candidate seed set and re-enter step 4.1.

[0048] A further approach is to use a linear threshold model as the propagation model and a Monte Carlo simulation method as the social network propagation method, and to test the propagation effect of the seed set selected in step 4 on the dataset.

[0049] A further proposed solution is that the propagation effect test includes:

[0050] At time t, the newly activated node u affects its neighboring node v with a success probability of p. If v has multiple neighboring nodes that are also newly activated, then these nodes will try to activate node v in any order.

[0051] If node v is successfully activated, then at time t+1, node v becomes active and will affect its adjacent inactive nodes; otherwise, the state of node v does not change at time t+1.

[0052] Repeat the above process until there are no more influential active nodes in the network, at which point the propagation process ends.

[0053] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention filters users in social networks and sorts them according to their influence and relevance to the current marketing content. Then, it performs dynamic deduplication one by one during the seed selection process. It introduces important factors that affect the spread of influence in social network product marketing in real life, such as multi-level composite evaluation model, joint filtering of feature vector centrality and relevance, and dynamic coverage deduplication algorithm. It better simulates the process of advertising on social networks in real life. Experimental results show that the dynamic coverage deduplication seed selection algorithm based on multi-level composite evaluation model can accurately reflect the behavior of advertising on social networks in real life, and can make the scope of influence more precise and extensive. Attached Figure Description

[0054] The following figures are for illustrative purposes only and are not intended to limit the scope of the invention, wherein:

[0055] Figure 1 This is a schematic diagram of a multi-level composite evaluation model.

[0056] Figure 2 This is a schematic diagram illustrating the dynamic deduplication algorithm;

[0057] Figure 3 This is a graph showing the results of a comparative experiment conducted on the Facebook dataset;

[0058] Figure 4 This is a graph showing the results of a comparative experiment conducted on the CA-GrQc dataset;

[0059] in, Figure 3 and Figure 4 CD-MCE in this application is the dynamic overlay deduplication seed selection algorithm. Detailed Implementation

[0060] To make the objectives, technical solutions, design methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0061] In this embodiment, there are two datasets: CA-GrQc and facebook, both of which are from the SNAP dataset on the Stanford website.

[0062] This application presents a dynamic coverage deduplication seed selection algorithm based on a multi-level composite evaluation model under budget constraints, such as... Figure 1 and Figure 2 As shown, the algorithm includes the following steps:

[0063] Step 1: Given budget L1, obtain the social network dataset, preprocess the dataset, and obtain the connection relationships between nodes in the social network;

[0064] Step 2: Calculate the influence of a single node (Inf(V)) on the processed dataset. i ) calculation;

[0065] The influence of the node as described in the above invention is Inf(V) i The specific technical solution for this embodiment includes the following steps in step 2:

[0066] Step 2.1: Derive the relationship matrix A based on the node relationships in the dataset. ijCalculate the eigenvector centrality P(V) of each node. i The calculation formula is as follows:

[0067]

[0068] in It is the characteristic constant, X j This represents the relation matrix A. ij The eigenvectors corresponding to the eigenvalues ​​λ.

[0069] Step 2.2: The overlap R between the domain of this marketing content and the domain of the historical posts of this node is set to a randomly generated fixed value in this experiment.

[0070] Step 2.3, β1 and β2 represent the weight coefficients of the corresponding parameters, respectively. The probability that the eigenvector centrality of a node can be used as a selection criterion is β1, and the probability that the overlap degree can be used as a selection criterion is β2, where β1 + β2 = 1. From this, the influence of a single node in the social network, Inf(V), can be obtained. i The calculation function for ) uses the following formula:

[0071] Inf(V i )=β1·P(V i )+β2·R,

[0072] In this embodiment, β1 and β2 are both 0.5, indicating that the eigenvector centrality EC and the overlap R each account for half of the influence, and the correlation R between the node and the current marketing content is also a determined random value.

[0073] Step 3, based on the node influence Inf(V) in the dataset i All nodes are evaluated hierarchically and then sorted according to the evaluation values ​​of the decision function.

[0074] Step 3.1, in the dataset, the Inf(V) of the node i The highest calculated value is 0.68, and the lowest value is 0.04. Filtering by one-third, the calculated value is 0.253. Therefore, the P(V) value in the CA-GrQc dataset is... i Nodes with a value lower than 0.253 are filtered out, and the remaining nodes are used as a candidate seed set.

[0075] Step 3.2, sort the candidate seed set according to Inf(V) i The values ​​are divided into three equal parts from highest to lowest. Large component nodes are classified into class S, medium component nodes into class P, and small component nodes into class D; the selection costs for class S, class P, and class D nodes are set as C, respectively. s C P C d .

[0076] Step 3.3, based on the Inf(V) of each node in the dataset i The decision function is as follows: (This includes the cost of selecting this node, and the subsequent sorting of the candidate seed set using a decision function.)

[0077]

[0078] In the formula, f represents the ranking decision value of the candidate seed set. Indicates the selection of node V i The required costs.

[0079] Step 4: Implement dynamic overlay and deduplication during the seed selection process.

[0080] Following the specific technical solutions provided in the above-described invention, step 4 in this embodiment includes the following steps:

[0081] Step 4.1: Determine if the current candidate seed set is empty and if the budget C has been used. Determine if the remaining budget L is greater than the cost of selecting node V1 (i.e., if the node belongs to class S, the cost is C). s If the node belongs to class P, then the cost is C. P If the node belongs to class D, then the cost is C. D If L is greater than the choice cost of V1, that is... Then add node V1 to the seed set and proceed to step 4.2; if L1-C is less than the selection cost of V1, proceed to step 4.3; if L is 0 or the candidate seed set is empty, the seed selection process ends.

[0082] Step 4.2: Perform K MC simulations on node V1, and find the intersection of the influence sets of the K simulations to obtain the high-frequency influence set of node V1. Will Remove the nodes from the candidate seed set, update the selected seed set and the candidate seed set, and proceed to step 4.1.

[0083] Step 4.3: Discard V1 from the candidate seed set and re-enter step 4.1.

[0084] Step 5: Using a linear threshold model as the propagation model and Monte Carlo simulation as the social network propagation method, the seed set selected in Step 4 is used as the initial set of active nodes. The propagation effect is tested on the dataset. The specific steps are as follows:

[0085] Step 5.1, the initial set of active nodes, Seeds.

[0086] Step 5.2: At time t, the newly activated node u affects its neighboring node v with a success probability of p. If v has multiple neighboring nodes that are also newly activated, these nodes will attempt to activate node v in any order.

[0087] Step 5.3: If node v is successfully activated, then at time t+1, node v becomes active and will affect its adjacent inactive nodes; otherwise, the state of node v does not change at time t+1.

[0088] This process is repeated until there are no more influential active nodes in the network, at which point the propagation process ends. In this embodiment, the impact result is the average of 10,000 Monte Carlo simulation results.

[0089] In this embodiment, publicly available datasets from Stanford University are used, namely Facebook.txt and CA-GrQc.txt.

[0090] The CA-GrQc dataset covers scientific collaborations between authors in the general relativity and quantum cosmology categories. If author i and author j co-authored a paper, the graph contains an undirected edge from i to j. If the paper is co-authored by k authors, a fully connected (sub)graph is generated at the k nodes.

[0091] The data covers papers published between January 1993 and April 2003 (124 months). It begins within a few months of the founding of arXiv, and therefore essentially represents a complete part of its history. Detailed parameters of this network dataset are shown in Table 1.

[0092] Table 1 CA-GrQc Network Parameters

[0093]

[0094] The Facebook dataset consists of “circles” (or “friends lists”) from Facebook. Facebook data was collected from survey participants using this Facebook application. The dataset includes node features (profiles), circles, and self-networks. The Facebook data has been anonymized by replacing each user’s Facebook internal ID with a new value. Using the anonymized data, it is possible to determine whether two users have the same political affiliation, but it is not possible to determine what their individual political affiliations represent. Detailed parameters of this network dataset are shown in Table 2.

[0095] Table 2 Facebook Network Parameters

[0096]

[0097]

[0098] The algorithm evaluation metric used in this embodiment is the influence maximization algorithm. In practical application scenarios, there are many requirements, among which the number of activated nodes, running time and scalability are all factors that need to be considered. This paper uses the number of activated nodes as the evaluation criterion for the influence maximization algorithm.

[0099] The comparison algorithm used in this embodiment is:

[0100] RandomHeuristic (RANDOM): This heuristic involves uniformly and randomly placing nodes into a seed set from the network. This process is repeated until the entire budget is exhausted. Previously, this heuristic has been used as a baseline method in many studies.

[0101] MaximumDegreeHeuristic(MAX_DEG): This is a popular centrality-based heuristic. In this method, nodes are sorted according to their degree, and selections are made from the sorted list until the entire budget is exhausted. This method has also been used as a baseline algorithm in many studies.

[0102] BalancedSeedSelectionAlgorithm (BSSA): This is a balanced seed set selection algorithm proposed by Han et al. (2014), which combines influential and cost-effective nodes in the selected seed set.

[0103] To evaluate the social network seed selection method proposed in this invention, experiments were conducted on the Facebook dataset and the CA-GrQc dataset. Comparisons were made with three other methods, and the experimental results are as follows: Figure 3 and Figure 4 As shown in the figure, in experiments comparing the proposed method with three other methods on two different datasets, the proposed method outperforms all three methods by a large margin, demonstrating its effectiveness and generality in maximizing social network influence within budget constraints.

[0104] This invention overcomes the limitations of previous methods that always performed preliminary influence estimation on nodes, which led to high inefficiency. It uses eigenvector centrality to quickly prune the network, thereby improving efficiency. Furthermore, by dynamically simulating the influence of seeds during seed selection, it can adjust the candidate seed set in real time, solving the problem that static covering deduplication methods based on topology are not suitable for real social networks. At the same time, for the influence maximization problem under budget constraints, it proposes a selection method that combines seed selection cost with its own characteristics, making the entire algorithm more closely aligned with the realistic influence maximization problem with budget constraints.

[0105] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model, characterized in that, Includes the following steps: Step 1: Given a marketing budget L1, obtain a social network dataset, preprocess the dataset, and obtain the connection relationships between nodes in the social network; Step 2: Assess the influence of individual nodes on the preprocessed dataset. calculate; Step 3: Based on the node influence in the dataset All nodes are used as a seed set and evaluated hierarchically, and then sorted according to the evaluation values ​​of the decision function. Step 4: Select the seed set (seeds) as the set of active nodes; The influence of a single node as described in step 2 The calculation function uses the following formula: ； in, The degree of overlap between the domain of the marketing content and the domain of the historical posts on that node. , Let represent the weight coefficients, and represent the probability that the eigenvector centrality of a node can be used as a selection criterion. ,in =1; In step 3, the hierarchical evaluation of the seed set includes: Step 3.1: with As a boundary, let , ,when or At that time, the nodes are filtered, and the nodes in the dataset are calculated. The highest and lowest values ​​are used to filter the low-order nodes between the highest and lowest values, and the remaining nodes are used as the candidate seed set. Step 3.2, sort the candidate seed set according to The values ​​are divided into three equal parts from high to low, with large component nodes assigned to class S, medium component nodes to class P, and small component nodes to class D; the selection costs for class S, class P, and class D nodes are set as follows: , , ; Step 3.3, based on the data of each node in the dataset And the cost of selecting this node, and sort the candidate seed set using a decision function, as follows: ； In the formula, This represents the ranking decision value of the candidate seed set. Indicates the selection of a node Required costs; The preprocessing in step 1 involves using a directed weighted graph G=(V,E) to represent the relationships between nodes in the social network. Here, V represents the set of nodes, each node represents a user in the social network, E represents the set of edges between nodes, each edge represents the influence relationship between connected vertices, and the weight of the edge represents the strength of the relationship between two users in the social network. This is used to find the seed set with the greatest influence in the social network, and the total cost required to select the seed set does not exceed the budget L1. The relation matrix is ​​derived from the directed weighted graph G=(V,E). Calculate the eigenvector centrality of each node. The calculation formula is as follows: ； in, It is a characteristic constant. This represents a relation matrix. The eigenvectors corresponding to the eigenvalues ​​λ.

2. The dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model according to claim 1, characterized in that, The method for selecting the seed set includes the following steps: Step 4.1: Determine if the current candidate seed set is empty and if the budget has been used. Determine if the remaining budget L is greater than the selected node. The cost, if L is greater than The choice cost, i.e. Then the node Add to the seed set, proceed to step 4.2; if it is L1- Less than If the selection cost is determined, proceed to step 4.3; if L is 0 or the candidate seed set is empty, the seed selection process ends. Step 4.2, for the nodes Perform K MC simulations, find the intersection of the influence sets of the K simulations, and obtain the nodes. High-frequency influence set ,Will Remove the nodes from the candidate seed set, update the selected seed set (seeds) and the candidate seed set, and proceed to step 4.1; Step 4.3, will Discard from the candidate seed set and re-enter step 4.

1.

3. The dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model according to claim 1, characterized in that, The algorithm also includes using a linear threshold model as the propagation model and a Monte Carlo simulation method as the social network propagation method, and testing the propagation effect of the seed set selected in step 4 on the dataset.

4. The dynamic deduplication seed selection algorithm based on a multi-level composite evaluation model according to claim 3, characterized in that, The propagation effect test includes: At time t, the newly activated node u affects its neighboring node v with a success probability of p. If node v has multiple neighboring nodes that are also newly activated, then these nodes will try to activate node v in any order. If node v is successfully activated, then at time t+1, node v becomes active, meaning node v is an active node, and the active node will affect its adjacent inactive nodes; otherwise, the state of node v does not change at time t+1. Repeat the above process until there are no active nodes in the network, at which point the propagation process ends.

Images