Influence maximization method, system and device in social network and storage medium

Abstract

The invention discloses an influence maximization method, system and device in a social network and a storage medium. The method belongs to the technical field of complex network analysis, and comprises the following steps: obtaining an embedded scale-free network of each node in a training network through structural similarity characteristics of the nodes; each node independently serves as a seed node set to conduct simulation information propagation, nodes affected by propagation of the single node are obtained, feature tags obtained by the nodes in the obtained training network are input into GAT to conduct feature reprocessing and training optimization of relevant parameters, and then influence prediction is conducted on a real network; and introducing a k-shell value and a degree centrality index to reset an influence score for each node, and selecting the first k nodes with high influence scores as a seed node set. According to the method provided by the invention, the structural similarity is better captured, the local influence degree is better aggregated, graph embedding and graph attention networks are introduced in solving the influence maximization problem, and the efficiency of mining seed nodes is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of complex network analysis, and relates to a method, system, device and storage medium for maximizing influence in a social network, and in particular to a method, system, device and storage medium for maximizing influence in a social network that captures structural similarity and aggregates local influence. Background Art

[0002] Complex social networks can model the complex relationships between people in the real world. It is of great significance to analyze useful information and unknown relationships in social networks. Information will spread in the network. This process of spreading has important practical and theoretical significance in the control of false information, viral marketing, reducing economic costs, social recommendation, recruitment, viral transmission control, selecting influential tweets, and pushing targeted advertisements. There are many such relationships that can be modeled into graph structures. In order to more effectively help the spread of information in the network, the influence maximization (IM) problem is proposed. The influence maximization problem is defined as identifying the most influential nodes in the network. The identified nodes have high propagation capabilities. A piece of information from them can reach the maximum extent in the network by triggering information diffusion. Identifying influential nodes is one of the hot research topics in the field of network science.

[0003] In recent years, the influence maximization problem has been widely studied in the field of social networks. It has been proven that the IM problem belongs to the category of NP-Hard, and methods such as greedy, heuristic, deep learning and community detection have been developed to solve it. Degree centrality evaluates the importance of a node by calculating the number of direct neighbors (1-hop neighbors) of the node, providing a measure of the local influence of the node. This method determines the importance of a node by counting the proportion of the shortest path between each pair of nodes passing through a specific node, thereby identifying those nodes in important positions. Kitsak et al. proposed K-shell centrality, assuming that nodes located in the core area of ​​the network are more conducive to information dissemination than peripheral nodes, and gradually assign K-shell values ​​to nodes from the periphery to the core. However, K-shell centrality assigns the same influence score to all nodes in the same shell layer, making it difficult to distinguish the actual diffusion capabilities of these nodes. To solve this problem, the K-shell values ​​of neighboring nodes are introduced to further distinguish the influence of nodes in the same shell. Although the algorithm based on node centrality has a low computational cost, it is not widely applicable to different types of social networks. Although the greedy algorithm has a theoretical performance guarantee, it is difficult to adapt to the needs of large-scale social networks due to its high computational complexity. Therefore, how to design an algorithm to improve the recognition efficiency of large-scale networks while ensuring that the identified nodes have a high influence is the breakthrough point of the current influence maximization method. This dilemma has prompted researchers to explore more efficient and generalizable methods, and graph neural networks have therefore entered the field of vision of researchers.

[0004] Graph neural networks (GNNs) are a type of model that can perform deep learning directly on graph structured data. In recent years, they have made significant progress in social network analysis, recommendation systems, knowledge graphs and other fields. The core idea of ​​GNN is to capture the dependencies between nodes and neighbors in the graph by iteratively updating the feature representation of nodes, thereby obtaining a global representation of each node. In the problem of maximizing influence, GNN can well model the local and global structural features of nodes and provide a more accurate representation of the influence of nodes. Among them, Keikha et al. proposed the DeepIM method, which extracts suitable information by mapping the seed set to the latent space and selects nodes with greater influence by combining the local and global structural features of bridge nodes. Subsequently, on this basis, network embedding and deep reinforcement learning were further used to solve the IM problem, achieving good results. Graph Attention Networks (GAT), as an important branch of GNN, can adaptively learn the influence weights of different neighbor nodes on the target node because it introduces the attention mechanism in the process of information propagation, further improving the expressive power of the model. In the GAT model, each node is assigned different weights according to the importance of neighbor nodes, so as to better capture the influence propagation characteristics between nodes. This mechanism fits well with the core challenge in the influence maximization problem, as it is able to evaluate the influence of nodes based on the social network structure and reduce the reliance on artificial feature selection. Summary of the invention

[0005] In view of the above problems, the present invention aims to propose a method, system, device and storage medium for maximizing influence in a social network with better scalability and flexibility.

[0006] The technical solution of the present invention is: a method for maximizing influence in a social network according to the present invention, the operation steps of which include:

[0007] Step (1): By taking into account the structural similarity characteristics of nodes in the network, the struc2vec graph embedding method is used to obtain the embedding representation of each node in the training network as the feature representation vector of the node. The training network is a scale-free network generated by the Barabási-Albert method (hereinafter referred to as the BA method);

[0008] Step (2): Take each node as a seed node set and simulate the propagation of information through the linear threshold (LT) diffusion model. The number of nodes affected by a single node after propagation is considered as the influence range of each node and serves as the node label.

[0009] Step (3): Input the feature representation vectors obtained by the struc2vec graph embedding method and the labels obtained by the LT linear threshold diffusion model into the graph attention network (GAT) for feature reprocessing and training and optimizing related parameters to obtain the optimal model, and then use the optimal model to predict the influence of the real network;

[0010] Step (4): Considering the topological position of the nodes, the k-shell value and degree centrality index are introduced to reset a final influence score for each node, and the top k nodes with high influence scores are selected as the seed node set. The seed node set composed of these k seed nodes can maximize the influence propagation.

[0011] Furthermore, in step (1), the generation of the training network is specifically to generate multiple scale-free networks by the BA method; the BA network is generated by preferential connection, and its generation process can be expressed as:

[0012]

[0013] Where P(i) represents the probability of a new node connecting to an existing node i, k i is the degree of node i, ∑ j k j is the sum of all existing node degrees.

[0014] Furthermore, in step (1), the specific steps of the graph embedding process are as follows:

[0015] Define G = (V, E) to represent a social network; where V = {v1, v2, …, v n} represents the set of nodes in the network, v n represents the nth node, and the set E represents the set of relationship edges between nodes in the network;

[0016] By using struc2vec to generate features on the network nodes, a feature vector of size 128 is obtained for each node in the network; let S be the node embedding generated by the struc2vec method of graph G with dimension d, and the obtained embedding is as follows:

[0017]

[0018] is the embedding set of all nodes in the network, so the embedding of node v can be expressed as:

[0019]

[0020] Among them, z vrepresents the generated embedding for each node v in the network; therefore, for each node in the network, the model obtains a feature vector.

[0021] Furthermore, in step (2), the method for obtaining the node label is as follows:

[0022] Use the LT information diffusion model to calculate the influence range of each node in the training network;

[0023] The LT model is a mathematical model used to simulate the spread of information, behavior or influence in social networks. Each node decides whether to be activated based on the influence of its neighbor nodes. The influence of neighbors is represented by the weight of the edge, and the sum does not exceed 1. The node has a preset threshold θ v , represents the critical value of the node v being activated, θ v Draw from uniform distribution U(0,1); each edge E(u,v) has a weight w uv , represents the influence of node on node u, and the total influence of node u’s neighbors on it can be expressed as:

[0024]

[0025] If the total influence exceeds the node threshold θ v ,Right now:

[0026]

[0027] Then node v is activated, and when the sum of the influence of its activated neighbors exceeds the threshold, the node will be activated to become the next propagation node; this process will be repeated until no new nodes are activated; finally, the influence range of each node I(v) refers to the number of all nodes it can activate, and the label L(v) is expressed as:

[0028] L(v)=I(v) (6)

[0029] The LT model can simulate phenomena in social networks, such as new product promotion, technology diffusion, and the spread of social behavior. It can be used to optimize marketing strategies, help companies identify and activate key influencers, maximize the spread of information on social platforms, and evaluate the impact of different network structures on information diffusion.

[0030] The calculated influence range forms the label used to train the regression task. For the selected information diffusion model (SDM), the label set (label) represents the calculated influence of each node in the graph (G) and can be expressed as:

[0031] label G =SDM(G(V,E)) (7)

[0032] Among them, V represents the number of nodes in the graph G, E represents the number of edges in the graph G, and the labels calculated by the LT model measure the influence of each node.

[0033] Furthermore, in step (3), the feature reprocessing method is specifically to input the feature representation vector obtained by the struc2vec graph embedding method and the label obtained by the LT linear threshold diffusion model of the node in the training network into the graph attention network (GAT) for feature reprocessing and training optimization of related parameters to obtain the optimal model, and then use the optimal model to predict the influence of the real network. The feature reprocessing method is as follows:

[0034] Input feature matrix: The 128-dimensional node feature vector generated by struc2vec is used as the input of the GAT model; assuming that the node feature matrix is ​​X, the size is N×128, where N is the number of nodes;

[0035] Calculate the attention coefficient: For a given node v, its embedding vector h v and the embedding vector h of the neighbor node v u The corresponding attention coefficient e will be calculated through the attention mechanism uv ; First, the correlation between node v and its neighbor u is calculated using the attention mechanism:

[0036] e ij =LeakyReLU(a T [Wh i ×Wh j ]) (8)

[0037] Among them, h v and h u are the feature vectors of v and u respectively; W is a trainable weight matrix used to map the node feature vector to the new feature space; α is a parameter vector used to calculate the attention coefficient; [] represents a vector concatenation operation, which concatenates the feature vectors of two nodes into one vector; LeakyReLU is the activation function;

[0038] Normalized attention coefficient: To ensure that the attention coefficient e uv The sum between all neighbor nodes is 1, which needs to be normalized by the Softmax function:

[0039]

[0040] Among them, Γ v Represents the set of neighbor nodes of node v;

[0041] Feature aggregation: Use the normalized attention coefficient to perform weighted summation on the features of neighboring nodes to obtain a new feature representation for each node:

[0042]

[0043] Where σ is the activation function;

[0044] Furthermore, the influence prediction described in step (3) is specifically as follows:

[0045] Input the feature vectors and labels of the nodes in the training network into the GAT model, and train and optimize the relevant parameters to obtain the optimal model. The steps are as follows:

[0046] Define the loss function: Calculate the loss function using the true label (propagation range) and predicted value of the node; for regression tasks, the loss function used is the mean square error (MSE):

[0047]

[0048] Where N is the number of nodes, y i is the true label of the node (propagation range), is the predicted value of the node;

[0049] Back propagation and parameter update: The gradient is calculated through the back propagation algorithm, and the model parameters (attention coefficient parameters and linear transformation matrix) are updated using the optimization algorithm (gradient descent) to minimize the loss function:

[0050]

[0051] Where η is the learning rate;

[0052] Iterative training: Repeat the above steps (forward propagation, loss calculation, back propagation and parameter update) until the optimal model is reached; then, the real network is passed through step (1) to obtain the embedded vector representation of the node, which is input into the regressor of the graph attention network (GAT) to predict the influence weight of the node in the real network.

[0053] Furthermore, the step (4) is specifically as follows: considering the topological position of the node in the network graph, the k-shell algorithm is introduced to perform shell decomposition operation on the network graph, and the k-shell value of each node v is obtained and recorded as K v , while paying attention to the structural characteristics of the node itself, the degree of node v is calculated and recorded as D v , the final impact score IF v The calculation of is as follows:

[0054]

[0055] Furthermore, the present invention also provides a social network influence maximization system, comprising:

[0056] The node embedding module is used to embed all nodes in the social network. By considering the structural similarity between nodes in the network, building a context graph (multi-layer weighted graph) and setting the random walk length to generate context for the node, the feature vector of the node is obtained.

[0057] The information diffusion module is used to set each node as a seed node set for information propagation, and obtain the influence range of each node through the linear threshold (LT) diffusion model as a label for supervised learning;

[0058] The optimization module is used to learn node feature representation in a supervised environment and evaluate the learning process using a loss function. The output representation is optimized through a graph-based loss function and the parameters are adjusted to optimize the learning efficiency of the model.

[0059] The prediction module uses the optimal model obtained after training and optimizing parameters to predict the influence of real network nodes, and then determines the seed nodes based on the final influence scores of the updated nodes.

[0060] Furthermore, the present invention also provides a computing device, comprising: one or more processors, one or more memories, and one or more programs, wherein the programs are stored in the memories and are configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the method for maximizing influence in a social network according to any of the above items are implemented.

[0061] Furthermore, the present invention also provides a storage medium, characterized in that the storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the steps of the method for maximizing influence in a social network according to any of the above items.

[0062] The beneficial effects of the present invention are as follows: the present invention uses struc2vec graph embedding to map the nodes in the graph to a low-dimensional space, so that the learned node representation can retain the structure and attribute information in the original graph as much as possible, and can capture the structural similarity between the nodes; the present invention uses the LT information diffusion model to obtain the influence range of the node as the label of the node, so that supervised learning is more suitable for the influence maximization task; the present invention uses the BA algorithm to generate a number of undirected scaling graphs for training, thereby improving the generalization of the model for the influence maximization task of networks of different scales; based on the topological structure of the nodes in the graph, the present invention introduces the k-shell algorithm and the degree index, reassigns influence scores to the nodes, reselects the seed nodes, reduces the complexity of the model, and significantly reduces the time of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is an algorithm flow chart of an embodiment of the present invention;

[0064] Figure 2 It is a schematic diagram of the overall framework of the embodiment of the present invention;

[0065] Figure 3 is a schematic diagram of the BA algorithm used in an embodiment of the present invention;

[0066] Figure 4 It is a diagram showing the effect of data propagation by the method in the embodiment of the present invention and other methods. DETAILED DESCRIPTION

[0067] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0068] As shown in the figure, a method for maximizing influence in a social network according to the present invention includes:

[0069] Step (1): By taking into account the structural similarity characteristics of nodes in the network, the struc2vec graph embedding method is used to obtain the embedding representation of each node in the training network as the feature representation vector of the node. The training network is a scale-free network generated by the Barabási-Albert method;

[0070] Step (2): Take each node as a seed node set and simulate the propagation of information through the linear threshold (LT) diffusion model. The number of nodes affected by a single node after propagation is considered as the influence range of each node and serves as the node label.

[0071] Step (3): Input the feature representation vectors obtained by the struc2vec graph embedding method and the labels obtained by the LT linear threshold diffusion model into the graph attention network (GAT) for feature reprocessing and training and optimizing related parameters to obtain the optimal model, and then use the optimal model to predict the influence of the real network;

[0072] Step (4): Considering the topological position of the nodes, the k-shell value and degree centrality index are introduced to reset a final influence score for each node, and the top k nodes with high influence scores are selected as the seed node set, which is believed to maximize the influence propagation.

[0073] Specifically, in step (1): by taking into account the structural similarity characteristics of nodes in the network, the struc2vec graph embedding method is used to obtain the embedded representation of each node in the training network as the feature representation vector of the node. The training network is a scale-free network generated by the Barabási-Albert method; Figure 3As shown, the BA method generates by preferential connection:

[0074]

[0075] Among them, P(i) represents the probability of a new node connecting to an existing node i, k i is the degree of node i, Σ j k j is the sum of all existing node degrees.

[0076] Traditional graph embedding algorithms (such as DeepWalk, node2vec, etc.) are mainly based on local neighborhood information between nodes for embedding. They tend to rely on the direct connection relationship between nodes and pay less attention to the structural role of nodes. This leads to poor performance in capturing global structural similarities, especially when topologically unconnected but structurally similar nodes cannot be embedded close together, the effectiveness of such methods is limited.

[0077] Struc2vec is optimized to address this limitation, focusing on capturing the structural similarity of nodes, that is, even if the nodes are not connected, if they play similar topological roles in the graph (such as similar degrees, connection patterns), they can still be mapped to similar positions in a low-dimensional space; the advantage of struc2vec is that it compares the local structural features of nodes at multiple scales, and builds a hierarchical representation that reflects structural similarity layer by layer, so that the algorithm retains topological information globally; compared with embedding methods based only on neighborhood information, struc2vec can better handle nodes in different regions but with similar structures, thereby enhancing the model's ability to find similar nodes in large and complex networks; at the same time, struc2vec uses weighted random walks in the complete graph to ensure that capturing structural similarity is not restricted by direct connections in the network, further improving the embedding quality.

[0078] Let G = (V, E) represent a social network; where V = {v1, v2, …, v n} represents the set of nodes in the network, and set E represents the set of relationship edges between nodes in the network;

[0079] First, by using struc2vec to generate features on the network nodes, a feature vector of size 128 is obtained for each node in the network; let S be the node embedding generated by the struc2vec method of graph G with dimension d; the obtained embedding is as follows:

[0080] Z=struc2vec(G,d) (2)

[0081] is the embedding set of all hosts in the network, so the embedding of node v can be expressed as:

[0082] z v= Z[v],v∈V (3)

[0083] Among them, z v represents the generated embedding for each node in the network, so for each node in the network, the model obtains a feature vector.

[0084] In step (2), each node is taken as a seed node set and the LT linear threshold diffusion model is used to simulate the propagation of information. The number of nodes affected by a single node after propagation is considered to be the influence range of each node and serves as the label of the node.

[0085] In order to better apply supervised learning to solve the problem of maximizing influence in social networks, the linear threshold (LT) diffusion model is used to calculate the influence range of each node in the training network; the LT model is a mathematical model used to simulate the spread of information, behavior or influence in social networks; each node decides whether to be activated based on the influence of its neighbor nodes, and the influence of the neighbors is represented by the weight of the edge, and the sum does not exceed 1; the node has a preset threshold θ v , represents the critical value of the node v being activated, θ v Draw from uniform distribution U(0,1); each edge E(u,v) has a weight w uv , represents the influence of node u on node u, and the total influence of node v’s neighbors on it can be expressed as:

[0086]

[0087] If the total influence exceeds the node threshold θ v ,Right now:

[0088]

[0089] Then node v is activated, and when the sum of the influence of its activated neighbors exceeds the threshold, the node will be activated to become the next propagation node; this process will be repeated until no new nodes are activated; finally, the influence range of each node refers to the number of all nodes it can activate, and the label is expressed as:

[0090] L(v)=I(v) (6)

[0091] The LT model can simulate phenomena in social networks, such as new product promotion, technology diffusion, and the spread of social behavior; it can be used to optimize marketing strategies, help companies identify and activate key influencers, maximize the spread of information on social platforms, and evaluate the impact of different network structures on information diffusion;

[0092] The calculated influence range forms the label used to train the regression task. For the selected information diffusion model (SDM), the label set (label) represents the calculated influence of each node in the graph (G) and can be expressed as:

[0093] label G =SDM(G(V,E)) (7)

[0094] Among them, V represents the number of nodes in the graph G, E represents the number of edges in the graph G, and the labels calculated by the LT model measure the influence of each node.

[0095] In step (3), the feature representation vectors obtained by the struc2vec graph embedding method and the labels obtained by the LT linear threshold diffusion model are input into the graph attention network (GAT) for feature reprocessing and training and optimizing related parameters to obtain the optimal model, and then the optimal model is used to predict the influence of the real network;

[0096] First, the feature matrix of the input network is used: the 128-dimensional node feature vector generated by struc2vec is used as the input of the GAT model. Assume that the node feature matrix is ​​X, and the size is N×128, where N is the number of nodes; then the attention coefficient is calculated: for a given node v, its embedding vector h v and the embedding vector h of the neighbor node v u The corresponding attention coefficient e will be calculated through the attention mechanism uv ; First, the correlation between node v and its neighbor u is calculated using the attention mechanism:

[0097] e ij =LeakyReLU(a T [Wh i ×Wh j ]) (8)

[0098] Among them, h v and h u are the feature vectors of v and u respectively, W is a trainable weight matrix used to map the node feature vector to the new feature space, and α is a parameter vector used to calculate the attention coefficient; Represents the vector concatenation operation, which concatenates the feature vectors of two nodes into one vector. LeakyReLU is the activation function.

[0099] Normalized attention coefficient: To ensure that the attention coefficient e uv The sum between all neighbor nodes is 1, which needs to be normalized by the Softmax function:

[0100]

[0101] Among them, Γ v Represents the set of neighbor nodes of node v;

[0102] Feature aggregation: Use the normalized attention coefficient to perform weighted summation on the features of neighboring nodes to obtain a new feature representation for each node:

[0103]

[0104] Among them, σ is the activation function step; followed by the influence prediction step, the feature vectors and labels of the nodes in the training network are input into the GAT model, and the training and optimization of relevant parameters are performed to obtain the optimal model. The steps are as follows:

[0105] Define the loss function: Calculate the loss function using the true label (propagation range) and predicted value of the node; for regression tasks, the loss function used in this article is the mean square error (MSE):

[0106]

[0107] Where N is the number of nodes, y i is the true label of the node (propagation range), is the predicted value of the node;

[0108] Back propagation and parameter update: The gradient is calculated through the back propagation algorithm, and the model parameters (attention coefficient parameters and linear transformation matrix) are updated using the optimization algorithm (gradient descent) to minimize the loss function:

[0109]

[0110] Where η is the learning rate;

[0111] Iterative training: Repeat the above steps (forward propagation, loss calculation, back propagation and parameter update) until the optimal model is reached; then, the real network is passed through step (1) to obtain the embedded vector representation of the node, which is input into the regressor of the graph attention network (GAT) to predict the influence weight W of the node in the real network. v .

[0112] In step (4), considering the topological position of the node in the network graph, the k-shell algorithm is introduced to perform shell decomposition on the network graph, and the k-shell value of each node v is recorded as K v , while paying attention to the structural characteristics of the node itself, the degree of node v is calculated and recorded as D v , the final impact score IF v The calculation of is as follows:

[0113]

[0114] Furthermore, this embodiment also provides a system for maximizing influence in a social network, including:

[0115] The node embedding module is used to embed all nodes in the social network. By considering the structural similarity between nodes in the network, building a context graph (multi-layer weighted graph) and setting the random walk length to generate context for the node, the feature vector of the node is obtained.

[0116] The information diffusion module is used to set each node as a seed node set for information propagation, and obtain the influence range of each node through the linear threshold (LT) diffusion model as a label for supervised learning;

[0117] The optimization module is used to learn node feature representation in a supervised environment and evaluate the learning process using a loss function. The output representation is optimized through a graph-based loss function and the parameters are adjusted to optimize the learning efficiency of the model.

[0118] The prediction module uses the optimal model obtained after training and optimizing parameters to predict the influence of real network nodes, and then determines the seed nodes based on the final influence scores of the updated nodes.

[0119] In this embodiment, the above-mentioned influence maximization method Str GAT is used to experiment on social networks. The specific data sets are all public data sets. The statistical information of the data sets is shown in Table 1, where |V| and |E| are the number of nodes and the number of edges in the data set, respectively, <k> represents the average degree value, and k m It represents the maximum value of the node degree in the data set, and C represents the clustering coefficient of the data set.

[0120] Table 1 Social network data

[0121]

[0122] The influence maximization method Str GAT in this embodiment captures the structural similarity and the degree of aggregated local influence, and processes the above social network data with other methods (degree centrality, betweenness centrality, eigenvector centrality, CELF), and the results are shown in Table 2 and Figure 4 shown.

[0123] Table 2 Comparison of the influence maximization method StrGAT in this embodiment with other methods

[0124]

[0125] Table 2 and Figure 4It shows that the influence maximization method StrGAT in this embodiment captures the structural similarity and the degree of aggregated local influence. Compared with other methods (degree centrality, betweenness centrality, eigenvector centrality, CELF), the influence maximization method in this embodiment is significantly better than other methods.

[0126] Furthermore, a computing device includes: one or more processors, one or more memories, and one or more programs, wherein the programs are stored in the memories and are configured to be executed by the processors, and when the programs are loaded into the processors, the steps of any one of the methods for maximizing influence in a social network are implemented.

[0127] Furthermore, this embodiment also provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the steps of any one of the methods for maximizing influence in a social network.

[0128] Table 3 is a variable table involved in the present invention. The present invention can be combined with a computer system to achieve maximum influence in the network.

[0129] Table 3: Variable table

[0130]

[0131]

[0132] The present invention uses struc2vec graph embedding to map the nodes in the graph to a low-dimensional space, so that the learned node representation can retain the structure and attribute information in the original graph as much as possible, and can capture the structural similarity between nodes; the present invention uses the LT information diffusion model to obtain the influence range of the node as the label of the node, so that supervised learning is more suitable for the influence maximization task; the present invention uses the BA algorithm to generate a number of undirected scaling graphs for training, which improves the generalization of the model for the influence maximization task of networks of different scales; based on the topological structure of the nodes in the graph, the present invention introduces the k-shell algorithm and the degree index, re-assigns influence scores to the nodes, re-selects the seed nodes, reduces the complexity of the model, and significantly reduces the time of model training.

Claims

1. A method for maximizing influence in a social network, characterized in that: The operation steps are as follows: Step (1): By taking into account the structural similarity characteristics of nodes in the network, the struc2vec graph embedding method is used to obtain the embedding representation of each node in the training network as the feature representation vector of the node. The training network is a scale-free network generated by the BA method; Step (2): Take each node as a seed node set and simulate the propagation of information through the LT diffusion model. The number of nodes affected by a single node after propagation is considered as the influence range of each node and serves as the label of the node. Step (3): Input the feature representation vectors obtained by the struc2vec graph embedding method and the labels obtained by the LT diffusion model into GAT for feature reprocessing and training related parameters to obtain the optimal model, and then use the optimal model to predict the influence of the real network; Step (4): Considering the topological position of the nodes, the k-shell value and degree centrality index are introduced to reset a final influence score for each node, and the top k nodes with high influence scores are selected as the seed node set, which is believed to maximize the influence propagation.

2. A method for maximizing influence in a social network according to claim 1, characterized in that: In step (1), the generation of the training network is specifically to generate multiple scale-free networks by using the BA method; The BA network is generated by preferential connection, and its generation process is expressed as: In the formula, P(i) represents the probability of a new node connecting to an existing node i, k i is the degree of node i, ∑ j k j is the sum of all existing node degrees.

3. A method for maximizing influence in a social network according to claim 1, characterized in that: In step (1), the specific steps of the struc2vec graph embedding are as follows: define G = (V, E) to represent a social network; Where V = {v1,v2,…,v n } represents the set of nodes in the network, v n represents the nth node, and the set E represents the set of relationship edges between nodes in the network; By using struc2vec to generate features on the network nodes, a feature vector of size 128 is obtained for each node in the network; Let S be the node embeddings generated by the struc2vec method of graph G with dimension d. The resulting embeddings are as follows: Z = struc2vec(G,d)(2) Z is the embedding set of all receptions in the network, so the embedding of node v is expressed as: with v =Z[v],v∈V (3) In the formula, z v represents the generated embedding of each node in the network; therefore, for each node in the network, the model obtains a feature vector.

4. A method for maximizing influence in a social network according to claim 1, characterized in that: In step (2), the specific steps of obtaining the node label are as follows: using the LT information diffusion model to calculate the influence range of each node in the training network; wherein the LT model is a mathematical model used to simulate the spread of information, behavior or influence in a social network; Each node decides whether to be activated based on the influence of its neighbor nodes. The influence of neighbors is represented by the weight of the edge, and the sum does not exceed 1; The node has a preset threshold θ v , represents the critical value of the node v being activated, θ v Draw from uniform distribution U(0,1), each edge E(u,v) has a weight w uv , represents the influence of node u on node u, and the total influence of node v’s neighbors on it is expressed as: If the total influence exceeds the node threshold θ v ,Right now: Then node v is activated, and when the sum of the influence of its activated neighbors exceeds the threshold, the node is activated to become the next propagation node; this process is repeated until no new nodes are activated; finally, the influence range I(v) of each node refers to the number of all nodes it can activate, and the label L(v) is expressed as: L(v)=I(v)(6) The LT model can simulate phenomena in social networks and is used to optimize marketing strategies, help companies identify and activate key influencers, maximize the effect of information dissemination on social platforms, and evaluate the impact of different network structures on information diffusion. The calculated influence range forms a label for training regression tasks. For the selected information diffusion model SDM, label represents the calculated influence of each node in the graph (G) as: lable G =SDM(G(V,E))(7) In the formula, V represents the number of nodes in graph G, E represents the number of edges in graph G, and the labels calculated by the LT model measure the influence of each node.

5. A method for maximizing influence in a social network according to claim 1, characterized in that: In step (3), the method for reprocessing features is as follows: Input feature matrix: The 128-dimensional node feature vector generated by struc2vec is used as the input of the GAT model; assuming that the node feature matrix is ​​X, the size is N×128, where N is the number of nodes; Calculate the attention coefficient: For a given node v, its embedding vector h v and the embedding vector h of the neighbor node v u The corresponding attention coefficient e will be calculated through the attention mechanism uv ; First, the correlation between node v and its neighbors u is calculated using the attention mechanism: e ij =LeakyReLU(aT[Wh i ×Wh j ]) (8) In the formula, h v and h u are the feature vectors of v and u respectively, W is a trainable weight matrix used to map the node feature vector to the new feature space, α is a parameter vector used to calculate the attention coefficient, [] represents the vector concatenation operation, which concatenates the feature vectors of two nodes into one vector, and LeakyReLU is the activation function; Normalized attention coefficient: To ensure the attention coefficient e uv The sum between all neighbor nodes is 1, which needs to be normalized by the Softmax function: In the formula, Γ v Represents the set of neighbor nodes of node v; Feature aggregation: Use the normalized attention coefficient to perform weighted summation on the features of neighboring nodes to obtain a new feature representation for each node: Where σ is the activation function.

6. A method for maximizing influence in a social network according to claim 1, characterized in that: In step (3), the influence prediction method is as follows: input the feature vectors and labels of the nodes in the training network into the GAT model, and train and optimize the relevant parameters to obtain the optimal model, wherein the steps are as follows: Define the loss function: Calculate the loss function using the true label and predicted value of the node; for regression tasks, the loss function used is the mean square error: Where N is the number of nodes, y i is the true label of the node, is the predicted value of the node; Backpropagation and parameter update: Gradients are calculated through the backpropagation algorithm and the model parameters are updated using an optimization algorithm to minimize the loss function: Where η is the learning rate; Iterative training: Repeat the above steps until the optimal model is achieved; then, the embedding vector representation of the node in the real network obtained through step (1) is input into the GAT regressor to predict the influence weight W of the node in the real network. v .

7. A method for maximizing influence in a social network according to claim 1, characterized in that: The specific meaning of step (4) is: considering the topological position of the node in the network graph, the k-shell algorithm is introduced to perform shell decomposition operation on the network graph, and the k-shell value of each node v is recorded as K v , while paying attention to the structural characteristics of the node itself, the degree of node v is calculated and recorded as D v , the final impact score IF v The calculation of is as follows:

8. A negative information dissemination suppression system, characterized in that: include: The node embedding module is used to embed all nodes in the social network. By considering the structural similarity between nodes in the network, building a context graph and setting the random walk length to generate context for the node, the feature vector of the node is obtained. The information diffusion module is used to set each node as a seed node set for information propagation, and obtain the influence range of each node through the LT diffusion model as a label for supervised learning; The optimization module is used to learn node feature representation in a supervised environment and evaluate the learning process using a loss function. The output representation is optimized through a graph-based loss function and the parameters are adjusted to optimize the learning efficiency of the model. The prediction module uses the optimal model obtained after training and optimizing parameters to predict the influence of real network nodes, and then determines the seed nodes based on the final influence scores of the updated nodes.

9. A computing device, characterized in that include: One or more processors, one or more memories, and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the method for maximizing influence in a social network according to any one of claims 1 to 7 are implemented.

10. A storage medium, characterized in that: The storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the method for maximizing influence in a social network according to any one of claims 1 to 7.

Images