Network node influence prediction model determination method, application method and related device

Through the graph active learning framework and deep learning model based on node relative entropy, the accuracy and efficiency of node impact prediction in complex networks are solved, and accurate prediction and efficient evaluation of key node impact influence are achieved.

CN118945065AInactive Publication Date: 2024-11-12BEIHANG UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411170555.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

It is difficult for the prior art to accurately identify and predict key nodes with strong propagation effects in large-scale networks, especially in complex networks. Traditional methods cannot effectively sort and quantitatively estimate node influence, and the simulation methods have high time complexity.

Method used

Adopting a graph active learning framework based on node relative entropy, through the establishment of similarity network and seepage attack, a representative node data collection is constructed, and a graph embedded neural network and multi-layer perceptron regression model is used to optimize the pre-training model, and the network node influence is obtained. Force prediction model.

Benefits of technology

Accurate prediction of node influence in different types of networks is achieved, the computational amount required for simulation is reduced, generalization ability is improved, and the key nodes with high influence are able to quickly identify high-impact and quantitatively predict their impact expectations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118945065A_ABST
    Figure CN118945065A_ABST
Patent Text Reader

Abstract

The invention discloses a network node influence prediction model determination method, an application method and a related device, and relates to the field of influence prediction.The network node influence prediction model determination method comprises the steps that a similarity network of real network nodes is established based on node relative entropy, seepage attack is conducted on the similarity network, and the network node influence prediction model is obtained; establishing a representative node data set; the representative node data set comprises node features of representative nodes, a network topology structure of a real network where the representative nodes are located and influence expectations corresponding to the representative nodes; and network parameters of the pre-training model are optimized through the established representative node data set, so that a network node influence prediction model can be obtained. According to the method, the network node influence can be accurately predicted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of influence prediction, and in particular to a method for determining a network node influence prediction model, an application method and related devices. Background Art

[0002] Changes in local characteristics in the network may spread to part or even the entire network. These sudden impacts may lead to explosive disasters, such as traffic jams, rumor spread, power grid collapse, disease outbreaks, etc. For example, overloaded transmission lines may trigger cascading failures in the power grid, leading to large-scale power outages and huge economic losses. If the influential nodes in the network can be identified in advance, fault warnings can be achieved by monitoring these key nodes, and network optimization can be achieved by adjusting the structure of these nodes. On the other hand, people also use this explosive influence for marketing and information dissemination. In order to mitigate the impact of cascading failures or optimize the impact of opinion propagation, it is crucial to identify key nodes with significant propagation influence in large-scale networks. However, accurately identifying these key structures with strong propagation influence remains challenging.

[0003] Identifying influential nodes requires an accurate assessment of the influence of individual nodes and their influence ranking among all nodes. Currently, the main methods for evaluating node influence are network structure analysis and dynamic simulation. Network structure analysis methods usually use global metrics, local metrics, or semi-local metrics to calculate the structural centrality of nodes and rank the influence of nodes according to their structural centrality. However, these methods may miss those nodes with greater influence, which are embedded with elusive structural features. Although dynamic simulation methods can accurately evaluate the influence of nodes, the computational time to identify key nodes in large-scale networks is immeasurable due to the probabilistic nature of the propagation model. Summary of the invention

[0004] The purpose of the present invention is to provide a method for determining a network node influence prediction model, an application method and related devices, which can accurately predict the influence of network nodes.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] In a first aspect, the present invention provides a method for determining a network node influence prediction model, the method comprising:

[0007] A similarity network of real network nodes is established based on node relative entropy; the node relative entropy is an indicator for measuring the structural similarity between nodes.

[0008] A percolation attack is performed on the similarity network to establish a representative node data set; the percolation attack is used to find a representative node that is similar to the most nodes based on the similarity between nodes; the representative node data set includes: node characteristics of the representative node, the network topology structure of the real network where the representative node is located, and the influence expectation corresponding to the representative node.

[0009] The node features of the representative nodes and the network topology structure of the real network where the representative nodes are located are input into the pre-training model to obtain the output of the pre-training model; the pre-training model is a deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model; the output of the pre-training model is the influence expectation corresponding to the predicted representative node.

[0010] A first loss value is determined according to the expected influence corresponding to the representative node, the output of the pre-trained model and the determined first loss function.

[0011] The network parameters of the pre-trained model are optimized according to the first loss value to obtain a network node influence prediction model.

[0012] In a second aspect, the present invention provides a network node influence prediction model application method, the network node influence prediction model application method comprising:

[0013] Get the node to be predicted.

[0014] According to the correlation between the node influence and the node index of the specified random diffusion model, the node characteristics of the node to be predicted are determined.

[0015] The node characteristics of the node to be predicted and the network topology structure of the real network where the node to be predicted is located are input into the network node influence prediction model to obtain the output of the network node influence prediction model; the network node influence prediction model is a model trained based on the network node influence prediction model determination method described in any one of claims 1-6; the output of the network node influence prediction model is the expected influence corresponding to the predicted node to be predicted.

[0016] In a third aspect, the present invention provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for determining a network node influence prediction model or the above-described method for applying a network node influence prediction model.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for determining a network node influence prediction model or the above-mentioned method for applying a network node influence prediction model.

[0018] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned methods for determining a network node influence prediction model or the above-mentioned method for applying a network node influence prediction model.

[0019] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0020] The present invention provides a method for determining a network node influence prediction model, an application method and related devices, by establishing a similarity network of real network nodes based on node relative entropy, and performing a percolation attack on the similarity network to establish a representative node data set; the representative node data set includes: node characteristics of representative nodes, network topology of the real network where the representative nodes are located, and influence expectations corresponding to the representative nodes; thereby optimizing the network parameters of the pre-trained model through the established representative node data set, and obtaining a network node influence prediction model. The active learning process adopted by the present invention only requires a small amount of simulation, and can produce excellent generalization capabilities in various types of networks (such as social networks, biological networks, and infrastructure networks), quantitatively predict the influence of nodes on different networks, and solve the problem that it is difficult to achieve accurate sorting and quantitative estimation of different network influences using traditional network influence analysis methods, and the time complexity of the simulation method is too high. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 This is an application environment diagram of a method for determining a network node influence prediction model in one embodiment of the present invention.

[0023] Figure 2 A flowchart of a method for determining a network node influence prediction model provided by an embodiment of the present invention.

[0024] Figure 3 A flowchart of a method for applying a network node influence prediction model provided by one embodiment of the present invention.

[0025] Figure 4 A flowchart of an influence prediction method for realizing accurate prediction of a personalized structure of a complex network is provided in accordance with an embodiment of the present invention.

[0026] Figure 5 A schematic diagram of the structure of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] With the development of deep learning, it has gradually been applied in the field of complex networks. Compared with traditional network analysis methods, deep learning methods are not only good at automatically extracting complex and abstract features from large amounts of data, but also can capture subtle patterns and relationships in network structures. They can also adapt to different network structures and dynamic changes through training. This makes them perform well in rapidly changing network environments such as social network analysis and Internet topology identification, while also having strong generalization capabilities.

[0030] The present invention mainly proposes a graph entropy-based graph active learning framework as a solution to the problem of how to accurately evaluate the influence of multiple network nodes. First, a graph neural network is used as the underlying predictor to effectively establish a mapping relationship between node features and influence, and then the personalized structural features of the target network (a representative subset of nodes) are sampled through an active learning strategy based on graph entropy. Finally, the simulated influence of the representative node fine-tunes the graph neural network predictor to achieve accurate impact assessment. The active learning process used in the present invention only requires a small amount of simulation to produce excellent generalization capabilities in various types of networks (such as social networks, biological networks, and infrastructure networks), and quantitatively predict the influence of nodes on different networks, solving the problem that it is difficult to use traditional network influence analysis methods to accurately sort and quantitatively estimate the influence of different networks, as well as the high time complexity of simulation methods.

[0031] Since different networks have the characteristics of personalized structure and differentiated communication, traditional influence evaluation indicators and methods are more about ranking the influence of network nodes from a single network indicator and a single dimension, ignoring the personalized structure of different networks and the multi-dimensional roles of nodes in the network, such as local, semi-local and global. At the same time, the influence value of the node cannot be accurately given according to the communication model, so the influence of the network node cannot be well evaluated. Therefore, the present invention constructs a quantitative evaluation method that can effectively realize the accurate and efficient influence of network nodes. The influence prediction method consists of a graph neural network module, an active learning module and an influence simulation module. It mainly constructs a transferable model that mines the personalized structural information of the network through active learning to realize the accurate evaluation of the influence of the real network.

[0032] In view of the defects existing in the prior art, the purpose of the present invention is to provide a network node influence prediction method based on graph entropy active learning. In view of the inconsistent distribution of network data, the active learning method based on graph entropy proposed in the present invention combines graph neural network with active learning method, and adopts the indicator of "node relative entropy" in complex network as the evaluation indicator of node similarity. By establishing a node similarity network and analyzing its percolation process, an active learning training node set is selected, which contains both high-information nodes and difficult nodes. The pre-trained model is fine-tuned using the data obtained from the node simulation of the set to obtain a model that can accurately predict the influence of the real network.

[0033] The network node influence prediction model determination method provided by the embodiment of the present invention can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send a representative node data set to the server 104. After the server 104 receives the representative node data set, for the representative node data set, the server 104 inputs the node characteristics of the representative node and the network topology structure of the real network where the representative node is located into the pre-training model to obtain the output of the pre-training model; the pre-training model is a deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model; the output of the pre-training model is the expected influence corresponding to the predicted representative node; according to the expected influence corresponding to the representative node, the output of the pre-training model and the determined first loss function, the first loss value is determined; according to the first loss value, the network parameters of the pre-training model are optimized to obtain a network node influence prediction model. The server 104 can feed back the obtained network node influence prediction model to the terminal 102. In addition, in some embodiments, the network node influence prediction model determination method can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly train the network node influence prediction model for the representative node data set, or the server 104 can obtain the representative node data set from the data storage system and train the network node influence prediction model for the representative node data set.

[0034] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones and tablet computers, IoT devices and portable wearable devices. IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0035] In an exemplary embodiment, Figure 2 As shown, a method for determining a network node influence prediction model is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In an embodiment of the present invention, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps A1 to A5. Among them:

[0036] A1: A similarity network of real network nodes is established based on node relative entropy; the node relative entropy is an indicator for measuring the structural similarity between nodes.

[0037] A2: Perform a percolation attack on the similarity network to establish a representative node data set; the percolation attack is used to find representative nodes that are similar to the most nodes based on the similarity between nodes; the representative node data set includes: node characteristics of the representative nodes, the network topology structure of the real network where the representative nodes are located, and the influence expectations corresponding to the representative nodes.

[0038] A3: Input the node features of the representative nodes and the network topology structure of the real network where the representative nodes are located into the pre-training model to obtain the output of the pre-training model; the pre-training model is a deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model; the output of the pre-training model is the expected influence corresponding to the predicted representative nodes.

[0039] A4: Determine a first loss value according to the expected influence corresponding to the representative node, the output of the pre-trained model and the determined first loss function.

[0040] A5: Optimize the network parameters of the pre-trained model according to the first loss value to obtain a network node influence prediction model.

[0041] Among them, training is based on the gradient descent method to obtain a network node influence prediction model with low prediction loss. The prediction loss is calculated by the mean square error between the influence predicted by the network node influence prediction model and the label.

[0042] Implementing the above steps A1 to A5 solves the shortcomings of traditional influence prediction methods, such as inaccurate node influence ranking and inefficient quantitative prediction of node influence expectations, and realizes personalized prediction of the network based on graph entropy theory. This influence prediction method can quickly identify key nodes with high influence and quantitatively predict their influence expectations, and has good practical application value.

[0043] In another exemplary embodiment of the present invention, in step A1, a similarity network of real network nodes is established based on node relative entropy, which specifically means: measuring the similarity of nodes according to the node relative entropy calculated from the real network structure, and establishing a similarity network of nodes with a node similarity matrix; specifically comprising the following steps:

[0044] A11: Calculate the node relative entropy of the real network.

[0045] The specific meaning of "node relative entropy" is: node relative entropy is an indicator to measure the structural similarity between nodes, which is calculated by Tsallis entropy between node probability sets. Node probability sets are calculated by indicators of the local structure formed by a node and its first-order neighbors.

[0046] The specific steps are as follows:

[0047] 1. Calculate the fractal dimension of the network based on the box covering algorithm: Among them, N(l B ) is the box size equal to l B The number of boxes, l B The range is 1 to the box size that can cover the entire network, d f It is the fractal dimension of the network.

[0048] 2. Calculate the local dimension of each node: Where N i (r) is the number of nodes whose shortest path to node i is less than or equal to r. The value of r ranges from 1 to max. j∈V d ij , where max j∈V d ij Represents the maximum distance from all nodes to node i.

[0049] 3. Calculate the probability set of nodes based on the local dimensions of each node and its neighboring nodes: First determine the length of the probability set m = D + 1, where D is the maximum degree of the network, and then calculate the probability set P of the node i =[p i (1),p i (2),…,p i (k),…,p i (m)], the specific formula is as follows:

[0050]

[0051] Among them, di is the degree of node i, is the local dimension of node k;

[0052] 4. Calculate the relative entropy between nodes based on probability set and fractal dimension:

[0053]

[0054] Where c is a constant, usually set to 1, P′(i) is the descending order of the probability set P(i) of node i, and m′=min(d i ,d j )+1.

[0055] A12: According to the node relative entropy, calculate the correlation between each pair of real network nodes to obtain a correlation matrix of the real network nodes.

[0056] The specific steps are as follows:

[0057] Calculate the correlation r of each pair of nodes i and j based on relative entropy ij =P RE(P′(i)||P′(j))+P RE (P′(j)||P′(i)), that is, the correlation matrix of the node is obtained

[0058]

[0059] A13: Standardize the elements in the correlation matrix of the real network nodes to obtain a similarity matrix of the real network nodes.

[0060] The specific steps are as follows:

[0061] Normalize the elements in the matrix to get the normalized similarity of the nodes And the corresponding similarity matrix The value range of sij is 0-1, and the larger the value, the more similar the nodes are.

[0062] A14: According to the similarity matrix of the real network nodes, a similarity network of the real network nodes is established based on the dichotomy method.

[0063] The specific approach is as follows: based on the similarity value between nodes and a given threshold, a topological structure of a similarity network is established, and there are edges between nodes whose similarity values ​​between nodes are higher than the threshold; the threshold is obtained by searching between 0 and 1 through binary search, so that the established similarity network satisfies connectivity while having the least number of edges.

[0064] In another exemplary embodiment of the present invention, in step A2, a percolation attack is performed on the similarity network to establish a representative node data set, which specifically means: selecting key nodes in the percolation attack process as representative nodes, annotating the influence of these nodes through simulation methods, and establishing a representative node data set of the real network; wherein, the specific meaning of "percolation attack" is: strategically removing or destroying key nodes in the network, thereby affecting the connectivity and function of the entire network. In this embodiment, the percolation attack refers to the continuous removal of the node with the largest degree in the network and its first-order neighbor nodes for a given network until there are no nodes in the network, and the percolation attack process results; this process aims to find representative nodes similar to the most nodes based on the similarity between nodes. The step A2 specifically includes the following steps:

[0065] A21: Calculate the degree of each node in the similarity network.

[0066] The specific steps are:

[0067] For node i in the similarity network G(V,E), the number of its first-order neighbor nodes j∈Ν(i) is counted as the degree of node i, and the degree values ​​of all nodes in the similarity network are calculated.

[0068] A22: Remove the node with the largest degree and the neighboring nodes of the node with the largest degree in the similarity network.

[0069] The specific approach is: based on the calculated degree values ​​of all nodes in the network, select the node with the largest degree value. If there are multiple nodes, randomly select a node and remove the node and its neighboring nodes. The node with the largest degree means that there are the most nodes similar to this node, that is, this node can represent the most nodes in the current network. Because its neighboring nodes are similar to this node, the model can learn the prediction of its similarity network neighboring nodes while learning the influence prediction of representative nodes.

[0070] A23: Determine whether there is a node in the similarity network and obtain a determination result.

[0071] A24: If the judgment result is yes, return to the step of calculating the degree of each node in the similarity network.

[0072] A25: If the judgment result is no, the node with the largest degree removed is taken as the representative node.

[0073] Iteratively remove the node with the largest degree in the network and its neighboring nodes, and record the node with the largest degree each time it is removed. The set of such nodes is similar to all the nodes in the network. Using such nodes as a training set can improve the overall prediction performance.

[0074] A26: Determine the node characteristics of the representative node based on the correlation between the node influence of the real network where the representative node is located and the representative node indicator.

[0075] A27: Perform multiple simulations on the representative nodes on a real network to obtain expected influence of the representative nodes.

[0076] The specific steps are:

[0077] Calculating the approximate contagion threshold of real networks based on mean field theory in <k>is the average degree of the network, <k 2 > is the homogeneity of the square of the node degree in the network, 1.5 times the infection threshold is used as the infection probability, each node in the representative node set is used as an infection node, and other nodes are susceptible nodes. Each node is simulated 1000 times to calculate the average number of nodes that are finally infected, which is the expected influence of the node.

[0078] A28: Establish a representative node data set based on the node characteristics of the representative nodes, the network topology structure of the real network where the representative nodes are located, and the influence expectations corresponding to the representative nodes.

[0079] In another exemplary embodiment of the present invention, in step A3, the method for determining the pre-training model includes:

[0080] A31: Initialize the deep learning network to obtain the initialized deep learning network. The specific meaning is: initialize the parameters of the deep learning network based on the graph embedding neural network model and the multi-layer perceptron regression model, and determine the node features of the network influence prediction model input.

[0081] Among them, the deep learning network is used as a network influence prediction model. The specific meaning of "network influence" is: in the process of random diffusion, each node in a network has its own influence ability. In the susceptible-infected-recovered infectious disease model, the influence I of node i is defined as the total number of nodes that are eventually infected in the network under the initial state that the node is an infected node and other nodes are susceptible nodes. Where V is the set of nodes in the network, and p(j|i) is the probability that node j is infected by node i. In the random diffusion process, the influence range of an active node is a random variable, and its influence is generally estimated by its expected value of multiple experiments.

[0082] The specific steps for "initializing the parameters of the deep learning network based on the graph embedding neural network model and the multi-layer perceptron regression model" are as follows:

[0083] The pre-trained model consists of two parts: a graph embedding neural network model and a multi-layer perceptron regression model. The graph embedding neural network model consists of three convolutional layers. Each convolutional layer is convolved by an 8-head GATv2 based on the Resnet framework. Except for the last layer, the Elu activation function is connected after the convolution operation. The multi-layer perceptron regression model consists of two fully connected layers, and the ReLu activation function is connected between the layers. The hidden layer dimension of the model is 32. Before training the pre-trained model, the parameters of the above neural network model need to be orthogonally initialized.

[0084] The specific steps for "determining node characteristics for network influence prediction model input" are as follows:

[0085] According to the correlation between the node influence and node indicators of the specified random diffusion model, the node structure characteristics used as input are determined; taking the susceptible-infected-recovered infectious disease diffusion model as an example, the node influence is highly correlated with node characteristics such as the node degree, the average degree of the node's first-order neighbors, the average clustering coefficient of the node's first-order neighbors, the number of internal and external edges of the node's egocentric network, the node's 1-3 order CI value, the node's VoteRank value, the node's eigenvector centrality, and the node's K-core value. Therefore, the above node characteristics will be used as the input of the model.

[0086] A32: Establish a pre-training set; the pre-training set includes: node characteristics of the pre-training nodes, the network topology of the specified random diffusion model where the pre-training nodes are located, and the influence expectations corresponding to the pre-training nodes.

[0087] A33: Input the node characteristics of the pre-trained nodes and the network topology structure of the specified random diffusion model where the pre-trained nodes are located into the initialized deep learning network to obtain the output of the deep learning network; the output of the deep learning network is the expected influence corresponding to the predicted pre-trained nodes.

[0088] A34: Determine a second loss value according to the expected influence corresponding to the pre-trained node, the output of the deep learning network and the determined second loss function.

[0089] A35: Optimize the network parameters of the deep learning network according to the second loss value to obtain a pre-trained model.

[0090] Among them, training is performed based on the gradient descent method to obtain a pre-trained model with lower prediction loss. The prediction loss is calculated by the mean square error between the influence predicted by the pre-trained model and the label.

[0091] The prediction loss is calculated by the mean squared error between the influence of the model prediction and the label.

[0092] In another exemplary embodiment of the present invention, in step A32, the following steps are specifically included:

[0093] A3201: Obtaining pre-trained nodes on small synthetic networks.

[0094] A3202: Determine node characteristics of pre-trained nodes based on the correlation between node influence and node metrics for a specified random diffusion model.

[0095] A3203: Perform multiple simulations on a small synthetic network to obtain the expected influence of pre-trained nodes.

[0096] The specific steps are as follows:

[0097] Generate several synthetic networks of size 50 and calculate the approximate infection threshold of each network based on mean field theory in <k>is the average degree of the network, <k 2 > is the homogeneity of the square of the node degree in the network, with 1.5 times the infection threshold as the infection probability, each node as an infectious node, and other nodes as susceptible nodes. Each node is simulated 1000 times to calculate the average number of nodes that are finally infected, which is the expected influence of the node.

[0098] A3204: Establish a pre-trained node data set based on the node characteristics of the pre-trained nodes, the network topology of the specified random diffusion model where the pre-trained nodes are located, and the influence expectations corresponding to the pre-trained nodes.

[0099] Through the above steps, the present invention can use the representative node data set of the calculated real network as the training set, fine-tune the model parameters of the pre-trained model, and obtain a personalized prediction model corresponding to the real network, which can accurately predict the influence of the real network. The present invention solves the shortcomings of traditional influence prediction methods in that node influence ranking is inaccurate and the expected value of node influence cannot be quantitatively predicted efficiently, and realizes personalized prediction of the network based on graph entropy theory. This influence prediction method can quickly identify key nodes with high influence, and can quantitatively predict their expected influence, and has good practical application value.

[0100] The present invention has the following innovative features:

[0101] 1. Quantitative method: The network node influence prediction method based on graph active learning adopted in the present invention firstly learns the precise mapping relationship between the node structure and the expected value of the node influence through a small-scale synthetic network, and then establishes the precise mapping relationship between the real network node structure and the expected value of the node influence through a small amount of fine-tuning training, thereby realizing the quantitative prediction of the influence value.

[0102] 2. Easy to transplant: The node influence prediction method designed by the present invention is applicable to many types of complex networks. When applied to different types of real networks, it is only necessary to "fine-tune the influence pre-training model using representative node data" according to the structural characteristics of the real network to quickly realize the transplantation of the influence prediction method proposed by the present invention. There is no need for large-scale data labeling and training. The prediction method has strong transplantability and reduces the amount of calculation of numerical simulation.

[0103] 3. More accurate: According to the network node influence prediction method based on graph active learning proposed in the present invention, the method combining active learning with graph learning can solve the problem of unstable prediction effect of traditional influence prediction methods on different complex networks. The personalized prediction model proposed in the present invention can achieve more accurate prediction across different types of networks.

[0104] In summary, this network node influence prediction method based on graph active learning provides a good solution for identifying high-impact key components of complex systems in engineering applications.

[0105] like Figure 3 As shown, an embodiment of the present invention further provides a network node influence prediction model application method, characterized in that the network node influence prediction model application method includes:

[0106] B1: Get the node to be predicted.

[0107] B2: Determine the node characteristics of the node to be predicted based on the correlation between the node influence and the node index of the specified random diffusion model.

[0108] B3: Input the node characteristics of the node to be predicted and the network topology structure of the real network where the node to be predicted is located into the network node influence prediction model to obtain the output of the network node influence prediction model; the network node influence prediction model is a model trained by the network node influence prediction model determination method described above; the output of the network node influence prediction model is the expected influence corresponding to the predicted node to be predicted.

[0109] The present invention also provides an application scenario, which applies the above-mentioned network node influence prediction model determination method and network node influence prediction model application method. Specifically: the network node influence prediction model determination method and network node influence prediction model application method provided in this embodiment can be applied in large-scale power network prediction scenarios. Large-scale power network prediction scenarios include a pre-trained model training link, a pre-trained model parameter fine-tuning link, and a real network influence prediction link; the pre-trained model is first trained on a small synthetic network, and then a representative node data set is established. The pre-trained model parameters are fine-tuned through the representative node data set to obtain a network node influence prediction model, and then the real network influence prediction can be achieved based on the network node influence prediction model.

[0110] The embodiment of the present invention takes a large power network as an example to illustrate the method of the present invention. Specifically, the power network contains 4941 nodes, 6594 edges, the standard deviation of the network degree distribution is 1.79, the maximum eigenvalue of the network is 6.23, the power law index of the degree distribution is 3689.25, the assortativity coefficient is 0, and the clustering coefficient is 0.1. Now it is necessary to predict the influence of the network.

[0111] In order to achieve the above-mentioned purpose, the technical solution adopted by the method of the present invention is: an influence prediction method for realizing accurate prediction of complex network personalized structure. The process is as follows Figure 4 Said:

[0112] Step S1: Initialize the deep learning network constructed based on the graph embedding neural network model and the multi-layer perceptron regression model.

[0113] Step S2: Pre-train the deep learning network on the small synthetic network to obtain a pre-trained model.

[0114] Step S3: Establish a real network node similarity network based on node relative entropy.

[0115] Step S4: Perform a percolation attack on the similarity network to establish a representative node data set.

[0116] Step S5: Fine-tune the pre-trained model using a representative node data set to achieve accurate prediction of real network influence.

[0117] Among them, in step S1, the deep learning network is a network influence prediction model. The specific meaning of "network influence" is: in the process of random diffusion, each node in a network has its own influence ability. In the susceptible-infected-recovered infectious disease model, the influence I of node i is defined as the total number of nodes that are eventually infected in the network under the initial state that the node is an infected node and other nodes are susceptible nodes. Where V is the set of nodes in the network, and p(j|i) is the probability that node j is infected by node i. In the random diffusion process, the influence range of an active node is a random variable, and its influence is generally estimated by its expected value of multiple experiments.

[0118] In step S1, "initializing a deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model" specifically means: initializing the parameters of the deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model, and determining the node features input to the network influence prediction model; including the following steps:

[0119] Step S11: Initialize the parameters of the pre-trained model.

[0120] Step S12: Determine the node characteristics input into the network influence prediction model.

[0121] The specific method of "initializing the parameters of the pre-trained model" in step S11 is as follows:

[0122] The pre-trained model consists of two parts: a graph embedding neural network model and a multi-layer perceptron regression model. The graph embedding neural network model consists of three convolutional layers. Each convolutional layer is convolved by an 8-head GATv2 based on the Resnet framework. Except for the last layer, the Elu activation function is connected after the convolution operation. The multi-layer perceptron regression model consists of two fully connected layers, and the ReLu activation function is connected between the layers. The hidden layer dimension of the model is 32. Before training the pre-trained model, the parameters of the above neural network model need to be orthogonally initialized.

[0123] The specific method of "determining the node features input to the network influence prediction model" in step S12 is as follows:

[0124] According to the correlation between the node influence and node indicators of the specified random diffusion model, the node structure characteristics used as input are determined; taking the susceptible-infected-recovered infectious disease diffusion model as an example, the node influence is highly correlated with the node degree, the average degree of the node's first-order neighbors, the average clustering coefficient of the node's first-order neighbors, the number of internal and external edges of the node's egocentric network, the node's 1-3 order CI value, the node's VoteRank value, the node's eigenvector centrality, the node's K-core value and other characteristics, so the above node characteristics will be used as the input of the model.

[0125] In step S2, "pre-training the deep learning network on the small synthetic network to obtain a pre-trained model" specifically means: pre-training the model initialized in step S1 by using the node influence labels simulated on the small synthetic network; including the following steps:

[0126] Step S21: Simulate multiple times on a small synthetic network to obtain the expected influence of the node.

[0127] Step S22: Pre-train the model using small synthetic network data.

[0128] The specific method of "simulating multiple times on a small synthetic network to obtain the expected influence of a node" in step S21 is as follows:

[0129] Generate several synthetic networks of size 50 and calculate the approximate infection threshold of each network based on mean field theory in <k>is the average degree of the network, <k 2 > is the homogeneity of the square of the node degree in the network, with 1.5 times the infection threshold as the infection probability, each node as an infectious node, and other nodes as susceptible nodes. Each node is simulated 1000 times to calculate the average number of nodes that are finally infected, which is the expected influence of the node.

[0130] The specific method of "pre-training the model using small synthetic network data" in step S22 is as follows:

[0131] Taking the expected influence of each node as a label, and the features of the node calculated in step S12 and the structure of the network where the node is located as data, the pre-trained model initialized in step S11 is trained based on the gradient descent method to obtain a prediction model with low prediction loss, where the prediction loss is calculated by the mean square error between the influence predicted by the model and the label.

[0132] In step S3, the "node relative entropy" has the following specific meanings: the node relative entropy is an indicator for measuring the structural similarity between nodes, which is calculated by the Tsallis entropy between node probability sets. The node probability set is calculated by the indicator of the local structure formed by the node and its first-order neighbors.

[0133] The "establishing a real network node similarity network based on node relative entropy" described in step S3 specifically means: measuring the similarity of nodes according to the node relative entropy calculated from the real network structure, and establishing a node similarity network with the node similarity matrix; comprising the following steps:

[0134] Step S31: Calculate the relative entropy of the real network nodes.

[0135] Step S32: Establish a node similarity matrix.

[0136] Step S33: Establish a node similarity network based on the dichotomy method.

[0137] Among them, in step S31, the specific method of "calculating the relative entropy of the real network node" is as follows:

[0138] 1. Calculate the fractal dimension of the network based on the box covering algorithm: Among them, N(l B ) is the box size equal to l B The number of boxes, l B The range is 1 to the box size that can cover the entire network, d f It is the fractal dimension of the network.

[0139] 2. Calculate the local dimension of each node: Where N i (r) is the number of nodes whose shortest path to node i is less than or equal to r. The value of r ranges from 1 to max. j∈V d ij , where max j∈V d ij Represents the maximum distance from all nodes to node i.

[0140] 3. Calculate the probability set of nodes based on the local dimensions of each node and its neighboring nodes: First determine the length of the probability set m = D + 1, where D is the maximum degree of the network, and then calculate the probability set P of the node i =[p i (1),p i (2),…,p i (k),…,p i (m)], the specific formula is as follows:

[0141]

[0142] Among them, di is the degree of node i, is the local dimension of node k;

[0143] 4. Calculate the relative entropy between nodes based on probability set and fractal dimension:

[0144]

[0145] Where c is a constant, usually set to 1, P′(i) is the descending order of the probability set P(i) of node i, and m′=min(d i ,d j )+1.

[0146] Among them, in step S32, the "establishing the node correlation matrix" is specifically done as follows:

[0147] Calculate the correlation r of each pair of nodes i and j based on relative entropy ij =P RE (P′(i)||P′(j))+P RE (P′(j)||P′(i)), that is, the correlation matrix of the node is obtained Then the elements in the matrix are standardized to obtain the standardized similarity of the nodes And the corresponding similarity matrix The value range of sij is 0-1, and the larger the value, the more similar the nodes are.

[0148] Among them, in step S33, the specific method of "establishing a minimum similarity network of nodes based on the dichotomy method" is as follows:

[0149] The topological structure of the similarity network is established based on the similarity value between nodes and a given threshold. There are edges between nodes whose similarity values ​​between nodes are higher than the threshold. The threshold is obtained by searching between 0 and 1 through binary search, so that the established similarity network satisfies connectivity while having the least number of edges.

[0150] The "percolation attack" described in step S4 specifically means: strategically removing or destroying key nodes in the network, thereby affecting the connectivity and function of the entire network. In this method, the percolation attack refers to, for a given network, continuously removing the node with the largest degree in the network and its first-order neighbor nodes until there are no nodes in the network, and the percolation attack process results; this process aims to find representative nodes that are similar to the most nodes based on the similarity between nodes.

[0151] Among them, in step S4, the "permeation attack on the similarity network to establish a representative node data set" specifically means: selecting key nodes in the permeation attack process as representative nodes, marking the influence of these nodes through simulation methods, and establishing a representative node data set of the real network; including the following steps:

[0152] Step S41: Calculate the degree of the nodes in the similarity network.

[0153] Step S42: remove the node with the largest degree in the similarity network and its neighboring nodes.

[0154] Step S43: Repeat steps S41 and S42 until there are no nodes in the similarity network, and record the node with the largest degree removed each time as the representative node set.

[0155] Step S44: Perform multiple simulations on the representative nodes on the real network to obtain the expected influence of the nodes and form a representative node data set.

[0156] Among them, in step S41, the specific method of "calculating the degree of nodes in the similarity network" is:

[0157] For node i in the similarity network G(V,E), the number of its first-order neighbor nodes j∈Ν(i) is counted as the degree of node i, and the degree values ​​of all nodes in the similarity network are calculated.

[0158] Among them, in step S42, the specific method of "removing the node with the largest degree in the similarity network and its neighboring nodes" is as follows:

[0159] According to the degree values ​​of all nodes in the network calculated in step S41, the node with the largest degree is selected. If there are multiple nodes, a node is randomly selected and the node and its neighboring nodes are removed. The node with the largest degree indicates that there are the most nodes similar to this node, that is, this node can represent the most nodes in the current network. Since its neighboring nodes are similar to this node, the model can learn the prediction of its similarity network neighboring nodes while learning the influence prediction of representative nodes.

[0160] Among them, in step S43, the specific method of "repeating steps S41 and S42 until there are no nodes in the similarity network, and recording the node with the largest degree removed each time as the representative node set" is as follows:

[0161] Iteratively remove the node with the largest degree in the network and its neighboring nodes, and record the node with the largest degree each time it is removed. The set of such nodes is similar to all the nodes in the network. Using such nodes as a training set can improve the overall prediction performance.

[0162] Among them, in step S44, the "simulating the representative nodes on the real network multiple times to obtain the expected influence of the nodes and form a representative node data set" is specifically done by: calculating the approximate infection threshold of the real network according to the mean field theory in <k>is the average degree of the network, <k 2 > is the homogeneity of the square of the node degree in the network, with 1.5 times the infection threshold as the infection probability, each node in the representative node set as an infectious node, and other nodes as susceptible nodes. Each node is simulated 1000 times to calculate the average number of nodes that are finally infected. The average value is the expected influence of the node, forming a data set of representative nodes.

[0163] The specific meaning of "using representative node data to fine-tune the influence prediction model to achieve accurate prediction of the real network influence" in step S5 is: using the representative node data set of the real network calculated in step S4 as the training set, fine-tuning the model parameters of the pre-trained model in step S2 to obtain a personalized prediction model corresponding to the real network, which can achieve accurate prediction of the real network influence; including the following steps:

[0164] Step S51: fine-tune the pre-trained model based on the representative node data set.

[0165] Step S52: Use the personalized prediction model to predict the influence of the real network nodes.

[0166] Among them, in step S51, the "fine-tuning of the pre-trained model based on the representative node data set" is specifically done as follows: using the expected influence of the representative node as a label, and the node features calculated according to step S12 and the real network structure as input, the pre-trained model in step S2 is trained based on the gradient descent method to obtain a personalized prediction model of the real network.

[0167] Among them, the specific method of "using a personalized prediction model to predict the influence of real network nodes" described in step S52 is: based on step S12, the characteristics of the nodes to be predicted on the real network are calculated, and the characteristics of the nodes to be predicted and the topological structure of the real network are input into the personalized model of the network, and the model outputs the influence value of the nodes to be predicted.

[0168] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store node data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a network node influence prediction model determination method or a network node influence prediction model application method is implemented.

[0169] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0170] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the above-mentioned method embodiments are implemented when the processor executes the computer program.

[0171] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.

[0172] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.

[0173] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0174] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0175] The database involved in each embodiment provided by the present invention may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided by the present invention may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0176] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0177] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.< / k> < / k> < / k> < / k>

Claims

1. A method for determining a network node influence prediction model, characterized in that: The network node influence prediction model determination method includes: A similarity network of real network nodes is established based on node relative entropy; the node relative entropy is an indicator for measuring the structural similarity between nodes; Performing a percolation attack on the similarity network to establish a representative node data set; the percolation attack is used to find a representative node that is similar to the most nodes based on the similarity between nodes; the representative node data set includes: node characteristics of the representative node, the network topology of the real network where the representative node is located, and the influence expectation corresponding to the representative node; Input the node features of the representative nodes and the network topology structure of the real network where the representative nodes are located into the pre-training model to obtain the output of the pre-training model; the pre-training model is a deep learning network constructed based on a graph embedding neural network model and a multi-layer perceptron regression model; the output of the pre-training model is the expected influence corresponding to the predicted representative node; Determine a first loss value according to the expected influence corresponding to the representative node, the output of the pre-trained model, and the determined first loss function; The network parameters of the pre-trained model are optimized according to the first loss value to obtain a network node influence prediction model.

2. The method for determining a network node influence prediction model according to claim 1, characterized in that: Based on the relative entropy of nodes, a similarity network of real network nodes is established, including: Calculate the node relative entropy of the real network; Calculate the correlation of each pair of real network nodes according to the node relative entropy to obtain a correlation matrix of the real network nodes; Standardizing the elements in the correlation matrix of the real network nodes to obtain a similarity matrix of the real network nodes; According to the similarity matrix of the real network nodes, a similarity network of the real network nodes is established based on the dichotomy method.

3. The method for determining a network node influence prediction model according to claim 1, characterized in that: Perform a percolation attack on the similarity network to establish a representative node data set, specifically including: Calculating the degree of each node in the similarity network; Remove the node with the largest degree and the neighboring nodes of the node with the largest degree in the similarity network; Determine whether there is a node in the similarity network and obtain a determination result; If the judgment result is yes, returning to the step of calculating the degree of each node in the similarity network; If the judgment result is no, the node with the largest degree removed is taken as the representative node; Determine the node characteristics of the representative node based on the correlation between the node influence of the real network where the representative node is located and the representative node index; Perform multiple simulations on the representative nodes on a real network to obtain the expected influence of the representative nodes; A representative node data set is established based on the node characteristics of the representative nodes, the network topology structure of the real network where the representative nodes are located, and the influence expectations corresponding to the representative nodes.

4. The method for determining a network node influence prediction model according to claim 1, characterized in that: The method for determining the pre-training model includes: Initialize the deep learning network to obtain an initialized deep learning network; Establishing a pre-training set; the pre-training set includes: node characteristics of the pre-training node, the network topology structure of the specified random diffusion model where the pre-training node is located, and the influence expectation corresponding to the pre-training node; Inputting node features of the pre-trained nodes and the network topology of the specified random diffusion model where the pre-trained nodes are located into the initialized deep learning network to obtain the output of the deep learning network; the output of the deep learning network is the expected influence corresponding to the predicted pre-trained nodes; Determine a second loss value according to the expected influence corresponding to the pre-trained node, the output of the deep learning network, and the determined second loss function; The network parameters of the deep learning network are optimized according to the second loss value to obtain a pre-trained model.

5. The method for determining a network node influence prediction model according to claim 4, characterized in that: Initialize the deep learning network to obtain the initialized deep learning network, which specifically includes: The three convolutional layers of the graph embedding neural network model and the two fully connected layers of the multi-layer perceptron regression model are orthogonally initialized to obtain the initialized deep learning network.

6. The method for determining a network node influence prediction model according to claim 4, characterized in that: The establishing of the pre-training set specifically includes: Get pre-trained nodes on a small synthetic network; Determine the node characteristics of the pre-trained nodes according to the correlation between the node influence and the node index of the specified random diffusion model; Multiple simulations were performed on a small synthetic network to obtain the expected influence of the pre-trained nodes. A pre-trained node data set is established based on the node characteristics of the pre-trained nodes, the network topology structure of the specified random diffusion model where the pre-trained nodes are located, and the influence expectations corresponding to the pre-trained nodes.

7. A network node influence prediction model application method, characterized in that: The network node influence prediction model application method includes: Get the node to be predicted; Determine the node characteristics of the node to be predicted according to the correlation between the node influence and the node index of the specified random diffusion model; The node characteristics of the node to be predicted and the network topology structure of the real network where the node to be predicted is located are input into the network node influence prediction model to obtain the output of the network node influence prediction model; the network node influence prediction model is a model trained based on the network node influence prediction model determination method described in any one of claims 1-6; the output of the network node influence prediction model is the expected influence corresponding to the predicted node to be predicted.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the network node influence prediction model determination method described in any one of claims 1 to 6 or the network node influence prediction model application method described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the network node influence prediction model determination method described in any one of claims 1 to 6 or the network node influence prediction model application method described in claim 7.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the network node influence prediction model determination method described in any one of claims 1 to 6 or the network node influence prediction model application method described in claim 7.

Citation Information

Patent Citations

  • Complex network node influence maximization method based on depth auto-encoder

    CN108510059A

  • Dynamic influence maximization method based on cohesion entropy

    CN112380456A

  • Graph neural network-based interpretable graph classification anti-attack method

    CN118015376A