Propagation suppression method and system based on influence node prediction and terminal equipment

Abstract

The invention discloses a propagation suppression method and system based on influence node prediction and terminal equipment, and the method comprises the steps: carrying out the time sequence slicing of a social network, obtaining a plurality of time slice networks, and extracting the network features of each time slice network and the features of an influence node; constructing Gaussian process regression based on the network features of each time slice network and the influence node features, designing a kernel function oriented to a time sequence periodic network, obtaining an influence node prediction model based on Gaussian process regression, and optimizing hyper-parameters of the influence node prediction model; and predicting influence nodes of the future time slice network based on the influence node prediction model, and activating the predicted influence nodes to perform propagation suppression. The method is applied to the field of social network analysis, the nodes change along with the periodic influence of the time sequence in the node prediction process, the prediction effect is enhanced, and the propagation inhibition efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of social network analysis, in particular to a propagation suppression method and system based on influence node prediction and a terminal device. BACKGROUND

[0002] In the social network environment, the convenient forwarding characteristics of induced information make it difficult to completely eliminate the negative impact once the information spreads and diffuses, even if the original propagation source is removed. Therefore, how to effectively identify a group of influential user groups, and spread positive information to cover the original induced information, is a problem that needs to be studied in current propagation suppression. Existing researches mainly focus on the identification of influence node set in single time slice network, and less consider the propagation suppression in time sequence network, and the influence nodes in single time slice are not necessarily the influence nodes in time sequence network. In addition, for the time sequence influence maximization problem, the existing methods usually rely on historical time slice information to find the influence nodes in the current time slice, and there are few methods to find the influence nodes in the future time slice. And in the identification idea of future influence nodes, the existing methods usually adopt the linear thinking of constant strength of strong nodes, without considering the periodic structural changes of time sequence network. SUMMARY

[0003] In view of the deficiencies in the prior art, the present application provides a propagation suppression method and system based on influence node prediction and a terminal device, which strengthens the prediction effect and improves the propagation suppression efficiency by considering the periodic influence changes of nodes with time sequence in the process of node prediction.

[0004] To achieve the above purpose, the present application provides a propagation suppression method based on influence node prediction, comprising the following steps:

[0005] Step 1, time sequence slicing is performed on the social network to obtain a plurality of time slice networks, and network features of each time slice network are extracted;

[0006] Step 2, the influence nodes in each time slice network are labeled to obtain the influence node features of each time slice network;

[0007] Step 3, a Gaussian process regression is constructed based on the network features and influence node features of each time slice network, a kernel function for time sequence periodic network is designed, an influence node prediction model based on Gaussian process regression is obtained, and the hyperparameters of the influence node prediction model are optimized;

[0008] Step 4, the influence nodes of the future time slice network are predicted based on the influence node prediction model, and the predicted influence nodes are activated for propagation suppression.

[0009] In one of the embodiments, in step 1, the Laplacian matrix is constructed as the network feature of each of the time-slice networks, specifically:

[0010] L = I - D -1 / 2 · A · D -1 / 2

[0011] wherein L is the Laplacian matrix of the time-slice network, I is the identity matrix, A is the adjacency matrix of the time-slice network, and D is the degree matrix of the time-slice network;

[0012]

[0013] wherein n is the number of nodes of the time-slice network, A ij is an element of the adjacency matrix A, used to represent whether the node v i is connected to the node v j ;

[0014] After the Laplacian matrix is constructed, the pairwise distance between any two nodes in the time-slice network is calculated based on the constructed Laplacian matrix, which is:

[0015]

[0016] wherein d ij is the pairwise distance between the node v i and the node v j , is the pseudo-inverse of the Laplacian matrix L, is the i-th row and i-th column element of the matrix , is the j-th row and j-th column element of the matrix , is the i-th row and j-th column element of the matrix , and T is the transpose of the matrix.

[0017] In one of the embodiments, in step 2, the process of obtaining the influence node feature is:

[0018] In step 201, the node set V and the edge set E of the time-slice network are obtained, as well as the propagation probability p v and the edge activation probability q v,u of each node, wherein v, u ∈ V, and the node u is the neighbor node of the node v, i.e., the network edge e vu ∈ E;

[0019] In step 202, the influence I v of each node is initialized to 0, the active set is initialized to V, and the candidate set S is initialized to V.

[0020] Step 203, calculate the gain of each node in the current candidate set S, which is:

[0021]

[0022] Step 203, select the node with the maximum gain

[0023] Step 204, update the influence of each node in the current candidate set S, which is:

[0024]

[0025] Where, A v is the neighbor node set of node v;

[0026] Step 205, let S * =S * ∪{v *}, S=S\{v *}, and then determine whether the iteration termination condition is met:

[0027] If yes, construct the influence node feature based on the current active set S * .

[0028] Otherwise, return to step 203.

[0029] In one embodiment, in step 205, the iteration termination condition is or ∑ v∈V I v >I max , where I max is a preset influence threshold.

[0030] In one embodiment, in step 3, the kernel function for the time-series periodic network is specifically:

[0031]

[0032] Where, k(x,x ' ) is the kernel function value between input point x and input point x ' in Gaussian process regression, x and x ' represent the feature vectors mapped in Gaussian process regression of the two time-slice networks respectively, σ is the radial basis kernel parameter, T is the period of the periodic kernel function, λ is the amplitude influence parameter of the periodic kernel, and δ xx' is a noise control term.

[0033] In one embodiment, in step 3, the process of calculating the influence node prediction model hyperparameters is specifically:

[0034] Obtain training samples (X, Y), wherein X={x1, x2, …, x N} is an input feature matrix, Y={y1, y2, …, y N} is a corresponding output vector;

[0035] Define a negative log marginal likelihood function Measure the difference between observed data and model prediction, which is:

[0036]

[0037] Wherein, θ=(σ, t, λ) is a hyperparameter to be calculated, K is a kernel matrix, and N is the number of samples;

[0038] Minimize the negative log marginal likelihood function Optimize the hyperparameter θ, and expand the negative log marginal likelihood function According to the expression of the kernel function, as follows:

[0039]

[0040] Wherein, The partial derivative of the negative log marginal likelihood function with respect to σ is The partial derivative matrix of the kernel matrix K with respect to σ is Tr(·), and the trace of the matrix is;

[0041] Obtain the derivative of the kernel function, which is:

[0042]

[0043] The gradient descent method is used to iteratively optimize the hyperparameter θ, and in each iteration step, the update process of the hyperparameter θ is:

[0044]

[0045] Wherein, θ new Is the updated hyperparameter, θ old Is the hyperparameter before updating, and α is the learning rate.

[0046] To achieve the above purpose, the application also provides a propagation suppression system based on influence node prediction, which adopts the above method for propagation suppression, and the propagation suppression system comprises:

[0047] A time sequence network construction unit is used to time sequence slice the social network to obtain a plurality of time slice networks;

[0048] A network feature extraction unit is used to extract network features of each time slice network;

[0049] A node feature extraction unit is configured to obtain the influence node features of each time slice network.

[0050] A prediction model construction unit is configured to construct a Gaussian process regression based on the network features and the influence node features of each time slice network, design a kernel function for the time sequence cycle network, obtain an influence node prediction model based on the Gaussian process regression, and calculate the hyperparameters of the influence node prediction model.

[0051] A propagation inhibition unit is configured to predict the influence nodes of a future time slice network according to the influence node prediction model, and activate the predicted influence nodes for propagation inhibition.

[0052] To achieve the above object, the application further provides a terminal device, wherein the terminal device is provided with:

[0053] A memory is configured to store a program.

[0054] A processor is configured to execute the program stored in the memory, and when the program is executed, the processor is configured to execute the method as described above.

[0055] Compared with the prior art, the application has the following beneficial technical effects:

[0056] The application learns the influence node and network structure correspondence relationship based on the Gaussian process regression, thereby effectively adapting to the nonlinear changes of the time sequence network structure. In addition, the cycle term is added in the Gaussian process regression cycle kernel function, thereby effectively extracting the time sequence information of the social network. In the process of node prediction, the node changes with the periodic influence of the time sequence, thereby strengthening the prediction effect and improving the propagation inhibition efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only show some embodiments of the application, and for those skilled in the art, other drawings can also be obtained according to the structures shown in these drawings without creative labor.

[0058] Figure 1 A flow chart of the propagation inhibition method based on influence node prediction in the embodiments of the application;

[0059] Figure 2 A sampling node and Gaussian process diagram in the embodiments of the application;

[0060] Figure 3 A structure block diagram of the propagation inhibition system based on influence node prediction in the embodiment 2 of the application;

[0061] Figure 4 This is a structural block diagram of the terminal device in Example 3 of the present invention.

[0062] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0065] While information dissemination promotes fair access to information, it can also lead to the spread of misleading information. The spread of misleading information can lead to cognitive biases in the public, disrupting group perception, judgment, and decision-making. Once misleading information spreads on a large scale, negative impacts are already in place, and removing the source of the information will not eliminate them. In practical applications, tracing the source of information is costly, and only vague estimates of its spread and probability distribution are available. Therefore, random distribution and fixed-probability activation methods can be used to simulate the source of misleading information.

[0066] For multi-point propagation source suppression, the random sampling and fixed probability method can be used to simulate multi-point propagation sources, transforming the propagation suppression problem into the problem of identifying the influential node set. In this case, the propagation range of the induced information depends only on the number of propagation source nodes. Specifically, in a given network G, it is assumed that the propagation source set of the induced information is R P , the influential node set is S P ,have:

[0067] B P (R P ,S P )=I(S P )+I(R P )

[0068] Among them, I(S P ) is the set of influential nodes S in the social network P The spread of positive information generated, I(R P) is a set of spreading sources in a social network P The propagation range of the induced information generated by B P (R P ,S P ) represents the suppression ability of the positive information influence node pair to the induced information.

[0069] Therefore, the target of the influence node prediction-based propagation suppression method in the embodiment is to determine an influence node set S P * such that the influence node set S P ) is maximized, that is, the maximization of the suppression range is achieved. Based on this assumption, the value of B P (R P ,S P ) only changes with the positive information propagation source. Under the premise that the propagation range of the induced information is relatively fixed, the suppression problem is converted into the problem of identifying the influence node set with effective suppression effect. The influence of public opinion can be guided through the propagation of positive information, so that I(S P ) is maximized, and the maximization suppression of B P (R P ,S P ) is achieved. For the above-mentioned uncertain source propagation suppression problem based on influence nodes, further considering the multi-point propagation source suppression problem under the time sequence network, there are the following problems: 1) in the current suppression method based on influence node identification, the propagation suppression under the time sequence network is less considered, and the influence node in a single time slice is not necessarily the influence node in the time sequence network; 2) for the time sequence influence maximization problem, the existing method usually relies on historical time slice information to find the influence node in the current time slice, and there are few methods to find the influence node in the future time slice, and in the identification idea of the future influence node, the existing method usually adopts the linear thinking of constant strength of strong nodes, and does not consider the time sequence network with periodic structural changes.

[0070] Based on this, the embodiment discloses an influence node prediction-based propagation suppression method, which mainly includes the following steps:

[0071] Step 1, time sequence slicing is performed on the social network to obtain a plurality of time slice networks, and network features of the time slice networks are extracted;

[0072] Step 2, the influence nodes in the time slice networks are labeled to obtain influence node features of the time slice networks;

[0073] Step 3: Based on the network characteristics and influential node characteristics of each time slice network, a Gaussian process regression is constructed, and a kernel function for the time series periodic network is designed to obtain an influential node prediction model based on Gaussian process regression, and the hyperparameters of the influential node prediction model are optimized;

[0074] Step 4: Predict the influential nodes of the future time slice network based on the influential node prediction model, and activate the predicted influential nodes to suppress propagation.

[0075] During the implementation of step 1, a temporal network is constructed by slicing the social network in time series. A temporal network is a dynamic network model that considers nodes and edges in the temporal dimension. Unlike static networks, the connections in a temporal network evolve over time, and the connections between nodes are timestamped, reflecting the temporal dependence of the network topology. Typical temporal models based on network snapshots can be considered extensions of static graphs, reflecting the evolution of the network in the temporal dimension. These models divide the entire time period under study into multiple network snapshots. Each snapshot can cover the state of the network nodes, edges, and their attributes at a specific time, or it can be a static aggregation of all node and edge changes that occurred within a specific time window.

[0076] Due to the difficulty of data collection in real situations, this embodiment adopts a multi-network snapshot modeling method to establish a time series network. Specifically: the time series network model can be expressed as represents the time slice network at time nodes 1, 2, ..., T0. Time Slicing Network The node set of Time Slicing Network The edge set, T0 is the time slice network For example, in a social media network, the nodes of the time-slicing network are social media users, and the edges of the time-slicing network are information interaction behaviors of social media users. For example, if the number of information interactions between two information interaction behaviors exceeds a threshold within a set time period, then the network edge corresponding to the two nodes exists in the time-slicing network, and the two nodes are neighbors. Otherwise, the network edge corresponding to the two nodes does not exist.

[0077] In this embodiment, the network characteristics of the time-slicing network are characterized by constructing a Laplacian matrix, and the specific implementation process is as follows:

[0078] For any time-slicing network, its degree matrix can represent the degree of the nodes in the time-slicing network, that is, the number of edges connected to the nodes. The degree matrix is ​​defined as:

[0079]

[0080] Where D is the degree matrix of the time slice network, n is the number of nodes in the time slice network, A is the adjacency matrix of the time slice network, and A ij is an element of the adjacency matrix A, used to represent the node v in the time slice network i With node v j whether they are connected;

[0081] The Laplace matrix L can be calculated from the degree matrix and the adjacency matrix as follows:

[0082] L=DA

[0083] After normalization:

[0084] L=ID -1 / 2 ·A·D -1 / 2

[0085] Where I is the identity matrix;

[0086] The Laplace matrix reflects both the connection relationship between nodes in the time-slicing network and the degree distribution of network nodes. For node connection relationships, the Laplace matrix reflects the topological structure of the network. For connections between adjacent nodes, the corresponding elements of the Laplace matrix are negative. For nodes without connections, the corresponding elements are 0. On the other hand, the Laplace matrix contains node degree information and local structural information, which can simulate local node correlations in the data. In addition, the Laplace matrix can also calculate the pairwise distance between any two nodes in the time-slicing network, which is:

[0087]

[0088] Among them, d ij is the node v in the time slicing network i With node v j The pairwise distance of is the pseudo-inverse of the Laplace matrix L, is a matrix The i-th row and i-th column element of is a matrix The j-th row and j-th column element of is a matrix The i-th row and j-th column element of is , where T is the transpose of the matrix. This shows that the pseudo-inverse of the Laplace matrix can measure the similarities and differences between network nodes and express the relationships between network nodes in Laplace space. Combining the structural features expressed in the matrix with node degree information, calculating node relationships provides the basic network topology input data for subsequent propagation suppression.

[0089] The Laplacian matrix distance extracts the network topology of each time-slice network. To establish the connection between the network topology and the influential nodes, the embodiment also labels the set of influential nodes in each time-slice network to form the influential node feature. Thus, the network topology and the influential node label can be used to determine the influential nodes in the future network structure, publish positive information, cover and suppress induced information. To this end, the embodiment adopts a cost-effective delay forwarding method to calculate the influential nodes in the time-slice network, divide the network nodes into influential nodes and non-influential nodes, assign different labels, and form the influential node feature based on the influential node coding. The implementation process includes the following steps:

[0090] Step 201, obtaining the node set V and the edge set E of the time-slice network, and the propagation probability p of each node v and the edge activation probability q v,u , where v, u ∈ V, and node u is the neighbor node of node v, i.e., the network edge e vu ∈ E, the propagation probability p v and the edge activation probability q v,u can be given rough empirical values according to the historical data of each node, or can be directly set;

[0091] Step 202, initializing the influence I v = 0, the active set and the candidate set S = V of each node;

[0092] Step 203, calculating the gain of each node in the current candidate set S, which is:

[0093]

[0094] Step 203, selecting the node with the maximum gain

[0095] Step 204, updating the influence of each node in the current candidate set S, which is:

[0096]

[0097] where A v is the neighbor node set of node v;

[0098] Step 205, S * = S * ∪ {v *}, S = S \ {v *}, and then determining whether the iteration termination condition is met, for example, determining whether or ∑ v∈V I v >I max is met, where Imax is a preset influence threshold:

[0099] If yes, feature encoding is performed based on the nodes in the current active set S * , thereby obtaining influence node features, such as encoding in a one-dimensional feature vector according to the order in which the nodes enter the active set S * .

[0100] Otherwise, return to step 203.

[0101] The core idea of the cost-effective delay forwarding method in this embodiment continues the idea of the greedy algorithm, and selects potential nodes. Compared with the greedy algorithm, the calculation efficiency is improved. That is, in each iteration, the node with the maximum influence gain is selected, and a priority queue is maintained according to the influence gain potential of the node The order of the candidate nodes is dynamically adjusted. The queue contains the influence gain of each node, and the maximum gain of the influence under the same cost is calculated. Compared with other heuristic algorithms, the dynamic updating mechanism makes the cost-effective delay forwarding method have more accurate results. This method uses the submodular property of the influence gain to find potential nodes. Through the candidate node sorting operation of the influence gain, the efficiency and accuracy of the identification of the influence nodes in the historical time slice can be improved.

[0102] The propagation suppression method based on influence node prediction in this embodiment mainly predicts the node influence at the future time point by observing the node influence at the historical time. One form of time series network is a plurality of network time slices arranged in time series, which is similar to the description of Gaussian process. Therefore, the influence node prediction model is constructed based on Gaussian process regression.

[0103] A Gaussian process regards a function as a set of random variables, and the values of a limited number of points follow a multi-dimensional Gaussian distribution. Gaussian process defines a distribution in the function space, not a model for a specific function. Taking the spatial domain as an example, the Gaussian process is the joint Gaussian distribution of a plurality of points on a continuous domain. The mean value of each point on the space and the covariance matrix between these points are used to represent the infinite-dimensional Gaussian distribution, i.e., the Gaussian process. Compared with one-dimensional Gaussian distribution, the mean value and the covariance matrix in the Gaussian process are changed from numbers to functions, i.e., the mean function and the covariance function, which are:

[0104] S GP (x1,x2,…,x n ,…)~N(μ(x),k(x,x′))

[0105] Where x and x' represent points, μ(x) is the mean function, and k(x,x') represents the covariance function.

[0106] Gaussian process regression is a supervised learning method that fits data into a Gaussian process in some scenarios and then makes predictions. In terms of principles, Gaussian process regression is a non-parametric regression method that models and predicts data by calculating the probability distribution of the data. That is, it is assumed that the relationship between data points follows a multi-dimensional Gaussian distribution. By learning the function values and the covariance information between functions in the training data, the function values of new data points can be predicted, and the uncertainty of the prediction results can be given. Corresponding to the Gaussian process, Gaussian process regression usually uses a mean function and a covariance function to describe the distribution characteristics of the data. Through the training data set, the parameters of the two functions can be inferred, and the new input data can be predicted. In particular, in graph data, Gaussian process regression can model and predict variables on nodes.

[0107] For Gaussian process regression, taking the spatial domain as an example, it is assumed that the coordinates of the i-th point in the space are x (i) , and the data output of the point is y (i) . The expression of the Gaussian process regression model is:

[0108] y (i) =f(x (i) )+∈ (i) i=1,2,…,n

[0109] where ∈ (i) is a noise variable, which follows N(0,σ 2 ). In Gaussian process regression, it is assumed that f(x (i) ) follows a Gaussian process, that is:

[0110] f(x)~GP(m(x),k(x,x′))

[0111] where m(x) is the mean function, and k(x,x') is the covariance function, that is, the kernel function.

[0112] In this embodiment, the network topology information (i.e., network features) and node label information (i.e., influential node features) of the time slicing network are used in combination with Gaussian process regression to solve the hyperparameters and design a suitable kernel function, so that the designed influential node prediction model can be applied to the problem of finding influential nodes in a time series network.

[0113] In Gaussian process regression, the future distribution is predicted by sampling multiple points, and in each time slice, the labeled influential node can be regarded as a single node sampled in multiple time slices. As shown in Figure 2 , two point samples of the Gaussian process are shown. This process can also be regarded as sampling the influence of two nodes in a time series network. Specifically, in n time points t1, t2,…, t n , an n-dimensional vector ξ1, ξ2,…, ξn all satisfy that it is an n-dimensional Gaussian distribution, then ξ i is a Gaussian process. As the nodes of each time slice sample, respectively, represent the state of the influence of the nodes at different times, the mean and variance of the future node influence can be given by the kernel function in Gaussian process regression, that is where ξ t and can be predicted.

[0114] The input variables of the influence node prediction model in this embodiment are the network features of the time slice network and the influence node features, and the output variables are the influence node set in the future time slice. The network features constructed in this embodiment can obtain the distance information of the node pairs, and the influence node features can obtain the influence node labels. According to the distance relationship between each node and other nodes in the network, the features of the influence nodes are extracted. The influence node prediction model is trained by the node information in multiple time slices. By using the training data and the covariance matrix, a probability model based on Gaussian distribution can be fitted to identify the nodes with higher probability to become influential. For example Figure 2 The two dashed lines in the figure can be used to calculate the hyperparameter training model according to the node influence at different times, predict the node influence at future time t, and predict the mean and variance interval by giving a new input variable.

[0115] In Gaussian process regression, the kernel function is one of the components that define the Gaussian process, which is used to measure the similarity or correlation between different points in the input space. The kernel function in this embodiment improves more considering the time series periodic network, fitting the time series periodic characteristics, and predicting the future time series network characteristics. At the same time, the kernel function affects the fitting degree, smoothness and generalization ability in the Gaussian process. Based on the results of the kernel function, the covariance matrix between the input variables can be calculated to obtain a complete Gaussian process model. Radial basis function is a commonly used kernel function that can handle non-linear separable data. Its principle is to calculate the similarity between features by using the distance between input data points and reference points, and introduce a nonlinear relationship. Radial basis function maps data to a higher dimensional feature space for classification. Radial basis function usually has the shape of a bell curve, where the weight of samples farther from the kernel center is smaller, and this shape helps to smooth the displacement and scaling of training samples. In addition, radial basis function measures the similarity between input vectors and center vectors to perform pattern recognition and other tasks. In Gaussian process regression, radial basis function can be used to construct a non-linear decision boundary, so that the model can handle complex sample distribution.

[0116] In the improved periodic kernel function of this embodiment, a Gaussian kernel function based on Euclidean distance is constructed. The Euclidean distance between reference points is defined as:

[0117]

[0118] where x, y are a given reference point x and a sample point y. By calculating the Euclidean distance, the radial basis function decreases with the increase of the distance between the sample point and the reference point. By calculating the Euclidean distance, the radial basis function decreases with the increase of the distance between the sample point and the reference point. Under the Euclidean distance, given an input vector x and a center vector x ' The calculation formula of the Gaussian kernel in the radial basis function is:

[0119]

[0120] An important parameter of the Gaussian kernel function is the radial basis kernel parameter σ, which is mainly used to control the influence range of the sample on the decision boundary, and different values of σ will affect the smoothness and generalization ability of the kernel function;

[0121] In order to obtain the internal correlation information in the time sequence network, the periodic term and the noise term are added in the kernel function in this embodiment for fitting, so the kernel function for the time sequence periodic network in this embodiment is:

[0122]

[0123] where k(x, x ' ) is the kernel function value between the input point x and the input point x ' in the Gaussian process regression, x and x ' respectively represent the feature vectors mapped in the Gaussian process regression of the two time slice networks, T is the period of the periodic kernel function, λ is the amplitude influence parameter of the periodic kernel, and δ xx' is the noise control term.

[0124] In the specific application process, λ is used to control the intensity of periodicity, σ is used to control the generalization ability of the kernel function, T is used to control the period, and σ xx' is responsible for controlling whether to increase the noise. The kernel function in this embodiment combines the periodic and Gaussian kernel two parts, and can select a specific noise function according to the characteristics of the data set to obtain the expression of the kernel function. If the influence propagation range of the node is obtained by multiple simulations of the IC propagation model and averaging, adding noise can avoid the influence of simulation error on the model performance. For example, Gaussian white noise can be added as a noise term in the kernel function, and the probability density function of the Gaussian white noise is:

[0125]

[0126] where p g (x) is the Gaussian white noise, and σ g is the standard deviation of the Gaussian distribution.

[0127] In multiple time slices, the timing network node has the consistency of node number, and the realistic meaning is that the user in the social network is consistent in different time IDs. In multiple time slices, the kernel function designed in the embodiment fits the main periodic trend in the timing network to achieve the purpose of predicting the influence node set.

[0128] To optimize the hyperparameters of the influence node prediction model, the embodiment adopts the negative log marginal likelihood function as the objective function, and the specific implementation process is as follows:

[0129] Obtain the training sample (X, Y), wherein X={x1,x2,…,x N} is the input feature matrix, and Y={y1,y2,…,y N} is the corresponding output vector;

[0130] Define the negative log marginal likelihood function The difference between the observed data and the model prediction is measured as:

[0131]

[0132] Wherein, θ=(σ,T,λ) is the hyperparameter to be calculated, K is the kernel matrix, and N is the number of samples, that is, the number of observed data;

[0133] Minimize the negative log marginal likelihood function As the objective to optimize the hyperparameter θ, according to the expression of the kernel function, the negative log marginal likelihood function is expanded as:

[0134]

[0135] Wherein, is the partial derivative of the negative log marginal likelihood function with respect to σ, is the partial derivative matrix of the kernel matrix K with respect to σ, and Tr(·) is the trace of the matrix;

[0136] Get the derivative of the kernel function, which is:

[0137]

[0138] Adopt the gradient descent method to iteratively optimize the hyperparameter θ, and in each iteration step, the update process of the hyperparameter θ is as follows:

[0139]

[0140] Wherein, θ new is the updated hyperparameter, θ old is the hyperparameter before updating, and α is the learning rate.

[0141] The hyperparameters directly affect the variance of the prediction distribution, i.e., the uncertainty of the prediction. Reasonable selection of the hyperparameters can improve the accuracy and reliability of the uncertainty evaluation. In the process of optimizing the hyperparameters, whether a node in the future time series network is an influential node is extracted by the law of the influence of the node on the distance. After the optimization of the hyperparameters is completed, the set of influential nodes of the time series network at a certain time point in the future can be predicted based on the influential node prediction model, for example, assuming that k(x i * ,x j * ) is the kernel function value between the training set sample x i * and the test set sample x j * , then the mean μ * and the variance σ *2 of the prediction result are:

[0142] μ * = K T (K+σ 2 I) -1 Y

[0143] σ *2 = k(x i * ,x j * )-K T (K+σ 2 I) -1 K

[0144] According to the predicted mean μ * and variance σ *2 , the part of the time series network with the optimal influence state at a certain time point in the future can be determined to form the set of influential nodes, so that the positive information is released, and the propagation of the induced information can be effectively inhibited.

[0145] It should be noted that although each step in the embodiment Figure 1 is displayed in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the embodiment may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be alternately executed with other steps or at least part of the sub-steps or stages of other steps.

[0146] Embodiment 2

[0147] Based on the propagation suppression method based on influence node prediction in Embodiment 1, this embodiment discloses a propagation suppression system based on influence node prediction, which refers to Figure 3 The propagation suppression system based on influence node prediction includes a time sequence network construction unit, a network feature extraction unit, a node feature extraction unit, a prediction model construction unit, and a propagation suppression unit, specifically:

[0148] The time sequence network construction unit is used to time sequence slice the social network to obtain a plurality of time slice networks.

[0149] The network feature extraction unit is used to extract network features of each time slice network.

[0150] The node feature extraction unit is used to obtain influence node features of each time slice network.

[0151] The prediction model construction unit is used to construct a Gaussian process regression based on the network features and the influence node features of each time slice network, design a kernel function for the time sequence cycle network, obtain an influence node prediction model based on the Gaussian process regression, and calculate the hyperparameters of the influence node prediction model.

[0152] The propagation suppression unit is used to predict the influence nodes of the future time slice network according to the influence node prediction model, and activate the predicted influence nodes for propagation suppression.

[0153] In this embodiment, the specific working process and working principle of the partial time sequence network construction unit, the network feature extraction unit, the node feature extraction unit, the prediction model construction unit, and the propagation suppression unit are the same as those in Embodiment 1, so this embodiment will not be described again. Each unit module can be realized by software, hardware, and their combination, and each unit module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each unit module.

[0154] Embodiment 3

[0155] As Figure 4 shown is a terminal device disclosed in this embodiment, which includes a transmitter, a receiver, a memory, and a processor. The transmitter is used to send instructions and data, the receiver is used to receive instructions and data, the memory is used to store computer execution instructions, and the processor is used to execute the computer execution instructions stored in the memory to realize the method in Embodiment 1.

[0156] It should be noted that the above memory can be independent or integrated with the processor. When the memory is independently arranged, the terminal device further includes a bus for connecting the memory and the processor.

[0157] The above merely describes the preferred embodiments of the present application and is not intended to limit the protection scope of the present application. Any equivalent structure transformation made under the inventive concept of the present application, or direct / indirect application in other related technical fields, is included in the protection scope of the present application.

Claims

1. A propagation suppression method based on influential node prediction, characterized in that: The steps include: Step 1: Time-series slicing of the social network to obtain a number of time-sliced ​​networks, and extracting network features of each of the time-sliced ​​networks; Step 2: Mark the influential nodes in each time slice network to obtain the influential node features of each time slice network; Step 3: constructing a Gaussian process regression based on the network characteristics and influential node characteristics of each time slice network, and designing a kernel function for the time series periodic network to obtain an influential node prediction model based on Gaussian process regression, and optimizing the hyperparameters of the influential node prediction model; Step 4: predict the influential nodes of the future time slice network based on the influential node prediction model, and activate the predicted influential nodes to suppress propagation.

2. The method for suppressing propagation based on influential node prediction according to claim 1, characterized in that: In step 1, a Laplacian matrix is ​​constructed as the network feature of each time slice network, specifically: L=I-D -1 / 2 ·A·D -1 / 2 Where L is the Laplace matrix of the time-slicing network, I is the identity matrix, A is the adjacency matrix of the time-slicing network, and D is the degree matrix of the time-slicing network; Where n is the number of nodes in the time slice network, A ij is an element of the adjacency matrix A, used to represent the node v in the time slice network i With node v j whether they are connected; After constructing the Laplace matrix, the pairwise distance between any two nodes in the time slice network is calculated based on the constructed Laplace matrix, which is: Among them, d ij is the node v in the time slicing network i With node v j The pairwise distance of is the pseudo-inverse of the Laplace matrix L, is a matrix The i-th row and i-th column element of is a matrix The j-th row and j-th column element of is a matrix The i-th row and j-th column element of , where T is the transpose of the matrix.

3. The method for suppressing propagation based on influential node prediction according to claim 1, characterized in that: In step 2, the process of obtaining the influential node features is as follows: Step 201: Obtain the node set V and edge set E of the time slice network, as well as the propagation probability p of each node. v and edge activation probability q v,u , where v,u∈V, and node u is the neighbor node of node v, that is, the network edge e vu ∈E; Step 202: Initialize the influence I of each node v =0, active set and candidate set S = V; Step 203: Calculate the gain of each node in the current candidate set S, which is: Step 203: Select the maximum gain node Step 204: Update the influence of each node in the current candidate set S to: Among them, A v is the set of neighbor nodes of node v; Step 205, let S * =S * ∪{v * }、S=S\{v * |After that, determine whether the iteration termination condition is met: If so, based on the current active set S * Construct influential node features; Otherwise, return to step 203.

4. The method for suppressing propagation based on influential node prediction according to claim 3, characterized in that: In step 205, the iteration termination condition is or ∑ v∈V I v >I max , where I max is the preset influence threshold.

5. The method for suppressing propagation based on influential node prediction according to claim 1, 2, 3 or 4, characterized in that: In step 3, the kernel function for the sequential periodic network is specifically: Among them, k(x,x ' ) is the input point x and the input point x in Gaussian process regression ' The kernel function value between x and x ' Respectively represent the eigenvectors of the two time-slicing networks mapped in the Gaussian process regression, σ is the radial basis kernel parameter, T is the period of the periodic kernel function, λ is the amplitude influence parameter of the periodic kernel, δ xx' is the noise control item.

6. The method for suppressing propagation based on influential node prediction according to claim 1, 2, 3 or 4, characterized in that: In step 3, the process of calculating the hyperparameters of the influential node prediction model is specifically as follows: Get training samples (X, Y), where X = {x1, x2, ..., x N } is the input feature matrix, Y={y1,y2,…,y N } is the input feature matrix and the corresponding output vector; Define the negative log marginal likelihood function A measure of the difference between the observed data and the model predictions, which is: Where θ = (σ, t, λ) is the hyperparameter to be calculated, K is the kernel matrix, and N is the number of samples; To minimize the negative log marginal likelihood function To optimize the hyperparameter θ for the target, expand the negative log marginal likelihood function according to the expression of the kernel function for: in, is the partial derivative of the negative log-marginal likelihood function with respect to σ, is the partial derivative matrix of the kernel matrix K with respect to σ, and Tr(·) is the trace of the matrix; Get the derivative of the kernel function, which is: The hyperparameter θ is iteratively optimized using the gradient descent method. In each iterative step, the update process of the hyperparameter θ is: Among them, θ new is the updated hyperparameter, θ old is the hyperparameter before updating, and α is the learning rate.

7. A propagation suppression system based on influential node prediction, characterized in that: Propagation suppression is performed using the method according to any one of claims 1 to 6, wherein the propagation suppression system comprises: A time-series network construction unit is used to perform time-series slicing on the social network to obtain a number of time-sliced ​​networks; A network feature extraction unit, configured to extract network features of each of the time slice networks; A node feature extraction unit, configured to obtain influential node features of each of the time slice networks; A prediction model construction unit is used to construct a Gaussian process regression based on the network characteristics and influential node characteristics of each of the time slice networks, design a kernel function for the time series periodic network, obtain an influential node prediction model based on Gaussian process regression, and calculate the hyperparameters of the influential node prediction model; The propagation suppression unit is used to predict the influential nodes of the future time slice network according to the influential node prediction model, and activate the predicted influential nodes to perform propagation suppression.

8. A terminal device, characterized in that: The terminal device is provided with: Memory, used to store programs; A processor is configured to execute the program stored in the memory, wherein when the program is executed, the processor is configured to execute the method according to any one of claims 1 to 6.

Images