Source tracing method for social network multi-source information propagation

Abstract

The invention discloses a source tracing method for social network multi-source information propagation, and the method comprises the steps: building a node set evaluation unit, employing a rough sampling-refining strategy of overall sampling, and effectively capturing the synergistic effect among multiple propagation sources; a Bayesian optimization framework fused with the graph convolutional network is provided, so that network topology characteristics and propagation dynamic characteristics can be integrated at the same time in the uncertainty modeling process, and the precision of multi-source positioning in the social network and the practical application robustness of the method are remarkably improved. Through a three-stage optimization mechanism of coarse-grained sampling-Fourier refining-probability evaluation, positioning of multiple information sources in a complex network information propagation scene is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of social network analysis, and in particular relates to a source tracing method for multi-source information dissemination in a social network. Background Art

[0002] With the rapid development of the Internet, social networks, based on social media platforms, have become the most widely used information dissemination medium in the community, greatly facilitating people's lives. Therefore, methods for tracing the source of information dissemination on social networks are extremely meaningful.

[0003] In recent years, various methods have been proposed from different perspectives to locate the source of message propagation. These methods can be roughly divided into three categories: complete graph-based methods, observation node-based methods, and propagation snapshot-based methods. Complete graph-based methods trace the source of message propagation within the complete network structure by utilizing techniques such as infection path tracing, graph centrality metrics, or likelihood maximization in epidemic diffusion models. Complete graph-based methods can achieve high accuracy when the network structure is fully observable. However, in practical scenarios, these methods are less applicable due to the large computational requirements and the difficulty of obtaining complete network observations. The core strategy of observation node-based methods is to deploy multiple nodes as observation points of propagation status to strategically monitor the dynamic changes of information diffusion. The collected observation data can be used to infer the source location based on minimum distance estimation or signal arrival time. However, their effectiveness is highly dependent on the selection of observation nodes, and different locations and numbers of observation nodes produce different traceability results. This characteristic makes them less reliable in dynamic or sparsely monitored networks. The most widely studied current methods are propagation snapshot-based methods, which perform the task of information source localization by analyzing snapshots of node status or information propagation captured at specific time intervals during the propagation process. In most cases, machine learning models are used to extract features from propagation snapshots to capture diffusion dynamics, which has better flexibility than the first two methods.

[0004] However, existing methods still have some shortcomings in source localization:

[0005] (1) The inherent uncertainty problem in information dissemination dynamics. During the information dissemination process, the same information source may produce different snapshots over time. Conversely, the same snapshot may come from different information sources. Therefore, the inherent uncertainty in the information dissemination process will seriously affect the accuracy and effectiveness of information source tracing.

[0006] (2) In multi-source communication, due to the complex interactions within the network, multiple information dissemination sources are not independent of each other during the dissemination process, but will produce interactions that are difficult to quantify. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this paper provides a method for tracing the origins of multi-source information dissemination in social networks. By establishing a node set evaluation unit and employing a coarse sampling-refinement strategy based on holistic sampling, this method effectively captures the synergy between multiple dissemination sources. A Bayesian optimization framework fused with a graph convolutional network is proposed, enabling the uncertainty modeling process to simultaneously integrate network topology characteristics and dissemination dynamics, significantly improving the accuracy of multi-source localization in social networks and the method's practical robustness. Through a three-stage optimization mechanism of "coarse-grained sampling-Fourier refinement-probabilistic evaluation," this method achieves the localization of multiple information sources in complex network information dissemination scenarios.

[0008] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0009] Step 1: Obtain the initial filtered node set from the social network propagation subgraph;

[0010] Step 2: Initialize the optimal candidate set of sources;

[0011] Step 3: Sample from the node set to obtain a sampling set;

[0012] Step 4: Perform Fourier domain refinement on the sample set;

[0013] Step 5: Training of the surrogate model using the Bayesian optimization framework;

[0014] Step 6: Update the optimal candidate set of the source;

[0015] Step 7: Determine whether the maximum number of iterations has been reached;

[0016] Step 8: Identify the information dissemination source node from the optimal candidate set.

[0017] Preferably, the step 1 is specifically:

[0018] According to different information propagation modes, all node sets in the social network propagation subgraph are Perform initial filtering to obtain the filtered node set in the social network propagation subgraph ;

[0019] The information propagation mode is divided into SI (Susceptible-Infected) propagation mode and SIR (Susceptible-Infected-Recovered) propagation mode. The specific filtering methods are as follows:

[0020] (1)

[0021] When the SI propagation mode is adopted, the set of uninfected nodes that do not participate in the propagation process is removed. Get the filtered node set ; When the SIR propagation mode is used, the number of connected edges of each node, i.e., the degree, is counted, and then the node set with a degree of 0 is removed Get the filtered node set .

[0022] Preferably, the step 2 is specifically as follows:

[0023] From the filtered node set Randomly sample several groups of node sets in the propagation source, and use these node sets as propagation sources to simulate propagation; compare the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph, and select the first several node sets with the highest similarity as the optimal candidate set X of the initial source; the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph is calculated as follows:

[0024] (2)

[0025] Where n represents the number of nodes, represents the number of infected nodes in the observed subgraph obtained by simulating the spread, represents the number of infected nodes in the actual observed subgraph, represents the number of infected nodes in the intersection of the observed subgraph obtained by simulated propagation and the actual observed subgraph, represents the observation subgraph obtained by simulating propagation, Represents the actual observation subgraph.

[0026] Preferably, the step 3 is specifically:

[0027] Node collection Using uniform distribution function Perform p random samplings to obtain a coarse sampling set containing p groups of node sets , , i=1,…,p; Represents the filtered node set The p groups of node sets obtained by random sampling are represented as the first group of node sets, the second group of node sets, and so on until the pth group of node sets; Indicates the minimum number of node sets, Indicates the maximum number of nodes in a set.

[0028] Preferably, the step 4 is specifically as follows:

[0029] Step 4-1: Fourier transform stage;

[0030] First, the coarse sampling set The optimal candidate set X of the source and the source are respectively subjected to Fourier transform. The graph domain node set is mapped to the Fourier domain through the graph Laplace matrix L to obtain the frequency domain feature representation. The Fourier transform process is as follows:

[0031]

[0032] (3)

[0033] in, represents the inverse matrix of the graph Laplacian matrix, Represents a coarse sampling set The result of Fourier transform is The result after Fourier transform of the optimal candidate set X of the source;

[0034] Step 4-2: Similarity screening stage;

[0035] In the Fourier domain, we calculate and The Euclidean distance similarity SIM(.) is used to select the q node sets with the highest similarity to obtain the fine sampling set D as the set of node sets obtained by the final sampling. <p;

[0036] Specifically, Each node set in , calculate its The minimum similarity of all samples in The node sets in are sorted in descending order by similarity score, and the first q node sets are selected. The similarity measure is calculated as follows:

[0037] (4)

[0038] Where m is The number of node sets in represents the set of q nodes with the highest similarity, express The jth sample in .

[0039] Preferably, the step 5 is specifically as follows:

[0040] Multi-layer graph convolutional networks (GCNs) are used as the proxy model for the Bayesian optimization framework. The hierarchical propagation rule of the multi-layer graph convolutional network is defined as:

[0041] (5)

[0042] (6)

[0043] in and denote the input and output of layer l respectively, is the graph convolution operator after symmetric normalization, represents the adjacency matrix of the network, represents the degree matrix of the network, σ(·) represents the ReLU activation matrix, is the matrix of trainable parameters;

[0044] The multi-layer propagation process of the proxy model is as follows:

[0045] (7)

[0046] Among them, softmax(·) represents the standard normalization function, Represents the final output of the proxy model, represents the weight matrix of the first layer of graph convolutional network, Represents the weight matrix of the second layer graph convolutional network;

[0047] In the Bayesian optimization framework, the proxy model is trained using the optimal candidate set X of the source. The trained proxy model uses the fine sampling set D obtained in step 4 as input. The proxy model calculates the score Y of each sample in the sampling set, and the sample with the highest score is selected as the optimal subset of the sampling set.

[0048] Preferably, the step 6 is specifically:

[0049] Assume that the optimal subset obtained in step 5 is the information propagation source set, simulate the propagation process, and calculate the similarity between the simulated propagation snapshot and the real propagation process to obtain the likelihood score of this optimal subset;

[0050] The optimal subset and its likelihood score are added to the optimal candidate set X of the source to update the optimal candidate set of the source.

[0051] Preferably, the step 7 is specifically as follows:

[0052] Check whether the preset maximum number of iterations has been reached. If so, proceed to step 8; if not, loop through steps 3 to 6.

[0053] Preferably, the step 8 is specifically as follows:

[0054] The node set with the largest likelihood score in the final optimal candidate set X of the source is selected and regarded as the final located information propagation source.

[0055] The beneficial effects of the present invention are as follows:

[0056] 1. In the process of evaluating whether a node is a source node, the present invention adopts an overall sampling mechanism that uses the node set as the basic evaluation unit, takes the interaction of multiple propagation sources in the diffusion process into consideration, and improves the traceability accuracy in multi-source propagation.

[0057] 2. By integrating the proxy model design of the graph convolutional network, the present invention enables the uncertainty modeling process to simultaneously utilize the network topology structure and propagation dynamics characteristics, thereby improving the accuracy and robustness of multi-source positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall flow chart of the method of the present invention;

[0059] Figure 2 It is the overall framework diagram of the method of the present invention;

[0060] Figure 3 The test results of the method of the present invention and the existing methods on different social networks under the SI propagation mode (HSGBO is the method of the present invention).

[0061] Figure 4 The test results of the method of the present invention and the existing methods on different social networks under the SIR propagation mode (HSGBO is the present method);

[0062] Figure 5(a) shows the parameter (hidden unitsizes) analysis of the surrogate model in the Bayesian framework on the synthetic network dataset SW.

[0063] Figure 5(b) shows the parameter (Hiddenunit sizes) analysis of the proxy model in the Bayesian framework on the real network dataset Cora.

[0064] Figure 6 (a) shows the comparison of the operating efficiency of the proposed method (HSGBO) and the existing methods on the synthetic network dataset SW;

[0065] Figure 6 (b) shows the comparison of the operating efficiency of the proposed method (HSGBO) and existing methods on the real network dataset Cora. DETAILED DESCRIPTION

[0066] The present invention will be further described below with reference to the accompanying drawings and examples.

[0067] This paper proposes a method for tracing information origins in scenarios where information propagates from multiple sources on social networks. Based on a Bayesian optimization framework, it constructs a probabilistic inference model that can, to a certain extent, characterize the uncertainty of propagation dynamics. This method uses a holistic assessment approach to account for the complex interactions in multi-source propagation, improving the accuracy of tracing information in multi-source localization scenarios.

[0068] like Figure 1 and Figure 2 As shown, a source tracing method for multi-source information dissemination in social networks is provided, and the technical solution is as follows:

[0069] Step 1: Obtain the initial filtered node set from the social network propagation subgraph;

[0070] According to different information propagation modes, all node sets in the social network propagation subgraph are Perform initial filtering to obtain the filtered node set in the social network propagation subgraph ;

[0071] The information propagation mode is divided into SI (Susceptible-Infected) propagation mode and SIR (Susceptible-Infected-Recovered) propagation mode. The specific filtering methods are as follows:

[0072] (1)

[0073] When the SI propagation mode is adopted, the set of uninfected nodes that do not participate in the propagation process is removed. Get the filtered node set ; When the SIR propagation mode is used, the number of connected edges of each node, i.e., the degree, is counted, and then the node set with a degree of 0 is removed Get the filtered node set .

[0074] Step 2: Initialize the optimal candidate set of sources;

[0075] From the filtered node set Randomly sample several groups of node sets in the propagation source, and use these node sets as propagation sources to simulate propagation; compare the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph, and select the first several node sets with the highest similarity as the optimal candidate set X of the initial source; the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph is calculated as follows:

[0076] (2)

[0077] Where n represents the number of nodes, represents the number of infected nodes in the observed subgraph obtained by simulating the spread, represents the number of infected nodes in the actual observed subgraph, represents the number of infected nodes in the intersection of the observed subgraph obtained by simulated propagation and the actual observed subgraph, represents the observation subgraph obtained by simulating propagation, Represents the actual observation subgraph.

[0078] Step 3: Sample from the node set to obtain a sampling set;

[0079] Node collection Using uniform distribution function Perform p random samplings to obtain a coarse sampling set containing p groups of node sets , , i=1,…,p; Represents the filtered node set The p groups of node sets obtained by random sampling are represented as the first group of node sets, the second group of node sets, and so on until the pth group of node sets; Indicates the minimum number of node sets, Indicates the maximum number of nodes in a set.

[0080] Step 4: Perform Fourier domain refinement on the sample set;

[0081] Step 4-1: Fourier transform stage;

[0082] First, the coarse sampling set The optimal candidate set X of the source and the source are respectively subjected to Fourier transform. The graph domain node set is mapped to the Fourier domain through the graph Laplace matrix L to obtain the frequency domain feature representation. The Fourier transform process is as follows:

[0083]

[0084] (3)

[0085] in, represents the inverse matrix of the graph Laplacian matrix, Represents a coarse sampling set The result of Fourier transform is The result after Fourier transform of the optimal candidate set X of the source;

[0086] Step 4-2: Similarity screening stage;

[0087] In the Fourier domain, we calculate and The Euclidean distance similarity SIM(.) is used to select the q node sets with the highest similarity to obtain the fine sampling set D as the set of node sets obtained by the final sampling. <p;

[0088] Specifically, Each node set in , calculate its The minimum similarity of all samples in The node sets in are sorted in descending order by similarity score, and the first q node sets are selected. The similarity measure is calculated as follows:

[0089] (4)

[0090] Where m is The number of node sets in represents the set of q nodes with the highest similarity, express The jth sample in .

[0091] Step 5: Training of the surrogate model using the Bayesian optimization framework;

[0092] Multi-layer graph convolutional networks (GCNs) are used as the proxy model for the Bayesian optimization framework. The hierarchical propagation rule of the multi-layer graph convolutional network is defined as:

[0093] (5)

[0094] (6)

[0095] in and denote the input and output of layer l respectively, is the graph convolution operator after symmetric normalization, represents the adjacency matrix of the network, represents the degree matrix of the network, σ(·) represents the ReLU activation matrix, is the matrix of trainable parameters;

[0096] The multi-layer propagation process of the proxy model is as follows:

[0097] (7)

[0098] Among them, softmax(·) represents the standard normalization function, Represents the final output of the proxy model, represents the weight matrix of the first layer of graph convolutional network, Represents the weight matrix of the second layer graph convolutional network;

[0099] In the Bayesian optimization framework, the proxy model is trained using the optimal candidate set X of the source. The trained proxy model uses the fine sampling set D obtained in step 4 as input. The proxy model calculates the score Y of each sample in the sampling set, and the sample with the highest score is selected as the optimal subset of the sampling set.

[0100] Step 6: Update the optimal candidate set of the source;

[0101] Assume that the optimal subset obtained in step 5 is the information propagation source set, simulate the propagation process, and calculate the similarity between the simulated propagation snapshot and the real propagation process to obtain the likelihood score of this optimal subset;

[0102] The optimal subset and its likelihood score are added to the optimal candidate set X of the source to update the optimal candidate set of the source.

[0103] Step 7: Determine whether the maximum number of iterations has been reached;

[0104] Check whether the preset maximum number of iterations has been reached. If so, proceed to step 8; if not, loop through steps 3 to 6.

[0105] Step 8: Identify the information dissemination source node from the optimal candidate set.

[0106] The node set with the largest likelihood score in the final optimal candidate set X of the source is selected and regarded as the final located information propagation source.

[0107] Finally, two core indicators are used to evaluate the multi-source localization performance.

[0108] (1) Use the distance error score (Serr) as the main evaluation criterion, which is calculated by calculating the set S of inferred information sources and the set of real information sources Minimize Dijkstra path distance between , and normalized based on the network diameter and set size, the mathematical formula is:

[0109]

[0110] Among them, P represents all possible permutations and combinations of S, and L takes S and The maximum value of the set size, d(·) calculation and arrangement The shortest path distance of the corresponding node in is the diameter of the network G, that is, the maximum value of the shortest path length between all pairs of nodes in the graph; the range of the distance error score indicator is [0,1], and the closer it is to 1, the higher the accuracy of the source localization algorithm.

[0111] (2) The average distance error (Derr) is used as the evaluation index, and the set of predicted source nodes to the real source nodes is calculated. The average value of the shortest path distance to the nearest node in the network can provide a more intuitive performance evaluation compared to the evaluation index (1). Its mathematical formula is:

[0112]

[0113] This indicator directly reflects the physical distance deviation between the predicted source node and the real source. The smaller the value, the more accurate the positioning.

[0114] like Figure 3 、 Figure 4 ,Figure 5(a),Figure 5(b),Evaluation index Average Distance Error (i.e. Figure 3 and Figure 4 The lower the value of Derr in Distance Error Score (i.e. Figure 3 and 4 The higher the Serr value in , the better the model effect.

[0115] Figure 6 (a) and Figure 6 (b) are the comparisons of the operating efficiency of the proposed method (HSGBO) and the existing methods on the synthetic network dataset SW and the real network dataset Cora, respectively.

[0116] Unlike traditional source localization methods based on single-node evaluation and existing probabilistic graphical modeling approaches, this paper establishes a node-set evaluation unit and innovatively employs a holistic sampling coarse-grained sampling-refinement strategy to effectively capture the synergy between multiple propagation sources. It also proposes a Bayesian optimization framework that integrates graph convolutional networks, enabling the uncertainty modeling process to simultaneously integrate network topology and propagation dynamics, significantly improving the accuracy and practical robustness of multi-source localization in social networks. Through a three-stage optimization mechanism of "coarse-grained sampling-Fourier refinement-probabilistic evaluation," this paper achieves the localization of multiple information sources in complex network information propagation scenarios.

Claims

1. A method for tracing the source of multi-source information dissemination in a social network, characterized by: The steps include: Step 1: Obtain the initial filtered node set from the social network propagation subgraph; Step 2: Initialize the optimal candidate set of sources; Step 3: Sample from the node set to obtain a sampling set; Step 4: Perform Fourier domain refinement on the sample set; Step 5: Training of the surrogate model using the Bayesian optimization framework; Step 6: Update the optimal candidate set of the source; Step 7: Determine whether the maximum number of iterations has been reached; Step 8: Identify the information dissemination source node from the optimal candidate set.

2. A method for tracing the source of multi-source information dissemination in a social network according to claim 1, characterized in that: The step 1 is specifically as follows: According to different information propagation modes, all node sets in the social network propagation subgraph are Perform initial filtering to obtain the filtered node set in the social network propagation subgraph ; The information propagation mode is divided into SI (Susceptible-Infected) propagation mode and SIR (Susceptible-Infected-Recovered) propagation mode. The specific filtering methods are as follows: (1) When the SI propagation mode is adopted, the set of uninfected nodes that do not participate in the propagation process is removed. Get the filtered node set ; When the SIR propagation mode is used, the number of connected edges of each node, i.e., the degree, is counted, and then the node set with a degree of 0 is removed Get the filtered node set .

3. A method for tracing the source of multi-source information dissemination in a social network according to claim 2, characterized in that: The step 2 is specifically as follows: From the filtered node set Randomly sample several groups of node sets in the propagation source, and use these node sets as propagation sources to simulate propagation; compare the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph, and select the first several node sets with the highest similarity as the optimal candidate set X of the initial source; the similarity between the observed subgraph obtained by simulated propagation and the actual observed subgraph is calculated as follows: (2) Where n represents the number of nodes, represents the number of infected nodes in the observed subgraph obtained by simulating the spread, represents the number of infected nodes in the actual observed subgraph, represents the number of infected nodes in the intersection of the observed subgraph obtained by simulated propagation and the actual observed subgraph, represents the observation subgraph obtained by simulating propagation, Represents the actual observation subgraph.

4. A method for tracing the source of multi-source information dissemination in a social network according to claim 3, characterized in that: The step 3 is specifically as follows: Node collection Using uniform distribution function Perform p random samplings to obtain a coarse sampling set containing p groups of node sets , , i=1,…,p; Represents the filtered node set The p groups of node sets obtained by random sampling are represented as the first group of node sets, the second group of node sets, and so on until the pth group of node sets; Indicates the minimum number of node sets, Indicates the maximum number of nodes in a set.

5. A method for tracing the source of multi-source information dissemination in a social network according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4-1: Fourier transform stage; First, the coarse sampling set Perform Fourier transform on the optimal candidate set X of the source and the source respectively; map the graph domain node set to the Fourier domain through the graph Laplace matrix L to obtain the frequency domain feature representation; The Fourier transform process is as follows: ； (3) in, represents the inverse matrix of the graph Laplacian matrix, Represents a coarse sampling set The result of Fourier transform is The result after Fourier transform of the optimal candidate set X of the source; Step 4-2: Similarity screening stage; In the Fourier domain, we calculate and The Euclidean distance similarity SIM(.) is used to select the q node sets with the highest similarity to obtain the fine sampling set D as the set of node sets obtained by the final sampling. <p; Specifically, Each node set in , calculate its The minimum similarity of all samples in The node sets in are sorted in descending order by similarity score, and the first q node sets are selected. The similarity measure is calculated as follows: (4) Where m is The number of node sets in represents the set of q nodes with the highest similarity, express The jth sample in .

6. A method for tracing the source of multi-source information dissemination in a social network according to claim 5, characterized in that: The step 5 is specifically as follows: Multi-layer graph convolutional networks (GCNs) are used as the proxy model for the Bayesian optimization framework. The hierarchical propagation rule of the multi-layer graph convolutional network is defined as: (5) (6) in and denote the input and output of layer l respectively, is the graph convolution operator after symmetric normalization, represents the adjacency matrix of the network, represents the degree matrix of the network, σ(·) represents the ReLU activation matrix, is the matrix of trainable parameters; The multi-layer propagation process of the proxy model is as follows: (7) Among them, softmax(·) represents the standard normalization function, Represents the final output of the proxy model, represents the weight matrix of the first layer of graph convolutional network, Represents the weight matrix of the second layer graph convolutional network; In the Bayesian optimization framework, the proxy model is trained using the optimal candidate set X of the source. The trained proxy model uses the fine sampling set D obtained in step 4 as input. The proxy model calculates the score Y of each sample in the sampling set, and the sample with the highest score is selected as the optimal subset of the sampling set.

7. A method for tracing the source of multi-source information dissemination in a social network according to claim 6, characterized in that: The step 6 is specifically as follows: Assume that the optimal subset obtained in step 5 is the information propagation source set, simulate the propagation process, and calculate the similarity between the simulated propagation snapshot and the real propagation process to obtain the likelihood score of this optimal subset; The optimal subset and its likelihood score are added to the optimal candidate set X of the source to update the optimal candidate set of the source.

8. A method for tracing the source of multi-source information dissemination in a social network according to claim 7, characterized in that: The step 7 is specifically as follows: Check whether the preset maximum number of iterations has been reached. If so, proceed to step 8; if not, loop through steps 3 to 6.

9. A method for tracing the source of multi-source information dissemination in a social network according to claim 8, characterized in that: The step 8 is specifically as follows: The node set with the largest likelihood score in the final optimal candidate set X of the source is selected and regarded as the final located information propagation source.

Images