A rumor propagation source tracing method and device

Abstract

The application discloses a rumor propagation source tracing method and device, and relates to the technical field of propagation dynamics. To solve the problem that the existing rumor propagation source tracing method has high complexity of users and propagation information, and imbalance between source nodes and non-source nodes. The method comprises the following steps: when it is determined that a user node in a social network is infected by a rumor, a social network snapshot of the social network is acquired, and a first feature vector represented by a user node is constructed based on the social network snapshot; according to the first feature vector of the user node and a trainable parameter matrix included in a graph attention network, an attention coefficient of each neighbor node around the user node and a first output feature vector of the user node through the graph attention network are determined; and based on a loss function, a training loss of a final output feature vector of the user node is determined, and when the training loss converges, a rumor propagation source is determined.

Description

Technical Field

[0001] This invention relates to the field of propagation dynamics technology, and more specifically to a method and apparatus for tracing the source of rumor propagation. Background Technology

[0002] With the development of society and technology, the emergence of various social media platforms has brought great convenience; at the same time, the spread of rumors has brought great obstacles to social development, and rumors that cause significant social impact are often caused by one or several sources of dissemination; therefore, it is of great significance to locate the source of rumors in a timely and effective manner and thus control the dissemination process.

[0003] Real-world transmission phenomena can be simulated using different transmission models; for example, infection-based models include the SI (Susceptible-Infected) model, the SIR (Susceptible-Infected-Recovery) model, and the SIS (Susceptible-Infected-Susceptible) model. Impact-based transmission models include the IC (Independent Cascade) model and the LT (Linear Threshold) model.

[0004] For non-machine learning methods, Pinto et al. proposed a method to locate the source of spread by setting up "observation points" in social networks to record the infection time. Specifically, they first selected a group of users in the social network as observation points, and recorded the infection time of these users after the rumor began to spread, recording it as an observation time vector. During the localization process, they traversed every user in the social network, assuming they were the source of spread and triggered the spread process. The time it took for the rumor to spread to each observation point formed another theoretical time vector. They used a multivariate normal probability density function to calculate the similarity between these two vectors, and the user with the highest similarity was considered the source of spread. Wang et al., based on the SourceCentrality Theory, proposed a label value iteration method to locate the source of spread by locally maximizing the label value of the user in the center of the infected area. Specifically, they stopped spreading when a certain proportion of users in the social network were infected, obtained a snapshot of the social network at this time, and assigned a label value to each user based on their state (+1 for infected users, -1 for uninfected users). They then iterated the label value for each user in the social network until the label value of each user in the social network no longer changed, and selected the user with the largest local label value as the source of spread.

[0005] For machine learning methods, Dong et al. first proposed a machine learning-based propagation source tracing method. Specifically, propagation stops when a certain percentage of users in a social network are infected, a snapshot of the social network at this point is obtained, user features are calculated using the LPSI algorithm, and the resulting feature vectors are input into a neural network model. The propagation source is then selected using classification in the final layer of the model. By learning the propagation process of information on social networks, Wang et al. proposed the IVGD method. Specifically, it first learns the propagation process of information on social networks, then learns the propagation source tracing process as the inverse of the propagation process, and finally introduces an error compensation mechanism among the calculated users to select the true propagation source.

[0006] In summary, existing methods for tracing the source of rumor dissemination suffer from problems such as high complexity of users and disseminated information, and class imbalance between source nodes and non-source nodes. Summary of the Invention

[0007] This invention provides a method and apparatus for tracing the source of rumor propagation, which addresses the problems of high complexity of users and propagation information and class imbalance between source nodes and non-source nodes in existing rumor propagation source tracing methods.

[0008] This invention provides a method for tracing the source of rumor propagation, including:

[0009] When it is determined that a user node in the social network has been infected by rumors, a social network snapshot of the social network is obtained, and a first feature vector represented by user nodes is constructed based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information.

[0010] Based on the first feature vector of the user node and the trainable parameter matrix included in the graph attention network, determine the attention coefficient of each neighbor node around the user node and the first output feature vector of the user node through the graph attention network.

[0011] Delete the first output feature vector using the Xdrop operation. The elements of % are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to the multi-head graph attention network to obtain the final output feature vector of the user node.

[0012] The training loss is determined based on the loss function to identify the final output feature vector of the user node. When the training loss converges, the source of the rumor propagation is determined.

[0013] Preferably, the first feature vector is:

[0014]

[0015] in, This represents the first feature vector of user node i. This represents the state information of user node i. This represents the status information of the neighboring nodes of user node i. This represents the social network structure information of user node i. Represents the rumor propagation information of user node i; the || symbol represents vector concatenation, R is a positive integer, and 0 <R<5。

[0016] Preferably, the neighbor node status information is:

[0017]

[0018] in, Y represents the state information of the neighboring nodes of user node i, and Y represents the user state in the social network snapshot. This indicates that user j was infected by the rumor, and This indicates that user J was not influenced by the rumors. This represents all neighboring nodes of user node i. This represents the number of all neighbors of user node i.

[0019] Preferably, the social network structure information is as follows:

[0020]

[0021] in, This represents the social network structure information of user node i. Indicates degree centrality. V represents the set of user nodes in the network.

[0022] Preferably, the user node status information is:

[0023]

[0024] The rumor propagation information of the user nodes:

[0025] in, This represents the state information of user node i. Y represents the rumor propagation information of user node i, and Y represents the state of user node i in the social network snapshot. This indicates that user node i has been infected by rumors. This indicates that user node i is not infected by the rumor, and T represents the timestamp vector of each user node i being infected. This indicates the time when user node i was infected by the rumor.

[0026] Preferably, the Xdrop operation is as follows:

[0027]

[0028] in, This represents the first output feature vector. , This represents the second output feature vector. This indicates that the feature vector is extracted from the first output feature vector. %, Represents a positive integer less than 10;

[0029] The step of concatenating and averaging the second output feature vectors based on the multi-head graph attention network to obtain the final output feature vector of the user node specifically includes:

[0030]

[0031]

[0032] in, Let represent the final output feature vector of user node i, and σ(·) represent the activation function. This represents all neighboring nodes of user node i. W represents the trainable parameter matrix included in the graph attention network, and a represents the weight vector. Let K represent all neighboring nodes of user node i, and K represent that the attention mechanism has K heads.

[0033] Preferably, the loss function is:

[0034]

[0035] Where L represents the cross-entropy loss, and V represents the set of user nodes in the network. Represents user node i on the network. Represents user node j on the network. This represents the loss value of user node i during this training. This represents the loss of user node j during this training session. This represents the 2-norm of the parameter matrix W. It equals 0.0005. This indicates the number of sources of rumor transmission. This indicates the number of non-sources involved in the spread of rumors.

[0036] This invention also provides a rumor source tracing device, comprising:

[0037] The construction unit is used to obtain a social network snapshot of the social network when it is determined that a user node in the social network is infected by rumors, and to construct a first feature vector represented by user nodes based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information.

[0038] The determining unit is used to determine the attention coefficients of each neighboring node around the user node and the first output feature vector of the user node through the graph attention network, based on the first feature vector of the user node and the trainable parameter matrix included in the graph attention network.

[0039] The first obtaining unit is used to delete features from the first output feature vector through the Xdrop operation. The elements of % are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to the multi-head graph attention network to obtain the final output feature vector of the user node.

[0040] The second unit is used to determine the training loss of the final output feature vector of the user node based on the loss function. When the training loss converges, the source of the rumor propagation is determined.

[0041] This invention also provides a computer device, which includes a processor and a memory; the memory is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the computer device executes the above-described rumor propagation tracing method.

[0042] This invention also provides a computer-readable storage medium including computer instructions that, when executed on a computer device, cause the computer device to perform the aforementioned rumor propagation tracing method.

[0043] This invention provides a method and apparatus for tracing the source of rumor propagation. The method includes: when it is determined that a user node in a social network is infected by a rumor, acquiring a snapshot of the social network, and constructing a first feature vector represented by the user node based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information; determining the attention coefficient of each neighbor node around the user node and the first output feature vector of the user node through the single-layer graph attention network based on the first feature vector of the user node and the trainable parameter matrix included in the single-layer graph attention network; and deleting the first output feature vector by performing an Xdrop operation. The second output feature vector is obtained by concatenating and averaging the elements of the first feature vector. The final output feature vector of the user node is obtained by using multiple independent attention mechanisms. The training loss of the final output feature vector of the user node is determined based on a loss function. When the training loss converges, the source of the rumor propagation is identified. This method fully considers the user node state information, neighbor node state information, social network structure information, and rumor propagation information in the social network. A first feature vector representing the user node is constructed based on a social network snapshot. The construction of the first feature vector considers the different characteristics of user nodes and rumor propagation information, thus solving the problem of high complexity of users and propagation information in reality. Furthermore, the first feature vector is trained through a multi-layer graph attention network to obtain the final output feature vector of the user node, effectively solving the multi-source localization problem. Further, a loss function is introduced to address the imbalance between existing source nodes and non-source nodes, balancing the imbalance between source nodes and non-source nodes corresponding to the final output feature vector of the user node, thus improving the accuracy of rumor propagation source localization on the social network. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of a method for tracing the source of rumor propagation provided in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the rumor propagation tracing method framework provided in an embodiment of the present invention;

[0047] Figure 3 A schematic diagram showing a comparison of test results between the rumor propagation tracing method provided in this embodiment of the invention and existing methods on different social networks;

[0048] Figure 4 A schematic diagram illustrating the test results of the rumor propagation tracing method provided in this embodiment of the invention under different propagation models;

[0049] Figure 5 A schematic diagram illustrating the test results of the rumor propagation tracing method provided in this embodiment of the invention and existing methods under different data noise conditions;

[0050] Figure 6 This is a schematic diagram of a rumor propagation source tracing device provided in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The source of a spread refers to the node or group of nodes that first initiate a spread process on the internet. Accurately locating the source of a spread on the internet not only allows for direct control of the spread process at its origin but also helps to reconstruct the entire process, thus aiding in a better understanding of online spread phenomena. In the real world, we face numerous issues requiring source tracing, such as rumors on social networks (Facebook or Twitter), computer viruses on the internet, and various infectious diseases, all of which require timely and accurate identification of their origins.

[0053] Existing rumor tracing methods all have the following shortcomings: non-machine learning methods have the problem of lower computational complexity but also relatively lower accuracy; machine learning methods generally face the problem of class imbalance.

[0054] To address the aforementioned issues, this invention proposes a heuristic rumor propagation source tracing framework based on Graph Neural Networks (GNNs) on social networks, which to some extent resolves the contradiction between accuracy and model complexity faced by existing methods.

[0055] Figure 1 An exemplary schematic diagram of a rumor source tracing method provided by an embodiment of the present invention is shown. This method can be applied to at least various source tracing methods, such as rumor source tracing, disease source tracing, etc.

[0056] This invention provides a method for tracing the source of rumor propagation. It proposes a heuristic framework for tracing the source of rumor propagation on social networks based on graph neural networks, which to some extent resolves the contradiction between accuracy and model complexity faced by existing methods. Specifically, the method includes the following steps:

[0057] Step 101: When it is determined that a user node in the social network has been infected by rumors, obtain a social network snapshot of the social network, and construct a first feature vector represented by user nodes based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information.

[0058] Step 102: Based on the first feature vector of the user node and the trainable parameter matrix included in the single-layer graph attention network, determine the attention coefficient of each neighboring node around the user node and the first output feature vector of the user node through the single-layer attention network.

[0059] Step 103: Delete features from the first output feature vector using the Xdrop operation. The elements of % are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to multiple independent attention mechanisms to obtain the final output feature vector of the user node.

[0060] Step 104: Determine the training loss of the final output feature vector of the user node based on the loss function. When the training loss converges, determine the source of the rumor propagation.

[0061] Figure 2 This is a schematic diagram of the rumor propagation tracing method framework provided in an embodiment of the present invention. The above four steps can be based on... Figure 2 The process is divided into three stages: the construction stage of the first feature vector, the training stage of the convolution module, and the training stage of the loss function. Step 101 introduces the construction stage of the first feature vector, steps 102 and 103 introduce the training stage of the convolution module, and step 104 introduces the training stage of the loss function.

[0062] Before introducing step 101, we need to first introduce social networks. Specifically, for a given social platform, abstracting each user as a node and the connections between users as edges, we can obtain a social network. ,in, Represents the set of nodes in the network. Represents user nodes in the network , Represents the set of edges.

[0063] In step 101, after the rumor outbreak, when it is determined that user nodes in the social network have been infected, a snapshot of the social network at that time can be obtained. Where V represents the set of nodes in the social network, E represents the set of edges, and T represents the set of times when users in the network were infected; furthermore, the social network snapshot also contains the topology of the social network, the state of each node in the social network (whether it has been infected by rumors), and the time when each user was infected.

[0064] In this embodiment of the invention, for any user node in a social network snapshot, the user node status information and the rumor propagation information of the user node can be directly obtained. The user node status information can be represented by the following formula (1):

[0065] (1)

[0066] In formula (1), Y represents the state information of user node i, and Y represents the state of user node i in the social network snapshot. This indicates that user node i has been infected by rumors. This indicates that user node i has not been infected by the rumor.

[0067] The information on the spread of rumors by user nodes is represented by the following formula (2):

[0068] (2)

[0069] In formula (2), This represents the rumor propagation information of user node i, where T represents the timestamp vector of each user node i being infected. This indicates the time when user node i was infected by the rumor.

[0070] In this embodiment of the invention, the feature vector also includes neighbor node status information, social network result information, and rumor propagation information.

[0071] In practical applications, existing methods mostly only include the number of infected neighbors when selecting neighbor node status information. However, according to the principle of source node centrality, a user with more infected neighbors is more likely to be the source of propagation. Therefore, the neighbor node status information provided in this embodiment of the invention can display both the number of infected and uninfected neighbor nodes, thus including more information. Specifically, the neighbor node status information is represented by the following formula (3):

[0072] (3)

[0073] In formula (3), Y represents the state information of the neighboring nodes of user node i, and Y represents the user state in the social network snapshot. This indicates that user j was infected by the rumor, and This indicates that user J was not influenced by the rumors. This represents all neighboring nodes of user node i. This represents the number of all neighbors of user node i.

[0074] Furthermore, the social network structure information of user nodes is represented by the following formula (4):

[0075] (4)

[0076] In formula (4), This represents the social network structure information of user node i. Indicates degree centrality. V represents the set of user nodes in the network.

[0077] It should be noted that, in this embodiment of the invention, in order to reduce the differences in the numerical values ​​in the feature vectors, the degree centrality included in the social network result information is normalized.

[0078] Furthermore, after obtaining user node state information, neighbor node state information, social network structure information, and rumor propagation information based on the social network snapshot, the first feature vector, represented by the user node, can be obtained by constructing each part of the feature vector, which is expressed by formula (5):

[0079] (5)

[0080] In formula (5), This represents the first feature vector of user node i. This represents the state information of user node i. This represents the status information of the neighboring nodes of user node i. This represents the social network structure information of user node i. Represents the rumor propagation information of user node i; the || symbol represents vector concatenation, R is a positive integer, and 0 <R<5。

[0081] In steps 102 and 103, the first feature vector of the user node is used as input parameter and fed into the Graph Attention Network (GAT) to obtain the first output feature vector of the GAT, which specifically includes the following:

[0082] For a user node i on a social network and any of its neighbor nodes j, the weight coefficient of neighbor node j with respect to user node i can be determined simultaneously by the following (6):

[0083] (6)

[0084] In formula (5), This represents the trainable parameter matrix for each layer in a graph attention network. Let a represent the first feature vector of user node i, and a be a weight vector.

[0085] Furthermore, once the weight coefficient of neighbor node j relative to user node i is determined, the proportion of a certain neighbor node j among all neighbor nodes can be determined by the following formula (7):

[0086] (7)

[0087] In formula (7), exp represents an exponential function with base e. Let represent a non-linear activation function that performs linear transformations for negative and positive numbers respectively, and let W be the trainable parameter matrix for each layer in the graph attention network. Let represent the first feature vector of user node i, a represent a weight vector, (·)T represent the transpose operation, and || represent the vector concatenation operation. This represents all neighboring nodes of user node i.

[0088] Furthermore, for each user node i, its first output feature vector after one convolution operation can be represented by the following formula (8):

[0089] (8)

[0090] In formula (8), σ(·) represents the activation function. This represents all neighboring nodes of user node i. W represents the trainable parameter matrix for each layer in the graph attention network. This represents the first feature vector of user node i.

[0091] It should be noted that, in this embodiment of the invention, to avoid overfitting, an Xdrop operation is performed on the first output feature vector of each layer of the graph attention network, that is, selecting from the first output feature vector of each layer of the graph attention network... The first output feature vector is deleted to obtain the second output feature vector, which is specifically represented by the following formula (9):

[0092] (9)

[0093] in, This represents the first output feature vector. , This represents the second output feature vector. This indicates that the feature vector is extracted from the first output feature vector. % quantity, It is a positive integer less than 10.

[0094] Furthermore, for the multi-head attention mechanism, if we consider that there are K heads, then K independent convolutional operations will be applied to the convolutional layer respectively and the results will be concatenated together, specifically expressed by the following formula (10):

[0095] (10)

[0096] Formula (10) yields Due to the concatenation operation, there are K times more elements than the initial feature vector of user node i. Therefore, an averaging operation is needed in the last layer of the model to return the number of elements to the initial number. This can be expressed by the following formula (11):

[0097] (11)

[0098] In formulas (10) and (11), Let be the final output feature vector of user node i, and σ(·) represent the activation function. This represents all neighboring nodes of user node i. W represents the trainable parameter matrix included in the graph attention network, X represents the first feature vector of each user node, and a represents the weight vector. Represents user node K represents the total number of multi-head attention nodes among all neighboring nodes.

[0099] After obtaining the final output feature vector of the user node, the final output vector of the user node can be trained using a loss function. When the training loss converges, the source of the rumor propagation can be determined. Specifically, in step 104, before training the loss on the final output vector of the user node, the final output vector (2-classification result) of the user node is first subjected to a softmax operation using formula (12):

[0100] (12)

[0101] In formula (12), Z represents the final output vector of the user node, Zn represents the nth element in vector Z, and Zm represents the mth element in vector Z. The softmax operation can normalize the elements in the final output vector and make the sum of all elements equal to 1.

[0102] In practical applications, the number of rumor sources is usually only a few, while the scale of social networks is tens of thousands, which leads to class imbalance. To address this problem, in this embodiment of the invention, the loss function is redesigned, specifically: the number of sources and non-sources in a rumor propagation are |S| and |V|-|S|, respectively. Since |V|-|S| is much larger than |S|, it is necessary to balance the number of these two types of samples by multiplying the non-source nodes by |S| / (|V|-|S|), that is, |S| = |S| / (|V|-|S|) * (|V|-|S|). Then the influence of the two types of samples is balanced, as shown in formula (13):

[0103] (13)

[0104] In formula (13), L represents the cross-entropy loss, which is L(x,y) = -log(x)×y for a training sample x and its label y. S represents the set of propagation sources, and V represents the set of nodes in the social network. Represents user node i on the social network. Let j represent user node j on the social network. This represents the loss value of user node i during this training. This represents the loss value of user node j during this training. This represents the 2-norm of the parameter matrix W. It equals 0.0005. This indicates the number of sources of rumor transmission. This indicates the number of non-sources involved in the spread of rumors.

[0105] Furthermore, loss training is performed on the results obtained by formula (13). When the loss training converges or the number of training times reaches the maximum set value, the loss training can be stopped, and the loss training model can be determined as the rumor propagation source model. Based on this rumor propagation source model, the propagation source corresponding to any input test sample can be determined.

[0106] Based on the rumor propagation source tracing method provided in this embodiment of the invention, it can be compared with existing methods to verify the effectiveness of the rumor propagation source model provided in this embodiment of the invention. The test method is the propagation source localization method mentioned in the first section. First, S is defined as the real propagation source set, and S' is the propagation source set obtained by localization. Here, we introduce several commonly used evaluation indicators for propagation source localization methods, namely Accuracy, Precision, Recall and F-Score; Accuracy represents the proportion of samples correctly classified among all samples, Precision represents the proportion of real propagation sources in the propagation source set obtained by localization, Recall represents the proportion of propagation sources correctly located in the real propagation source set, and F-Score combines Precision and Recall. The calculation formula is shown in equation (14), where γ is taken as 1.

[0107] (14)

[0108] Test Result Analysis: The results obtained by this method are not sensitive to propagation models and data noise, proving that this method has high applicability in reality. At the same time, the F-Score obtained by this method is 15%-25% higher than that of the source node centrality-based method. For machine learning methods that take into account the propagation process, the results obtained by this method are 5%-15% higher than those of the source node centrality-based method.

[0109] Furthermore, Figure 3 The results of testing methods for tracing the source of rumor spread on different social networks were presented. Figure 3 In the example of HFSD, which represents the rumor propagation source tracing method provided in this invention, the accuracy of HFSD is higher than that of the comparison algorithm in eight real-world networks. Figure 4 The test results of the rumor source tracing method under different propagation models are shown, such as Figure 4 As shown, the accuracy of the proposed method for tracing the source of rumor propagation is above 0.9 under different models, which proves that the rumor propagation source tracing method provided in this embodiment of the invention has high robustness. Figure 5 The results of the rumor source tracing method and existing methods under different data noise conditions are presented. Figure 5 In the figure, HFSD represents the rumor source tracing method provided in the embodiments of the present invention. Under different intensities of data noise, the accuracy of the rumor source tracing method provided in the embodiments of the present invention is higher than that of other comparative methods, which proves that the rumor source tracing method provided in the embodiments of the present invention has high robustness.

[0110] In summary, this invention provides a method and apparatus for tracing the source of rumor propagation. The method includes: when it is determined that a user node in a social network is infected by a rumor, acquiring a snapshot of the social network, and constructing a first feature vector represented by the user node based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information; determining the attention coefficient of each neighbor node around the user node and the first output feature vector of the user node through the single-layer graph attention network based on the first feature vector of the user node and the trainable parameter matrix included in the single-layer graph attention network; and deleting the first output feature vector by performing an Xdrop operation. The second output feature vector is obtained by concatenating and averaging the elements of the first feature vector. The final output feature vector of the user node is obtained by using multiple independent attention mechanisms. The training loss of the final output feature vector of the user node is determined based on a loss function. When the training loss converges, the source of the rumor propagation is identified. This method fully considers the user node state information, neighbor node state information, social network structure information, and rumor propagation information in the social network. A first feature vector representing the user node is constructed based on a social network snapshot. The construction of the first feature vector considers the different characteristics of user nodes and rumor propagation information, thus solving the problem of high complexity of users and propagation information in reality. Furthermore, the first feature vector is trained through a multi-layer graph attention network to obtain the final output feature vector of the user node, effectively solving the multi-source localization problem. Further, a loss function is introduced to address the imbalance between existing source nodes and non-source nodes, balancing the imbalance between source nodes and non-source nodes corresponding to the final output feature vector of the user node, thus improving the accuracy of rumor propagation source localization on the social network.

[0111] Based on the same inventive concept, this invention provides a device structure for tracing the source of rumor propagation. Since the principle of this device in solving the technical problem is similar to that of the method for tracing the source of rumor propagation, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be described again.

[0112] Figure 6 This is a schematic diagram of a rumor source tracing device provided in an embodiment of the present invention, as shown below. Figure 6 As shown, the device mainly includes: a construction unit 201, a determination unit 202, a first obtaining unit 203, and a second obtaining unit 204.

[0113] The construction unit 201 is used to obtain a social network snapshot of the social network when it is determined that a user node in the social network is infected by rumors, and to construct a first feature vector represented by user nodes based on the social network snapshot. The first feature vector includes user node status information, neighbor node status information, social network structure information and rumor propagation information.

[0114] The determining unit 202 is used to determine the attention coefficients of each neighboring node around the user node and the first output feature vector of the user node through the graph attention network based on the first feature vector of the user node and the trainable parameter matrix included in the graph attention network.

[0115] The first obtaining unit 203 is used to delete features from the first output feature vector through the Xdrop operation. The elements of % are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to the multi-head graph attention network to obtain the final output feature vector of the user node.

[0116] The second unit 204 is used to determine the training loss of the final output feature vector of the user node based on the loss function. When the training loss converges, the source of the rumor propagation is determined.

[0117] Preferably, the first feature vector is:

[0118]

[0119] in, This represents the first feature vector of user node i. This represents the state information of user node i. This represents the status information of the neighboring nodes of user node i. This represents the social network structure information of user node i. Represents the rumor propagation information of user node i; the || symbol represents vector concatenation, R is a positive integer, and 0 <R<5。

[0120] Preferably, the neighbor node status information is:

[0121]

[0122] in, Y represents the state information of the neighboring nodes of user node i, and Y represents the user state in the social network snapshot. This indicates that user j was infected by the rumor, and This indicates that user J was not influenced by the rumors. This represents all neighboring nodes of user node i. This represents the number of all neighbors of user node i.

[0123] Preferably, the social network structure information is as follows:

[0124]

[0125] in, This represents the social network structure information of user node i. Indicates degree centrality. V represents the set of user nodes in the network.

[0126] Preferably, the user node status information is:

[0127]

[0128] The information spread by the rumor:

[0129] in, This represents the state information of user node i. Y represents the rumor propagation information of user node i, and Y represents the state of user node i in the social network snapshot. This indicates that user node i has been infected by rumors. This indicates that user node i is not infected by the rumor, and T represents the timestamp vector of each user node i being infected. This indicates the time when user node i was infected by the rumor.

[0130] Preferably, the Xdrop operation is as follows:

[0131]

[0132] in, This represents the first output feature vector. , This represents the second output feature vector. This indicates that the feature vector is extracted from the first output feature vector. %, Represents a positive integer less than 10;

[0133] The first obtaining unit is specifically used for:

[0134]

[0135]

[0136] in, Let represent the final output feature vector of user node i, and σ(·) represent the activation function. This represents all neighboring nodes of user node i. W represents the trainable parameter matrix included in the graph attention network, and a represents the weight vector. Let K represent all neighboring nodes of user node i, and K represent that the attention mechanism has K heads.

[0137] Preferably, the loss function is:

[0138]

[0139] Where L represents the cross-entropy loss, and V represents the set of user nodes in the network. Represents user node i on the network. Represents user node j on the network. This represents the loss value of user node i during this training. This represents the loss of user node j during this training session. This represents the 2-norm of the parameter matrix W. It equals 0.0005. This indicates the number of sources of rumor transmission. This indicates the number of non-sources involved in the spread of rumors.

[0140] It should be understood that the units included in the above rumor source tracing device are only a logical division based on the functions implemented by the device. In actual applications, the above units can be superimposed or split. Furthermore, the functions implemented by the rumor source tracing device provided in this embodiment correspond one-to-one with the rumor source tracing method provided in the above embodiments. The more detailed processing flow implemented by the device has been described in detail in the above method embodiment one, and will not be described in detail here.

[0141] Another embodiment of the present invention also provides a computer device, the computer device including: a processor and a memory; the memory is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the rumor propagation source tracing method in the method flow shown in the above method embodiment.

[0142] Another embodiment of the present invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform each step of the rumor propagation source tracing method in the method flow shown in the above method embodiment.

[0143] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks, SSDs).

[0144] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for tracing the source of rumor dissemination, characterized in that, include: When it is determined that a user node in the social network has been infected by rumors, a social network snapshot of the social network is obtained, and a first feature vector represented by user nodes is constructed based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information. Based on the first feature vector of the user node and the trainable parameter matrix included in the graph attention network, determine the attention coefficient of each neighbor node around the user node and the first output feature vector of the user node through the graph attention network. Delete the first output feature vector using the Xdrop operation. The elements are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to the multi-head graph attention network to obtain the final output feature vector of the user node. The training loss is determined based on the loss function to determine the final output feature vector of the user node. When the training loss converges, the source of rumor propagation is determined. The first feature vector is: The Xdrop operation is as follows: The step of concatenating and averaging the second output feature vectors based on the multi-head graph attention network to obtain the final output feature vector of the user node specifically includes: in, Represents user node The first eigenvector, Represents user node Status information, Represents user node The status information of neighboring nodes, Represents user node Social network structure information, Represents user node The spread of rumors; || symbol represents vector concatenation, R is a positive integer, and 0 <R<5, Represents user node The first output feature vector, , Represents user node The second output feature vector, This indicates that the feature vector is extracted from the first output feature vector. %, Represents positive integers less than 10. Represents user node The final output feature vector, This represents the activation function. Represents user node All neighboring nodes, W represents the trainable parameter matrix included in the graph attention network. Represents the weight vector. Represents user node All neighboring nodes, This represents the total number of heads receiving multi-head attention.

2. The method for tracing the source of rumor propagation as described in claim 1, characterized in that, The status information of the neighboring nodes is as follows: in, Represents user node The status information of neighboring nodes, This indicates the user's state in a social network snapshot. Indicates user Infected by rumors, and Indicates user Not infected by rumors Represents user node All neighboring nodes, Represents user node The number of all neighbors.

3. The method for tracing the source of rumor propagation as described in claim 1, characterized in that, The social network structure information is as follows: in, Represents user node Social network structure information, Indicates degree centrality. ; This represents the set of user nodes in the network.

4. The method for tracing the source of rumor propagation as described in claim 1, characterized in that, The user node status information is as follows: The rumor propagation information of the user nodes: in, Represents user node Status information, Represents user node The spread of rumors, where Y represents a user node in a social network snapshot. state, Represents user node Infected by rumors, Represents user node Not infected by rumors, T represents each user node. The infected timestamp vector, Represents user node The time spent being infected by rumors.

5. The method for tracing the source of rumor propagation as described in claim 1, characterized in that, The loss function is: in, Represents cross-entropy loss, Represents the set of user nodes in the network. Represents user nodes on the network , Represents user nodes on the network , This indicates the user node used in this training. The loss value, This indicates the user node used in this training. The loss, Represents the parameter matrix The 2-norm, It equals 0.0005. This indicates the number of sources of rumor transmission. This indicates the number of non-sources involved in the spread of rumors.

6. A device for tracing the source of rumor dissemination, characterized in that, include: The construction unit is used to obtain a social network snapshot of the social network when it is determined that a user node in the social network is infected by rumors, and to construct a first feature vector represented by user nodes based on the social network snapshot. The first feature vector includes user node state information, neighbor node state information, social network structure information, and rumor propagation information. The determining unit is used to determine the attention coefficients of each neighboring node around the user node and the first output feature vector of the user node through the graph attention network, based on the first feature vector of the user node and the trainable parameter matrix included in the graph attention network. The first obtaining unit is used to delete features from the first output feature vector through the Xdrop operation. The elements of % are used to obtain the second output feature vector. The second output feature vector is then concatenated and averaged according to the multi-head graph attention network to obtain the final output feature vector of the user node. The second unit is used to determine the training loss of the final output feature vector of the user node based on the loss function. When the training loss converges, the source of the rumor is determined. The first feature vector is: The Xdrop operation is as follows: The first obtaining unit is specifically used to: concatenate and average the second output feature vector according to the multi-head graph attention network to obtain the final output feature vector of the user node. in, Represents user node The first eigenvector, Represents user node Status information, Represents user node The status information of neighboring nodes, Represents user node Social network structure information, Represents user node The spread of rumors; || symbol represents vector concatenation, R is a positive integer, and 0 <R<5, Represents user node The first output feature vector, , Represents user node The second output feature vector, This indicates that the feature vector is extracted from the first output feature vector. %, Represents positive integers less than 10. Represents user node The final output feature vector, This represents the activation function. Represents user node All neighboring nodes, W represents the trainable parameter matrix included in the graph attention network. Represents the weight vector. Represents user node All neighboring nodes, This represents the total number of heads receiving multi-head attention.

7. A computer device, characterized in that, The computer device includes a processor and a memory; the memory is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the computer device executes the rumor propagation source tracing method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The method includes computer instructions that, when executed on a computer device, cause the computer device to perform the rumor propagation source tracing method as described in any one of claims 1 to 5.

Images