A method for constructing a G-rumor GAT model for network rumor detection

Abstract

The application discloses a kind of network rumor detection G-rumorGAT model construction method, specifically includes the following steps: S1.G-rumorGAT model is constructed;S2.in the data set about speech information, sample is divided into training set, verification set and test set, sample in training set, verification set and test set are input into G-rumorGAT model and are trained, after training set, verification set and test set training, the parameter of G-rumorGAT model is revised;S3.G-rumorGAT model after revision parameter is evaluated.The application can capture the complex interaction in the process of rumor propagation, so as to realize effective detection to rumor.

Description

Technical Field

[0001] This invention relates to a method for constructing a rumor detection model, specifically a method for constructing a G-rumorGAT model for detecting online rumors. Background Technology

[0002] With the development of technology and the widespread use of social media, the level of interaction on the internet has reached an unprecedented level. Everyone can easily express their opinions and viewpoints on social platforms. However, this convenience has also facilitated the widespread dissemination of rumors, which have experienced an unprecedented explosive growth on various social media platforms. Simultaneously, in today's fast-paced world where there is a general lack of professional judgment, people are very likely to believe and spread information, even if the information is intentionally misleading. Rumors have a significant negative impact on the stability and security of human society. Identifying rumors on various social media platforms can not only promote the healthy development of cyberspace but also protect the public from being misled. Therefore, building an effective automatic rumor detection method is becoming increasingly important and necessary.

[0003] Early research on rumor detection was based on blog post content. However, as rumor detection technology has evolved, the ability of rumors to disguise themselves has also changed dynamically. For example, rumor publishers often learn the writing characteristics and style of authentic information, deliberately imitating and forging information to evade detection. Therefore, methods based solely on blog post content are not effective in identifying rumors. Compared to blog post content, the dissemination patterns of rumors often differ significantly from those of authentic information, and these differences are difficult to conceal. Based on this, the analysis and exploration of information dissemination structures has become a current research hotspot in rumor detection.

[0004] To explore the characteristics of rumor propagation, some rumor detection methods primarily rely on statistical propagation patterns, manually constructing features, and then using traditional machine learning methods to identify rumors. These methods depend on cumbersome feature engineering, are very time-consuming, require significant human resources, and the manually constructed features are highly subjective and lack high-order feature representation. Recently, researchers have proposed many effective rumor detection methods using deep learning models. Recent approaches based on graph models utilize graph neural networks to model propagation tree structure features, transforming the rumor detection task into a graph classification task, and have also achieved some success.

[0005] However, these methods only focus on the explicit interaction relationships between blog posts during the dissemination process, such as forwarding (or commenting) relationships, ignoring the potential interaction relationships between blog posts, or failing to fully explore the potential interaction relationships between blog posts, making it difficult to capture the complex and diverse characteristics of the dissemination structure and limiting the performance of rumor detection. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a G-rumorGAT model for detecting online rumors, so as to solve the problem that existing technologies cannot effectively detect rumors.

[0007] The objective of this invention is achieved as follows:

[0008] A method for constructing a G-rumorGAT model for detecting online rumors includes the following steps:

[0009] S1. Construct the G-rumorGAT model;

[0010] S2. Divide the samples in the dataset of speech information into training set, validation set and test set. Input the samples in training set, validation set and test set into G-rumorGAT model through loss function for training. After training with training set, validation set and test set, revise the parameters of G-rumorGAT model.

[0011] S3. Evaluate the G-rumorGAT model with revised parameters;

[0012] Furthermore, the method for inputting samples from the training set, validation set, and test set into the G-rumorGAT model for training in step S2 is as follows:

[0013] S2-1. Input the samples into the first part of the G-rumorGAT model for preprocessing to obtain the initial feature vector x of the root node. r The initial feature vectors x1,…,x of each of the other nodes. n－1 And the adjacency matrix of the edges, where n is the total number of nodes, x r ,x1-x n－1 ∈R d0 d0 is the dimension of the initialized feature vector;

[0014] Based on the initial feature vectors of nodes and the adjacency matrix of edges, a heterogeneous relational graph G is constructed along the information propagation direction. TD Heterogeneous graph G relating to the direction of information diffusion BU The heterogeneous graph G representing the relationship in the direction of information propagation TD G is a heterogeneous relational graph in the top-down information transmission direction, representing the heterogeneous relational graph G in the information diffusion direction. BU A heterogeneous relational graph representing the bottom-up direction of information transmission;

[0015] S2-2. The heterogeneous graph G representing the relationships along the information propagation direction... TDThe input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information propagation direction output by the second part of the G-rumorGAT model. This is then used to construct the heterogeneous relational graph G in the information propagation direction. BU The input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information diffusion direction of the second part of the G-rumorGAT model;

[0016] S2-3. Input the feature vector of each node in the information propagation direction output by the second part of the G-rumorGAT model into the third part of the G-rumorGAT model for aggregation to obtain the feature vector in the information propagation direction after aggregation.

[0017] The aggregated feature vectors in the information propagation direction and the feature vectors in the information diffusion direction are pooled and concatenated. The concatenated feature vectors are then processed through an FC layer and a softmax function to calculate the probabilities of belonging to true rumors, false rumors, unverified rumors, and non-rumors, respectively, thus obtaining the rumor detection results of the sample.

[0018] Furthermore, the method for preprocessing the samples in step S2-1 to obtain the adjacency matrix of the edges includes:

[0019] The explicit relationships between nodes in the sample are identified, and a community detection algorithm is used to find nodes with potential relationships, obtaining the explicit or potential relationships of edges, denoted as R, the direction of information propagation. k TD ={e k,ij |i,j=0,…,n-1} and R of the information diffusion direction k BU ={e k,ij |i,j=0,…,n-1}, where k∈{obv,pot}, i and j are nodes, obv represents explicit interaction relationship, and pot represents potential interaction relationship;

[0020] By traversing each node, we obtain the adjacency matrix A of the information propagation direction. TD Adjacency matrix A in the direction of information diffusion BU :

[0021] Among them, A k,ij TD For A TD The element in, A k,ij BU For ABU Elements in;

[0022] Using DropEdge to analyze the adjacency matrix A of the information propagation direction TD Adjacency matrix A in the direction of information diffusion BU After performing random deactivation, the adjacency matrix A of the information propagation direction is obtained. TD' Adjacency matrix A in the direction of information diffusion BU' .

[0023] Furthermore, step S2-2 includes the following sub-steps:

[0024] S2-21. The heterogeneous graph G representing the relationship along the information propagation direction... TD The input is fed into the first layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the first layer RGAT along the information propagation direction. The heterogeneous relation graph G along the information diffusion direction BU The input is fed into the first layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the information diffusion direction in the first layer RGAT.

[0025] S2-22. The feature vector of each node in the first layer RGAT in the information propagation direction. Feature enhancement is performed to obtain the enhanced feature vector of each node in the first layer RGAT along the information propagation direction. For each node in the information diffusion direction, the eigenvector of the first layer RGAT Feature enhancement is performed to obtain the enhanced feature vector of each node in the first layer RGAT along the information diffusion direction.

[0026] S2-23. Enhance the feature vector of each node in the information propagation direction in the first layer RGAT. The input is fed into the second layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the second layer RGAT along the information propagation direction. The enhanced feature vector of each node in the information diffusion direction in the first layer of RGAT is used to enhance the feature vector. The input is fed into the second RGAT layer of the G-rumorGAT model to obtain the feature vector of each node in the second RGAT layer along the information diffusion direction.

[0027] S2-24. The feature vector of each node in the information propagation direction in the second layer RGAT. Feature enhancement is performed to obtain the feature vector of each node in the information propagation direction of the second part of the G-rumorGAT model output. For each node in the information diffusion direction, the feature vector of the second layer RGAT Feature enhancement is performed to obtain the feature vector of each node in the information diffusion direction of the second part of the G-rumorGAT model output.

[0028] Furthermore, for each node in the information propagation direction, the feature vector of the first layer RGAT is... And the enhanced feature vector of each node in the first layer RGAT in the direction of information propagation. Methods for feature enhancement include:

[0029] Based on the product of the appeal of node i in the propagation tree structure between node i and node j and the average distance of information carrying, the first key node corresponding to each node is determined:

[0030]

[0031] Where j represents the nodes that node i can connect to, and d ij The distance between node i and node j is represented by , count(i) represents the number of all other nodes connected to node i, and outdegree(i) represents the appeal of node i. The same nodes have the same key nodes in the direction of information propagation and information diffusion.

[0032] The initial feature vector of the root node, the initial feature vector of the first key node corresponding to each node, and the feature vector of each node in the first layer of RGAT in the information propagation direction are concatenated to obtain the enhanced feature vector of each node in the first layer of RGAT in the information propagation direction.

[0033]

[0034] Where, x r The initial feature vector for the root node, x key(i) The initial feature vector of the first key node corresponding to each node; concate represents the vector concatenation operation.

[0035] The initial feature vector of the root node, the initial feature vector of the first key node corresponding to each node, and the feature vector of each node in the first layer of RGAT along the information diffusion direction are concatenated to obtain the enhanced feature vector of each node in the first layer of RGAT along the information diffusion direction.

[0036] Furthermore, for each node in the information propagation direction, the feature vector of the second-layer RGAT is... And the feature vector of each node in the first layer RGAT along the information diffusion direction. Methods for feature enhancement include:

[0037] Based on the location of the first critical node corresponding to each node, determine the second critical node of each node in the first layer RGAT in the information propagation direction and the third critical node of each node in the first layer RGAT in the information diffusion direction.

[0038] The feature vectors of the root node in the information propagation direction in the first layer RGAT, the feature vectors of the second key node corresponding to each node in the information propagation direction in the first layer RGAT, and the feature vectors of each node in the information propagation direction in the second layer RGAT are concatenated to obtain the feature vectors of each node in the information propagation direction output by the second part of the G-rumorGAT model.

[0039]

[0040] in, The root node in the information propagation direction is the feature vector of the first layer RGAT. The feature vector of the second key node corresponding to the first layer RGAT for each node in the information propagation direction;

[0041] The feature vectors of the root node in the information diffusion direction in the first layer RGAT, the feature vectors of the third key node corresponding to each node in the information diffusion direction in the first layer RGAT, and the feature vectors of each node in the information diffusion direction in the second layer RGAT are concatenated to obtain the feature vectors of each node in the information diffusion direction output by the G-rumorGAT model.

[0042]

[0043] in, The root node in the information diffusion direction is the feature vector of the first layer RGAT. The feature vector of the second key node corresponding to each node in the first layer RGAT in the information diffusion direction.

[0044] Furthermore, the method for aggregating the feature vectors of each node in the information propagation direction and the feature vectors of each node in the information diffusion direction of the second part output of the G-rumorGAT model in steps S2-3 includes:

[0045] The feature matrix is ​​formed by combining the feature vectors of each node in the information propagation direction output by the second part of the G-rumorGAT model. The feature matrix is ​​formed by combining the feature vectors of each node in the information diffusion direction output from the second part of the G-rumorGAT model. in, This is the feature vector of the root node in the transmission direction output by the second part of the G-rumorGAT model. This represents the feature vectors of other nodes along the transmission direction output by the second part of the G-rumorGAT model. This represents the feature vector of the root node in the information diffusion direction output by the second part of the G-rumorGAT model. The feature vectors of other nodes in the information diffusion direction output by the second part of the G-rumorGAT model;

[0046] Aggregate the feature matrices along the information propagation direction to obtain Aggregating the feature matrices along the information diffusion direction yields... in, aggergate represents aggregate functions.

[0047] Furthermore, in steps S2-21, the feature vector of each node in the information propagation direction in the first layer RGAT... And the feature vector of each node in the first layer RGAT along the information diffusion direction. The calculation formulas are as follows:

[0048]

[0049] in, d1 is the dimension of the feature vector output by the first layer RGAT, σ represents the linear function that can be selected, and R k It is a collection of relation types. The attention coefficient represents the direction of information dissemination. The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of each node in the direction of information propagation. Let be the set of neighboring nodes of each node in the direction of information diffusion. This is the initial feature vector for each node in the set of neighboring nodes along the information propagation direction. This is the initial feature vector for each node in the set of neighboring nodes along the information diffusion direction.

[0050] Furthermore, in steps S2-23, the feature vector of each node in the information propagation direction in the second layer RGAT... The feature vector of each node in the second layer RGAT along the information diffusion direction. The calculation formulas are as follows:

[0051]

[0052] in, d1 is the dimension of the feature vector output by the second-layer RGAT, and σ represents the linear function that can be selected. The attention coefficient represents the direction of information dissemination. The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of each node in the direction of information propagation. Let be the set of neighboring nodes of each node in the direction of information diffusion. For each node in the neighborhood node set along the information propagation direction, this is the enhanced feature vector of the first layer RGAT. This is the enhanced feature vector of each node in the set of neighboring nodes in the information diffusion direction in the first layer of RGAT.

[0053] Furthermore, the loss function is:

[0054]

[0055] Where C is the set of samples in the training set, validation set, and test set, C = {c1, c2, ..., c3} |c|}, It is the true label vector representation of the i-th sample. is the detection result obtained by the G-rumorGAT model for the i-th sample, β is the hyperparameter representing the magnitude of L2 regularization, and θ is all the parameters that the model needs to learn.

[0056] This invention is based on the graph G-rumorGAT model. It does not rely on heavy feature engineering, saving a lot of manpower. It can also extract higher-order representations more conveniently and effectively. Compared with many existing graph G-rumorGAT models, the new method for constructing potential interaction relationships and the new key node enhancement strategy of this invention can more effectively mine the potential features in the rumor propagation process and more effectively capture the complex interaction relationships in the rumor propagation process.

[0057] Previous techniques utilize root node features and the maximum forwarding node in the current propagation path to aid in learning node feature representations. However, this key node enhancement strategy overlooks the fact that the influence of the maximum forwarding node diminishes as rumors spread through multiple levels. Therefore, for some nodes, their key node is no longer the maximum forwarding node. The key node enhancement strategy proposed in this invention is more effective than previous solutions. It identifies a more suitable key node for each node, more accurately and comprehensively models the potential influence of key nodes on information dissemination, thereby improving the network platform's ability to detect rumors and enhancing its management capabilities. Attached Figure Description

[0058] Figure 1 This is a flowchart of the detection method of the present invention. Detailed Implementation

[0059] The present invention will now be described in further detail.

[0060] like Figure 1 As shown, the method for constructing the G-rumorGAT model for detecting online rumors provided by this invention specifically includes the following steps:

[0061] S1. Construct the G-rumorGAT model.

[0062] The G-rumorGAT model consists of three parts: the first part preprocesses the input data (samples), the second part performs convolution and feature enhancement on the preprocessed data, and the third part aggregates and computes the data.

[0063] S2. Divide the samples in the dataset of speech information into training set, validation set and test set. Input the samples in the training set, validation set and test set into the G-rumorGAT model for training. After training with the training set, validation set and test set, revise the parameters of the G-rumorGAT model.

[0064] S2-1 inputs the samples into the first part of the G-rumorGAT model for preprocessing, obtaining the initial feature vector x of the root node. r The initial feature vectors x1,…,x of each of the other nodes. n－1 And the adjacency matrix of the edges, where x r This represents the initial feature vector of the source blog post, x1-x n－1 x represents the initial feature vector of the forwarded blog post. r ,x1-x n－1 ∈R d0 d0 is the dimension of the feature vector;

[0065] Based on the initial feature matrix of each node and the adjacency matrix of the edges, a top-down (TD) relational heterogeneous graph is constructed along the information propagation direction. TD Heterogeneous graph of relationships in the direction of information diffusion (bootom-up, BU)G BU .

[0066] Among them, the heterogeneous graph G of relationships in the direction of information propagation TD G is a heterogeneous graph of relationships in the top-down information transmission direction and a heterogeneous graph of relationships in the information diffusion direction. BU This is a heterogeneous graph of relationships based on the bottom-up direction of information transmission.

[0067] To obtain the initial vector representation of a node, the top 5000 most frequent words in the dataset are extracted based on word frequency. Initial feature vectors for each node are obtained using TF-IDF features. Based on these initial feature vectors, an initial feature matrix for the node is derived, and potential interaction relationships are constructed. In social networks, a post forwarded earlier may have a propagation impact on subsequent posts. To capture these potential relationships, a community detection algorithm is used to find nodes in the propagation tree that may have potential interaction relationships. First, based on the chronological order of forwarding, candidate directed temporal edges are constructed between the nodes corresponding to the forwarded posts of each post, serving as potential interaction relationships.

[0068] The explicit relationships between nodes in the dataset are identified, and a community detection algorithm is used to identify potential relationships between nodes that have no interaction. Formally, for each type of edge, R is denoted as the direction of information propagation. k TD ={e k,ij |i,j=0,…,n-1} and R of the information diffusion direction k BU ={e k,ij |i,j=0,…,n-1}, where k∈{obv,pot}, i and j are nodes, obv represents explicit interaction relationship, and pot represents potential interaction relationship;

[0069] By traversing each node, we obtain the adjacency matrix A of the information propagation direction. TD Adjacency matrix A in the direction of information diffusion BU :

[0070]

[0071] Among them, A k,ij TD For A TD The element in, A k,ij BU For A BUThe elements in.

[0072] To alleviate the overfitting problem in graph convolutional networks, the DropEdge method is adopted. The idea of ​​this method is to randomly deactivate edges in the input graph with a certain probability, thereby achieving the purpose of mitigating overfitting.

[0073] Random deactivation processing is applied to the adjacency matrix based on the deactivation probability η, affecting the adjacency matrix A along the information propagation direction. TD Random sampling N e ×η edges, forming A drop TD Adjacency matrix A in the direction of information diffusion BU Random sampling N e ×η edges, forming A drop BU .

[0074] Based on A TD and A drop TD The adjacency matrix A of the processed information propagation direction is obtained. TD' Based on A BU and A drop BU The adjacency matrix A of the processed information diffusion direction is obtained. BU' :

[0075] A TD' =A TD -A drop TD

[0076] A BU' =A BU -A drop BU

[0077] The adjacency matrix of the processed information propagation direction includes the submatrix A of explicit relations. TD' obv Submatrix A of potential relationships TD' pot The adjacency matrix in the direction of information diffusion includes the submatrix A of explicit relations. BU' obv Submatrix A of potential relationships BU' pot .

[0078] A heterogeneous relation graph G is constructed based on the adjacency matrix and eigenvectors along the information propagation direction. TD Based on the adjacency matrix and eigenvectors in the information diffusion direction, a heterogeneous relation graph G is constructed in the information diffusion direction. BU .

[0079] Step S2-2. Transform the heterogeneous graph G of relationships along the information propagation direction. TD The input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information propagation direction of the second part of the G-rumorGAT model, which is then used to construct the heterogeneous graph G of the relationship in the information propagation direction. BU The input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information diffusion direction of the second part of the G-rumorGAT model.

[0080] The heterogeneous graph G representing the relationship along the information propagation direction TD The heterogeneous relational graph GBU along the information propagation direction is input into the first layer of the Relational Graph Attention Network (RGAT) in the second part of the G-rumorGAT model, thereby obtaining the hidden layer feature vector. TD The information transfer formula in the first layer of RGAT (Graph Convolutional Network) is:

[0081]

[0082] in, The feature vector x in the direction of information propagation i The feature vector after the first RGAT layer d1 is the dimension of the latent vector features of the first-layer RGAT graph convolutional network (i.e., the feature vector output by the first-layer RGAT), σ represents the linear function that can be chosen, and R... k It is a collection of relationship types, including both explicit and implicit interaction relationships. The attention coefficient represents the direction of information propagation. In the heterogeneous graph attention mechanism, each relationship has an independent attention coefficient. Let be the set of neighboring nodes of each node in the direction of information propagation. Let be the initial feature vector of each node in the set of neighboring nodes in the direction of information propagation.

[0083] Heterogeneous graph G of relationships in the direction of information diffusion TD The information transmission formula for the first-level RGAT is:

[0084]

[0085] in, The feature vector x in the direction of information diffusion i The feature vector after the first RGAT layer The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of node i in the direction of information diffusion. This is the initial feature vector for each node in the set of neighboring nodes along the information diffusion direction.

[0086] In the process of rumor dissemination, content posted by many users plays a role in facilitating its spread. Taking forwarding relationships as an example, to explore the potential interactions between source posts and key posts with other nodes during information dissemination, we enhance the features of the current node using two types of key nodes in the propagation tree, thereby learning more comprehensive node features. For any node in the propagation tree, the two corresponding key nodes are its root node and the node with the greatest influence on it on the current path. The root node represents the source post, containing rich information about the source of the rumor, which helps in learning a more accurate node representation.

[0087] The key node with the greatest influence on the current node in the current propagation path: This is the node with the greatest influence on the current node in the current propagation path. This node often has a significant impact on the current node during the propagation of information.

[0088] This invention employs a novel method for identifying key nodes, describing node importance from both global and local perspectives. The influence of node i on node j along the current propagation path is defined as the product of node i's appeal within the propagation tree structure between node i and node j and the average distance the information carries. That is:

[0089]

[0090] Where j represents the nodes that node i can connect to, and d ij `i` represents the distance between node `i` and node `j`; `count(i)` represents the number of all other nodes connected to node `i`. `outdegree(i)` is the appeal of node `i`. The main focus is on considering the influence of nodes from a local perspective;

[0091] Calculate the node with the greatest influence on the current node along the current propagation path, and calculate the corresponding key node for each node, using this node as the key node. Nodes with the same information propagation direction and information diffusion direction have the same first key node, and the initial feature vectors of the root nodes corresponding to the information propagation direction and information diffusion direction are the same.

[0092] For the feature vector of each node in the transmission direction output by the first-layer RGAT, feature enhancement is performed. This is done by concatenating the initial feature vector of the first key node and the initial feature vector of the root node with the feature vector representation of each node in the transmission direction after passing through the first-layer RGAT. This yields the enhanced feature vector of each node in the information propagation direction after the first-layer RGAT.

[0093]

[0094] Where, x r The initial feature vector for the root node, x key(i) The initial feature vector of the first key node corresponding to each node, and concate represents the vector concatenation operation.

[0095] For the feature vector of each node in the information diffusion direction output by the first layer RGAT, the initial feature vector of the first key node and the initial feature vector of the root node corresponding to each node are concatenated with the feature vector representation of each node in the information diffusion direction after passing through the first layer RGAT. This yields the enhanced feature vector of each node in the information diffusion direction after the first layer RGAT.

[0096]

[0097] After performing feature augmentation on the feature vector output by the first-layer RGAT, the augmented feature vector of each node in the information propagation direction is then used in the first-layer RGAT. The information transfer formula for the second layer RGAT of the G-rumorGAT model is as follows:

[0098]

[0099] in, d1 is the dimension of the latent vector features of the second-layer graph convolutional network (i.e., the feature vector output by the second-layer RGAT). The attention coefficient in the direction of information dissemination. Let be the set of neighboring nodes of each node in the direction of information propagation. This is the enhanced feature vector of each node in the set of neighboring nodes in the information propagation direction in the first layer of RGAT.

[0100] The enhanced feature vector of each node in the information diffusion direction in the first layer of RGAT is used to enhance the feature vector. The information transfer formula for the second layer RGAT of the G-rumorGAT model is as follows:

[0101]

[0102] in, The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of each node in the direction of information diffusion. This is the enhanced feature vector of each node in the set of neighboring nodes in the information diffusion direction in the first layer of RGAT.

[0103] After outputting the feature vectors of each node in the information diffusion and transmission directions in the second layer, feature enhancement is performed on the feature vectors of the nodes output in the second layer. Based on the position of the first key node, the second key node corresponding to the first layer RGAT for each node in the transmission direction and the third key node corresponding to the first layer RGAT for each node in the information diffusion direction are determined. This leads to the determination of the feature vectors of the second key node corresponding to the first layer RGAT for each node in the transmission direction and the third key node corresponding to the first layer RGAT for each node in the information diffusion direction.

[0104] The feature vectors of the root node in the information propagation direction in the first layer of RGAT, the feature vectors of the second key node corresponding to each node in the information propagation direction, and the feature vectors of each node in the information propagation direction in the second layer of RGAT are concatenated to obtain the feature vectors of each node in the information propagation direction output by the second part of the G-rumorGAT model.

[0105] in, The root node in the information propagation direction is the feature vector of the first layer RGAT. The feature vector of the second key node corresponding to the first layer RGAT for each node in the information propagation direction;

[0106] The feature vectors of the root node in the information diffusion direction in the first layer of RGAT, the feature vectors of the third key node corresponding to each node in the information diffusion direction in the second layer of RGAT, and the feature vectors of each node in the information diffusion direction in the second layer of RGAT are concatenated to obtain the feature vectors of each node in the information diffusion direction output by the second part of the G-rumorGAT model.

[0107]

[0108] in, The root node in the information diffusion direction is the feature vector of the first layer RGAT. This is the feature vector of the second key node corresponding to each node in the information diffusion direction.

[0109] The output of the second layer is concatenated with the feature vector of the key node and used as the input of the rumor classifier. Through two different graph convolution operations, local neighborhood features under different interaction relationships can be effectively accumulated.

[0110] Step S2-3. Input the feature vector in the information diffusion direction output by the second part of the G-rumorGAT model into the third part of the G-rumorGAT model for aggregation to obtain the feature vector in the information diffusion direction after aggregation.

[0111] The aggregated feature vectors in the information propagation direction and the feature vectors in the information diffusion direction are pooled and concatenated to form the overall feature vector of the event. The overall feature vector is then used to calculate the probability of belonging to true rumors, false rumors, unverified rumors, and non-rumors through an FC layer and a softmax function, respectively, to obtain the rumor detection result of the sample.

[0112] Finally, after the second part of the G-rumorGAT model, the heterogeneous graph G based on the information propagation direction... TD Heterogeneous graph G in the direction of information diffusion BU The feature vectors of each node in the propagation tree are learned, and the node feature matrix in the information propagation direction is obtained. Node feature matrix in the direction of information diffusion

[0113]

[0114] in, This is the feature vector of the root node in the transmission direction output by the second part of the G-rumorGAT model, which is the enhanced feature vector of the root node in the transmission direction in the second layer of RGAT. This refers to the feature vectors of other nodes in the transmission direction output by the second part of the G-rumorGAT model, which are the enhanced feature vectors of other nodes in the transmission direction in the second layer of RGAT. This refers to the feature vector of the root node in the information diffusion direction output by the second part of the G-rumorGAT model, which is the enhanced feature vector of the root node in the information diffusion direction in the second layer of RGAT. The second part of the G-rumorGAT model outputs the feature vectors of other nodes in the information diffusion direction, which are the enhanced feature vectors of other nodes in the second layer of RGAT in the information diffusion direction.

[0115] The feature vectors of each node in the top-down information propagation direction and the bottom-up information diffusion direction of the second part of the G-rumorGAT model are aggregated separately to obtain the feature vector C of the information propagation direction after the propagation tree is aggregated. TD The eigenvector C of the information diffusion direction BU :

[0116]

[0117] in, aggergate represents an aggregation function, and average pooling is applied to the aggregated feature vector to obtain a pooled feature vector representing the direction of information propagation. Feature vectors of information diffusion direction

[0118] Finally, the top-down and bottom-up feature vectors are concatenated to form the final feature vector of the sample:

[0119] This invention transforms the rumor detection task into a graph classification problem. Based on the rumor feature vector representation, the probability of a sample belonging to each category is calculated through an FC layer and a softmax function.

[0120]

[0121] Among them, W c and b c These are parameters that need to be learned.

[0122] During training, the G-rumorGAT model is trained by minimizing the cross-entropy loss. The training, validation, and test sets contain samples C = {c1, c2, ..., c...}. |c| Loss function for:

[0123]

[0124] in, It represents the true label vector representation of the i-th sample. is the recognition result obtained by the G-rumorGAT model for the i-th sample, β is the hyperparameter representing the size of L2 regularization, and Θ is all the parameters that the model needs to learn.

[0125] Step S3. Evaluate the G-rumorGAT model after revising the parameters: This invention selects rumor detection methods based on feature engineering, rumor detection methods based on kernel functions, and rumor detection methods based on deep learning models as baseline methods for comparison with this invention.

[0126] This paper will test the proposed method on three public datasets: Twitter15, Twitter16, and PHEME.

[0127] Table 1: Statistical Information of the Rumor Detection Dataset

[0128]

[0129] As shown in Table 1, the Twitter15 and Twitter16 datasets were created by Ma et al., collecting rumor information from the well-known international social networking platform Twitter at different times, containing 1490 and 818 samples respectively. Following the guidelines of Zubiaga et al. and Ma et al., based on the veracity tags of articles on debunking websites (such as snopes.com, Emergent.info, etc.), each sample was labeled with one of four tags: True-Rumor (TR), False-Rumor (FR), Unverified-Rumor (UR), and Non-Rumor (NR).

[0130] The PHEME dataset, created by Zubiaga et al., collected 2,402 rumors around nine events. These rumors were categorized into three classes: true rumors, false rumors, and unverified rumors. The dataset was divided in accordance with baseline methods.

[0131] The baseline methods for the Twitter15 and Twitter16 datasets include:

[0132] DTC.Castillo et al. constructed a decision tree classifier to obtain information credibility based on manually designed global statistical features; SVM-RBF.Yang et al. constructed statistical features based on blog content and built a support vector machine classifier based on the RBF kernel function to identify rumors; SVM-TS.Ma et al. constructed a linear support vector machine classifier based on temporal context features to classify rumors; SVM-TK.Ma et al. used a propagation tree-based kernel function to extract structural features and used a support vector machine classifier to complete the classification; GRU-RNN.Ma et al. learned the feature vector representation of rumors by modeling the sequence structure of related posts based on recurrent neural networks; RvNN.Ma Some researchers used two recurrent neural network models to model the information propagation direction and information diffusion direction respectively, and learned the feature vector representation of the propagation tree; StA-PLAN.Khoo et al. used the Transformer model to mine the long-distance interaction features between blog posts in the propagation process, learned the feature vector representation of rumors, and completed the classification task based on the feature vector representation; Bi-GCN.Bian et al. constructed a relation graph attention network model based on the information propagation direction and information diffusion direction of the propagation tree, and transformed the rumor classification task into a graph classification task; RumorGCN.Hu et al. proposed a rumor detection method based on multi-relation propagation trees based on text content and propagation structure information.

[0133] For the PHEME dataset, this invention is compared with five representative baseline methods that currently offer good performance:

[0134] NileTMRG.Enayet et al. proposed a bag-of-words model to obtain the vector representation of blog posts and use a support vector machine classifier to complete the classification; RvNN.Ma et al. proposed an RNN-based rumor detection model; BranchLSTM.Kochkina et al. used the sequence model LSTM to detect rumors and adopted a multi-task learning approach to jointly train rumor detection and stance recognition tasks; Bi-GCN.Bian et al. proposed a graph-based rumor detection method; RumorGCN.Hu et al. proposed a rumor detection method based on multiple relation propagation trees based on text content and propagation structure information.

[0135] The rumor detection problem of this invention is essentially a classification problem. Therefore, this invention uses a classification-based evaluation index to evaluate the rumor detection performance.

[0136] For the Twitter15 and Twitter16 datasets, this invention uses accuracy (Accuracy, Acc) and the F1 score for each class as evaluation metrics:

[0137]

[0138] Among them, TP (true positive) is a true positive sample, which is a positive sample that is correctly predicted by the model; FP (false positive) is a false positive sample, which is a negative sample that is correctly predicted by the model; FN (false negative) is a false negative sample, which is a positive sample that is incorrectly predicted by the model; and TN (true negative) is a true negative sample, which is a negative sample that is incorrectly predicted by the model.

[0139] For the PHEME dataset, referencing baseline methods, this invention selects accuracy and macro-averaging F1 (macro-F1). Macro-F1 involves first calculating the statistical index value for each class, and then taking the arithmetic mean over all classes. The calculation method is as follows:

[0140]

[0141] Where n represents the number of predicted categories. Considering the balance of samples across categories in this dataset, this invention also compares the weighted-averaging F1 (weighted-F1) values. First, a statistical indicator value is calculated for each category, and then a weighted average is calculated for all categories. The calculation method is as follows:

[0142]

[0143] Where, γ i The weight is the proportion of each category in the sample.

[0144] The experiments of this invention on the Twitter15 and Twitter16 datasets are based on the open-source code proposed by Tsinghua University.

[0145] Table 2: Experimental results on the Twitter15 dataset

[0146]

[0147] Table 3: Experimental results on the Twitter16 dataset

[0148]

[0149]

[0150] As shown in Tables 2 and 3, the present invention outperforms the baseline methods on both the Twitter15 and Twitter16 datasets. For the Twitter15 dataset, compared to the best baseline method, the present invention improves accuracy by 1.2 percentage points and F1 score by a maximum of 1.2 percentage points; for the Twitter16 dataset, the present invention improves accuracy by 1.3 percentage points and F1 score by a maximum of 1.5 percentage points. These results demonstrate that the present invention has higher rumor detection performance than other baseline methods.

[0151] Based on the experimental results in Tables 2 and 3, the specific analysis is as follows:

[0152] All deep learning-based methods (this invention, GRU-RNN, RumorGCN, Bi-GCN, StA-PLAN, and RvNN) outperform rumor detection methods based on manually constructed features (DTC, SVM-RBF, SVM-TS, SVM-TK) in rumor detection performance. This phenomenon confirms that this invention outperforms manually constructed feature-based rumor detection methods in the rumor detection task. The main advantage of deep learning models is that such models can learn the latent feature vector representation of rumors. DTC, SVM-RBF, SVM-TS, and SVM-TK use manually constructed features to identify rumors, and the extracted features are highly subjective, lacking latent feature representations of rumors, and cannot effectively identify rumors in social networks. In contrast, this invention is a graph-based deep learning model that can more fully learn the latent feature representations of rumors and the structural features in the rumor propagation process.

[0153] RvNN models the propagation tree using a recurrent neural network, but it struggles to capture long-range interactions within a sequence, thus limiting its rumor detection performance. StA-PLAN, utilizing a Transformer structure, effectively alleviates this problem, achieving superior detection performance compared to RvNN. Compared to RvNN and StA-PLAN, graph-based rumor detection methods (this invention, RumorGCN, and Bi-GCN) outperform all deep learning models, demonstrating that graph models can more fully learn the latent feature representations of rumors and the structural features of rumor propagation, enabling them to more effectively capture complex interactions.

[0154] Compared to the best baseline method RumorGCN, the present invention outperforms the method on both Twitter datasets, demonstrating its effectiveness in the rumor detection task.

[0155] This invention believes that the performance improvement is mainly due to two aspects:

[0156] During the spread of rumors, when a user forwards a source post, they are not only influenced by the source post itself but also by other posts that have forwarded it, resulting in complex and multifaceted propagation paths. Compared to baseline models, this invention can more fully learn the latent features within the propagation tree structure and more effectively capture the complex interactions during rumor propagation. This demonstrates that fully learning the latent interactions within the propagation tree can provide more effective information for rumor detection.

[0157] RumorGCN utilizes root node features and the maximum forwarding node in the current propagation path to assist in learning node feature representations. However, this key node enhancement strategy ignores the fact that the influence of the maximum forwarding node diminishes as the rumor spreads through multiple levels. Therefore, for some nodes, their key node is no longer the maximum forwarding node. The key node enhancement strategy proposed in this invention is more effective than the one used in RumorGCN. It can identify a more suitable key node for each node and more accurately and comprehensively model the potential influence of key nodes on information propagation, thereby improving the detection performance of the rumor model.

[0158] Table 4: Experimental results of the present invention and the baseline method in comparison on the PHEME dataset.

[0159] NileTMRG 36 29.7 RvNN 34.1 26.4 BranchLSTM 31.4 25.9 Bi-GCN 49.2 46.7 63.2 RumorGCN 66.2 54.6 73.5 This invention 69.5 57.1 74.5

[0160] As shown in Table 4, the present invention was tested based on the open-source code of Bi-GCN. The results show that, compared with the optimal baseline method, the present invention improves the accuracy by 3.3 percentage points, the macro-average F1 score by 2.5 percentage points, and the weighted average F1 score by 1 percentage point.

[0161] Based on the experimental results in Table 4, the specific analysis is as follows:

[0162] RvNN models the propagation tree using a recurrent neural network model, but this approach struggles to capture long-range interactions within the propagation sequence. This limitation restricts the model's rumor detection performance. Among various models, graph-based rumor detection methods perform better, demonstrating the ability of graph models to capture complex interactions.

[0163] The BranchLSTM model improves classification accuracy by leveraging the correlation between tasks and auxiliary tasks, particularly in rumor detection and stance classification, through a multi-task learning approach. However, since this model models the propagation structure based on LSTM, it cannot fully extract structural information from the propagation tree compared to graph-based rumor detection models. This limitation significantly restricts the model's rumor detection performance.

[0164] Compared to baseline methods, this invention achieves superior rumor detection results on the PHEME dataset, further demonstrating that it can more fully learn the latent feature representations of rumors and the structural features in the rumor propagation process, and more effectively capture complex interaction relationships. The main reason for the improved rumor detection performance of this invention is that it can more fully learn the latent interaction features in the propagation tree structure, while RumorGCN does not fully consider the latent interaction relationships in the propagation tree structure. This also shows that fully learning the latent interaction relationships in the propagation tree can provide more effective information for rumor detection. Furthermore, the important node enhancement strategy applied in this invention is more effective than the important node enhancement strategy in RumorGCN, which uses root node features and the largest forwarding node in the current propagation path to assist in learning node feature representations. It can more accurately and fully model the potential influence of key nodes on information propagation, thereby improving rumor detection performance.

Claims

1. A method for constructing a G-rumorGAT model for detecting online rumors, characterized in that, Includes the following steps: S1. Construct the G-rumorGAT model; S2. Divide the samples in the dataset of speech information into training set, validation set and test set. Input the samples in training set, validation set and test set into G-rumorGAT model through loss function for training. After training with training set, validation set and test set, revise the parameters of G-rumorGAT model. S3. Evaluate the G-rumorGAT model with revised parameters; The method for inputting samples from the training set, validation set, and test set into the G-rumorGAT model for training in step S2 is as follows: S2-1. Input the samples into the first part of the G-rumorGAT model for preprocessing to obtain the initial feature vector x of the root node. r The initial feature vectors x1,…,x of each of the other nodes. n－1 And the adjacency matrix of the edges, where n is the total number of nodes, x r ,x1-x n－1 ∈R d0 d0 is the dimension of the initialized feature vector; Based on the initial feature vectors of each node and the adjacency matrix of the edges, construct a heterogeneous relation graph G along the information propagation direction. TD Heterogeneous graph G relating to the direction of information diffusion BU ; S2-2. Transform the heterogeneous graph G of relationships along the information propagation direction. TD The input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information propagation direction of the second part of the G-rumorGAT model. This is then used to construct the heterogeneous relational graph G in the information propagation direction. BU The input is fed into the second part of the G-rumorGAT model to obtain the feature vector of each node in the information diffusion direction of the second part of the G-rumorGAT model; S2-3. Input the feature vector of each node in the information propagation direction output by the second part of the G-rumorGAT model into the third part of the G-rumorGAT model for aggregation to obtain the feature vector in the information propagation direction after aggregation. The aggregated feature vectors in the information propagation direction and the feature vectors in the information diffusion direction are pooled and concatenated. The concatenated feature vectors are then passed through an FC layer and a softmax function to calculate the probability that a rumor belongs to a true rumor, a false rumor, an unverified rumor, or a non-rumor, respectively, thus obtaining the detection result of the sample.

2. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 1, characterized in that, The methods for preprocessing the samples in step S2-1 to obtain the adjacency matrix of the edges include: The explicit relationships between nodes in the sample are identified, and a community detection algorithm is used to find nodes with potential relationships, obtaining the explicit or potential relationships of edges, denoted as R, the direction of information propagation. k TD ={e k,ij |i,j=0,…,n-1} and R of the information diffusion direction k BU ={e k,ij |i,j=0,…,n-1}, where k∈{obv,pot}, i and j are nodes, obv represents explicit interaction relationship, and pot represents potential interaction relationship; By traversing each node, we obtain the adjacency matrix A of the information propagation direction. TD Adjacency matrix A in the direction of information diffusion BU : Among them, A k,ij TD For A TD The element in, A k,ij BU For A BU Elements in; Using DropEdge to analyze the adjacency matrix A of the information propagation direction TD Adjacency matrix A in the direction of information diffusion BU After performing random deactivation, the adjacency matrix A of the information propagation direction is obtained. TD 'Adjacency matrix A in the direction of information diffusion BU '.

3. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 1, characterized in that, Step S2-2 includes the following sub-steps: S2-21. The heterogeneous graph G representing the relationship along the information propagation direction... TD The input is fed into the first layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the first layer RGAT along the information propagation direction. The heterogeneous relation graph G along the information diffusion direction BU The input is fed into the first layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the information diffusion direction in the first layer RGAT. S2-22. The feature vector of each node in the first layer RGAT in the information propagation direction. Feature enhancement is performed to obtain the enhanced feature vector of each node in the first layer RGAT along the information propagation direction. For each node in the information diffusion direction, the eigenvector of the first layer RGAT Feature enhancement is performed to obtain the enhanced feature vector of each node in the first layer RGAT along the information diffusion direction. S2-23. Enhance the feature vector of each node in the information propagation direction in the first layer RGAT. The input is fed into the second layer RGAT of the G-rumorGAT model to obtain the feature vector of each node in the second layer RGAT along the information propagation direction. The enhanced feature vector of each node in the information diffusion direction in the first layer of RGAT is used to enhance the feature vector. The input is fed into the second RGAT layer of the G-rumorGAT model to obtain the feature vector of each node in the second RGAT layer along the information diffusion direction. S2-24. The feature vector of each node in the information propagation direction in the second layer RGAT. Feature enhancement is performed to obtain the feature vector of each node in the information propagation direction of the second part of the G-rumorGAT model output. For each node in the information diffusion direction, the feature vector of the second layer RGAT Feature enhancement is performed to obtain the feature vector of each node in the information diffusion direction of the second part of the G-rumorGAT model output.

4. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 3, characterized in that, For each node in the information propagation direction, the feature vector of the first layer RGAT And the enhanced feature vector of each node in the first layer RGAT in the direction of information propagation. Methods for feature enhancement include: Based on the product of the appeal of node i in the propagation tree structure between node i and node j and the average distance of information carrying, the first key node corresponding to each node is determined: Where j represents the nodes that node i can connect to, and d ij The distance between node i and node j is represented by , count(i) represents the number of all other nodes connected to node i, and outdegree(i) represents the appeal of node i. The same nodes have the same key nodes in the direction of information propagation and information diffusion. The initial feature vector of the root node, the initial feature vector of the first key node corresponding to each node, and the feature vector of each node in the first layer of RGAT in the information propagation direction are concatenated to obtain the enhanced feature vector of each node in the first layer of RGAT in the information propagation direction. Where, x r The initial feature vector for the root node, x key(i) The initial feature vector of the first key node corresponding to each node; concate represents the vector concatenation operation. The initial feature vector of the root node, the initial feature vector of the first key node corresponding to each node, and the feature vector of each node in the first layer of RGAT along the information diffusion direction are concatenated to obtain the enhanced feature vector of each node in the first layer of RGAT along the information diffusion direction.

5. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 4, characterized in that, For each node in the information propagation direction, the feature vector of the second layer RGAT And the feature vector of each node in the first layer RGAT along the information diffusion direction. Methods for feature enhancement include: Based on the location of the first critical node corresponding to each node, determine the second critical node of each node in the first layer RGAT in the information propagation direction and the third critical node of each node in the first layer RGAT in the information diffusion direction. The feature vectors of the root node in the information propagation direction in the first layer RGAT, the feature vectors of the second key node corresponding to each node in the information propagation direction in the first layer RGAT, and the feature vectors of each node in the information propagation direction in the second layer RGAT are concatenated to obtain the feature vectors of each node in the information propagation direction output by the second part of the G-rumorGAT model. in, The root node in the information propagation direction is the feature vector of the first layer RGAT. The feature vector of the second key node corresponding to the first layer RGAT for each node in the information propagation direction; The feature vectors of the root node in the information diffusion direction in the first layer RGAT, the feature vectors of the third key node corresponding to each node in the information diffusion direction in the first layer RGAT, and the feature vectors of each node in the information diffusion direction in the second layer RGAT are concatenated to obtain the feature vectors of each node in the information diffusion direction output by the G-rumorGAT model. in, The root node in the information diffusion direction is the feature vector of the first layer RGAT. The feature vector of the second key node corresponding to each node in the first layer RGAT in the information diffusion direction.

6. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 5, characterized in that, The method for aggregating the feature vectors of each node in the information propagation direction and the feature vectors of each node in the information diffusion direction of the second part output of the G-rumorGAT model in steps S2-3 includes: The feature matrix is ​​formed by combining the feature vectors of each node in the information propagation direction output by the second part of the G-rumorGAT model. The feature matrix is ​​formed by combining the feature vectors of each node in the information diffusion direction output from the second part of the G-rumorGAT model. in, This is the feature vector of the root node in the transmission direction output by the second part of the G-rumorGAT model. This represents the feature vectors of other nodes along the transmission direction output by the second part of the G-rumorGAT model. This represents the feature vector of the root node in the information diffusion direction output by the second part of the G-rumorGAT model. The feature vectors of other nodes in the information diffusion direction output by the second part of the G-rumorGAT model; Aggregate the feature matrices along the information propagation direction to obtain Aggregating the feature matrices along the information diffusion direction yields... in, aggergate represents aggregate functions.

7. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 3, characterized in that, In step S2-21, the feature vector of each node in the first layer RGAT along the information propagation direction... And the feature vector of each node in the first layer RGAT along the information diffusion direction. The calculation formulas are as follows: in, d1 is the dimension of the feature vector output by the first layer RGAT, σ represents the linear function that can be selected, and R k It is a collection of relation types. The attention coefficient represents the direction of information dissemination. The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of each node in the direction of information propagation. Let be the set of neighboring nodes of each node in the direction of information diffusion. This is the initial feature vector for each node in the set of neighboring nodes along the information propagation direction. This is the initial feature vector for each node in the set of neighboring nodes along the information diffusion direction.

8. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 3, characterized in that, In steps S2-23, the feature vector of each node in the information propagation direction in the second layer RGAT The feature vector of each node in the second layer RGAT along the information diffusion direction. The calculation formulas are as follows: in, d1 is the dimension of the feature vector output by the second-layer RGAT, and σ represents the linear function that can be selected. The attention coefficient represents the direction of information dissemination. The attention coefficient represents the direction of information diffusion. Let be the set of neighboring nodes of each node in the direction of information propagation. Let be the set of neighboring nodes of each node in the direction of information diffusion. For each node in the neighborhood node set along the information propagation direction, this is the enhanced feature vector of the first layer RGAT. This is the enhanced feature vector of each node in the set of neighboring nodes in the information diffusion direction in the first layer of RGAT.

9. The method for constructing the G-rumorGAT model for detecting online rumors according to claim 1, characterized in that, The loss function is: Where C is the set of samples in the training set, validation set, and test set, C = {c1, c2, ..., c3} |c| }, It is the true label vector representation of the i-th sample. is the detection result obtained by the G-rumorGAT model for the i-th sample, β is the hyperparameter representing the magnitude of L2 regularization, and θ is all the parameters that the model needs to learn.

Images