Rumor Detection Method Based on Bidirectional Graph Neural Network

Abstract

This invention discloses a rumor detection method based on a bidirectional graph neural network, belonging to the fields of natural language processing and information dissemination technology. The specific steps include: extracting word features from user-generated short text data using deep learning methods; establishing a Heterogeneous Information Network (HIN) framework using the extracted word features to integrate additional information and capture the rich relationships between the text and the additional information, thereby reducing text sparsity. This invention proposes a flexible HIN framework for modeling short texts, which can integrate any type of additional information and capture their relationships to address semantic sparsity. Based on this, a Heterogeneous Graph Attention Network (HGAT) based on a two-level attention mechanism is proposed. This network performs text classification by embedding HIN, possessing two levels of attention mechanisms: node-level attention and type-level attention.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information dissemination technology, specifically a rumor detection method based on bidirectional graph neural networks. Background Technology

[0002] The concept of rumors has existed since ancient times. However, it truly became a popular research topic and entered the public eye during the 2016 US presidential election, with its multi-party social media war. This necessitates introducing the concept of social media. With the emergence of platforms like Twitter and Weibo, traditional print media has gradually transformed into social media. We, the users, have shifted from being mere recipients of information to becoming both recipients and creators of information. This has led some users to use these rumors as weapons in their struggles to attract attention and gain personal benefit.

[0003] A widely accepted definition of a rumor is "an unverified and instrumental statement of information circulating in the public domain." This unverified information may eventually prove to be true, or partially or wholly false. In today's interconnected world, rumors can emerge and spread with lightning speed thanks to social media platforms, and they are not only false but also misleading and dangerous to the public. Therefore, timely tracking and debunking of such rumors is crucial.

[0004] Journalists and fact-checking websites (such as snopes.com) strive to track and detect rumors. However, this effort is manual, making it prone to poor coverage and slow speed. Feature-based methods have achieved some success by employing holistic statistics (e.g., number of retweets, propagation time, etc.) from message content, user profiles, and diffusion patterns. However, this approach is overly simplistic and fails to consider both message content and the overall spread of the information.

[0005] Therefore, considering the multi-source heterogeneity of social media data, and utilizing user information, user-generated content, and information dissemination, a more accurate rumor detection algorithm is constructed, achieving more accurate results than single data streams. Furthermore, a rumor detection method based on bidirectional graph neural networks is invented to address the new characteristics of social network data. Summary of the Invention

[0006] In view of the problems existing in the above and / or existing rumor detection methods based on bidirectional graph neural networks, the present invention is proposed.

[0007] Therefore, the purpose of this invention is to provide a rumor detection method based on a bidirectional graph neural network, which can solve the aforementioned existing problems.

[0008] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0009] The rumor detection method based on bidirectional graph neural networks includes the following specific steps:

[0010] S1: Use deep learning methods to extract word features from user-generated short text data, and use the extracted word features to build a heterogeneous information network (HIN) framework, which can integrate additional information and capture the rich relationship between text and additional information, thereby reducing the sparsity of text.

[0011] S2: A heterogeneous graph attention network HGAT with a two-level attention mechanism is used to embed HIN in it for extracting features from short texts;

[0012] S3: For the user's forwarded data, a bidirectional graph neural network is used to construct two graphs, one top-down and one bottom-up. The feature vectors representing the propagation and diffusion information are extracted by concatenating the two graphs. The specific steps are as follows:

[0013] S31: Based on the relationship of propagation and diffusion, for rumor events c i Constructing the propagation structure G<V,E> , making Let X be the adjacency matrix of the rumor event c. i Feature matrix based on rumor propagation tree;

[0014] Use DropEdge to avoid overfitting, the formula is as follows:

[0015] A' = AA drop (1)

[0016] Here, it is assumed that the total number of edges in graph A is N. e Let the random deletion rate be p, and A drop Random sampling N from the original edge set e *p edges are then used to construct the structure;

[0017] In each training session, according to equation (1), select N. e Edges with a ratio of p are removed from the original graph G to form A', thus avoiding overfitting. Furthermore, a Bi-GCN model can be constructed using A' and X.

[0018] S32: After the DropEdge operation, the TD-GCN graph neural network and the BU-GCN graph neural network can be used to obtain the top-down propagation features and bottom-up propagation features, respectively. The hidden layer representation of the two layers of the TD-GCN is as follows:

[0019]

[0020] in and This represents the hidden layer features of the two layers of TD-GCN. and This is the convolution kernel parameter matrix of TD-GCN, where ReLU is used as the activation function;

[0021] By analogy with equations (2) and (3), the two hidden layer features of BU-GCN can be obtained. and

[0022] S33: For the k-th layer GCL of TD-GCN, concatenate the hidden feature vector of each node with the hidden feature vector of the root node of the (k-1)-th layer GCL to form a new feature matrix:

[0023]

[0024] in Utilizing the enhanced hidden layer features of the root node Replace the formula in (2) In this way The formula is as follows:

[0025]

[0026] Equations (5) and (6) can be calculated by analogy to obtain the following results. and

[0027] S34: Aggregate the node representations in TD-GCN and BU-GCN respectively to obtain propagation and diffusion representations. Use pooling operations to aggregate these two sets of node representations, in the following form:

[0028]

[0029] Then, the representations of propagation and diffusion are concatenated to fuse the information from both:

[0030] S = concat(S) TD ,S BU (9);

[0031] S4: Combine the text information representation obtained using HGAT with the propagation and diffusion representation obtained using a bidirectional graph neural network and use it to predict the label, while using cross-entropy to train the model.

[0032] As a preferred embodiment of the rumor detection method based on bidirectional graph neural networks described in this invention, the deep learning method in S1 is as follows:

[0033] Step 1: Use the BERT model to convert the text into a vector matrix. This method enables the extraction and aggregation of data text vector features.

[0034] Step 2: Use BERTopic to extract topics. BERTopic is based on pre-trained BERT, uses UMAP for dimensionality reduction and HDBSCAN for clustering, and uses c-TD-IDF for topic representation.

[0035] As a preferred embodiment of the rumor detection method based on bidirectional graph neural networks described in this invention, wherein: the user-generated graph G = (V, E) in S1 contains short texts D = {d1, ..., d...} m}, topic T = {t1, ..., t} k}, entity E = {e1, ..., e n As a node, V = D∪T∪E.

[0036] As a preferred embodiment of the rumor detection method based on bidirectional graph neural networks described in this invention, the specific steps of S2 are as follows:

[0037] S21: First, we use GCN to perform graph convolution. For a graph G = (V, E), V and E are the set of nodes and the set of edges, respectively, while X ∈ R. |V|*q It is a matrix containing the eigenvectors of all nodes, the adjacency matrix A' = A + I contains self-joins, and the degree matrix is ​​a diagonal matrix M, where M... i,i =∑ j A' i,j The rules for inter-layer propagation are as follows:

[0038]

[0039] in W is a normalized adjacency matrix representation. (l) It is a trainable transformation matrix for a specific layer, σ(·) is the representation of the activation function, and H (l) H is the implicit representation of the node at level l. 0 =X;

[0040] However, due to the heterogeneity of short texts, the original convolution is modified based on general convolution to obtain and use heterogeneous graphs for convolution. The rules for inter-layer propagation are as follows:

[0041]

[0042] in yes A submatrix, whose rows represent all nodes, columns represent the nodes' τ-type neighbors, and the node's representation is H. (l+1)It is achieved by using transformation matrices of different types τ. Node embedding for each type of τ system Then, aggregation is performed to obtain the transformation matrix, which considers the differences between different feature spaces and projects them onto the implicit common space. Finally, during initialization, set...

[0043] Next, based on the generation results of the BERTopic model, edges are established between the text and the topic. Then, entities in the text are identified and mapped to the encyclopedia document. If the text contains entities, an edge is established between the text and the entity. If the similarity between entities is higher than a predefined threshold δ, an edge is also established between them.

[0044] S22: Secondly, a two-layer attention mechanism is used to simultaneously capture the different importance of node-level and type-level nodes;

[0045] In type-level attention, given a specific node v, type-level attention learns the weights of neighbors of different categories, and the type τ of node v is embedded in the sub-element representation. It is the set of neighbors of type τ for v, and then the current h is embedded based on node v. v and h τ The attention score formula for node v with respect to type τ is calculated as follows:

[0046]

[0047] Where || denotes connection, μ τ It is an attention vector of type τ, with parameters shared across all nodes, and σ(·) is the representation of the activation function;

[0048] Then, by normalizing the attention scores for all types using the softmax function, the formula for the type-level attention weights is as follows:

[0049]

[0050] In node-level attention, given a specific node v of type τ and its neighbors v'∈N of type τ',... v ,

[0051] Using embedded h v and h τ and type-level attention score α τ The formula for calculating the attention score of node v is as follows:

[0052] b vv′ =σ(v T·α τ′ [h v ||h v′ (14)

[0053] Where v is the attention vector, and the standardized formula for the node-level attention score obtained by using the softmax function is as follows:

[0054]

[0055] Finally, substituting the two-layer attention mechanism, which includes type-level and node-level attention, into equation (11) and introducing it into heterogeneous graph convolution, the propagation formula between each layer is now as follows:

[0056]

[0057] Among them, B τ It is the attention matrix, (B τ ) i,j =β i,j .

[0058] As a preferred embodiment of the rumor detection method based on bidirectional graph neural networks described in this invention, the Bi-GCN model consists of two graph neural networks, namely TD-GCN and BU-GCN. For TD-GCN, its adjacency matrix is ​​A. TD =A', while for BU-GCN, its adjacency matrix is ​​A. BU =A 'T Both use the same feature matrix X.

[0059] As a preferred embodiment of the rumor detection method based on bidirectional graph neural networks described in this invention, the specific steps of S4 are as follows:

[0060] S41: The text information representation obtained by HGAT processing is concatenated with the propagation and diffusion representation obtained by Bi-GCN processing to fuse the information from both:

[0061] F = concat(H) (L) ,S)(17)

[0062] Where H (L) This represents the node representation of the last layer in HGAT;

[0063] S42: Finally, a softmax layer is used for prediction:

[0064] Z = softmax(F)(18).

[0065] Compared with existing technologies:

[0066] 1. The method of this invention proposes a flexible HIN (Heterogeneous Information Network) framework for modeling short texts. It can integrate any type of additional information and capture their relationships to address semantic sparsity. Based on this, a Heterogeneous Graph Attention Network (HGAT) based on a two-level attention mechanism is proposed. This network performs text classification by embedding HIN, and it possesses two levels of attention mechanisms: node-level attention and type-level attention. The attention mechanism can learn the importance of different neighboring nodes and the importance of different node (information) types to the current node.

[0067] 2. The method of this invention takes into account both the top-down and bottom-up propagation directions of rumors. The model uses three Generative Networks (GCNs), two of which focus on information about the direction of rumor propagation. In these two GCNs, one uses a directed graph of the top-down propagation of rumors to learn the propagation pattern; the other uses an inverse directed graph to capture the dispersion pattern of rumors. Furthermore, each layer of the GCN utilizes information from the root node posts to enhance the influence of the rumor root cause and improve the efficiency of rumor detection.

[0068] 3. The method of the present invention can effectively detect rumors and can be extended to other multi-data-stream or multi-modal domains, with good portability, robustness, effectiveness and practicality. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0071] This invention provides a rumor detection method based on a bidirectional graph neural network. Please refer to [link / reference]. Figure 1 The specific steps are as follows:

[0072] S1: Use deep learning methods to extract word features from user-generated short text data, and use the extracted word features to build a Heterogeneous Information Network (HIN) framework, which can integrate additional information and capture the rich relationship between text and additional information, thereby reducing the sparsity of text.

[0073] in:

[0074] The deep learning methods in S1 are as follows:

[0075] Step 1: Use the BERT (Bidirectional Encoder Representation from Transformers) model to convert the text into a vector matrix. This method enables the extraction and aggregation of data text vector features.

[0076] Step 2: Extract topics using BERTopic, which is based on pre-trained BERT and uses UMAP (Uniform Manifold Approximation and Projection) for dimensionality reduction.

[0077] HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) is used for clustering, and c-TD-IDF (class-based version of Term Frequency-Inverse Document Frequency) is used for topic representation;

[0078] In addition, user-generated graph G = (V, E) in S1 contains short text D = {d1, ..., d2}. m}, topic T = {t1, ..., t} k}, entity E = {e1, ..., e n As nodes, V = D∪T∪E;

[0079] S2: Heterogeneous Graph Attention Networks (HGAT) uses a two-layer attention mechanism. To extract features from short texts, HIN is embedded in it. HGAT uses heterogeneous graph convolution to consider the heterogeneity of different types of information. Moreover, HGAT's two-layer attention mechanism is conducive to capturing the importance of different neighboring nodes and exploring the importance of different node (information) types to specific nodes.

[0080] The specific steps of S2 are as follows:

[0081] S21: First, we use GCN (Graph Convolutional Network) to perform graph convolution. For a graph G = (V, E), V and E are the set of nodes and the set of edges, respectively, while X ∈ R. |V|* q is a matrix containing the eigenvectors of all nodes, and the adjacency matrix A. ' =A+I then includes self-joins, and the degree matrix is ​​a diagonal matrix M, where M i,i =∑ jThe rules for inter-layer propagation of A'i,j are as follows:

[0082]

[0083] in W is a normalized adjacency matrix representation. (l) It is a trainable transformation matrix for a specific layer, and σ(·) is the representation of the activation function, such as ReLU, H (l) H is the implicit representation of the node at level l. 0 =X;

[0084] However, due to the heterogeneity of short texts, the original convolution is modified based on general convolution to obtain and use heterogeneous graphs for convolution. The rules for inter-layer propagation are as follows:

[0085]

[0086] in yes A submatrix, whose rows represent all nodes, columns represent the nodes' τ-type neighbors, and the node's representation is H. (l+1) It is achieved by using transformation matrices of different types τ. Node embedding for each type of τ system Then, aggregation is performed to obtain the transformation matrix, which considers the differences between different feature spaces and projects them onto the implicit common space. Finally, during initialization, set...

[0087] Next, based on the generation results of the BERTopic model, edges are established between the text and the topic. Then, entities in the text are identified and mapped to the encyclopedia document. If the text contains entities, an edge is established between the text and the entity. If the similarity between entities is higher than a predefined threshold δ, an edge is also established between them.

[0088] S22: Secondly, a two-layer attention mechanism is used to simultaneously capture the different importance of node-level and type-level nodes;

[0089] In type-level attention, given a specific node v, type-level attention learns the weights of neighbors of different categories, and the type τ of node v is embedded in the sub-element representation. It is the set of neighbors of type τ for v, and then the current h is embedded based on node v. v and h τ The attention score formula for node v with respect to type τ is calculated as follows:

[0090]

[0091] Where || denotes connection, μ τ It is an attention vector of type τ, with parameters shared across all nodes, and σ(·) is a representation of the activation function, such as Leaky ReLU;

[0092] Then, by normalizing the attention scores for all types using the softmax function, the formula for the type-level attention weights is as follows:

[0093]

[0094] In node-level attention, given a specific node v of type τ and its neighbors v'∈N of type τ',... v ,

[0095] Using embedded h v and h τ and type-level attention score α τ The formula for calculating the attention score of node v is as follows:

[0096] b vv′ =σ(v T ·α τ′ [h v ||h v′ (14)

[0097] Where v is the attention vector, and the standardized formula for the node-level attention score obtained by using the softmax function is as follows:

[0098]

[0099] Finally, the two-layer attention mechanism, which includes type-level and node-level attention, is substituted into equation (2) and introduced into heterogeneous graph convolution. Now the propagation formula between each layer is as follows:

[0100]

[0101] Among them, B τ It is the attention matrix, (B τ ) i,j =β i,j ;

[0102] S3: For the user's forwarded data, a bidirectional graph neural network is used to construct two graphs, one from top to bottom and one from bottom to top. The feature vectors representing the propagation and diffusion information are extracted by splicing the two graphs.

[0103] The specific steps of S3 are as follows:

[0104] S31: Based on the relationship of propagation and diffusion, for rumor events c i Constructing the propagation structure G<V,E> , making Let X be the adjacency matrix of the rumor event c. i Feature matrix based on rumor propagation tree;

[0105] Use DropEdge to avoid overfitting, the formula is as follows:

[0106] A' = AA drop (1)

[0107] Here, it is assumed that the total number of edges in graph A is N. e Let the random deletion rate be p, and A drop Random sampling N from the original edge set e *p edges are then used to construct the structure;

[0108] In each training session, according to equation (1), select N. e Edges with a ratio of p are removed from the original graph G to form A', thus avoiding overfitting. Furthermore, a Bi-GCN (Bi-Directional Graph Convolutional Networks) model can be constructed using A' and X. The Bi-GCN model consists of two graph neural networks: TD-GCN (Top-Down Graph Convolutional Networks) and BU-GCN (Bottom-Up Graph Convolutional Networks). For TD-GCN, its adjacency matrix is ​​A. TD =A', while for BU-GCN, its adjacency matrix is ​​A. BU =A 'T Both use the same feature matrix X;

[0109] S32: After the DropEdge operation, the TD-GCN graph neural network and the BU-GCN graph neural network can be used to obtain the top-down propagation features and bottom-up propagation features, respectively. Taking TD-GCN as an example, the hidden layer representation of its two layers is as follows:

[0110]

[0111]

[0112] in and This represents the hidden layer features of the two layers of TD-GCN. and This is the convolution kernel parameter matrix of TD-GCN, where ReLU is used as the activation function;

[0113] By analogy with equations (2) and (3), the two hidden layer features of BU-GCN can be obtained. and

[0114] S33: For the k-th layer GCL of TD-GCN, concatenate the hidden feature vector of each node with the hidden feature vector of the root node of the (k-1)-th layer GCL to form a new feature matrix:

[0115]

[0116] in Utilizing the enhanced hidden layer features of the root node Replace the formula in (2) In this way The formula is as follows:

[0117]

[0118] The calculations of equations (5) and (6) can be derived by analogy. and

[0119] S34: Aggregate the node representations in TD-GCN and BU-GCN respectively to obtain propagation and diffusion representations. Use pooling operations to aggregate these two sets of node representations, in the following form:

[0120]

[0121] Then, the representations of propagation and diffusion are concatenated to fuse the information from both:

[0122] S = concat(S) TD ,S BU (9);

[0123] S4: Combine the text information representation obtained using HGAT with the propagation and diffusion representation obtained using a bidirectional graph neural network and predict its label, while using cross-entropy to train the model.

[0124] The specific steps of S4 are as follows:

[0125] S41: The text information representation obtained by HGAT processing is concatenated with the propagation and diffusion representation obtained by Bi-GCN processing to fuse the information from both:

[0126] F = concat(H) (L),S)(17)

[0127] Where H (L) This represents the node representation of the last layer in HGAT;

[0128] S42: Finally, a softmax layer is used for prediction:

[0129] Z = softmax(F)(18).

[0130] The above methods can effectively detect rumors and have good portability, robustness, effectiveness and practicality.

[0131] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A rumor detection method based on a bidirectional graph neural network, characterized in that, The specific steps are as follows: S1: Use deep learning methods to extract word features from user-generated short text data, and use the extracted word features to build a heterogeneous information network (HIN) framework, which can integrate additional information and capture the rich relationship between text and additional information. S2: A heterogeneous graph attention network HGAT with a two-level attention mechanism is used to embed HIN in it for extracting features from short texts; S3: For the user's forwarded data, a bidirectional graph neural network is used to construct two graphs, one top-down and one bottom-up. The feature vectors representing the propagation and diffusion information are extracted by concatenating the two graphs. The specific steps are as follows: S31: Based on the relationship of propagation and diffusion, for rumor events c i Constructing the propagation structure G<V,E> , making Let X be the adjacency matrix of the rumor event. i Feature matrix based on rumor propagation tree; Use DropEdge to avoid overfitting, the formula is as follows: A'=A-A drop (1) Here, it is assumed that the total number of edges in graph A is N. e Let the random deletion rate be p, and A drop Random sampling N from the original edge set e *p edges are then used to construct the structure; In each training session, according to equation (1), select N. e Edges with a ratio of p are removed from the original graph G to form A', thus avoiding overfitting. Furthermore, a Bi-GCN model can be constructed using A' and X. S32: After the DropEdge operation, the TD-GCN graph neural network and the BU-GCN graph neural network can be used to obtain the top-down propagation features and bottom-up propagation features, respectively. The hidden layer representation of the two layers of the TD-GCN is as follows: in and This represents the hidden layer features of the two layers of TD-GCN. and This is the convolution kernel parameter matrix of TD-GCN, where ReLU is used as the activation function; By analogy with equations (2) and (3), the two hidden layer features of BU-GCN can be obtained. and S33: For the k-th layer GCL of TD-GCN, concatenate the hidden feature vector of each node with the hidden feature vector of the root node of the (k-1)-th layer GCL to form a new feature matrix: in Utilizing the enhanced hidden layer features of the root node Replace the formula in (2) In this way The formula is as follows: Equations (5) and (6) can be calculated by analogy to obtain the following results. and S34: Aggregate the node representations in TD-GCN and BU-GCN respectively to obtain propagation and diffusion representations. Use pooling operations to aggregate these two sets of node representations, in the following form: Then, the representations of propagation and diffusion are concatenated to fuse the information from both: S=concat(S TD ,S BU )(9); S4: Combine the text information representation obtained using HGAT with the propagation and diffusion representation obtained using a bidirectional graph neural network and use it to predict the label, while using cross-entropy to train the model.

2. The rumor detection method based on a bidirectional graph neural network according to claim 1, characterized in that, The deep learning method in S1 is as follows: Step 1: Use the BERT model to convert the text into a vector matrix. This method enables the extraction and aggregation of data text vector features. Step 2: Use BERTopic to extract topics. BERTopic is based on pre-trained BERT, uses UMAP for dimensionality reduction and HDBSCAN for clustering, and uses c-TD-IDF for topic representation.

3. The rumor detection method based on a bidirectional graph neural network according to claim 1, characterized in that, The user-generated graph G = (V, E) in S1 contains short text D = {d1, ..., d2}. m }, topic T = {t1, ..., t} k }, entity E = {e1, ..., e n As a node, V = D∪T∪E.

4. The rumor detection method based on a bidirectional graph neural network according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: First, we use GCN to perform graph convolution. For a graph G = (V, E), V and E are the set of nodes and the set of edges, respectively, while X ∈ R. |V|* q is a matrix containing the eigenvectors of all nodes, the adjacency matrix A' = A + I contains self-joins, and the degree matrix is ​​a diagonal matrix M, where M i,i =∑ j A' i,j The rules for inter-layer propagation are as follows: in W is a normalized adjacency matrix representation. (l) It is a trainable transformation matrix for a specific layer, σ(·) is the representation of the activation function, and H (l) H is the implicit representation of the node at level l. 0 =X; However, due to the heterogeneity of short texts, the original convolution is modified based on general convolution to obtain and use heterogeneous graphs for convolution. The rules for inter-layer propagation are as follows: in yes A submatrix, whose rows represent all nodes, columns represent the nodes' τ-type neighbors, and the node's representation is H. (l+1) It is achieved by using transformation matrices of different types τ. Node embedding for each type of τ system Then, aggregation is performed to obtain the transformation matrix, which considers the differences between different feature spaces and projects them onto the implicit common space. Finally, during initialization, set... Next, based on the generation results of the BERTopic model, edges are established between the text and the topic. Then, entities in the text are identified and mapped to the encyclopedia document. If the text contains entities, an edge is established between the text and the entity. If the similarity between entities is higher than a predefined threshold δ, an edge is also established between them. S22: Secondly, a two-layer attention mechanism is used to simultaneously capture the different importance of node-level and type-level nodes; In type-level attention, given a specific node v, type-level attention learns the weights of neighbors of different categories, and the type τ of node v is embedded in the sub-element representation. It is the set of neighbors of type τ for v, and then the current h is embedded based on node v. v and h τ The attention score formula for node v with respect to type τ is calculated as follows: Where || denotes connection, μ τ It is an attention vector of type τ, with parameters shared across all nodes, and σ(·) is the representation of the activation function; Then, by normalizing the attention scores for all types using the softmax function, the formula for the type-level attention weights is as follows: In node-level attention, given a specific node v of type τ and its neighbors v'∈N of type τ',... v , Using embedded h v and h τ and type-level attention score α τ The formula for calculating the attention score of node v is as follows: b vv′ =σ(v T ·a τ′ [h v ||h v′ ]) (14) Where v is the attention vector, v'∈N v Then, by using the softmax function, the standardized formula for the node-level attention score is as follows: Finally, substituting the two-layer attention mechanism, which includes type-level and node-level attention, into equation (11) and introducing it into heterogeneous graph convolution, the propagation formula between each layer is now as follows: Among them, B τ It is the attention matrix, (B τ ) i,j =β i,j .

5. The rumor detection method based on a bidirectional graph neural network according to claim 1, characterized in that, The Bi-GCN model consists of two graph neural networks, TD-GCN and BU-GCN. For TD-GCN, its adjacency matrix is ​​A. TD =A', while for BU-GCN, its adjacency matrix is ​​A. BU =A' T Both use the same feature matrix X.

6. The rumor detection method based on a bidirectional graph neural network according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: The text information representation obtained by HGAT processing is concatenated with the propagation and diffusion representation obtained by Bi-GCN processing to fuse the information from both: F=concat(H (L) ,S)(17) Where H (L) This represents the node representation of the last layer in HGAT; S42: Finally, a softmax layer is used for prediction: Z = softmax(F)(18).

Images