A multi-view rumor detection method based on attention mechanism

Abstract

The present invention discloses a multi-view rumor detection method based on attention mechanism, which includes the following steps: multi-view modeling, including content propagation view and user propagation view; node feature initialization; graph convolution layer, two views are respectively subjected to two layers of graph convolution layer to obtain content propagation features and user propagation features; attention mechanism and classification, introducing attention mechanism to fuse the features of multiple views, and calculating through a classifier composed of a fully connected network layer to obtain a classification result. The advantages of this method are: adopting multi-view modeling to form a more complete and three-dimensional propagation view; adopting SBERT pre-training model to initialize the features of each node in the view, improving the traditional sentence embedding method; adopting a dual graph convolution layer network, by aggregating the neighborhood information of the nodes in the graph, the features of the graph can be effectively captured; adopting the attention mechanism, the weights of different types of features can be adaptively estimated.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and graph convolutional networks, and in particular to a multi-view rumor detection method based on an attention mechanism. Background Art

[0002] Nowadays, social networks play an important role in people's lives. Therefore, the detection and governance of rumors is imperative.

[0003] In order to reduce the heavy manual work, more and more methods use deep neural networks to automatically identify rumor information, and have achieved great improvements in recognition accuracy compared with traditional methods. In order to better capture the characteristics of propagation structure, researchers model the event propagation process as a graph structure, with posts as nodes and interactive relationships such as forwarding and comments as edges. However, obtaining a vector representation of the graph is a challenging task. Recently, graph neural networks (GNNs) have been widely adopted, using the content and structure information of the graph to represent them as feature vectors. At present, many rumor detection methods use graph neural networks to extract the propagation structure characteristics of events, and have achieved better results than existing methods. However, most existing methods that consider propagation structure only consider a single view when constructing a graph model. These modeling methods are incomplete, and ignoring any view will lead to the loss of feature information.

[0004] In response to the above problems, how to construct multiple views to represent rumor events, and on this basis use graph neural networks to extract features of multiple views and effectively fuse them is an issue that needs to be solved urgently. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a multi-view rumor detection method based on attention mechanism.

[0006] A multi-view rumor detection method based on attention mechanism includes the following steps:

[0007] (1) Multi-view modeling: construct two different views from the perspective of content propagation and the perspective of user propagation. The views include content propagation view and user propagation view.

[0008] (2) Node feature initialization: Initialize the features of each node in the view through the SBERT pre-trained model, including content propagation view feature initialization and user propagation view feature initialization;

[0009] (3) Graph convolution layer: After feature initialization, the content propagation view and user propagation view are transformed into content propagation features and user propagation features by extracting structural information from the view through a two-layer graph convolution network.

[0010] (4) Attention mechanism and classification: The attention mechanism is introduced to integrate the content propagation characteristics and user propagation characteristics, and the classification results are obtained through classifier calculation. The network rumor detection model is trained and used to detect network rumors.

[0011] In step (1), the detailed steps of multi-view modeling are as follows:

[0012] (1-1) Content propagation view construction: Each microblog content is regarded as a node, and the interactive relationships such as forwarding and commenting between microblog contents are regarded as edges. The propagation of events diverges downward from the source microblog, forming a gradually large tree structure. Each event is modeled as a A directed graph with root node in is a set of nodes, each node represents a microblog in the event propagation process, n i is the total number of microblogs involved in event i;

[0013] is an edge set, if two microblogs If there is a relationship of forwarding or commenting between them, then there is an edge Defining the adjacency matrix

[0014]

[0015] The element in row s and column t of The definition is as follows:

[0016]

[0017] That is, when there is When The corresponding element value of the s-th row and t-th column is 1, otherwise it is 0.

[0018] (1-2) User propagation view construction: The user view of each event i is modeled as a directed graph in is a set of nodes, each node represents a user participating in event propagation, m i is the total number of users involved in event i;

[0019] is an edge set, if two users If there is a forwarding or reply relationship between them, then there is an edge definition is the adjacency matrix, Indicates that the matrix size is m i Line, m i List:

[0020]

[0021] Each of these elements The value is user Retweeted or commented on the user The number of times.

[0022] In step (2), the detailed steps of node feature initialization are as follows:

[0023] (2-1) Content propagation view feature initialization: the node set is V w ={w0,w1,w2,…,w n-1}, each node w j is a microblog, and the corresponding text content is Extract the initialization features of each node through the SBERT pre-training model Indicates that the feature vector size is 1 row and d columns:

[0024]

[0025] d is the dimension of the initialized feature vector, and the set V w All node features in are combined to obtain the initial feature matrix T of the content propagation view node set w :

[0026]

[0027] (2-2) Initialization of user propagation view features: the node set is V u ={u0,u1,u2,…,u m-1}, each node u j is a user, and all text content published by the user for a certain event is Use the SBERT pre-trained model to extract text content features

[0028]

[0029] SBERT is a pre-trained model for extracting text content features. This method uses twin and three-level network structures to fine-tune BERT. It improves the traditional sentence embedding method and achieves very good results in multiple application scenarios. In Python, we directly call the SentenceTransformer method and use the text content As input, we can get the text content features

[0030] Collection V uAll node features in are combined to obtain the initial feature matrix T of the user propagation view node set u :

[0031]

[0032] In step (3), the two-layer graph convolutional network extracts structural information in the view, specifically including:

[0033] (3-1) The first independent two-layer graph convolutional network Content-GCN: In the content propagation view, the calculation formula of Content-GCN is as follows:

[0034]

[0035]

[0036] in I is the identity matrix, They represent the outputs of the first and second graph convolutional layers (GCL) of Content-GCN, i.e., the hidden states, n is the total number of nodes, v1 is the dimension of the output vector of the first layer, and v2 is the dimension of the output vector of the second layer. and is the parameter matrix in Content-GCN. σ(·) is the activation function, and we use the ReLU function here. Indicates that the matrix size is n rows and v1 columns, Indicates that the matrix size is n rows and v2 columns, It means the matrix size is d rows and v1 columns. It means the matrix size is v1 rows and v2 columns.

[0037] (3-2) The second independent two-layer graph convolutional network User-GCN: The calculation method is similar to Content-GCN. The specific calculation formula is as follows:

[0038]

[0039]

[0040] in I is the identity matrix, They represent the outputs of the first and second graph convolutional layers (GCL) of User-GCN, i.e., the hidden states, respectively. m is the total number of nodes, v1 is the dimension of the output vector of the first layer, and v2 is the dimension of the output vector of the second layer. and is the parameter matrix in User-GCN. σ(·) is the activation function, and we use the ReLU function here. Indicates that the matrix size is m rows and v1 columns. Indicates that the matrix size is m rows and v2 columns. It means the matrix size is d rows and v1 columns. It means the matrix size is v1 rows and v2 columns.

[0041] (3-3) Average pooling: Through average pooling, the final features of the two views are obtained, namely, the content propagation feature S w and user propagation characteristics S u .

[0042]

[0043]

[0044] Among them, meanpooling represents the average pooling operation, that is, averaging each column of the matrix to form a new vector.

[0045] In step (4), the attention mechanism is introduced to fuse the content propagation features and the user propagation features, and the classification results are obtained through classifier calculation, and the network rumor detection model is trained, which specifically includes:

[0046] (4-1) Feature fusion: Attention weights are introduced for each event, and the calculation formula for fusing the features of two views is as follows:

[0047]

[0048] Where type∈{w,u} represents the content propagation view and user propagation view respectively. α t , α i is a learnable parameter that needs to be estimated during the training process. S is the final feature of the event, Indicates that the feature size is v2-dimensional.

[0049] (4-2) Rumor detection: Use S as the feature vector and multi-layer perceptron as the classifier to obtain the classification result and train the network rumor detection model.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] 1. Multi-view modeling is used to form a more complete and three-dimensional propagation view, which facilitates more accurate detection of rumors.

[0052] 2. The SBERT pre-trained model is used to initialize the features of each node in the view, improving the traditional sentence embedding method.

[0053] 3. A dual-graph convolutional layer network is used to effectively capture the features of the graph by aggregating the neighborhood information of the nodes in the graph.

[0054] 4. The attention mechanism can adaptively estimate the weights of different types of features. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flowchart of a multi-view rumor detection model based on attention mechanism;

[0056] Figure 2 Schematic diagram of multi-view modeling, where (a) is an example of microblog event propagation; (b) is the constructed content propagation view; (c) is the constructed user propagation view;

[0057] Figure 3 Schematic diagram of average pooling. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings.

[0059] Reference Figure 1 , a multi-view rumor detection model based on attention mechanism, including the following steps:

[0060] (1) Multi-view modeling: Construct two different views from the perspective of content dissemination and the perspective of user dissemination, and establish a graph model based on this.

[0061] (2) Node feature initialization: The text content features are used as the initialization features of the nodes, and the SBERT pre-trained model is used to extract the text content features.

[0062] (3) Graph convolutional layer: Aggregates the neighborhood information of nodes in the graph and effectively captures the features of the graph.

[0063] (4) Attention mechanism and classification: The attention mechanism is introduced to adaptively estimate the weights of different types of features and perform prediction and classification based on this.

[0064] In step (1), the detailed steps of multi-view modeling are as follows:

[0065] (1-1) Content propagation view construction: Each microblog content is regarded as a node, and the interactive relationships such as forwarding and commenting between microblog contents are regarded as edges. The propagation of events diverges downward from the source microblog, forming a gradually large tree structure. From this perspective, we model each event as a tree with A directed graph with root node in is a set of nodes, each node represents a microblog in the event propagation process, n i is the total number of microblogs involved in event i. is an edge set, if two microblogs If there is a relationship of forwarding or commenting between them, then there is an edge For example, if for retweet, then there is a directed edge, Right now Defining the adjacency matrix

[0066]

[0067] The element in row s and column t of The definition is as follows:

[0068]

[0069] That is, when there is When The corresponding element value of the sth row and tth column is 1, otherwise it is 0 (see Figure 2 (case shown in (b)).

[0070] (1-2) User propagation view construction: We model the user view of each event i as a directed graph in is a set of nodes, each node represents a user participating in event propagation, m i is the total number of users involved in event i. is an edge set, if two users If there is a forwarding or reply relationship between them, then there is an edge For example, if the user Retweeted user Weibo, then there is a directed edge, Right now definition is the adjacency matrix:

[0071]

[0072] Each of these elements The value is user Retweeted or commented on the user The number of times (see Figure 2 (c) shows the case).

[0073] In step (2), the detailed steps of node feature initialization are as follows:

[0074] (2-1) Content propagation view feature initialization: the node set is V w={w0,w1,w2,…,w n-1}, each node w j is a microblog, and the corresponding text content is Extract the initialization features of each node through the SBERT model

[0075]

[0076] d is the dimension of the initialized feature vector, and the set V w All node features in are combined to obtain the initial feature matrix T of the content propagation view node set w :

[0077]

[0078] (2-2) Initialization of user propagation view features: the node set is V u ={u0,u1,u2,…,u m-1}, each node u j is a user, and all text content published by the user for a certain event is Use the SBERT pre-trained model to extract text content features

[0079]

[0080] SBERT is a pre-trained model for extracting text content features. This method uses twin and three-level network structures to fine-tune BERT. It improves the traditional sentence embedding method and achieves very good results in multiple application scenarios. In Python, we directly call the SentenceTransformer method and use the text content As input, we can get the text content features

[0081] Collection V u All node features in are combined to obtain the initial feature matrix T of the user propagation view node set u :

[0082]

[0083] In step (3), the detailed steps of using a two-layer graph convolutional network to extract structural information in the view are as follows:

[0084] (3-1) The first independent two-layer graph convolutional network Content-GCN: In the content propagation view, the calculation formula of Content-GCN is as follows:

[0085]

[0086]

[0087] in I is the identity matrix, They represent the outputs of the first and second graph convolutional layers (GCL) of Content-GCN, i.e., the hidden states, n is the total number of nodes, v1 is the dimension of the output vector of the first layer, and v2 is the dimension of the output vector of the second layer. and is the parameter matrix in Content-GCN. σ(·) is the activation function, and we use the ReLU function here.

[0088] (3-2) The second independent two-layer graph convolutional network User-GCN: The calculation method is similar to Content-GCN. The specific calculation formula is as follows:

[0089]

[0090]

[0091] in I is the unit matrix, and the dimensions of each operator in the formula are consistent with Content-GCN.

[0092] (3-3) Average pooling: Through average pooling, we obtain the final features of the two views, namely, the content propagation feature S w and user propagation characteristics S u .

[0093]

[0094]

[0095] Among them, meanpooling represents the average pooling operation, that is, averaging each column of the matrix to form a new vector. Figure 3 Case shown.

[0096] In step (4), the detailed steps of introducing the attention mechanism to fuse features and classification are as follows:

[0097] (4-1) Feature fusion: Attention weights are introduced for each event, and the calculation formula for fusing the features of two views is as follows:

[0098]

[0099] Where type∈{w,u} represents the content propagation view and user propagation view respectively. αt is a learnable parameter that needs to be estimated during the training process. S is the final feature of the event,

[0100] (4-2) Rumor detection: Using S as the feature vector and a multi-layer perceptron as the classifier, a network rumor detection model is trained.

[0101] In order to prove the effectiveness of the present invention, we selected a series of data based on Weibo platform events compiled by Song et al. and used in the paper. Weibo is one of the most popular social networking sites in China. The events in the data set have two labels, rumor (R) and non-rumor (N). Each event has rich information, such as source text, comments, and user information. Table 1 shows the statistical information of the data set.

[0102] Table 1

[0103]

[0104]

[0105] The present invention divides the data set into a training set, a validation set, and a test set in a ratio of 8:1:1. The present invention uses accuracy, precision, recall, and F1 value as evaluation indicators. The model is optimized by the Adam algorithm, and the learning rate is set to 0.002. The training process is iterated 200 times, and the training is terminated early when the validation loss stops decreasing for 10 iterations.

[0106] There are 7 baseline methods used for comparison in the present invention, namely HAN, Text-CNN, GRU, RvNN, EANN, GAN-GRU, and GLAN. The method of the present invention is named MVGCN. The comparative experimental results are shown in Table 2 below:

[0107] Table 2

[0108]

[0109] As can be seen from Table 2, the model MVGCN proposed in this invention has a significant improvement over other baseline models, with an accuracy rate of 92.9%. Other indicators are also better than traditional methods, which proves that considering multiple views can improve the accuracy of rumor detection tasks. The performance of GRU, Text-CNN, and HAN is inferior to other methods. The main reason is that they only consider the text content of Weibo and ignore the propagation structure characteristics. GLAN and MVGCN outperform other methods because they consider user views compared to other methods, indicating that user views can help identify rumors; in addition, GLAN models users and Weibo as a heterogeneous graph, while our method MVGCN models them as two views respectively, and introduces an attention mechanism to learn the weights of each view, so it is better than GLAN in accuracy.

Claims

1. A multi-view rumor detection method based on attention mechanism, characterized in that: The steps include: (1) Multi-view modeling: construct two different views from the perspective of content propagation and the perspective of user propagation. The views include content propagation view and user propagation view. (2) Node feature initialization: Initialize the features of each node in the view through the SBERT pre-trained model, including content propagation view feature initialization and user propagation view feature initialization; (3) Graph convolution layer: After feature initialization, the content propagation view and user propagation view are extracted through a two-layer graph convolutional network to obtain content propagation features and user propagation features. (4) Attention mechanism and classification: The attention mechanism is introduced to integrate content propagation features and user propagation features, and the classification results are obtained through classifier calculation. The network rumor detection model is trained and used to detect network rumors. In step (1), the content propagation view is constructed by taking each microblog content as a node, and the forwarding and comment interaction relationships between microblog contents as edges. The propagation of events diverges downward from the source microblog to form a gradually large tree structure. Each event is modeled as a A directed graph with root node in is a set of nodes, each node represents a microblog in the event propagation process, n i is the total number of microblogs involved in event i; is an edge set, if two microblogs If there is a relationship of forwarding or commenting between them, then there is an edge Defining the adjacency matrix The element in row s and column t of The definition is as follows: That is, when there is When The corresponding element value of the sth row and tth column is 1, otherwise it is 0; In step (1), the user propagation view is constructed: the user view of each event i is modeled as a directed graph in is a set of nodes, each node represents a user participating in event propagation, m i is the total number of users involved in event i; is an edge set, if two users If there is a forwarding or reply relationship between them, then there is an edge definition is the adjacency matrix: Each of these elements The value is user Retweeted or commented on the user The number of times; In step (2), the content propagation view feature is initialized: the node set is V w ={w0,w1,w2,…,w n-1 }, each node w j is a microblog, and the corresponding text content is Extract the initialization features of each node through the SBERT pre-training model d is the dimension of the initialized feature vector, and the set V w All node features in are combined to obtain the initial feature matrix T of the content propagation view node set w : In step (2), the user propagation view feature is initialized: the node set is V u ={u0,u1,u2,…,u m-1 }, each node u j is a user, and all text content published by the user for a certain event is Use the SBERT pre-trained model to extract text content features Collection V u All node features in are combined to obtain the initial feature matrix T of the user propagation view node set u : In step (3), the two-layer graph convolutional network extracts structural information in the view, specifically including: (3-1) The first independent two-layer graph convolutional network Content-GCN: In the content propagation view, the calculation formula of Content-GCN is as follows: in I is the identity matrix, They represent the outputs of the first and second graph convolutional layers GCL of Content-GCN, i.e., the hidden states. n is the total number of nodes. v1 is the dimension of the output vector of the first layer. v2 is the dimension of the output vector of the second layer. and is the parameter matrix in Content-GCN, σ(·) is the activation function; (3-2) The second independent two-layer graph convolutional network User-GCN is calculated as follows: in I is the identity matrix, They represent the outputs of the first and second graph convolutional layers GCL of User-GCN, i.e., the hidden states. m is the total number of nodes. v1 is the dimension of the output vector of the first layer. v2 is the dimension of the output vector of the second layer. and is the parameter matrix in User-GCN, σ(·) is the activation function; (3-3) Average pooling: Through average pooling, the final features of the two views are obtained, namely, the content propagation feature S w and user propagation characteristics S u ; 2. The multi-view rumor detection method based on attention mechanism according to claim 1 is characterized in that: In step (4), the attention mechanism is introduced to fuse the content propagation features and the user propagation features, and the classification results are obtained through classifier calculation, and the network rumor detection model is trained, which specifically includes: (4-1) Feature fusion: Attention weights are introduced for each event, and the calculation formula for fusing the features of two views is as follows: Where type∈{w,u} represents the content propagation view and user propagation view respectively. α t , α i is a learnable parameter, S is the final feature of the event, (4-2) Rumor detection: Use S as the feature vector and multi-layer perceptron as the classifier to obtain the classification result and train the network rumor detection model.

Images