A fake news detection method based on interactive graph layered pooling

Abstract

The present invention belongs to the field of fake news detection, and specifically relates to a fake news detection method based on hierarchical pooling of interaction graphs. The method comprises: constructing a comment graph and a propagation graph based on user comment interaction information, wherein each sentence of the news to be detected, the tweets of the news to be detected, and each comment of the tweets are used as nodes of the comment graph, and the text semantic features of each node are used as node attributes; the news to be detected, each original tweeter of the news to be detected, and each forwarding user of each original tweeter are used as nodes of the propagation graph, and the social profile of each user is used as user node attributes; using a node selection-based pooling method to perform hierarchical pooling on the comment graph, with each pooling layer being used to retain key content nodes; using a node clustering-based pooling method to perform hierarchical pooling on the propagation graph, with each pooling layer being used to capture the characteristics of the propagation group; and using the pooling results to evaluate the authenticity of the news to be detected. The present invention can effectively improve the speed and accuracy of fake news detection.

Description

Technical Field

[0001] The present invention belongs to the field of fake news detection, and more specifically, relates to a fake news detection method based on interactive graph layered pooling. Background Art

[0002] Traditional media, such as newspapers and radio, suffer from significant lags in information dissemination. However, social media, benefiting from the rapid development of the internet, boasts low costs, high speed, wide reach, and convenient communication, gradually replacing traditional media in the dominant position of news dissemination. While social media facilitates news dissemination, it also creates opportunities for the spread of false news.

[0003] Fake news detection, also known as rumor detection or fact detection, aims to verify the authenticity of news. The interaction graphs generated by news sources and information dissemination processes possess numerous features, such as news text, comments, and dissemination processes. These features significantly differentiate real news from fake news, providing a promising opportunity for detecting fake news. Existing fake news detection methods fall into two categories: news content-based detection and social content-based detection.

[0004] News content-based detection methods typically only consider the source text of the news, including its textual content and visual information, ignoring the impact of news dissemination on social platforms. They attempt to extract features solely from the news source to detect the authenticity of its content. In terms of news text, existing fake news detection methods can be further categorized as knowledge-based and text style-based. Social platforms offer a rich array of interactive features. In addition to posting information on social platforms, users can also engage in interactive behaviors such as likes, comments, and reposts on content posted by themselves or others. These behaviors constitute the social content of news and provide a sufficient basis for assessing its authenticity. Social content-based fake news detection focuses on the characteristics or influence of the dissemination process, analyzing the potential characteristics of fake news. Currently, such methods can be broadly categorized as social text-based and dissemination structure-based.

[0005] The amount of information contained in news social media content far exceeds the news itself, including text and images. Therefore, most existing fake news detection models are based on social media content. However, existing models based on social media content have many flaws, including the following:

[0006] (1) In terms of comment feature extraction, text semantic feature extraction mostly follows the traditional method of combining word vectors and recurrent neural networks, or uses the BERT model. The traditional semantic extraction method combining word vectors and recurrent neural networks processes text in both forward and reverse order and then splices it together. The bidirectional semantics remain at a shallow level; BERT's text semantic extraction method lags behind the existing RoBERTa model.

[0007] (2) Since comments themselves have a reply structure, the reply relationship contains the content relevance between the comment texts, which is a very important feature. Existing models rarely consider structural information, or only use tree-structured recurrent neural networks for processing. In addition to the backward semantic feature extraction, this method requires traversing the comment tree from top to bottom or bottom to obtain the feature vector of the comment tree, and all comment information will be integrated into it. However, in the real comment environment, there are many meaningless comments, such as advertisements and spam comments. These are irrelevant to the news and have little effect on evaluating the authenticity of the news. The tree-structured recurrent neural network processing method cannot eliminate these redundant comments.

[0008] (3) In the extraction of propagation network features, the method of manually designing features requires a relatively complete understanding of the data and has great limitations. The method of extracting features using graph convolution processes the entire propagation network in a unified manner and fails to capture the multiple group characteristics existing in it. Summary of the Invention

[0009] In response to the shortcomings of the existing technology and the need for improvement, the present invention provides a fake news detection method based on interactive graph layered pooling, which aims to improve the speed and accuracy of fake news detection.

[0010] To achieve the above objectives, according to one aspect of the present invention, a fake news detection method based on interactive graph layered pooling is provided, comprising:

[0011] Based on user comment interaction information, a comment graph and a communication graph are constructed. Each sentence of the news to be tested, the tweets of the news to be tested, and the comments on the tweets are used as nodes in the comment graph, and the text semantic features of each node are used as node attributes. The news to be tested, the original tweeters of the news to be tested, and the retweets of each original tweeter are used as nodes in the communication graph, and the social profile of each user is used as the user node attributes.

[0012] The latent features of the nodes of the comment graph and the communication graph are learned, and the comment graph is layered pooled using a pooling method based on node selection, where each layer of pooling is used to retain key content nodes, thereby obtaining a multi-layer first feature graph; the communication graph is layered pooled using a pooling method based on node clustering, where each layer of pooling is used to capture the characteristics of the communication group, thereby obtaining a multi-layer second feature graph;

[0013] The multi-layer first feature map and the multi-layer second feature map are used to evaluate the authenticity of the news to be detected, thereby completing false news detection.

[0014] Furthermore, the method of using the multi-layer first feature map and the multi-layer second feature map is:

[0015] For the first feature map of each layer and the second feature map of each layer, the global average pooling and global maximum pooling results are concatenated to obtain the feature vector of each feature map;

[0016] The feature vectors of all first feature maps are concatenated to obtain the feature vector of the comment map; the feature vectors of all second feature maps are concatenated to obtain the feature vector of the propagation map;

[0017] The feature vector of the comment graph and the feature vector of the propagation graph are concatenated to obtain final feature data for evaluating the authenticity of the news.

[0018] Furthermore, the social profile includes information about the creator, user creation time, name length, description length, number of favorite tweets, number of followers, number of friends, number of tweets, number of statuses, and authentication information.

[0019] Furthermore, the text semantic features are extracted using a pre-trained RoBERTa model.

[0020] Furthermore, GCN or GAT is used to learn the latent features of the nodes of the comment graph and the propagation graph.

[0021] Furthermore, in constructing the comment graph, the sentence nodes of the news are connected in sequence according to the order of the original news text, the tweets and their comment nodes are connected according to the reply relationship, and when the semantic similarity between two nodes between a news sentence and a tweet, between tweets, or between a comment with a comment level greater than L and its tweet is greater than a threshold, an edge is constructed for the two nodes, where L is a positive integer.

[0022] Furthermore, the value of L is 3.

[0023] The present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the above-mentioned fake news detection method based on interactive graph hierarchical pooling.

[0024] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0025] (1) Since comments themselves have a reply structure, the reply relationship contains the content relevance between the comment texts, which is a very important feature. The present invention introduces both the comment graph and the propagation graph in the detection of fake news. The comment graph is processed by layered pooling to effectively extract key feature information in the graph, evaluate the importance of nodes, remove redundant comments, and retain key semantic nodes in the comment graph; the propagation graph is processed by layered pooling to cluster user nodes and capture the potential group feature information of the propagating user group; features at different pooling levels are spliced ​​to reduce information loss. Therefore, using graph features at multiple pooling levels to detect news authenticity has a higher detection rate and accuracy.

[0026] (2) When constructing the comment graph, the present invention proposes that content associations should also be considered when constructing edges. Deep semantic extraction is integrated with content associations to effectively improve the semantic representation capabilities of the model: news sentences, tweets, and comments are organized into graph-structured data, with edges representing comment-reply relationships or semantic similarity relationships. The pre-trained RoBERTa model is used to extract text semantics, effectively extracting text features while preserving the content associations between texts as much as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A diagram of a fake news detection framework based on interactive graph layered pooling provided by an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of a review graph feature learning module provided in an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of a comment graph pooling operator provided by an embodiment of the present invention;

[0030] Figure 4 A schematic diagram of a propagation graph feature learning module provided in an embodiment of the present invention;

[0031] Figure 5 A schematic diagram of layered pooling provided by an embodiment of the present invention;

[0032] Figure 6 Schematic diagram of graph reading and fully connected layer classification provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0034] Example 1

[0035] A fake news detection method based on interactive graph hierarchical pooling, comprising:

[0036] S1. Construct a comment graph and a communication graph based on user comment interaction information. Each sentence of the news to be tested, the tweets of the news to be tested, and the comments on the tweets are used as nodes in the comment graph, and the text semantic features of each node are used as node attributes. The news to be tested, the original tweeters of the news to be tested, and the retweets of each original tweeter are used as nodes in the communication graph, and the social profile of each user is used as the user node attributes.

[0037] S2. Learn the node potential features of the comment graph and the communication graph, and use the node selection-based pooling method to perform layered pooling on the comment graph. Each layer of pooling is used to retain key content nodes, resulting in a multi-layer first feature map. Use the node clustering-based pooling method to perform layered pooling on the communication graph. Each layer of pooling is used to capture the characteristics of the communication group, resulting in a multi-layer second feature map.

[0038] S3. Use multi-layer first feature maps and multi-layer second feature maps to evaluate the authenticity of the news to be detected and complete fake news detection.

[0039] This example uses user interactions to construct a comment graph and a communication graph, whose node attributes are comments and user social profiles, respectively. After learning the latent features of the graph nodes, the comment graph is layered and pooled using a node selection-based approach to retain key content nodes. For the communication graph, a node clustering-based approach is used to layer and pool the graph to capture the characteristics of the communication group. These two graph features are combined to assess news authenticity.

[0040] About the comment graph. The comment graph of this embodiment is mainly composed of comments, combined with news sources and tweets related to comments, and organized into graph structure data according to the relationship between comment replies and the similarity of text content. Each news sentence is regarded as a node in the comment graph, and is connected in sequence with undirected edges according to the order of sentences in the news. Tweets and comments are regarded as single nodes in the comment graph, and edges are reply relationships. The semantic features of node text can be extracted using pre-trained RoBERTa. The graph contains the main content of the news event and the views of the user group. Through the content correlation between the nodes connected by edges, the key content in the graph can be extracted for evaluating the authenticity of the news.

[0041] That is, we use the interaction information of user comments to construct a comment graph. The nodes in the graph are mainly divided into three categories: news sentences, comments, and tweets related to comments. Since the news content is long, the news is divided into sentences, and each sentence is regarded as a node in the comment graph. The comments and tweets are short and are directly regarded as nodes in the comment graph. The text of the comment graph nodes is cleaned to remove the content that interferes with the semantic feature extraction. The pre-trained RoBERTa model is used to extract the text semantics of the nodes in the comment graph and use it as the attribute of the comment graph node. According to the semantic similarity of the nodes, new edges are added to increase the connectivity of the comment graph, and finally a connected graph G is obtained. c = <V c ,E c ,X c >, where v c,i ∈V c is the i-th node in the comment graph, which is a news sentence, tweet or comment; e c,ij ∈E c is the undirected edge connecting nodes i and j in the review graph; x i ∈X is the text feature vector of the i-th node.

[0042] In order to better illustrate the construction of the comment graph, the following details are given:

[0043] Perform data cleaning on news content, comments, and tweets to remove irrelevant or interfering fields, such as URLs, special characters, and redundant symbols. The cleaning rules are:

[0044] Table 1 Data cleaning rules

[0045]

[0046] For a given news original text S, firstly process it into sentences and get S=[s1,s1,...,s ls ], where ls represents the number of sentences in the news, s i =[w i,1 ,w i,2 ,...,w i,li ] represents the i-th sentence of the original news, where w ij is the jth word in the i-th sentence, and li represents the length of the i-th sentence. Each news sentence is considered as a node in the comment graph, and the pre-trained RoBERTa is used to extract the semantics. The sentences are connected in order in the news with undirected edges as the subgraph structure G of the news content in the comment graph. s = <V s ,E s ,X s >, where V s represents a set of news sentences, E srepresents the sentence edge set, X s Represents the sentence semantic feature matrix.

[0047] Tweets and comments are considered as single nodes in the comment graph, and the connecting edges between nodes are added according to the reply relations saved in the comment file, obtaining the subgraph set G of comment replies in the comment graph. r ={g r1 ,g r2 ,...,g rn}, where g r,i = <V r,i ,E r,i ,X r,i > represents the i-th subgraph, V r,i , E r,i , X r,i is the feature matrix composed of the corresponding node set, edge set, tweets and their comment text features. In particular, v r,i0 ∈V r,i Represents subgraph g r,i The tweets in v r,ij ∈V r,i , (j>0) is the tweet v r,i0 Relevant comments generated.

[0048] The aforementioned comment graph structure is divided into several independent subgraphs, resulting in a relatively sparse network structure and low correlation between subgraphs. Within the same news event, tweets and news articles are highly correlated in content. Furthermore, content across tweets also shares commonalities. When a tweet contains multiple levels of comments, the content between each level of comments and the tweet is also correlated.

[0049] Based on the content association characteristics between news sentences, tweets, and comments, the text semantic similarity between nodes is used to construct new edges, which can improve the connectivity of the network and facilitate the aggregation of network features. In terms of node semantic similarity calculation, the cosine value is used for evaluation. Among them, x1 and x2 are the text semantic feature vectors of the review graph nodes.

[0050] New edges are added between two nodes whose semantic similarity is greater than a threshold of Φ = 0.5. These edges are divided into the following three categories: (1) between news sentences and tweets; (2) between tweets; and (3) between comments and tweets with a comment level greater than L. The value of L is determined by the degree of relevance between the comment itself and the tweet, and can be set to three levels.

[0051] After processing according to the above rules, the comment graph is finally transformed into a connected graph G c = <V c ,E c ,X c>, where v c,i ∈V c is the i-th node in the comment graph, which is a news sentence, tweet or comment; e c,ij ∈E c is the undirected edge connecting nodes i and j in the review graph; x i ∈X is the text feature vector of the i-th node.

[0052] Regarding the diffusion graph. The diffusion graph of this embodiment is mainly composed of original push nodes and forwarding nodes, combined with news sources to construct the diffusion graph. The user's information on the social platform is used as the node attribute of the diffusion graph, thereby retaining the user group characteristics of the entire diffusion graph as the basis for evaluating the authenticity of the news.

[0053] That is, we use the user's communication interaction timeline information and reconstruct the communication network structure based on the user's circle of friends to obtain a communication graph. The nodes in the graph are divided into three categories: news sources, original push nodes, and forwarding nodes. The user's social information is used as the attributes of these nodes. The attribute of the original news node is the average value of the attributes of all user nodes. Finally, we get a connected graph G. p = <V p ,E p ,X p >, where v p,i ∈V p is the i-th user node participating in the communication in the graph, e p,ij ∈E p Represents that there is a propagation relationship between nodes i and j, x p,i ∈X p Represents the feature vector of the i-th propagating user.

[0054] In order to better illustrate the construction of the propagation map, the following details are given:

[0055] Based on the historical forwarding records of tweets and the relationships in the circle of friends, a communication graph is constructed, and the user's information on the social platform is used as the node attributes of the communication graph. The dimensions of the user feature vector are normalized: where x i Represents the value of a certain dimension of the vector, x i,min with x i,min is the maximum and minimum value of this dimension. The user feature table is as follows:

[0056] Table 2 User characteristics of nodes in the propagation graph

[0057]

[0058] These are the only user profiles that can be used, and the more attributes there are, the more comprehensive the user characteristics are. Therefore, all attributes in the table must be used, and then the neural network will directly extract the key attributes.

[0059] Based on the constructed comment graph and propagation graph, the embodiment method enters the graph feature learning stage. Specifically:

[0060] The graph feature learning of the comment graph includes two steps: comment graph feature learning and propagation graph feature learning. Figure 2 As shown. First, use the graph neural network to learn the potential feature representation of the nodes in the graph. Preferably, the graph neural network can choose GCN or GAT. If GCN is used, the mathematical description is: Among them H (l) is the input of the l-th layer GCN network, W is the weight parameter to be learned by the neural network, If GAT is used, the mathematical description is: Where δ represents the activation function, x′ i is the attention result of the i-th node.

[0061] The comment graph pooling operator is as follows Figure 3 As shown, the importance of the node is first calculated, including the structural importance score S1 and the attribute importance score S2, and finally the two are weighted as the final node score S final .

[0062] S1 represents the importance score of the node topology structure, which aims to calculate the importance score of each review graph node based on the structural information of the review graph, corresponding to Figure 3 The GCN part in the news comment graph G c Graph structure information is an important component, which contains the reply relationship or text content similarity between nodes. Since GCN takes structural information into account when learning graph node representation, the comment graph pooling operator uses GCN to learn the structural importance of nodes, which is mathematically described as in X c where α is the adjacency matrix and identity matrix of the pooled input review graph, σ is the node degree matrix and identity matrix, and σ is the activation function. α is a hyperparameter representing the weight ratio between node structure and attribute scores. The smaller the value, the more dependent the final score is on the attribute score, and vice versa.

[0063] S2 represents the node attribute importance score. In addition to the topological information in the network, the comment graph node also contains its own semantic features. In the comment graph, this feature represents the potential stance information of news sentences, tweets, and comments, and is the most direct and important feature representing the node. The node attribute importance score directly uses the semantic information of the comment graph node itself to learn the importance score of the node attribute. Specifically, a fully connected neural network is used to calculate the importance of the node text semantics. The mathematical expression is S2 = σ(MLP(X c )). Where X c The feature matrix of the review graph is input to the pooling.

[0064] After the previous two steps, we get the importance score of the node in the graph structure and the node's own semantics. By weighted summing the two scores, we get the final importance score S of the review graph node. final , which corresponds to Figure 3 The SFTL part in the mathematical description is S final =αS1+(1-α)S2.

[0065] After calculating the importance score of each comment graph node, the TopK nodes are retained according to the ranking as the pooled node set, that is, the TopK key content nodes, and then the GAT module is used to learn the pooled node features. Since GAT considers the features of direct neighbor nodes when learning the potential features of nodes, this feature learning method can improve the utilization rate of unselected nodes and reduce the problem of graph feature loss caused by discarding the remaining nodes due to the TopK restriction. After the TopK key nodes are selected and the node features are learned, the edges of the pooled graph are obtained based on the retained nodes. For the edge e in the original graph ij ∈E c , if nodes i and j are both retained after pooling, then the edge is retained after pooling, otherwise it is removed.

[0066] Regarding the pooling of the review graph, in general, we use a node-selection-based pooling method to evaluate the importance of each node and select the top K key nodes as the pooled result. The purpose of the review graph pooling topology learning is to use the topological information of the review graph to calculate an importance score for each node. This topology learning is divided into three parts:

[0067] (1) Structure-based topology learning: Calculate the importance of nodes based on the structure, which is mathematically described as Where S1 represents the importance score of the node topology structure, X c are the adjacency matrix and identity matrix of the pooled input review graph, the node degree matrix and identity matrix, and the node feature matrix, respectively. σ is the activation function.

[0068] (2) Attribute-based topology learning: The importance of nodes is calculated mathematically based on the nodes themselves, which is described mathematically as S2 = σ(MLP(X c )), where X c The feature matrix of the pooled input comment graph is used, and MLP is a fully connected layer.

[0069] (3) Structure-attribute based topological learning: integrating node structure and mathematical importance, mathematically described as S final =αS1+(1-α)S2, where α is a hyperparameter representing the weight ratio of node structure and attribute score. The smaller the value, the more the final score depends on the attribute score, and vice versa.

[0070] The graph feature learning method of the propagation graph is as follows Figure 4 As shown, it is also divided into two sub-steps: graph node potential feature learning and graph pooling. p,i is regarded as the central node of the local cluster, and the local cluster is recorded as c h (v p,i ), and accordingly the node v p,i The nodes within the h-hop range of and regarded as a member node in the cluster, that is, And remember For node v p,i The cluster c is the center h (v p,i ) feature representation. The node cluster assignment matrix of the propagation graph is Among them S p,ij Represents node v p,i With cluster c h (v p,j ) and the member strength between them, S p,ij The value range of is [0,1]. The larger the value, the more likely the node belongs to the cluster. The characteristics of a cluster are determined by the characteristics of the nodes and the strength of the members in the cluster. The stronger the node member strength, the more representative it is of the cluster. The characteristics of the nodes in the cluster can be regarded as the cluster characteristics by taking the weighted sum of their member strengths.

[0071] In the process of learning the weights between elements, the self-attention mechanism is a common method, but this mechanism is used to evaluate the correlation between elements and does not effectively utilize the cluster information. In order to better represent the cluster characteristics, the pooling operator modifies the self-attention mechanism. Specifically, it introduces the main query Used to represent all nodes in the same cluster, its mathematical description is: m i =f m (x′ p,j |{v p,j ∈c h (v p,i)}), where x′ p,j represents the potential features of the jth propagation node learned by the GNN module before pooling, which contains the node's own attributes and structural information in the cluster. m Is the main function, used to convert and merge nodes v p,j ∈c h (v p,i ) features and generate the main query m i .f m Using the max main function From this we can see that m i It uses the information of all nodes in the cluster and has the meaning of representing the cluster. i The additional cluster information is used to process all nodes in the cluster and calculate the membership strength of each node in the cluster:

[0072]

[0073] Among them, || represents the splicing operation, and W are the weight parameters to be learned, α i,j It is the attention score of each node, and also represents the membership strength between each node and the cluster, that is, S p,ij =α i,j . Using the attention score, the cluster features can be expressed as:

[0074]

[0075] For a given cluster, cluster characteristics are one of the important bases for evaluating its rationality. In addition, the greater the difference between a cluster and its neighboring clusters, the more obvious the group characteristics of the cluster are. Therefore, the two can be combined to evaluate the rationality of the cluster. Specifically, the rationality score of the cluster is calculated as follows:

[0076]

[0077] in represents the neighbor nodes of the i-th propagation node, W1, W2 and W3 are the weight parameters to be learned, σ represents the activation function, is the cluster node adjacency matrix, which is equal to the graph node adjacency matrix before pooling. The value is the edge weight. The initial default weight of all edges is 1. Combining the rationality scores of all clusters can get the vector where |V p | is the number of nodes in the propagation graph, which also represents the number of clusters. Multiplying Φ by the cluster feature matrix can obtain the new feature matrix of the cluster under the consideration of rationality:

[0078]

[0079] Where ⊙ represents the Hadamard product under the broadcast mechanism, that is, each row of the column vector Φ is first expanded to N dimensions according to the row header value to obtain an N×N rationality matrix, and then Performs a Hadamard product of matrices.

[0080] According to the rationality score of the cluster, the TopK mechanism is used to retain the top cluster with the largest score. clusters, where N is the number of nodes in the propagation graph, which also represents the number of clusters, k is the pooling rate, and its value range is [0,1]. The cluster index set It can be expressed as: Propagation Map G p The pooled graph G′ p By the retained The original image G p Middle node and G′ p The allocation matrix between clusters and G′ p Cluster node feature matrix It can be expressed as:

[0081]

[0082]

[0083] in Indicates that the number belongs to the set Column vector of , Indicates that the number belongs to the set A row vector of .

[0084] In the cluster feature learning phase, node clustering, cluster feature learning, and key cluster retention are completed. Graph pooling converts the original graph into a smaller subgraph with clustered features, so it is also necessary to construct the adjacency matrix of the cluster nodes. in From this we can know the edge between clusters i and j That is, if cluster i and cluster j have common nodes in the original graph, then cluster i and cluster j are neighbor nodes after pooling, and the connection weights between clusters are determined by the internal nodes through the distribution matrix and edge weights Decide.

[0085] Regarding the pooling of the communication graph, in general, a node clustering-based pooling method is used. Users are clustered according to their potential characteristics, divided into multiple groups, and the authenticity of news is tested based on the characteristics of the groups. In the communication graph, neighboring nodes with edge connections are much more likely to belong to the same social group than other nodes. Therefore, when clustering nodes, the nodes can be divided based on their neighbor information. In addition, a user node may belong to multiple groups, so the nodes are soft-partitioned. The process includes:

[0086] (1) Cluster feature learning and screening: Using the idea of ​​local clustering, first calculate the main query m containing cluster information for each local cluster. i =f m (x′ p,j |{v p,j ∈c h (v p,i )}), the main query contains the cluster information. The main query is used to calculate the importance score of each node in the cluster. The importance score also represents the strength of the node's membership in the cluster. Then the rationality score of each cluster is calculated. The calculation rule is in represents the neighbor nodes of the i-th propagation node, W1, W2 and W3 are the weight parameters to be learned, σ represents the activation function, is the cluster node adjacency matrix, which is equal to the graph node adjacency matrix before pooling. The value is the edge weight. The initial default weight of all edges is 1.

[0087] (2) Cluster node connectivity: with the help of node cluster allocation matrix And the cluster connection matrix of the original image and Equal to the node adjacency matrix A of the original graph p , the mathematical description of cluster node edge construction is: in

[0088] After graph feature learning, we enter the graph reading and fully connected classification stage. Specifically:

[0089] like Figure 5 As shown in Figure 2, for a given graph G, potential feature learning and pooling operations can be performed three times to obtain shallow, mid-level, and deep graph features. In particular, for the review graph G c , where Graph Pooling is pooling based on node selection, for the propagation graph G p ,Graph Pooling is pooling based on node clustering.

[0090] Since the feature map after pooling is still graph structure data, it cannot be directly input into the fully connected layer for classification tasks. Therefore, it is necessary to use a graph readout mechanism, which is a graph data representation method that can convert the graph into a feature vector. Specifically, the global average pooling and global maximum pooling results are spliced ​​together:

[0091]

[0092] Graph readout and fully connected layer classification modules such as Figure 6 As shown, the comment graph and the propagation graph are read out separately, and the feature vectors after splicing the features of each pooling layer are recorded as Among them, Nc and Np are the hidden layer dimensions of the comment graph and the propagation graph node in the graph feature learning stage respectively. The value 6 includes three pooling levels and two graph readout methods. In order to avoid the classifier being affected by the large difference in the dimensions of the two spliced ​​vectors, the feature vector X is first c With X p Perform linear mapping and compress to obtain Where Ncp is the dimension of the compressed feature vector.

[0093] In the classifier, the number of nodes in the first fully connected hidden layer is equal to 2Ncp, the number of nodes in the second hidden layer is Ncp, and the number of nodes in the last layer is 2, corresponding to the fake news binary classification task. As for the loss function, the negative logarithmic loss function is used:

[0094]

[0095] where y n is the true label of the news, 0 represents fake news, 1 represents real news, and p represents the model's predicted news label y n probability.

[0096] That is, the comment graph and the communication graph are pooled three times each to obtain shallow, middle, and deep pooled graphs. Then, the graph features of the three pooled levels are obtained by graph reading. The reading method uses the global average pooling and global maximum pooling splicing. where x i The feature vector of the i-th node in the graph. The three pooling features are concatenated to obtain the comprehensive features of the two graphs, and then the two comprehensive features are concatenated and sent to the fully connected layer for classification.

[0097] It should be noted that, after analysis by this application, the existing false news detection does not consider the preservation of comment structure and the removal of redundant comments at the same time. However, since the comments themselves have a reply structure, the reply relationship contains the content relevance between the comment texts, which is a very important feature. Existing models rarely consider structural information, or only use tree-like recurrent neural networks for processing. In addition to the backward semantic feature extraction, this method requires traversing the comment tree from top to bottom or bottom to obtain the feature vector of the comment tree, and all comment information will be integrated. However, in the real comment environment, there are many meaningless comments, such as advertisements, spam comments, etc. These are irrelevant to the news and have little effect on evaluating the authenticity of the news. The tree-like recurrent neural network processing method cannot eliminate these redundant comments. In addition, in the feature extraction of the propagation network, the method of manually designing features requires a relatively complete understanding of the data and has great limitations. The method of extracting features using graph convolution processes the entire propagation network in a unified manner, and does not capture the characteristics of multiple groups existing therein. This application proposes to simultaneously introduce comment graphs and communication graphs into fake news detection. When constructing the comment graph, the heterogeneous interaction graph of news on social platforms is utilized to construct a comment graph dominated by comments, where nodes represent text content such as news, comments, and tweets, and edges represent comment replies or content similarity relationships. In addition, fake news detection uses graph neural networks to learn the latent features of comment and communication graph nodes, and uses a node selection-based method to pool the comment graph to retain the key semantic nodes in the comment graph. The communication graph is also pooled using a node clustering-based method to capture the characteristics of the communication group, and the graph features of multiple pooling levels are used to detect the authenticity of news. Using the node selection-based pooling method, an importance score is first calculated for each node based on the node's attribute characteristics and structural information, and the top K important nodes are retained as the pooled node set. The edges of the pooled graph are then generated based on the original graph structure information to maintain the connectivity of the resulting graph. Therefore, pooling the comment graph can retain the key semantic nodes in the comment graph and remove redundant comment nodes. Studies have shown that fake news easily spreads within specific social circles, forming an echo chamber effect. In a propagation graph, neighbor nodes with edge connections are much more likely to belong to the same social group than other nodes. Therefore, when clustering nodes, they can be divided according to the node's neighbor information; in addition, a user node may belong to multiple groups, so soft division of the node is more reasonable than hard division. The present invention introduces a pooling operator for propagation graph processing, which adopts the idea of ​​local clustering to softly divide the nodes in the graph, and can capture local subgraph information in layers to learn global features with better edge connectivity in the pooled graph. Pooling the propagation graph can capture group information in the propagation graph.Furthermore, since pooling cannot avoid information loss while aggregating features, splicing features from different pooling levels can reduce information loss. Therefore, the method of the present invention uses the splicing of graph features from multiple pooling levels to detect news authenticity.

[0098] In summary, the present invention combines graph convolution and graph pooling technology, starting from the characteristics of comments and dissemination on news social platforms, to more effectively detect the authenticity of news. It includes: designing a false news detection framework, defining the model structure, the form of input and output data, to obtain the authenticity evaluation of news content; using the social heterogeneous interaction graph of news to construct isomorphic comment graph and dissemination graph, and using comments and user social information as node attributes of the comment graph and dissemination graph respectively, and using the pre-trained RoBERTa model to extract comment semantics; using graph convolution or graph pooling to learn the potential features of graph nodes; for the comment graph, adopting a pooling strategy based on node selection to retain key comments; for the dissemination graph, using a pooling strategy based on node clustering to capture the characteristics of the dissemination group; combining the layered pooling method to read and fuse graph features as the basis for news authenticity evaluation.

[0099] Example 2

[0100] A computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the processor controls the device where the storage medium is located to execute a fake news detection method based on interactive graph layered pooling as described in Example 1.

[0101] The relevant technical solutions are the same as those in Example 1 and will not be described again here.

[0102] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fake news detection method based on interactive graph layered pooling, characterized by: include: Based on user comment interaction information, a comment graph and a communication graph are constructed. Each sentence of the news to be tested, the tweets of the news to be tested, and the comments on the tweets are used as nodes in the comment graph, and the text semantic features of each node are used as node attributes. The news to be tested, the original tweeters of the news to be tested, and the retweets of each original tweeter are used as nodes in the communication graph, and the social profile of each user is used as the user node attributes. The latent features of the nodes of the comment graph and the communication graph are learned, and the comment graph is layered pooled using a pooling method based on node selection, where each layer of pooling is used to retain key content nodes, thereby obtaining a multi-layer first feature graph; the communication graph is layered pooled using a pooling method based on node clustering, where each layer of pooling is used to capture the characteristics of the communication group, thereby obtaining a multi-layer second feature graph; The multi-layer first feature map and the multi-layer second feature map are used to evaluate the authenticity of the news to be detected, thereby completing false news detection.

2. The fake news detection method according to claim 1, characterized in that: The method of using the multi-layer first feature map and the multi-layer second feature map is: For the first feature map of each layer and the second feature map of each layer, the global average pooling and global maximum pooling results are concatenated to obtain the feature vector of each feature map; The feature vectors of all first feature maps are concatenated to obtain the feature vector of the comment map; the feature vectors of all second feature maps are concatenated to obtain the feature vector of the propagation map; The feature vector of the comment graph and the feature vector of the propagation graph are concatenated to obtain final feature data for evaluating the authenticity of the news.

3. The fake news detection method according to claim 1, characterized in that: The social profile includes information about the creator, user creation time, name length, description length, number of favorite tweets, number of followers, number of friends, number of tweets, number of statuses, and authentication information.

4. The fake news detection method according to claim 1, characterized in that: The text semantic features are extracted using the pre-trained RoBERTa model.

5. The fake news detection method according to claim 1, characterized in that: GCN or GAT is used to learn the node potential features of the comment graph and the propagation graph.

6. The fake news detection method according to any one of claims 1 to 5, characterized in that: In constructing the comment graph, the sentence nodes of the news are connected in sequence according to the order of the original news text, the tweets and their comment nodes are connected according to the reply relationship, and when the semantic similarity between two nodes is greater than a threshold L, between a news sentence and a tweet, between tweets, or between a comment with a comment level greater than 3 and its tweet, an edge is constructed for the two nodes, where L is a positive integer.

7. The fake news detection method according to claim 6, characterized in that: The value of L is 3.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, the processor controls the device where the storage medium is located to execute the fake news detection method based on interactive graph hierarchical pooling according to any one of claims 1 to 7.

Images