A rumor detection method based on information propagation structure

Abstract

A rumor detection method based on information propagation structure is proposed. This method constructs a tweet propagation tree from social network data according to retweet relationships. The propagation tree is then converted into a user retweet subgraph using a node replacement strategy, and a global user retweet network is constructed by merging nodes and edge attributes. The local and global structural features of the user retweet network are extracted using the "Depth-Walk Spanning Tree" algorithm and the WL algorithm, respectively, and these two are concatenated to obtain a complete structural encoding. Simultaneously, a pre-trained large language model is used to extract tweet content features to obtain content encoding. Algorithms such as Laplacian matrix and affinity ranking are used to calculate positional encoding information on the user retweet network. These encodings are then concatenated and input into a Graph Transformer model to capture the correlation between various features and the authenticity of the propagated tweets. Finally, an MLP classifier is used to determine whether a tweet is a rumor. This provides an efficient and reliable solution for rumor detection in social networks.

Description

A rumor detection method based on information propagation structure Technical Field

[0001] This invention relates to the field of rumor detection technology, specifically a rumor detection method based on information dissemination structure. Background Technology

[0002] With the rapid development of internet and mobile internet technologies, online social media platforms have gradually become important channels for information dissemination and acquisition in the new media environment. In recent years, the number of users and content on social platforms such as Twitter and Sina Weibo has surged, and online social networking services have become increasingly sophisticated. Compared with traditional media such as television and newspapers, online social media platforms have certain advantages in terms of real-time performance, freedom, speed of dissemination, and user interaction, and have received widespread attention and use from users.

[0003] The first challenge is the early detection of rumors. If potential rumors can be detected early, and their high click-through rates predict a significant impact on the public in the near future, platforms can promptly flag them, alerting netizens to exercise caution and allowing for timely intervention and control to prevent negative social consequences. Therefore, effective early detection of rumors is crucial.

[0004] The second challenge is insufficient extraction of propagation features. The spread of rumors often relies on social relationships and interaction patterns among users, such as forwarding and commenting. Some users may act as "propagators," rapidly spreading rumors to a wider audience, while others may inhibit their spread through comments or forwards. However, if these complex interaction relationships cannot be effectively modeled, the detection model may ignore key social network features, thus reducing detection performance. These methods typically treat each tweet and its propagation tree as an independent graph structure, ignoring the connections between different propagation trees. Furthermore, existing interaction feature extraction methods struggle to effectively model the relationships between tweets posted and forwarded by the same user across multiple propagation trees. This independent processing prevents the model from capturing cross-propagation tree informational connections, thereby reducing the accuracy of rumor detection. Insufficient extraction of interaction features directly impacts rumor detection performance. First, the behavior of users posting and forwarding tweets across multiple propagation trees often contains rich information, which is crucial for determining their motivations and behavioral patterns. If these relationships cannot be effectively modeled, the model will not be able to fully utilize users' historical behavioral data, thus reducing the robustness of the detection algorithm. Secondly, the connections between different propagation trees can reveal the underlying patterns and network structure characteristics of rumor propagation, which is of great value for identifying complex rumor propagation mechanisms. Furthermore, insufficient interaction features may prevent the model from distinguishing between real information and rumors, especially when faced with highly ambiguous or concealed rumors. This underscores the importance of deeply exploring the characteristics of rumor propagation.

[0005] The third challenge is the limitation of propagation structure modeling. Current graph structure representation learning techniques face certain limitations in rumor detection tasks, especially in capturing the structural features of rumor propagation. Early dimensionality reduction-based methods, while capable of handling non-linear data, lack scalability and are prone to overfitting. Matrix factorization-based methods, while providing a global perspective, are insufficient in distinguishing local structural differences, failing to meet the sensitivity requirements of rumor detection for local features. Furthermore, classic random walk-based methods are inefficient in handling graph structures composed of propagation trees, have limited discriminative power for node and edge information, and struggle to explore deeper levels of graph structure. In contrast, while deep learning-based methods can handle complex graph data, they still have significant shortcomings in capturing long-distance dependencies between nodes, utilizing edge attribute information, and handling large-scale directed graph data. Early methods primarily described these interactions by constructing propagation trees, but this approach struggles with complex network structures and diverse data types. Moreover, while traditional RvNNs (Random Walk Neural Networks) or GCNs (Graph Convolutional Networks) can capture some structural features, they exhibit high computational complexity and limited adaptability when dealing with large-scale, dynamically changing social networks. To solve this problem, it is necessary to effectively model and uncover the distinguishing features between rumors and true information during the dynamic dissemination process.

[0006] Rumor detection on online social networks, following the development of research methods, can be mainly divided into two categories: detection methods based on traditional machine learning and detection methods based on deep learning. Traditional machine learning research primarily focuses on feature engineering, specifically mining static attributes of the tweet itself and its propagation process. Castillo et al., in their research on information credibility on the X platform, systematically proposed four feature-based approaches: text content-based, user attribute-based, topic-based, and propagation-based, making pioneering contributions to information credibility research. In the same year, Qazvinian proposed features related to the URL and hashtags associated with the tweet, such as calculating the distribution of topic hashtags in normal information and rumors. Wu modeled the information propagation process as a tree structure, using a random walk graph kernel algorithm to model the similarity between propagation trees. Based on this, and combining message features, user features, and forwarding features, he implemented rumor detection using methods such as SVM.

[0007] Deep learning-based methods primarily focus on modeling the context of information dissemination and mining the dynamic semantic representation of text. Graves et al. proposed the LSTM model to alleviate the long-term dependency problem that RNN networks struggle to solve. In 2014, Chung et al. simplified the gating units in LSTM, proposing the GRU model, which significantly reduced the number of parameters and improved computational speed. Currently, LSTM and GRU models are the most widely used RNN sequence models. In 2016, Ma et al., at the international conference IJCAI 2016, simplified the information dissemination process into a simple linear process based on temporal relationships, thus simulating the context of information dissemination. They were the first to use an LSTM deep model to model this information sequence, capturing the information dependencies between adjacent time points and performing rumor detection. Building on this, Ling et al. proposed a multi-task rumor detection model. On one hand, it uses an LSTM model to detect each user's stance and attitude towards the original tweet; on the other hand, it combines user text content and user information to detect rumors. RNN-based rumor detection methods neglect the interactive relationships of information in the actual dissemination process. Ling et al. improved the masking mechanism of Transformer to handle graph-structured data simulating information propagation processes, and the parallel mechanism of Transformer greatly improved computational efficiency. Building on the research of Ma et al., Bian et al. used Bi-GCN technology to simultaneously handle information propagation and diffusion processes in two directions, ultimately fusing information from different directions to achieve rumor detection. Based on this, Yang et al. considered the above relationships simultaneously, constructing a heterogeneous network with users and tweets as nodes, and achieving rumor detection on this network through adversarial learning. In addition, some studies have introduced Large Language Models (LLMs) for application in tasks.

[0008] In summary, traditional machine learning-based detection methods typically rely on researchers' prior knowledge, making it difficult to guarantee feature completeness. From a content feature perspective, simple statistical features struggle to represent underlying high-level semantic features and abstract information such as opinions and sentiments. Furthermore, at a higher level, different types of features are difficult to interact with, hindering effective fusion of different feature types. Deep learning-based detection methods, whether RNN-based, GNN-based, or based on large language models, have consistently focused on capturing the dynamics of tweets within their own propagation structure, never considering the impact of changes in the tweet structure on the detection results from a global perspective.

[0009] The differences between this application and the prior art are as follows:

[0010] Technical comparison with patent CN202210025558.X, "A rumor detection method and system based on multi-layer coding network".

[0011] Patent CN202210025558.X proposes a rumor detection method and system based on a multi-layer coding network. The method includes the following steps: acquiring all text to be detected and preprocessing the text; embedding word fragments with a labeled vocabulary into the preprocessed text, converting words in the text into tokenized words, and then performing vector encoding to obtain word vectors corresponding to each text; processing all word vectors to obtain input vectors; inputting the input vectors into a pre-trained multi-layer coding network to generate output vectors; processing the output vectors to obtain hidden state vectors; and feeding the hidden state vectors into a hidden layer and a classifier to obtain the probability that the text to be detected is classified as a rumor, with the category with the highest probability being the detection result for that text. Compared with existing technologies, this patent not only models the features of tweet text but also focuses on temporal information and the interaction relationships between users. By designing a user forwarding network, it achieves cross-propagation tree user relationship modeling. The two differ fundamentally in their implementation methods and technical approaches.

[0012] Technical comparison with patent CN119166907A "Multimodal rumor detection method and system integrating multi-granularity features"

[0013] Patent CN119166907A proposes a multimodal rumor detection method and system that integrates multi-granularity features. The method includes the following steps: collecting multimedia posts from social media, extracting text, images, and comments from the posts, and labeling the posts with authenticity tags to construct a training dataset; constructing a multimodal rumor detection model that integrates multi-granularity features, extracting multi-granularity features from the posts, deeply fusing text and image features, and resolving text-image ambiguity through text-image similarity weights; fully mining information from comments, utilizing comments from different angles, and finally capturing similar information from similar social media posts through comparative learning; training the multimodal rumor detection model using the training dataset; and inputting the text, images, and comments of undetected multimedia posts into the trained multimodal rumor detection model to obtain authenticity tags for the multimedia posts. Unlike other multimodal data sources, this patent focuses on collecting text data and user information, which has higher practical significance in real-world applications. Furthermore, this patent delves into the propagation structure of rumors and constructs a user forwarding network based on the original tweet propagation tree to better capture user relationships across the propagation tree. The two inventions differ fundamentally in their content.

[0014] To address the problem of early detection of rumors on online social networks, this invention proposes a novel structural information extraction algorithm from the perspective of rumor propagation structure. To overcome the insufficient extraction of propagation features, this invention proposes a "deep walk spanning tree" graph structure representation algorithm based on Prim's ideas. This method uses a spanning tree-based walking approach to replace traditional random walks and neighbor aggregation methods to perceive local structure. Combined with a tree coding algorithm, each seed node can obtain its encoded representation of the structure of its surrounding neighbors, enabling better capture of local propagation structure differences in subgraphs. To address the limitations of propagation structure modeling, this invention constructs a user retweet network that incorporates tweet propagation relationships. This network construction method better solves the problem that a single propagation tree cannot model the interaction relationships between information and other information during propagation. It can structurally model the propagation behavior of different tweets from the same user, thereby better capturing the impact of tweet interaction behavior on the information propagation structure from the user's perspective. This model not only improves the full mining and effective utilization of propagation information in rumor detection but also makes rumor detection more accurate, enabling timely capture of information features such as tweet content, retweet relationships, and user attributes, thus providing strong support for relevant decision-making. Summary of the Invention

[0015] To address the above problems, this invention proposes a rumor detection method based on information propagation structure. By using the ordered walk of a spanning tree, it can more systematically encode the contribution of each node to the surrounding neighborhood structure, thus more accurately reflecting the local propagation characteristics of the subgraph, rather than the disordered random walk. At the same time, it effectively integrates multiple propagation paths and user behaviors, revealing the connections formed between different information through common users, thereby more comprehensively reflecting the complexity of information propagation, rather than the simple accumulation of local propagation information of a single propagation tree, effectively improving the accuracy and richness of rumor detection.

[0016] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0017] A rumor detection method based on information propagation structure, characterized by the following steps:

[0018] (1) Data Analysis Module:

[0019] The original datasets PHEME and Misinfdect were cleaned and preprocessed to remove irrelevant information. The Chinese text was segmented using the Jieba word segmentation tool. A tweet propagation tree was constructed based on the relationship between the original tweets and their retweet tweets.

[0020] (2) Network construction module:

[0021] Based on the tweet propagation tree data obtained in step (1), a user retweet network is constructed. First, for each propagation tree, the tweet nodes are replaced with the user nodes that posted the tweet, preserving the original structure of the propagation tree, and the user retweet subgraph is obtained; then, all user retweet subgraphs are merged, and the same... Only one user node is retained, and only one edge in the same direction between two identical users is retained. The node attributes remain unchanged, and the edge attributes are merged to form a complete user forwarding network.

[0022] (3) Structure perception module:

[0023] Based on the user forwarding network obtained in step (2), and combined with the forwarding attributes such as the number of forwarding edges provided by the propagation tree, the structure code is extracted. First, the global structure code is extracted using the WL algorithm. Then, the local structure information from the user node's perspective is captured by the "depth walk spanning tree" algorithm and represented in the form of a tree structure. The tree structure is then converted into a binary tree and local structure code is performed. Finally, the local structure code and the global structure code are concatenated to obtain the complete structure code representation.

[0024] (4) Propagation Mode Module:

[0025] Based on the text content information obtained in step (1), the user forwarding network obtained in step (2), and the complete structure encoding obtained in step (3), the propagation node representation is extracted. First, using the pre-processed text content information, the pre-trained large language model BERT is used to extract text content features to obtain content encoding. Then, based on the user forwarding network, global, local, and relative position encodings are calculated using the Laplacian matrix, proximity relationship, and shortest path algorithm, respectively. The three encodings are concatenated to obtain the complete position encoding. Finally, the content encoding, position encoding, and structure encoding are concatenated and input into the Graph Transformer model to extract the propagation node representation.

[0026] (5) Rumor Detection Module:

[0027] The propagation node representation obtained in step (4) is input into the multilayer perceptron (MLP) model, and then... Activation function and The regression function is used to obtain the authenticity label of the tweet.

[0028] As a further improvement of the present invention, in step (1), after preprocessing and word segmentation of the original dataset, a tweet propagation tree is constructed, which specifically includes the following steps;

[0029] (1-1) Web links, user mentions, and platform tags in the data are converted into a unified format using regular expression matching, and the text content is filtered based on a stop word list. Then, for Chinese data, the Jieba word segmentation tool is used to segment the Chinese text, filter out the effective information, and store it as an event set. ,in Information representing each event, Represented as events The text content matrix, user information set, and tweet posting time set;

[0030] (1-2) For the preprocessed information, relying on the propagation tree set consisting of the original tweet and its retweet tweets, firstly, the original tweet... and all retweeted tweets As nodes in the graph, they form a set of nodes. Next, based on the forwarding relationship, edge connections are established. Finally, the propagation tree is constructed as ,in, This represents a set of nodes, including the original tweet and all retweeted tweets. Denotes the set of edges, edges Indicates the first The user forwarded the first one Each tweet is added to the event collection, and the propagation tree is ultimately obtained as tweet event data. .

[0031] As a further improvement of the present invention, step (2) specifically includes:

[0032] (2-1) To explore the relationships between multiple propagation trees and solve the problem that traditional propagation structures cannot represent the same user in different propagation trees, it is necessary to construct a user forwarding network that can unify multiple propagation trees, based on the event set obtained in step (1). ,in , These consist of a propagation tree, a text content matrix, a set of user information, and a set of tweet posting times, respectively. For each propagation tree... ,in For a set of nodes, Construct a user push subgraph for the set of edges. Traverse each propagation tree Each node Extract its tweet nodes The attributes of the publishing user , will tweet node Replace with the user node that posted the tweet. This constitutes a set of user nodes. At the same time, the tweet forwarding relationships in the original propagation tree are preserved, and the edge set is... Tweet nodes Replace with the user node that posted the tweet. This constitutes the user edge set. This forms a user referral subgraph. Finally, we obtain the user referral subgraph set. .

[0033] (2-2) Construct a set of user referral subgraphs Next, merge all user push subgraphs. The first step is to process the nodes in the user push subgraphs and construct a node set. Iterate through each user's referral subgraph in turn. For each node Check if the user ID of this node exists. If it does not exist, then in Create new nodes Initialize the node's attribute list. and Then, the text content of the current node. and time Add to the corresponding node Attribute list and middle;

[0034] The second step is to process the edges in the user push subgraph and construct the edge set. Iterate through each user's referral subgraph in turn. For each edge Get the edge starting node and termination node ,exist Find the corresponding and The user nodes are represented as follows: Check the edge Does it exist in If it does not exist, then create it. The attribute list will Add the attribute to the list and initialize. of The attribute is 1; if it exists, then... Add attributes In the attribute list, and of Increment the attribute value by 1; finally, you get the complete user retweet network. .

[0035] As a further improvement of the present invention, step (3) specifically includes:

[0036] (3-1) Based on the user forwarding network obtained in step (2) Combining the forwarding edge count and other attributes provided by the propagation tree, the structural encoding is extracted. The first step is to extract the global structural encoding of the user forwarding network using the WL (Weisfeiler-Lehman test) algorithm. The specific steps are as follows:

[0037] First, for each node in the graph Initialize node labels, use a hash function to assign colors based on node attributes, and encode the node's original attributes into an initial color value, represented as:

[0038]

[0039] in, Represents a node The label at iteration 0. Then, iteration begins, for each node... Calculate the label set of its neighboring nodes. That is, all of them The current label set of directly connected nodes. Then update the node labels and remove duplicates using a hash function or other feature mapping function. To combine nodes Current tag The set of labels of its neighboring nodes :

[0040]

[0041] in, Indicates the node in the graph The set of neighboring nodes on the map. To avoid label inflation, it is necessary to ensure that for two nodes with the same label, their new labels after mapping are also the same:

[0042]

[0043] The iterative process terminates when any of the following conditions are met:

[0044] First, when the node label does not change in two consecutive iterations, that is, for all... have:

[0045] Secondly, when the number of iterations... Reaching the preset maximum value ;

[0046] Finally, when the iteration stops, the labels of all nodes have converged. The final labels are combined into a matrix form of global structural feature encoding, denoted as . :

[0047]

[0048] (3-2) The second step is to capture the local structural information from the user node's perspective using the "depth walk spanning tree" algorithm and represent it as a tree structure. Then, the tree structure is serialized using a tree structure representation method based on traversal algorithms. Finally, the Skip-gram model is used to embed the sequence to obtain the local structure encoding. The specific steps are as follows:

[0049] First, based on the user forwarding network obtained in step (2) For each node to be represented along the way Create an empty tree Used to record the traversal results, creating an empty set of reachable nodes. Record the nodes that have been visited and an empty set of reached nodes. Record the nodes that can be reached in the next step during the traversal phase, and initialize the set of reached nodes. This represents a visited node and initializes the edge set. , used to store edge information in the tree;

[0050] For each node conduct Each independent depth-first walk. Each walk visits a maximum of [number missing] times. A new node is created, and the starting point for the traversal is set from the current node. Start, initialize the step counter ;

[0051] Then begin the traversal, starting from the current node in each step. Departure, Calculation All nodes in The node that can be reached by one hop is According to the walk parameters and calculate The transition probability of the middle node is used to sample nodes based on the transition probability distribution. ,in It is a node The neighbor set of nodes Add to collection :

[0052]

[0053] Simultaneously generate edges and according to the connecting edges Add it to the tree When already visited When encountering a new node or being unable to continue, terminate the current traversal and eventually obtain the node. The Spanning Tree ;

[0054] When the relative order of the child nodes of each node in a multi-way tree is fixed, the tree can be uniquely represented as a binary tree, and this process is reversible. Therefore, after obtaining... Tree containing A multi-way tree with nodes Then, the multi-way tree is converted into a binary tree using the "left child, right sibling" method. Then, perform preorder and inorder traversals on the binary tree to obtain the results of the preorder and inorder traversals. and Encoded sequence representation assembled into a tree structure , length is ;

[0055] Subsequently, the Skip-gram model was used to characterize the coding sequence. Embed structural features and encode the extracted node local structure. ;

[0056] (3-3) The third step is to encode the global structure. and local structure encoding The complete structure code is obtained by splicing. :

[0057]

[0058] in This indicates the concatenation of eigenvectors.

[0059] As a further improvement of the present invention, step (4) specifically includes:

[0060] (4-1) Based on the text content information obtained in step (1) for the tweet set , Represents the number of tweets in the set, and for any tweet The posting time is no later than the tweet. , It can also be expressed as ,in Tweet The first in One word, For tweets To determine the number of words in a tweet, the tweet is first segmented at the sentence level and then fed into a pre-trained BERT model to obtain its word vector representation, with each sentence corresponding to a word vector matrix. Next, the word-level word vectors are fed into a Bi-LSTM model to learn their contextual relationships. Because Bi-LSTM performs the LSTM process in both forward and reverse order, it can learn the connections between words in both directions while preserving their long-range dependency characteristics. In a forward-order LSTM, the previous word The semantic information it carries will be transmitted to it, and through learning... After considering the meaning of the preceding words, some less important information is discarded before being passed on to the next word. In reverse LSTM, What was not obtained in the forward sequence process Information about the subsequent word sequence will also be transmitted through The information is passed to it in a decaying manner, allowing Bi-LSTM to retain some global dependency information while focusing on local sequence information. Its formal representation is:

[0061]

[0062]

[0063]

[0064] in, and Each word Vector representations of words in forward and inverse LSTM hidden layers. Final expression Positive representation Reverse representation The concatenation of words in the sentence is then fed into the Attention layer, where the importance of each word is modeled, its attention coefficient is calculated, and the vector representation of the entire sentence is obtained by weighted summation with the content vector. For the already calculated first word... Hidden vectors of layers The next layer of words Location attention coefficient The calculation method is as follows:

[0065]

[0066] Then, the attention coefficients are normalized using a Softmax layer to calculate the word. Attention weights :

[0067]

[0068] Tweets calculated using weighted summation The final expression :

[0069]

[0070] Finally, the text content encoding matrix is ​​obtained. .

[0071] (4-2) Based on the user forwarding network obtained in step (2) First, the user forwarding network is transformed into the corresponding complete Laplace matrix to obtain the global location code. The formula is as follows:

[0072]

[0073] in, For the complete network adjacency matrix, The degree matrix of the complete network, and These correspond to eigenvalues ​​and eigenvectors, respectively.

[0074] Then, since the user retweet network is formed by fusing the propagation subgraphs transformed from the original propagation tree, the original propagation tree naturally forms a local cluster. Therefore, this local cluster, i.e., the user retweet network, is used to rank the affinity, i.e.:

[0075]

[0076] in, Indicates for the target node Other nodes will be arranged according to their relationship with The calculated intimacy scores are sorted in descending order to obtain the intimacy distance between two nodes, i.e., the intimacy matrix, from which the local location encoding is obtained. .

[0077] Subsequently, the posting time attribute of the tweets was used to determine the propagation direction relationship between nodes, and the relative position code was calculated using the Floyd shortest path algorithm. ,Right now:

[0078]

[0079] in, To in the complete network adjacency matrix The shortest distance matrix between nodes is calculated by performing the Floyd algorithm.

[0080] Finally, the global position is encoded. Local location coding Concatenate relative position codes To obtain the complete position code :

[0081]

[0082] in This indicates the concatenation of eigenvectors.

[0083] (4-3) Based on the structure code obtained in step (3) The text content encoding obtained in step (4-1) and the location code obtained in step (4-2) The content encoding of each tweet node in the text content encoding is represented as follows: The features of nodes and variables are mapped to d dimensions through a linear transformation. The specific calculation formula is as follows:

[0084] ,

[0085] in, For nodes Features For nodes and nodes Features of the edges between them and Let be the projection matrix. and The bias parameter for the linear mapping;

[0086] Similarly, the position and structure encodings are linearly mapped to the same dimension and added to the node features:

[0087]

[0088]

[0089] in, The feature encoding of the position encoding matrix, The feature encoding of the structure encoding matrix, and Let be the projection matrix. and The bias parameter for the linear mapping. This serves as the initial feature representation for the final node;

[0090] The final initial node representation sequence is input into the Graph Transformer network, and the hidden state after the last iteration is used as the propagation node representation. Each layer of the Graph Transformer mainly includes the following parts: self-attention mechanism, feature aggregation, normalization, and residual connections. The specific formula for updating nodes in a certain layer is:

[0091]

[0092] in, Indicates a connection. The attention weight is represented by the following formula:

[0093]

[0094] The output is then passed to a single-layer feedforward neural network (FNN), where residual connections and normalization are calculated to obtain the node embedding representations.

[0095]

[0096]

[0097]

[0098] For the output of the last layer We call this the propagation node representation. ,Right now:

[0099]

[0100] As a further improvement of the present invention, step (5) specifically includes:

[0101] Based on the propagation node representation obtained in step (4) The authenticity of tweets is scored using a multilayer perceptron (MLP). The specific calculation formula is as follows:

[0102]

[0103] in, and Here is the weight matrix of the MLP. This represents the final node. for Activation function. For The category label corresponding to the item with the highest authenticity score is the authenticity label of the predicted information.

[0104] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0105] (1) To address the issue of differences in propagation structure, this invention designs a local propagation feature modeling method based on a depth-walking spanning tree. This method systematically encodes the contribution of each node to the structure of its surrounding neighbors through the ordered walk of the spanning tree, thus more accurately reflecting the local propagation characteristics of the subgraph, rather than using an unordered random walk. This method can better capture the differences in the local propagation structure of the subgraph, significantly improving the accuracy of rumor detection.

[0106] (2) To address the limitations of information interaction modeling, this invention constructs a user retweet network that combines tweet propagation relationships. By integrating multiple propagation paths and user behaviors, it reveals the connections formed between different pieces of information through common users, thus more comprehensively reflecting the complexity of information propagation, rather than simply accumulating local propagation information from a single propagation tree. This method can effectively improve the full mining and effective utilization of propagation information, making rumor detection richer and more accurate;

[0107] (3) To address the issue of rumor detection effectiveness, this invention proposes a rumor detection method based on information propagation structure. By systematically encoding node neighbor relationships and integrating multi-dimensional information features, it comprehensively models information such as tweet content, forwarding relationships, and user attributes, thereby more accurately expressing the laws of information propagation and user behavior characteristics. This method can capture multifaceted information features in a timely manner, providing more comprehensive data support for rumor detection and making related decisions more scientific and reliable. Attached Figure Description

[0108] Figure 1 is a general framework diagram of the present invention;

[0109] Figure 2 is a schematic diagram of the tweet propagation tree structure;

[0110] Figure 3 is a schematic diagram of the network building modules;

[0111] Figure 4 is a schematic diagram of the global structure encoding;

[0112] Figure 5 is a schematic diagram of the depth-walking algorithm.

[0113] Figure 6 is a schematic diagram of local structure encoding.

[0114] Figure 7 is a schematic diagram of text content encoding.

[0115] Figure 8 is a schematic diagram of the propagation node representation based on the Graph Transformer network. Detailed Implementation

[0116] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0117] The following is only one embodiment of the present invention. The present invention has many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention. All such corresponding changes and modifications should fall within the protection scope of the appended claims.

[0118] As shown in Figure 1, a rumor detection method based on information propagation structure according to the present invention includes the following steps:

[0119] 1. Data Analysis Module

[0120] The original datasets PHEME and Misinfdect were cleaned and preprocessed to remove irrelevant information. The Chinese text was segmented using the Jieba word segmentation tool. A tweet propagation tree was constructed based on the relationship between the original tweets and their retweet tweets.

[0121] (1) The web page links, user mentions and platform tags in the data are converted into a unified format by regular expression matching, and the text content is filtered based on the stop word list. Then, for Chinese data, the Jieba word segmentation tool is used to segment the Chinese text, filter out the effective information and store it by event, which is represented as an event set. ,in Information representing each event, Represented as events The text content matrix, user information set, and tweet posting time set;

[0122] (2) For the preprocessed information, relying on the propagation tree set consisting of the original tweet and its retweet tweets, firstly, the original tweet is... and all retweeted tweets As nodes in the graph, they form a set of nodes. Next, based on the forwarding relationship, edge connections are established. Finally, the propagation tree is constructed as ,in, This represents a set of nodes, including the original tweet and all retweeted tweets. Denotes the set of edges, edges Indicates the first The user forwarded the first one The propagation tree structure of each tweet is shown in Figure 2. The propagation tree is added to the event set to obtain the final tweet event data. .

[0123] 2. Network Construction Module

[0124] Based on the tweet propagation tree data obtained in step (1), a user forwarding network is constructed. First, for each propagation tree, the tweet node is replaced with the user node that posted the tweet, while preserving the original structure of the propagation tree, and a user forwarding subgraph is obtained. Then, all user forwarding subgraphs are merged, keeping only one user node with the same ID, keeping only one edge in the same direction between two users, keeping the node attributes unchanged, and merging the edge attributes to form a complete user forwarding network. The network construction module is shown in Figure 3.

[0125] (1) In order to explore the relationship between multiple propagation trees and solve the problem that traditional propagation structures cannot represent the same users in different propagation trees, it is necessary to construct a user forwarding network that can unify multiple propagation trees, based on the event set obtained in step (1). ,in , These consist of a propagation tree, a text content matrix, a set of user information, and a set of tweet posting times, respectively. For each propagation tree... ,in For a set of nodes, Construct a user push subgraph for the set of edges. Traverse each propagation tree Each node Extract its tweet nodes The attributes of the publishing user , will tweet node Replace with the user node that posted the tweet. This constitutes a set of user nodes. At the same time, the tweet forwarding relationships in the original propagation tree are preserved, and the edge set is... Tweet nodes Replace with the user node that posted the tweet. This constitutes the user edge set. This forms a user referral subgraph. Finally, we obtain the user referral subgraph set. .

[0126] (2) Construct a set of user referral subgraphs Next, merge all user push subgraphs. The first step is to process the nodes in the user push subgraphs and construct a node set. Iterate through each user's referral subgraph in turn. For each node Check if the user ID of this node exists. If it does not exist, then in Create new nodes Initialize the node's attribute list. and Then, the text content of the current node. and time Add to the corresponding node Attribute list and middle;

[0127] The second step is to process the edges in the user push subgraph and construct the edge set. Iterate through each user's referral subgraph in turn. For each edge Get the edge starting node and termination node ,exist Find the corresponding and The user nodes are represented as follows: Check the edge Does it exist in If it does not exist, then create it. The attribute list will Add the attribute to the list and initialize. of The attribute is 1; if it exists, then... Add attributes In the attribute list, and of Increment the attribute value by 1; finally, you get the complete user retweet network. .

[0128] 3. Structure perception module

[0129] Based on the user forwarding network obtained in step (2), and combined with the forwarding attributes such as the number of forwarding edges provided by the propagation tree, the structure code is extracted. First, the global structure code is extracted using the WL algorithm. Then, the local structure information from the user node's perspective is captured by the "depth walk spanning tree" algorithm and represented in the form of a tree structure. The tree structure is then converted into a binary tree and local structure code is performed. Finally, the local structure code and the global structure code are concatenated to obtain the complete structure code representation.

[0130] (1) Based on the user forwarding network obtained in step (2) Combining the forwarding edge count and other attributes provided by the propagation tree, the structural encoding is extracted. The first step is to extract the global structural encoding of the user forwarding network using the WL (Weisfeiler-Lehman test) algorithm, as shown in Figure 4. The specific steps are as follows:

[0131] First, for each node in the graph Initialize node labels, use a hash function to assign colors based on node attributes, and encode the node's original attributes into an initial color value, represented as:

[0132]

[0133] in, Represents a node The label at iteration 0. Then, iteration begins, for each node... Calculate the label set of its neighboring nodes. That is, all of them The current label set of directly connected nodes. Then update the node labels and remove duplicates using a hash function or other feature mapping function. To combine nodes Current tag The set of labels of its neighboring nodes :

[0134]

[0135] in, Indicates the node in the graph The set of neighboring nodes on the map. To avoid label inflation, it is necessary to ensure that for two nodes with the same label, their new labels after mapping are also the same:

[0136]

[0137] The iterative process terminates when any of the following conditions are met:

[0138] First, when the node label does not change in two consecutive iterations, that is, for all... have:

[0139] Secondly, when the number of iterations... Reaching the preset maximum value ;

[0140] Finally, when the iteration stops, the labels of all nodes have converged. The final labels are combined into a matrix form of global structural feature encoding, denoted as . :

[0141]

[0142] (2) The second step is to capture the local structural information from the user node’s perspective through the “depth walk spanning tree” algorithm and represent it as a tree structure. The tree structure is then serialized using a tree structure representation method based on traversal algorithm. Finally, the Skip-gram model is used to embed the sequence to obtain the local structure encoding. The specific steps are as follows:

[0143] First, based on the user forwarding network obtained in step (2) For each node to be represented along the way Create an empty tree Used to record the traversal results, creating an empty set of reachable nodes. Record the nodes that have been visited and an empty set of reached nodes. Record the nodes that can be reached in the next step during the traversal phase, and initialize the set of reached nodes. This represents a visited node and initializes the edge set. , used to store edge information in the tree;

[0144] For each node conduct Each independent depth-first walk. Each walk visits a maximum of [number missing] times. A new node is created, and the starting point for the traversal is set from the current node. Start, initialize the step counter ;

[0145] Then begin the traversal, starting from the current node in each step. Departure, Calculation All nodes in The node that can be reached by one hop is According to the walk parameters and calculate The transition probability of the middle node is used to sample nodes based on the transition probability distribution. ,in It is a node The neighbor set of nodes Add to collection :

[0146]

[0147] Simultaneously generate edges and according to the connecting edges Add it to the tree When already visited When encountering a new node or being unable to continue, terminate the current traversal and eventually obtain the node. The Spanning Tree The "depth-walk spanning tree" algorithm is shown in Figure 5;

[0148] When the relative order of the child nodes of each node in a multi-way tree is fixed, the tree can be uniquely represented as a binary tree, and this process is reversible. Therefore, after obtaining... Tree containing A multi-way tree with nodes Then, the multi-way tree is converted into a binary tree using the "left child, right sibling" method. Then, perform preorder and inorder traversals on the binary tree to obtain the results of the preorder and inorder traversals. and Encoded sequence representation assembled into a tree structure , length is ;

[0149] Subsequently, the Skip-gram model was used to characterize the coding sequence. Embed structural features and encode the extracted node local structure. As shown in Figure 6;

[0150] (3) The third step is to encode the global structure. and local structure encoding The complete structure code is obtained by splicing. :

[0151]

[0152] in This indicates the concatenation of eigenvectors.

[0153] 4. Propagation Mode Module

[0154] Based on the text content information obtained in step (1), the user forwarding network obtained in step (2), and the complete structure encoding obtained in step (3), the propagation node representation is extracted. First, using the pre-processed text content information, the pre-trained large language model BERT is used to extract text content features to obtain content encoding. Then, based on the user forwarding network, global, local, and relative position encodings are calculated using the Laplacian matrix, proximity relationship, and shortest path algorithm, respectively. The three encodings are concatenated to obtain the complete position encoding. Finally, the content encoding, position encoding, and structure encoding are concatenated and input into the Graph Transformer model to extract the propagation node representation.

[0155] (1) Based on the text content information obtained in step (1), the tweet set , Represents the number of tweets in the set, and for any tweet The posting time is no later than the tweet. , It can also be expressed as ,in Tweet The first in One word, For tweets To determine the number of words in a tweet, the tweet is first segmented at the sentence level and then fed into a pre-trained BERT model to obtain its word vector representation, with each sentence corresponding to a word vector matrix. Next, the word-level word vectors are fed into a Bi-LSTM model to learn their contextual relationships. Because Bi-LSTM performs the LSTM process in both forward and reverse order, it can learn the connections between words in both directions while preserving their long-range dependency characteristics. In a forward-order LSTM, the previous word The semantic information it carries will be transmitted to it, and through learning... After considering the meaning of the preceding words, some less important information is discarded before being passed on to the next word. In reverse LSTM, What was not obtained in the forward sequence process Information about the subsequent word sequence will also be transmitted through The information is passed to it in a decaying manner, allowing Bi-LSTM to retain some global dependency information while focusing on local sequence information. Its formal representation is:

[0156]

[0157]

[0158]

[0159] in, and Each word Vector representations of words in forward and inverse LSTM hidden layers. Final expression Positive representation Reverse representation The concatenation of words in the sentence is then fed into the Attention layer, where the importance of each word is modeled, its attention coefficient is calculated, and the vector representation of the entire sentence is obtained by weighted summation with the content vector. For the already calculated first word... Hidden vectors of layers The next layer of words Location attention coefficient The calculation method is as follows:

[0160]

[0161] Then, the attention coefficients are normalized using a Softmax layer to calculate the word. Attention weights :

[0162]

[0163] Tweets calculated using weighted summation The final expression :

[0164]

[0165] Finally, the text content encoding matrix is ​​obtained. The content encoding is shown in Figure 7.

[0166] (2) Based on the user forwarding network obtained in step (2) First, the user forwarding network is transformed into the corresponding complete Laplace matrix to obtain the global location code. The formula is as follows:

[0167]

[0168] in, For the complete network adjacency matrix, The degree matrix of the complete network, and These correspond to eigenvalues ​​and eigenvectors, respectively.

[0169] Then, since the user retweet network is formed by fusing the propagation subgraphs transformed from the original propagation tree, the original propagation tree naturally forms a local cluster. Therefore, this local cluster, i.e., the user retweet network, is used to rank the affinity, i.e.:

[0170]

[0171] in, Indicates for the target node Other nodes will be arranged according to their relationship with The calculated intimacy scores are sorted in descending order to obtain the intimacy distance between two nodes, i.e., the intimacy matrix, from which the local location encoding is obtained. .

[0172] Subsequently, the posting time attribute of the tweets was used to determine the propagation direction relationship between nodes, and the relative position code was calculated using the Floyd shortest path algorithm. ,Right now:

[0173]

[0174] in, To in the complete network adjacency matrix The shortest distance matrix between nodes is calculated by performing the Floyd algorithm.

[0175] Finally, the global position is encoded. Local location coding Concatenate relative position codes To obtain the complete position code :

[0176]

[0177] in This indicates the concatenation of eigenvectors.

[0178] (3) Based on the structure code obtained in step (3) The text content encoding obtained in step (4) and position encoding The content encoding of each tweet node in the text content encoding is represented as follows: The features of nodes and variables are mapped to d dimensions through a linear transformation. The specific calculation formula is as follows:

[0179]

[0180]

[0181] in, For nodes Features For nodes and nodes Features of the edges between them and Let be the projection matrix. and The bias parameter for the linear mapping;

[0182] Similarly, the position and structure encodings are linearly mapped to the same dimension and added to the node features:

[0183]

[0184]

[0185] in, The feature encoding of the position encoding matrix, The feature encoding of the structure encoding matrix, and Let be the projection matrix. and The bias parameter for the linear mapping. This serves as the initial feature representation for the final node;

[0186] The final initial node representation sequence is input into the Graph Transformer network, as shown in Figure 8. The hidden state after the last iteration is taken as the propagation node representation. Each layer of the Graph Transformer mainly includes the following parts: self-attention mechanism, feature aggregation, normalization, and residual connections. The specific formula for updating nodes in a certain layer is:

[0187]

[0188] in, Indicates a connection. The attention weight is represented by the following formula:

[0189]

[0190] The output is then passed to a single-layer feedforward neural network (FNN), where residual connections and normalization are calculated to obtain the node embedding representations.

[0191]

[0192]

[0193]

[0194] For the output of the last layer We call this the propagation node representation. ,Right now:

[0195]

[0196] 5. Rumor Detection Module

[0197] Based on the propagation node representation obtained in step (4) The authenticity of tweets is scored using a multilayer perceptron (MLP). The specific calculation formula is as follows:

[0198]

[0199] in, and Here is the weight matrix of the MLP. This represents the final node. for Activation function. For The category label corresponding to the item with the highest authenticity score is the authenticity label of the predicted information.

[0200] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A rumor detection method based on information propagation structure, characterized in that: Includes the following steps: (1) Data analysis module: Cleaning and preprocessing the Chinese and English hot topic event dataset, removing irrelevant information, and using the Jieba word segmentation tool to segment the Chinese text; Based on the relationship between the original tweet and its retweet, construct a tweet propagation tree; (2) Network construction module: Based on the tweet propagation tree data obtained in step (1), construct a user retweet network. First, for each propagation tree, replace the tweet node with the user node that posted the tweet, retain the original structure of the propagation tree, and obtain the user retweet subgraph; then, merge all user retweet subgraphs, and so on. Only one user node is retained, and only one edge in the same direction between two identical users is retained. The node attributes remain unchanged, and the edge attributes are merged to form a complete user forwarding network; (3) Structure perception module: Based on the user forwarding network obtained in step (2), and combined with the forwarding attributes such as the number of forwarding edges provided by the propagation tree, the structure code is extracted. First, the global structure code is extracted using the WL algorithm; then, the local structure information from the user node's perspective is captured by the "depth walk spanning tree" algorithm and represented in the form of a tree structure. The tree structure is then converted into a binary tree and local structure code is performed; finally, the local structure code and the global structure code are spliced ​​together to obtain a complete structure code representation; (4) Propagation mode module: Based on the text content obtained in step (1) Information, the user forwarding network obtained in step (2) and the complete structure encoding obtained in step (3) are used to extract the propagation node representation. First, the pre-processed text content information is used to extract the text content features using the pre-trained large language model BERT to obtain the content encoding. Then, based on the user forwarding network, the global, local and relative position encodings are calculated using the Laplacian matrix, proximity relationship and shortest path algorithm respectively. The three encodings are concatenated to obtain the complete position encoding. Finally, the content encoding, position encoding and structure encoding are concatenated and input into the GraphTransformer model to extract the propagation node representation. (5) Rumor detection module: The propagation node representation obtained in step (4) is input into the multilayer perceptron MLP model and then processed. Activation function and The regression function is used to obtain the authenticity label of the tweet.

2. The rumor detection method based on information propagation structure according to claim 1, characterized in that: In step (1), after preprocessing and segmenting the original dataset, a tweet propagation tree is constructed, specifically including the following steps: (1-1) The web page links, user mentions, and platform tags in the data are converted into a unified format through regular expression matching, and the text content is filtered based on the stop word list. Then, for the Chinese data, the Jieba word segmentation tool is used to segment the Chinese text, filter out the effective information, and store it by event, representing it as an event set. ,in Information representing each event, Represented as events The text content matrix, user information set, and tweet posting time set; (1-2) For the preprocessed information, relying on the propagation tree set composed of the original tweets and their retweet tweets, firstly, the original tweets... and all retweeted tweets As nodes in the graph, they form a set of nodes. Next, based on the forwarding relationship, edge connections are established. Finally, the propagation tree is constructed as ,in, This represents a set of nodes, including the original tweet and all retweeted tweets. Denotes the set of edges, edges Indicates the first The user forwarded the first one Each tweet is added to the event collection, and the propagation tree is ultimately obtained as tweet event data. 。 3. The rumor detection method based on information propagation structure according to claim 1, characterized in that: Step (2) specifically includes: (2-1) To explore the relationship between multiple propagation trees, a unified user forwarding network for multiple propagation trees is constructed, based on the event set obtained in step (1). ,in ， These consist of a propagation tree, a text content matrix, a set of user information, and a set of tweet posting times. For each propagation tree... ,in For a set of nodes, Construct a user push subgraph for the set of edges. Traverse each propagation tree Each node Extract its tweet nodes The attributes of the publishing user , will tweet node Replace with the user node that posted the tweet. This constitutes a set of user nodes. At the same time, the tweet forwarding relationships in the original propagation tree are preserved, and the edge set is... Tweet nodes Replace with the user node that posted the tweet. This constitutes the user edge set. This forms a user referral subgraph. Finally, we obtain the user referral subgraph set. (2-2) Construct a set of user referral subgraphs. Next, merge all user push subgraphs. The first step is to process the nodes in the user push subgraphs and construct a node set. Iterate through each user's push subgraph in turn. For each node Check if the user ID of this node exists. If it does not exist, then in Create new nodes Initialize the node's attribute list. and Then, the text content of the current node. and time Add to the corresponding node Attribute list and The second step is to process the edges in the user push subgraph and construct an edge set. Iterate through each user's push subgraph in turn. For each edge Get the edge starting node and termination node ,exist Find the corresponding and The user nodes are represented as follows: Check the edge Does it exist in If it does not exist, then create it. The attribute list will Add the attribute to the list and initialize. of The attribute is 1; if it exists, then... Add attributes In the attribute list, and of Increment the attribute value by 1; finally, you get the complete user retweet network. 。 4. The rumor detection method based on information propagation structure according to claim 1, characterized in that: Step (3) specifically includes: (3-1) Based on the user forwarding network obtained in step (2) Combining the forwarding edge count and other attributes provided by the propagation tree, the structural encoding is extracted. The first step is to extract the global structural encoding of the user forwarding network using the WL algorithm. The specific steps are as follows: First, for each node in the graph... Initialize node labels, use a hash function to assign colors based on node attributes, and encode the node's original attributes into an initial color value, represented as: ;in, Represents a node The label at iteration 0, then the iteration begins, for each node. Calculate the label set of its neighboring nodes. That is, all of them The current label set of directly connected nodes is used to update the node labels and remove duplicates, using a hash function or other feature mapping function. To combine nodes Current tag The set of labels of its neighboring nodes ： ;in, Indicates the node in the graph The set of neighboring nodes on the map. To avoid label inflation, it is necessary to ensure that for two nodes with the same label, their new labels after mapping are also the same: The iteration process terminates when either of the following conditions is met: First, the node label remains unchanged in two consecutive iterations, i.e., for all... have: Secondly, when the number of iterations... Reaching the preset maximum value Finally, when the iteration stops, the labels of all nodes have converged. The final labels are combined into a matrix form of global structural feature encoding, denoted as . ： (3-2) The second step is to capture the local structural information from the user node's perspective using the "depth walk spanning tree" algorithm and represent it as a tree structure. Then, the tree structure is serialized using a tree structure representation method based on traversal algorithms. Finally, the Skip-gram model is used to embed the sequence to obtain the local structure encoding. The specific steps are as follows: First, based on the user forwarding network obtained in step (2), For each node to be represented along the way Create an empty tree Used to record the traversal results, creating an empty set of reachable nodes. Record the nodes that have been visited and an empty set of reached nodes. Record the nodes that can be reached in the next step during the traversal phase, and initialize the set of reached nodes. This represents a visited node and initializes the edge set. It is used to store edge information in the tree; for each node conduct Each independent depth-first traversal, with a maximum of [number] visits per traversal. A new node is created, and the starting point for the traversal is set from the current node. Start, initialize the step counter Then begin the traversal, starting from the current node in each step. Departure, Calculation All nodes in The node that can be reached by one hop is According to the walk parameters and calculate The transition probability of the middle node is used to sample nodes based on the transition probability distribution. ,in It is a node The neighbor set of nodes Add to collection ： Simultaneously generate connecting edges. and according to the connecting edges Add it to the tree When already visited When encountering a new node or being unable to continue, terminate the current traversal and eventually obtain the node. The Spanning Tree When the relative order of the child nodes of each node in a multi-way tree is fixed, the multi-way tree is uniquely and definitively represented as a binary tree, and this process is reversible. Therefore, after obtaining... Tree containing A multi-way tree with nodes Then, convert the multi-way tree into a binary tree. Then, perform preorder and inorder traversals on the binary tree to obtain the results of the preorder and inorder traversals. and Encoded sequence representation assembled into a tree structure , length is Subsequently, the Skip-gram model was used to characterize the encoded sequence. Embed structural features and encode the extracted node local structure. (3-3) The third step is to encode the global structure. and local structure encoding The complete structure code is obtained by splicing. ： ;in This indicates the concatenation of eigenvectors.

5. The rumor detection method based on information propagation structure according to claim 1, characterized in that: Step (4) specifically includes: (4-1) For the set of tweets, based on the text content information obtained in step (1) ， Represents the number of tweets in the set, and for any tweet The posting time is no later than the tweet. ， It is also expressed as ,in Tweet The first in One word, For tweets To determine the number of words in a tweet, the tweet is first segmented at the sentence level and then fed into a pre-trained BERT model to obtain its word vector representation; each sentence corresponds to a word vector matrix. Next, the word-level word vectors are fed into a Bi-LSTM model to learn their contextual relationships. In a forward-order LSTM, the previous word The semantic information it carries will be transmitted to it, and through learning... After considering the meaning of the preceding words, some less important information is discarded before being passed on to the next word. In reverse pair LSTM, What was not obtained in the forward sequence process Information about the subsequent word sequence will also be transmitted through The information is passed to it in a decaying manner, allowing Bi-LSTM to retain some global dependency information while focusing on local sequence information. Its formal representation is as follows: ； ； ;in, and Each word Vector representations of words in forward and inverse LSTM hidden layers. Final expression Positive representation Reverse representation The concatenation of words in the sentence is then fed into the Attention layer, where the importance of each word is modeled, its attention coefficient is calculated, and the vector representation of the entire sentence is obtained by weighted summation with the content vector. For the already calculated first word... Hidden vectors of layers The next layer of words Location attention coefficient The calculation method is as follows: Then, the attention coefficients are normalized using a Softmax layer to calculate the word... Attention weights ： Tweets are calculated by weighted summation. The final expression ： Finally, the text content encoding matrix is ​​obtained. (4-2) Based on the user forwarding network obtained in step (2) First, the user forwarding network is transformed into the corresponding complete Laplace matrix to obtain the global location code. The formula is as follows: ;in, For the complete network adjacency matrix, The degree matrix of the complete network, and These correspond to eigenvalues ​​and eigenvectors, respectively. Then, since the user referral network is formed by fusing the propagation subgraphs transformed from the original propagation tree, the original propagation tree naturally forms a local cluster. Therefore, this local cluster, i.e., the user referral network, is used to rank the affinity, i.e.: ;in, Indicates for the target node Other nodes will be arranged according to their relationship with The calculated intimacy scores are sorted in descending order to obtain the intimacy distance between two nodes, i.e., the intimacy matrix, from which the local location encoding is obtained. Subsequently, the posting time attribute of tweets was used to determine the propagation direction relationship between nodes, and the relative position encoding was calculated using the Floyd shortest path algorithm. ,Right now: ;in, To in the complete network adjacency matrix The shortest distance matrix between nodes is calculated using the Floyd-Warshall algorithm; finally, the global position is encoded. Local location coding Concatenate relative position codes To obtain the complete position code ： ;in This represents the concatenation of feature vectors; (4-3) based on the structural encoding obtained in step (3). The text content encoding obtained in step (4-1) and the location code obtained in step (4-2) The content encoding of each tweet node in the text content encoding is represented as follows: The features of nodes and variables are mapped to d dimensions through a linear transformation. The specific calculation formula is as follows: ， ;in, For nodes Features For nodes and nodes Features of the edges between them and For the projection matrix, and The bias parameter is used for the linear mapping; similarly, the position and structure encodings are linearly mapped to the same dimension and added to the node features: ； ;in, The feature encoding of the position encoding matrix, The feature encoding of the structure encoding matrix, and For the projection matrix, and The bias parameter for the linear mapping. The final initial feature representation of the nodes is used; the final initial node representation sequence is input into the Graph Transformer network, and the hidden state after the last iteration is taken as the propagation node representation. Each layer of the Graph Transformer mainly includes the following parts: self-attention mechanism, feature aggregation, normalization, and residual connections. For the node update of a certain layer, the specific formula is: ;in, Indicates a connection. The attention weight is represented by the following formula: The output is then passed to a single-layer feedforward neural network, where residual connections and normalization are calculated to obtain the node embedding representations. ； ； For the output of the last layer This is the representation of the propagation node. ,Right now: 。 6. The rumor detection method based on information propagation structure according to claim 1, characterized in that: Step (5) specifically includes: the propagation node representation obtained in step (4). The authenticity of tweets is scored using a multilayer perceptron (MLP). The specific calculation formula is as follows: ;in, and Here is the weight matrix of the MLP. This represents the final node. for Activation function, for The category label corresponding to the item with the highest authenticity score is the authenticity label of the predicted information.

Images