Rumor detection method and system based on heterogeneous propagation structure and dynamic feature fusion

Abstract

The invention discloses a rumor detection method and system based on heterogeneous propagation structure and dynamic feature fusion, and belongs to the field of computer technology, artificial intelligence and social network analysis. The method comprises the following steps: extracting static features of a user and a post, calculating dynamic features of the user and the post, and fusing the static and dynamic features to obtain a dynamic fusion feature; and constructing a heterogeneous graph sequence through dynamic fusion features, performing convolution updating by using a heterogeneous graph convolution network, and outputting a heterogeneous graph feature matrix. And processing the directed propagation tree by using a graph attention network to obtain propagation cascade global representation. Performing cross-time-step interaction on the heterogeneous graph features and the propagation cascade features to generate heterogeneous graph representation and propagation cascade representation; and finally, splicing to obtain a final representation, and inputting a rumor classification model to obtain a prediction result. The excellent performance of the method in various complex social scenes verifies the feasibility and effectiveness of the method as an efficient and accurate rumor detection method with excellent early warning capability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer technology, artificial intelligence and social network analysis, and particularly relates to a rumor detection method and system based on fusion of heterogeneous propagation structure and dynamic characteristics. BACKGROUND

[0002] With the development of mobile social network platforms, social platforms such as microblog have become the main platform for people to quickly obtain information, express opinions and interact. However, the convenience and immediacy of information dissemination also exacerbate the spread of rumors in social networks. Therefore, advanced technology capable of automatically detecting rumors on social media quickly and accurately has become a current research hotspot. Existing rumor detection technology has mainly experienced a development process from traditional machine learning based on artificial feature engineering to deep learning based on deep learning. Deep learning technology can automatically learn high-order, non-linear deep feature representation from data, thereby more accurately capturing the complex patterns of rumors in content semantics and propagation structure, and effectively avoiding the bottleneck of feature engineering. However, in existing methods, most methods usually model social network information as homogeneous, ignoring the multiple types of nodes and relationships in the network, losing rich heterogeneous semantic information, making it difficult for the model to understand complex interaction patterns. In addition, rumor propagation is a highly dynamic process over time. Existing methods often extract static features or simply treat time as a dimension, failing to deeply capture the structural evolution rules of events at different time steps, resulting in the inability to reflect the explosive or corrective characteristics of rumors in the early or later stages of propagation.

[0003] Currently, a large number of researches have been carried out on rumor detection methods at home and abroad, and some progress has been made. Early work uses a text-based fusion neural network for rumor detection, proposes a new data processing algorithm for accurate extraction of context features, and builds a fusion neural network including bidirectional gated recurrent unit, convolution model and attention mechanism. However, this method only processes the text features of social network rumors, ignoring other features, thus limiting the detection effect. Then, the work considers the rumor detection method based on topic classification and content understanding, which can extract features from different scale sub-data sets, consider the relevance and difference of different topics, and combine the discourse and word meaning to expand the model's understanding of text and enhance the detection ability of malicious rumor spreading traps. However, this method usually only focuses on maintaining the previous state without considering the subsequent context information. In addition, some methods consider using graph convolution networks, but mainly focus on the reconstruction and representation of static structure, and fail to effectively capture the dynamic changes of structure over time in the rumor spreading process, resulting in insufficient use of time sequence features. In order to capture the propagation pattern, a new conversion method converts the complex forwarding tree into a more easily processed binary tree while not losing reconstruction information. Then, the forwarding tree is serialized into a corpus of meta-tree paths, each of which can preserve a basic substructure. Subsequently, a deep neural network is used to integrate all meta-trees and generate global structure embeddings. However, the process of converting the propagation tree into a binary tree may lose some of the original topological information, and this method tends to treat the propagation structure as an independent feature, failing to model the user interaction process as a whole in a unified dynamic framework. The recurrent neural network model (RvNN-DB) can introduce an attention mechanism to adjust the weight of the clues when extracting deep propagation features, enabling the model to focus on key clues that are beneficial to rumor detection. When extracting broad propagation features, a one-dimensional convolutional neural network is first used to obtain the fine-grained relationships between words within the comments, and then the semantic interaction between comments is obtained to supplement the deep features. However, the method based on recurrent neural network can only capture the dependency relationship between parent and child nodes, and it is difficult to capture the global graph structure features, and the computational efficiency is low when dealing with deep tree structure, and the scalability is limited when facing large-scale social data. In order to fuse multiple content features, the hierarchical aggregation graph neural network method (HAGNN) focuses on capturing different granularities of high-level representations of text content and fusing rumor propagation structure. It applies graph convolution network (GCN) combined with rumor propagation graph to learn text granularity representation in event propagation. However, this method mainly aggregates features on static graphs, ignoring the fact that rumor propagation is a highly time-sensitive process, making it difficult to capture dynamic evolution features in the event development process. The DDCA-GCN rumor detection model uses a content-based data compensation method to solve the problem of data sparsity.Secondly, the method proposes a method for constructing network structure to solve the problem of user association in early theme field, and converts the theme field into a multi-user relationship network with non-Euclidean structure. Then the non-Euclidean data is processed by using graph convolution network (GCN), and good results are obtained in early detection. However, although the method considers user association, it focuses on static fusion, often ignoring the evolution law of propagation structure over time, and cannot fully utilize the dynamic interaction information on the time series. The proposal of the double dynamic graph convolution network method (DDGCN) realizes the modeling of the message dynamics in the propagation and the background knowledge dynamics in the knowledge graph in a unified framework. Specifically, two graph convolution networks are used to capture the above two structure information at different time stages, and then they are combined with the time fusion unit. This allows learning dynamic event representation in a more fine-grained way and gradually aggregating them to capture cascading effects, thereby better detecting rumors. However, this method only models homogeneous information and fails to align and deeply fuse dynamic data from different sources. SUMMARY

[0004] The present application is directed to the problems existing in the prior art, and provides a rumor detection method and system based on heterogeneous propagation structure and dynamic feature fusion, which can improve the feature extraction capability of heterogeneous interaction mode and propagation cascade structure, thereby effectively capturing the dynamic evolution law of rumor events, adapting to the complex heterogeneous information environment in social media, and ensuring accurate identification and rapid response in the early stage of rumor propagation.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a rumor detection method based on heterogeneous propagation structure and dynamic feature fusion, comprising the following steps:

[0006] S1, constructing a preprocessing and comprehensive feature extraction module: preprocessing the social network data, extracting user and post static features for the preprocessed data, and calculating user and post dynamic features based on a time window, and fusing static and dynamic features to obtain dynamic fusion features;

[0007] S2, constructing a heterogeneous interaction graph learning module: constructing a heterogeneous graph sequence through dynamic fusion features, and updating the heterogeneous graph sequence by using a heterogeneous graph convolution network to output a heterogeneous graph feature matrix H k ;

[0008] S3, constructing a propagation cascade module: first constructing a directed propagation tree, processing the directed propagation tree by using a graph attention network GAT to obtain a propagation cascade global representation P k ;

[0009] S4, constructing a spatiotemporal dynamic fusion module: based on the heterogeneous graph features H kand propagation cascade feature P k Deep fusion is performed on the heterogeneous graph features H k and propagation cascade feature P k Fusion is performed to generate a heterogeneous graph representation and propagation cascade representation ;

[0010] S5, a rumor classification model based on a multi-layer perception is constructed: the heterogeneous graph representation H and the propagation cascade representation P are spliced to obtain a final representation, which is input into a rumor classification model based on a multi-layer perception, and a cross-entropy loss function is combined to obtain a prediction result of a rumor or non-rumor.

[0011] Further, the foregoing step S1 includes the following sub-steps:

[0012] S1.1, for the post content in the social network, a pre-trained BERT model is used to obtain the static feature of the post by the following formula:

[0013] , ;

[0014] In the formula, denotes the static feature of the post, denotes the original content of the post, denotes a pre-trained language model for converting text content into a vector embedding; denotes the original attribute information of the user, denotes a pre-trained encoding model for extracting high-dimensional features of the user;

[0015] S1.2, the rumor propagation period is divided into k time windows, and the dynamic features of the user nodes and the post nodes are calculated in each time window respectively to obtain user dynamic features and post dynamic features ;

[0016] S1.3, for each user and post in each time window, the static and dynamic features are fused by the formula to obtain the dynamic fusion feature of the time window, wherein denotes a vector splicing operation, denotes the dynamic fusion feature of the i-th time window user or post, denotes the static feature of the user or the static feature of the post, denotes the dynamic feature of the user or the post, denotes the user node when i=u, and denotes the post node when i=p.

[0017] ​Further, the aforementioned step S2 comprises the following sub-steps:

[0018] S2.1, constructing a heterogeneous graph sequence by dynamic fusion features , wherein represents a heterogeneous graph snapshot of the i-th time window, including two types of nodes: user nodes and post nodes ; and edge relationships: user-user , post-post and user-post ;

[0019] S2.2, using a heterogeneous graph convolution network for heterogeneous relationship perception, and introducing a relationship perception propagation operator , different types of edge relationships are assigned learnable attention weights, and the propagation operator is calculated by the formula: to aggregate neighbor information of different relationships and output a heterogeneous graph feature matrix sequence { H1, H2,…, H k.}, , , respectively represent the normalized heterogeneous edge matrix and the corresponding attention weight parameters.

[0020] Further, the aforementioned step S3 is specifically: taking the source post as the root node and the forwarding / comment relationship as the edge to construct a directed propagation tree sequence , using a graph attention network GAT to process the propagation tree, automatically focusing on key propagation paths that contribute to rumor discrimination by calculating attention coefficients between nodes, and obtaining global representations of each time window through multi-layer graph attention network GAT aggregation and pooling operations .

[0021] Further, the aforementioned step S4 comprises the following sub-steps:

[0022] S4.1, setting a historical accumulation feature representation as the fusion feature representation of the previous i time windows, when performing heterogeneous graph feature fusion, let , when performing propagation cascade feature fusion, let , and the cross-attention score is calculated by the following formula:

[0023] ;

[0024] wherein , , represent the query matrix, key matrix and value matrix respectively, is the dimension of the key vector, used to scale the dot product, is a normalization function; Q α is the heterogeneous graph feature representation of the kth time window or the propagation cascade representation is multiplied by W q to obtain the query matrix Q, and is the historical accumulated feature representation of the k-1th time window is multiplied by W k and W v to obtain the key matrix K and the value matrix V, respectively, W q , W k , W v denote the linear mapping weight matrix.

[0025] S4.2, the feature vector of the time window k or is obtained as Q, and the historical accumulated feature vector of the time window k-1 is obtained as K and V, to identify the key interaction patterns throughout the time window, and to obtain the final heterogeneous graph representation and the propagation cascade representation .

[0026] Further, in the aforementioned step S5, the final representation is obtained by the following formula:

[0027] ,

[0028] wherein, denotes the concatenation operation.

[0029] Further, in the aforementioned step S5, the cross-entropy loss function formula is as follows:

[0030] ,

[0031] wherein, denotes the true label, denotes the predicted label, N denotes the number of samples, and log() denotes the natural logarithm operation.

[0032] Another aspect of the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described in the present application.

[0033] The present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of any of the methods described in the present application.

[0034] Compared with the prior art, the beneficial technical effects of the above technical scheme of the present application are as follows: the present application proposes a rumor detection method based on heterogeneous propagation structure and dynamic feature fusion, which can accurately and timely identify rumors based on social network information. First, by constructing a heterogeneous graph structure, the more complex interaction relationship in the social network is modeled, the user and post nodes and their various interaction relationships are clearly distinguished, more social network features are focused on, and more rich semantic interaction representations are obtained. At the same time, a propagation cascade module is constructed to capture the hierarchical sequence information of information propagation and better track the public opinion propagation situation. Second, a cross-attention mechanism is used to construct a space-time fusion module to cross-time step fusion of different time window sequences, which enables the model to not only see the current propagation state but also understand its evolution history, thereby more accurately capturing the essential features of rumors changing over time. Finally, multiple feature vectors are fused, and a multilayer perceptron is used to classify and predict rumors, and the model will update parameters according to the loss function to enhance the ability to distinguish rumors. In summary, the method of the present application based on the double-branch dynamic modeling framework and the cross-attention fusion module can extract deep representations using various information of the social network and significantly improve the ability of the model to distinguish rumors, especially in the early rumor detection task, it can make timely judgment on rumor information in the early stage of public opinion spread, and provide rapid and favorable support for subsequent blocking and suppression work. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is the overall model architecture of the present application.

[0036] Figure 2 is a cross-attention mechanism diagram in the space-time dynamic fusion module of the present application.

[0037] Figure 3 is a training loss fold line graph in the model training process of the present application.

[0038] Figure 4 is the early detection performance comparison test result graph of the present application, (a) is the early detection effect graph on the microblog data set, and (b) is the early detection effect graph on the PHEME data set. DETAILED DESCRIPTION

[0039] In order to better understand the technical content of the present application, specific embodiments are described below with the aid of the accompanying drawings.

[0040] Aspects of the present application are described in this disclosure with reference to the accompanying drawings, of which numerous illustrative embodiments are shown. The embodiments of this application are not limited to the attached drawings but extend to any number of embodiments comprehended by the claims and equivalents thereof, as described in this disclosure. It should be understood that the application is realized by any of the various constructions and embodiments introduced above, as well as any of the constructions and embodiments detailed below, since the constructions and embodiments disclosed in this disclosure are not limited to any of the embodiments. In addition, some aspects disclosed in this disclosure can be used alone or in any suitable combination with other aspects disclosed in this disclosure.

[0041] As Figure 1 shown, which shows the overall architecture of the present application, the whole process of rumor detection is described in detail. The process is mainly divided into four stages: data preprocessing and feature construction, double-branch dynamic feature extraction, spatiotemporal dynamic fusion and classification prediction. First, in the data input stage, the model receives the original data in the social network, including the source post, user attributes and comment forwarding relationship. The pre-trained BERT model is used to extract the text semantic features, and the activity and influence indicators of the user in different time windows are counted to construct the dynamic feature vector. Secondly, the model enters the double-branch parallel processing stage. The upper branch is the "heterogeneous interaction graph learning module", which constructs the heterogeneous graph sequence containing user and post nodes, and uses the heterogeneous graph convolution network (HGCN) to capture the complex interaction semantics between users and content; the lower branch is the "propagation cascade module", which constructs the directed propagation tree sequence from the forwarding and comment relationship, and uses the graph attention network (GAT) to capture the diffusion path and key propagation nodes of the information. Subsequently, the outputs of the two branches enter the "spatiotemporal dynamic fusion module", which aggregates the information of the historical time steps through the cross-attention mechanism. Finally, the fused heterogeneous graph representation and propagation cascade representation are spliced and input into the classifier for true or false discrimination. This architecture effectively integrates semantic, structural and temporal information, laying a foundation for high-precision rumor detection. The specific implementation steps of this framework are as follows:

[0042] In the data preprocessing and feature construction stage, the preprocessing and comprehensive feature extraction module is realized by step S1, S1, the social network data is preprocessed, the user and post static features are extracted respectively for the preprocessed data, and the user and post dynamic features are calculated respectively based on a time window, and the dynamic fusion features are obtained by fusing the static and dynamic features.

[0043] The double-branch dynamic feature extraction stage is realized by the heterogeneous interaction graph learning module and the propagation cascade module constructed by steps S2 and S3 respectively,

[0044] S2, construct the heterogeneous interaction graph learning module: construct the heterogeneous graph sequence through the dynamic fusion features, and use the heterogeneous graph convolution network to update the convolution of the heterogeneous graph sequence, output the heterogeneous graph feature matrix H k ;

[0045] S3, constructing a propagation cascade module: first, a directed propagation tree is constructed, a graph attention network GAT is used to process the directed propagation tree, and a propagation cascade global representation P is obtained k ;

[0046] After the double-branch dynamic feature extraction stage, enter the spatio-temporal dynamic fusion stage:

[0047] S4, constructing a spatio-temporal dynamic fusion module: based on the heterogeneous graph feature H k and the propagation cascade feature P k , deep fusion is performed, and the heterogeneous graph features H k and the propagation cascade features P k of different time windows are fused to generate heterogeneous graph representations H and propagation cascade representations P ;

[0048] Finally, enter the classification prediction stage:

[0049] S5, constructing a rumor classification model based on a multilayer perceptron: the heterogeneous graph representation H and the propagation cascade representation P are spliced to obtain a final representation, which is input into a rumor classification model based on a multilayer perceptron, combined with a cross-entropy loss function, to obtain a prediction result of rumor / non-rumor.

[0050] As a preferred embodiment of the present application, step S1 includes the following sub-steps:

[0051] S1.1, for the post content in the social network, using a pre-trained BERT model, the static feature of the post is obtained by the following formula:

[0052] ,

[0053] In the formula, denotes the static feature of the post, denotes the original content of the post, denotes a pre-trained language model for converting text content into vector embedding; denotes the original attribute information of the user, such as user id, number of fans, number of attention, registration time, etc., denotes a pre-trained encoding model for extracting high-dimensional features of the user.

[0054] S1.2, divide the rumor propagation period into k time windows, and calculate the dynamic features of the user nodes and post nodes in each time window respectively to obtain user dynamic features and post dynamic features ;

[0055] S1.3, for each user and post in each time window, the formula Fusion of static and dynamic features to obtain dynamic fusion features of time steps, wherein, represents a vector splicing operation, represents the dynamic fusion features of the user or post at the i-th time step, represents the static features of the user or the static features of the post, represents the dynamic features of the user or the post, wherein , when i=u, it represents a user node, and when i=p, it represents a post node.

[0056] As a preferred embodiment of the present application, step S2 constructs a heterogeneous interaction graph learning module, including the following sub-steps:

[0057] S2.1, construct a heterogeneous graph sequence by dynamic fusion features , wherein represents the i-th time window of the heterogeneous graph snapshot, including two types of nodes: user nodes and post nodes ; and edge relationships: user--user , post--post and user--post ;

[0058] S2.2, adopt a heterogeneous graph convolution network for heterogeneous relationship perception, and introduce a relationship perception propagation operator , assign a learnable attention weight to different types of edge relationships, and calculate the propagation operator by the formula: to aggregate neighbor information of different relationships and output a heterogeneous graph feature matrix sequence . , , , respectively represent the normalized heterogeneous edge matrix and the corresponding attention weight parameter.

[0059] As a preferred embodiment of the present application, step S3 constructs a propagation cascade module, which is specifically: taking the source post as the root node and the forwarding / comment relationship as the edge, constructing a directed propagation tree sequence , using a graph attention network GAT to process the propagation tree, calculating the attention coefficient between nodes to automatically focus on key propagation paths that contribute to rumor discrimination, and after multi-layer graph attention network GAT aggregation and pooling operation, obtaining the propagation cascade representation sequence of each time window .

[0060] ​Figure 2 The calculation process of the cross-attention mechanism, the core component of the spatio-temporal dynamic fusion module in the present application, is shown.

[0061] Through the cross-attention mechanism, the heterogeneous graph feature representation of different time windows and the propagation cascade representation of different time windows are fused respectively. The historical accumulation feature representation is set as the fusion feature representation of the previous i time windows, so that when the heterogeneous graph feature is fused, and when the propagation cascade feature is fused. The cross-attention score is calculated by the following formula:

[0062] ;

[0063] wherein , , represent the query matrix, the key matrix and the value matrix respectively, is the dimension of the key vector, used to scale the dot product to prevent the result after the dot product from being too large, is a normalization function used to convert the attention score into a probability distribution; Q α is the heterogeneous graph feature representation or the propagation cascade representation of the kth time window multiplied by W q to obtain the query matrix, and are the historical accumulation feature representation of the k-1th time window multiplied by W k and W v to obtain the key matrix K and the value matrix V, respectively, and W q , W k , W v represent linear mapping weight matrices.

[0064] The feature vector or of the time window k is taken as Q, the historical accumulation feature vector obtained from the k-1th time window is taken as K and V, the key interaction mode running through the time windows is identified, and the final heterogeneous graph representation and the propagation cascade representation are obtained.

[0065] This module aims to solve the problem that the traditional method is difficult to capture long-distance time sequence dependence. For the feature representation of the current time window , the model maps it to the query matrix . At the same time, the feature representation of the previous time window (or historical memory) is mapped to the key matrix value matrix The attention weight matrix is obtained by calculating the dot product of the transpose of and , dividing by the scaling factor , and then normalizing by the Softmax function. This weight matrix intuitively reflects which parts of the historical state are most critical to the judgment of the current state. Finally, the weight matrix is multiplied by the value matrix to generate a current time representation that integrates historical evolution information. Through this mechanism, the model can effectively identify malicious propagation patterns throughout the time axis, such as sustained bot coordinated attacks or specific topic evolution trends, significantly enhancing the robustness of the features.

[0066] As a preferred embodiment of the present application, in step S5, the heterogeneous graph representation and the propagation cascade representation are spliced to obtain the final representation by the following formula, wherein, represents the splicing operation.

[0067] The rumor classification model of the input multi-layer perceptron is combined with the cross-entropy loss function:

[0068] to obtain the prediction result of rumor / non-rumor.

[0069] In the formula, represents the true label, represents the predicted label, N represents the number of samples, and log() represents the natural logarithm operation.

[0070] Table 1 below shows the accuracy benchmark test comparison results of the present method and existing mainstream rumor detection algorithms (such as RvNN, Bi-GCN, DDGCN, etc.) on public data sets. The table shows the performance of each model on the accuracy index through different colored lines or bar graphs. The experimental results show that the present method has achieved the highest detection accuracy on both the microblog and PHEME real data sets. Specifically, compared with RvNN and Bi-GCN, which only use propagation structure, the present application can more fully utilize the background information of user nodes due to the introduction of heterogeneous graph modeling, thereby significantly improving the accuracy of the judgment; compared with DDGCN, which also uses dynamic graphs, the present application achieves deeper spatiotemporal feature fusion through the cross-attention mechanism, thus achieving further performance surpassing. This result fully proves the superiority of the heterogeneous propagation structure and dynamic feature fusion strategy in complex social scenarios.

[0071] Table 1

[0072]

[0073] Another aspect of the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of the embodiments when executing the computer program.

[0074] The present application also provides a computer readable storage medium having stored thereon a computer program, wherein the computer program implements the steps of the method according to any one of the embodiments when executed by a processor.

[0075] Figure 4 The comparative test results of the present application in early detection performance are shown. Figure (a) shows the early detection effect on the microblog dataset, and figure (b) shows the early detection effect on the PHEME dataset. Early detection refers to the true or false discrimination according to limited propagation data within a very short time (such as 2 hours, 4 hours, 6 hours) after the rumor is published. The horizontal coordinate in the figure represents the time delay after the rumor is published, and the vertical coordinate represents the detection accuracy. As can be clearly seen from the figure, with the passage of time and the increase of information, the accuracy of all models shows an upward trend. However, the method of the present application (red curve) is significantly better than other comparative methods at each time point. In particular, in the initial 2 hours and 4 hours stage, the present application still maintains a high accuracy, and the gap with other methods is most obvious. This is due to the fact that the heterogeneous graph module can use user portraits and other static features to make up for the defects of incomplete early propagation structure, and the attention mechanism can capture key early clues. This feature makes the present application have high application value in the actual public opinion monitoring system, and can suppress rumors in the embryonic state.

[0076] In summary, through comprehensive comparison and analysis on two real large-scale social network datasets, the rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features proposed by the present application shows significant advantages. The present application constructs a double-branch dynamic modeling framework, innovatively combines heterogeneous graph convolution network (HGCN) and graph attention network (GAT), and introduces cross-attention mechanism for spatio-temporal feature fusion. This design makes the model significantly better than existing homogeneous graph models and other dynamic graph models in processing complex heterogeneous interaction relationships and dynamic evolution of propagation cascade structure. These improvements not only improve the mining ability of deep semantic and structural features, but more importantly, significantly enhance the performance of the model in early detection tasks, which can make accurate discrimination under the condition of limited information in the early stage of rumor propagation. Therefore, the excellent performance of the present application in various complex social scenarios verifies the feasibility and effectiveness of the present application as an efficient, accurate and excellent early warning rumor detection method.

[0077] While the application has been described by way of example with reference to certain embodiments thereof, it is to be understood that the application is not limited to the embodiments described above, but intrinsically extends to various modifications and changes in form and details. Therefore, the scope of the application should be determined not by the embodiments described above, but by the scope of the appended claims and their equivalents.

Claims

1. A rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features, characterized in that, Includes the following steps: S1. Construct a preprocessing and comprehensive feature extraction module: preprocess the social network data, extract static features of users and posts from the preprocessed data, and based on... Calculate the dynamic features of users and posts in each time window, and then fuse the static and dynamic features to obtain the dynamic fused features. S2. Constructing a heterogeneous interaction graph learning module: A heterogeneous graph sequence is constructed by dynamically fusing features. This sequence is then updated using a heterogeneous graph convolutional network, outputting a heterogeneous graph feature matrix H. k ; S3. Constructing the Propagation Cascade Module: First, construct a directed propagation tree, then use a graph attention network (GAT) to process the directed propagation tree and obtain the global representation P of the propagation cascade. k ; S4. Construct a spatiotemporal dynamic fusion module: based on the heterogeneous graph features H from step S2. k and propagation cascade features P k Deep fusion is performed to combine the heterogeneous graph features H from different time windows. k and propagation cascade features P k The graphs are fused to generate heterogeneous graph representations. and propagation cascade representation ; S5. Construct a rumor classification model based on a multilayer perceptron: Represent heterogeneous graphs. and propagation cascade representation The final representation is obtained by splicing the data together. This representation is then input into a multilayer perceptron rumor classification model. Combined with the cross-entropy loss function, the prediction result of whether the data is rumor or not is obtained.

2. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. For the content of posts in social networks, a pre-trained BERT model is used to obtain the static features of the posts using the following formula: ， ； In the formula, This indicates the static characteristics of a post. This indicates the original content of the post. This represents a pre-trained language model used to convert text content into vector embeddings; Represents the user's original attribute information. This represents a pre-trained coding model used to extract high-dimensional features from users; S1.2 Divide the rumor propagation cycle into k time windows, and calculate the dynamic characteristics of user nodes and post nodes in each time window to obtain user dynamic characteristics. and post dynamic features ; S1.

3. For each user and post within each time window, the formula is used... By fusing static and dynamic features, dynamic fused features at the time step are obtained, where... This represents a vector concatenation operation. Indicates the first Dynamic fusion features of users or posts within a time window This represents the static characteristics of a user or the static characteristics of a post. Indicates the dynamic characteristics of a user or post. When i=u, it represents a user node; when i=p, it represents a post node.

3. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Constructing heterogeneous graph sequences through dynamic feature fusion ,in A snapshot of the heterogeneous graph representing the i-th time window, comprising two types of nodes: user nodes. and post nodes And edge relationships: user--user Posts -- Posts And users -- posts ; S2.

2. Use heterogeneous graph convolutional networks for heterogeneous relation awareness, and introduce a relation awareness propagation operator. Learnable attention weights are assigned to different types of edge relationships using the formula: The propagation operator is computed to aggregate neighbor information with different relationships, and the heterogeneous graph feature matrix sequence { H1, H2, ..., H} is output. k. }, , , Represent the normalized heterogeneous edge matrix respectively The corresponding attention weight parameters.

4. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, Step S3 specifically involves constructing a directed propagation tree sequence with the source post as the root node and the forwarding / commenting relationships as edges. The propagation tree is processed using a graph attention network (GAT) by calculating the attention coefficients between nodes. It automatically focuses on key propagation paths that contribute significantly to rumor detection, and obtains a global representation of the propagation cascade for each time window through multi-layer graph attention network (GAT) aggregation and pooling operations. .

5. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.1 Setting up historical accumulation feature representation Let be used as the fusion feature representation for the first i time windows. When performing heterogeneous graph feature fusion, let During the propagation cascade feature fusion, let The cross-attention score is calculated using the following formula: ； in , , These represent the query matrix, key matrix, and value matrix, respectively. It is the dimension of the key vector, used to scale the dot product. It is a normalization function; Q α It is the heterogeneous graph feature representation of the k-th time window. Or propagation cascade characterization With W q The query matrix Q obtained by multiplication, and It represents the historical accumulation features of the (k-1)th time window. respectively with W k and W v The key matrix K and value matrices V and W obtained by multiplication q W k W v Represents the linear mapping weight matrix; S4.2, extract the feature vector of time window k. or As Q, the historical accumulated feature vector obtained in the (k-1)th time window Using K and V as key interactions across the time window, we obtain the final heterogeneous graph representation. and propagation cascade representation .

6. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, In step S5, the final representation is obtained using the following formula: ， in, This indicates a splicing operation.

7. The rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features according to claim 1, characterized in that, In step S5, the cross-entropy loss function is formulated as follows: ， In the formula, Indicates the true label, This represents the predicted label, N represents the number of samples, and log() represents the natural logarithm operation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images