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

Abstract

The application discloses a rumor detection method and system based on heterogeneous propagation structure and dynamic feature fusion, and belongs to the fields of computer technology, artificial intelligence and social network analysis. The method steps are as follows: static features of users and posts are extracted, dynamic features of users and posts are calculated, and dynamic fusion features are obtained by fusing static and dynamic features. A heterogeneous graph sequence is constructed through the dynamic fusion features, convolution updating is carried out by using a heterogeneous graph convolution network, and a heterogeneous graph feature matrix is output. Then, a directed propagation tree is processed by using a graph attention network, and global representation of a propagation cascade is obtained. The heterogeneous graph features and the propagation cascade features are cross-time-step interacted, heterogeneous graph representation and propagation cascade representation are generated, and finally, the final representation is obtained by splicing, a rumor classification model is input, and a prediction result is obtained. The excellent performance of the application in various complex social scenes verifies the feasibility and effectiveness of the application as an efficient, accurate and excellent early warning rumor detection method.

Description

Technical Field

[0001] This invention belongs to the fields of computer technology, artificial intelligence and social network analysis, and specifically relates to a rumor detection method and system based on the fusion of heterogeneous propagation structure and dynamic features. Background Technology

[0002] With the development of mobile social networking platforms, platforms like Weibo have become major platforms for people to quickly obtain information, express opinions, and interact. However, the convenience and immediacy of information dissemination also exacerbate the spread of rumors on social networks. Therefore, advanced technologies that can quickly and accurately detect rumors on social media have become a current research hotspot. Existing rumor detection technologies have mainly evolved from traditional machine learning based on manual feature engineering to deep learning-based technologies. Deep learning technology can automatically learn high-order, nonlinear deep feature representations from data, thereby more accurately capturing the complex patterns of rumors in terms of content semantics and propagation structure, effectively avoiding the bottleneck of feature engineering. However, most existing methods typically perform homogeneous modeling of social network information, ignoring the various 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 that changes over time. Existing methods often extract static features or simply treat time as a dimension, failing to deeply capture the structural evolution of events at different time steps, resulting in an inability to reflect the explosive nature of rumors in the early stages of propagation or the corrective features in the later stages.

[0003] Currently, extensive research has been conducted both domestically and internationally on rumor detection methods, yielding some progress. Early work used text-based fusion neural networks for rumor detection, proposing a novel data processing algorithm to accurately extract contextual features and constructing a fusion neural network including bidirectional gated recurrent units, convolutional models, and attention mechanisms. However, this method only processes the textual features of social network rumors, ignoring other features, thus limiting its detection performance. Subsequent work considered rumor detection methods based on topic classification and content understanding. This approach can extract features from subsets of datasets at different scales, comprehensively considering the relevance and differences of different topics, and combining discourse and word meaning to expand the model's ability to understand text, enhancing its ability to detect malicious rumor traps. However, this method typically only focuses on maintaining the previous state without considering subsequent contextual information. Furthermore, some methods consider utilizing graph convolutional networks, but mainly focus on the reconstruction and representation of static structures, failing to effectively capture the dynamic changes in structure over time during rumor propagation, resulting in insufficient utilization of temporal features. To capture propagation patterns, novel transformation methods convert complex forwarding trees into more easily processed binary trees without losing reconstruction information. Then, the forwarding tree is serialized into a corpus of metatree paths, with each metatree retaining a basic substructure. A deep neural network is then used to integrate all metatrees and generate global structural embeddings. However, the process of converting the propagation tree into a binary tree may lose some original topological information, and this method tends to treat the propagation structure as an independent feature, failing to model the user interaction process holistically within a unified dynamic framework. The Recurrent Neural Network (RvNN-DB) model introduces an attention mechanism to adjust the weights of cues when extracting deep propagation features, enabling the model to focus on key cues beneficial for rumor detection. When extracting broad propagation features, a one-dimensional convolutional neural network is first used to obtain fine-grained relationships between words within comments, and then semantic interaction relationships between comments are obtained to supplement deep features. However, recurrent neural network-based methods typically only capture dependencies between parent and child nodes, making it difficult to capture global graph structure features, and they are computationally inefficient when processing deep tree structures, limiting their scalability when dealing with large-scale social data. To fuse multiple content features, the Hierarchical Aggregating Graph Neural Network (HAGNN) method focuses on capturing high-level representations of text content at different granularities and fusing rumor propagation structures. It combines Graph Convolutional Networks (GCNs) with a rumor propagation graph to learn text-granular representations of event propagation. However, this method primarily aggregates features on a static graph, neglecting the highly time-sensitive nature of rumor propagation and failing to capture the dynamic evolutionary features of the event's development. The DDCA-GCN rumor detection model addresses the data sparsity problem by employing a content-based data compensation method.Secondly, this method proposes a network structure construction approach to address the problem of user association within early-stage topic domains, transforming the topic domains into multi-user relationship networks with non-Euclidean structures. Graph Convolutional Networks (GCNs) are then used to process non-Euclidean data, achieving good results in early detection. However, although this method considers user association, it focuses on static fusion, often neglecting the evolution of the propagation structure over time and failing to fully utilize dynamic interaction information in time series. The proposed Dual Dynamic Graph Convolutional Network (DDGCN) method models the dynamics of messages in propagation and the dynamics of background knowledge in knowledge graphs within a unified framework. Specifically, two graph convolutional networks are used to capture the aforementioned two structural information at different time stages, and then combined with a temporal fusion unit. This allows for learning dynamic event representations in a finer-grained manner and progressively aggregating them to capture cascading effects, thus better detecting rumors. However, this method only models homogeneous information and fails to align and deeply fuse dynamic data from different sources. Summary of the Invention

[0004] This invention addresses the problems existing in the prior art by providing a rumor detection method and system based on the fusion of heterogeneous propagation structure and dynamic features. It can improve the feature extraction capability of heterogeneous interaction patterns and propagation cascade structures, 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 stages of rumor spread.

[0005] To address the above technical problems, this invention provides the following technical solution: a rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features, comprising the following steps:

[0006] 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.

[0007] 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 ;

[0008] 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 ;

[0009] S4. Construct a spatiotemporal dynamic fusion module: based on the heterogeneous graph features H from step S2. kand 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 ;

[0010] 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.

[0011] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0012] 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:

[0013] , ;

[0014] 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;

[0015] 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 ;

[0016] S1.3. For each user and post within each time window, the formula is used... By fusing static and dynamic features, dynamic fused features of the time window 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.

[0017] Furthermore, the aforementioned step S2 includes the following sub-steps:

[0018] 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 ;

[0019] 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.

[0020] Furthermore, step S3 described above 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. .

[0021] Furthermore, step S4 as described above includes the following sub-steps:

[0022] 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:

[0023] ;

[0024] 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;

[0025] S4.2, extract the feature vectors of time window k. or As Q, the historical accumulated feature vector obtained through time window k-1 Using K and V as key interactions across the time window, we obtain the final heterogeneous graph representation. and propagation cascade representation .

[0026] Furthermore, in step S5 above, the final representation is obtained using the following formula:

[0027] ,

[0028] in, This indicates a splicing operation.

[0029] Furthermore, in step S5 above, the cross-entropy loss function is formulated as follows:

[0030] ,

[0031] 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.

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

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

[0034] Compared to existing technologies, the beneficial technical effects of this invention using the above technical solution are as follows: This invention proposes a rumor detection method based on the fusion of heterogeneous propagation structure and dynamic features, which can accurately and timely identify rumors based on social network information. First, by constructing a heterogeneous graph structure, more complex interaction relationships in social networks are modeled, clearly distinguishing between users and post nodes and their diverse interaction relationships, focusing on more social network features, and thus obtaining richer semantic interaction representations. Simultaneously, a propagation cascade module is constructed to capture the hierarchical sequence information of information propagation, better tracking the spread of public opinion. Second, a spatiotemporal fusion module is constructed using a cross-attention mechanism to fuse sequences from different time windows across time steps. This allows the model to not only see the current propagation state but also understand its evolutionary history, thereby more accurately capturing the essential characteristics of rumors changing over time. Finally, multiple feature vectors are fused, and a multilayer perceptron is used to classify and predict rumors. The model updates its parameters according to the loss function, enhancing its ability to learn and distinguish rumors. In summary, the method of this invention, based on a bi-branch dynamic modeling framework and a cross-attention fusion module, can extract deep representations using various information from social networks and significantly improve the model's ability to distinguish rumors. In particular, in the early stages of rumor detection, it can promptly identify rumor information in the early stages of public opinion dissemination, providing rapid and favorable support for subsequent blocking and suppression efforts. Attached Figure Description

[0035] Figure 1 This is the overall model architecture diagram of the present invention.

[0036] Figure 2 This is a schematic diagram of the cross-attention mechanism in the spatiotemporal dynamic fusion module of the present invention.

[0037] Figure 3 This invention provides a training loss line graph during model training.

[0038] Figure 4 The figures show the results of the early detection performance comparison test of the present invention. In the figure, (a) is the early detection effect on the Weibo dataset, and (b) is the early detection effect on the PHEME dataset. Detailed Implementation

[0039] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0040] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0041] like Figure 1 As shown, this demonstrates the overall architecture of the invention and details the entire process of rumor detection. The process is mainly divided into four stages: data preprocessing and feature construction, dual-branch dynamic feature extraction, spatiotemporal dynamic fusion, and classification prediction. First, in the data input stage, the model receives raw data from social networks, including source posts, user attributes, and comment / forward relationships. A pre-trained BERT model is used to extract textual semantic features, while simultaneously calculating user activity and influence indicators at different time windows to construct dynamic feature vectors. Second, the model enters a dual-branch parallel processing stage. The upper branch is the "Heterogeneous Interaction Graph Learning Module," which constructs a heterogeneous graph sequence containing user and post nodes, using a Heterogeneous Graph Convolutional Network (HGCN) to capture the complex semantic interactions between users and content. The lower branch is the "Propagation Cascade Module," which constructs a directed propagation tree sequence from forward and comment relationships, using a Graph Attention Network (GAT) to capture the information diffusion path and key propagation nodes. Subsequently, the outputs of both branches enter the "Spatiotemporal Dynamic Fusion Module," which aggregates information from historical time steps through a cross-attention mechanism. Finally, the fused heterogeneous graph representation and the propagation cascade representation are concatenated and input into the classifier for true / false detection. This architecture effectively integrates semantic, structural, and temporal information, laying the 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, a preprocessing and comprehensive feature extraction module is constructed through step S1. S1 involves preprocessing the social network data, extracting static features of users and posts from the preprocessed data, and then... The dynamic features of users and posts are calculated separately for each time window, and the static and dynamic features are then merged to obtain dynamic fused features.

[0043] The dual-branch dynamic feature extraction stage is specifically implemented through the heterogeneous interaction graph learning module and the propagation cascade module constructed in steps S2 and S3, respectively.

[0044] 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 ;

[0045] 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 ;

[0046] After the dual-branch dynamic feature extraction stage, the spatiotemporal dynamic fusion stage begins:

[0047] 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 ;

[0048] Finally, we proceed to the classification prediction stage:

[0049] S5. Construct a rumor classification model based on a multilayer perceptron: Represent heterogeneous graphs. and propagation cascade representation The final representation is obtained by concatenating the data and inputting it into a multilayer perceptron rumor classification model. Combined with the cross-entropy loss function, the prediction result of rumor / non-rumor is obtained.

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

[0051] 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:

[0052] ,

[0053] 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; This represents the user's original attribute information, such as user ID, number of followers, number of followings, registration time, etc. This represents a pre-trained encoding model used to extract high-dimensional features from users.

[0054] 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 ;

[0055] S1.3. For each user and post within each time window, the formula is used... Static and dynamic features are fused to obtain dynamic fused features at a time step, wherein... This represents a vector concatenation operation. Indicates the first Dynamic fusion features of users or posts at each time step This represents the static characteristics of a user or the static characteristics of a post. Indicates the dynamic characteristics of a user or post, where When i=u, it represents a user node; when i=p, it represents a post node.

[0056] In a preferred embodiment of the present invention, step S2, which constructs a heterogeneous interaction graph learning module, includes the following sub-steps:

[0057] 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 ;

[0058] 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 is output. . , , , Represent the normalized heterogeneous edge matrix respectively The corresponding attention weight parameters.

[0059] In a preferred embodiment of the present invention, step S3, constructing the propagation cascade module, specifically involves: constructing a directed propagation tree sequence with the source post as the root node and the forwarding / commenting relationship as the edge. The propagation tree is processed using a graph attention network (GAT) by calculating the attention coefficients between nodes. The system automatically focuses on key propagation paths that contribute significantly to rumor detection. Through multi-layer graph attention network (GAT) aggregation and pooling operations, it obtains the propagation cascade representation sequence for each time window. .

[0060] Figure 2 The calculation process of the cross-attention mechanism, a core component of the spatiotemporal dynamic fusion module in this invention, is demonstrated.

[0061] Through a cross-attention mechanism, heterogeneous graph feature representations from different time windows and propagation cascade representations from different time windows are fused. Historical accumulated feature representations are then established. Let the fusion feature representation of the first i time windows be used, and when performing heterogeneous graph feature fusion, let Similarly, when fusion of propagation cascade features, let The cross-attention score is calculated using the following formula:

[0062] ;

[0063] in , , These represent the query matrix, key matrix, and value matrix, respectively. This is the dimension of the key vector, used to scale the dot product to prevent the result from becoming too large. It is a normalization function used to convert attention scores into a probability distribution; 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 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 This represents the linear mapping weight matrix.

[0064] 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 .

[0065] This module aims to address the problem that traditional methods struggle to capture long-range temporal dependencies. For the current time window... The model maps the feature representation to a query matrix. At the same time, the previous time window will be... The feature representation (or historical memory) is mapped to a key matrix. Sum matrix Through calculation and The dot product of the transpose of the product, divided by the scaling factor. After normalization using the Softmax function, the attention weight matrix is ​​obtained. This weight matrix intuitively reflects which parts of the historical states are most critical for judging the current state. Finally, the weight matrix is ​​compared with the value matrix. Multiplication generates a representation of the current moment that incorporates historical evolutionary information. Through this mechanism, the model can effectively identify malicious propagation patterns across the timeline, such as continuous coordinated online attacks or the evolutionary trends of specific topics, significantly enhancing the robustness of the features.

[0066] In a preferred embodiment of the present invention, in step S5, the heterogeneous diagram is represented. and propagation cascade representation The final representation is obtained by concatenating the following formulas. ,in, This indicates a splicing operation.

[0067] Input a multilayer perceptron rumor classification model, combined with the cross-entropy loss function:

[0068] This allows you to obtain predictions of whether something is a rumor or not.

[0069] 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.

[0070] Table 1 below shows the accuracy benchmark test results of the method of this invention and existing mainstream rumor detection algorithms (such as RvNN, Bi-GCN, DDGCN, etc.) on public datasets. The table uses line charts or bar charts of different colors to illustrate the performance of each model in terms of accuracy. Experimental results show that the method of this invention achieves the highest detection accuracy on both the Weibo and PHEME real-world datasets. Specifically, compared to RvNN and Bi-GCN, which only utilize the propagation structure, this invention, by introducing heterogeneous graph modeling, can more fully utilize the background information of user nodes, thus significantly improving the accuracy of the judgment; compared to DDGCN, which also utilizes dynamic graphs, this invention achieves deeper spatiotemporal feature fusion through a cross-attention mechanism, thus achieving further performance superiority. This result fully demonstrates 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 invention provides an electronic device, including a memory, a processor, and a computer program stored in 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 embodiments.

[0074] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described in the embodiments.

[0075] Figure 4 The comparative test results of the present invention in early detection performance are presented. Figure (a) shows the early detection effect on the Weibo dataset, and (b) shows the early detection effect on the PHEME dataset. Early detection refers to the determination of authenticity based on limited dissemination data within a very short period of time (e.g., within 2 hours, 4 hours, or 6 hours) after the rumor is published. The horizontal axis in the figure represents the time delay after the rumor is published, and the vertical axis represents the detection accuracy. It can be clearly seen from the figure that the accuracy of all models increases with the time and the amount of information. However, the method of the present invention (red curve) is significantly better than other comparative methods at all time points. Especially in the initial 2-hour and 4-hour stages, the present invention still maintains a high accuracy, and the gap with other methods is most obvious. This is due to the heterogeneous graph module's ability to use static features such as user profiles to make up for the defects of incomplete early dissemination structure, and the attention mechanism's keen capture of key early clues. This feature makes the present invention highly valuable in actual public opinion monitoring systems, and can curb rumors in their infancy.

[0076] In summary, through comprehensive comparison and analysis on two real-world large-scale social network datasets, the rumor detection method proposed in this invention, based on heterogeneous propagation structure and dynamic feature fusion, demonstrates significant advantages. This method innovatively combines Heterogeneous Graph Convolutional Networks (HGCN) and Graph Attention Networks (GAT) by constructing a dual-branch dynamic modeling framework, and introduces a cross-attention mechanism for spatiotemporal feature fusion. This design enables the model to achieve significantly higher detection accuracy than existing isomorphic graph models and other dynamic graph models when dealing with complex heterogeneous interaction relationships and dynamically evolving propagation cascade structures. These improvements not only enhance the ability to mine deep semantic and structural features, but more importantly, they significantly enhance the model's performance in early detection tasks, enabling accurate discrimination even with limited information in the early stages of rumor propagation. Therefore, the superior performance of this invention in various complex social scenarios verifies its feasibility and effectiveness as an efficient, accurate rumor detection method with excellent early warning capabilities.

[0077] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A rumor detection method based on a heterogeneous propagation structure and dynamic feature fusion, characterized in that, The method comprises the following steps: S1, constructing a pre-processing and comprehensive feature extraction module: pre-processing the social network data, extracting user and post static features for the pre-processed data, and calculating user and post dynamic features based on a time window respectively, and fusing the static and dynamic features to obtain dynamic fusion features; S1, constructing a pre-processing and comprehensive feature extraction module: pre-processing the social network data, extracting user and post static features for the pre-processed data, and calculating user and post dynamic features based on a time window respectively, and fusing the static and dynamic features to obtain dynamic fusion features; S2, construct a heterogeneous interaction graph learning module: construct a heterogeneous graph sequence by dynamically fusing features, update the heterogeneous graph sequence by using a heterogeneous graph convolution network, and output a heterogeneous graph feature matrix H k ; specifically comprising the following sub-steps: S2.1, Constructing a heterogeneous graph sequence by dynamically fusing 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 ; S2.2, adopt heterogeneous graph convolution network for heterogeneous relationship perception, set to introduce relationship perception propagation operator , assign learnable attention weights to different types of edge relationships, through the formula: Calculate the propagation operator, aggregate neighbor information of different relationships, and output the heterogeneous graph feature matrix sequence { H1, H2,…, H k.}, , , respectively represent the normalized heterogeneous edge matrix corresponding attention weight parameters; S3, build a propagation cascade module: first build a directed propagation tree, use graph attention network GAT to process the directed propagation tree, and obtain the global representation P of the propagation cascade k ; Specifically: taking the source post as the root node and the forwarding / comment relationship as the edge, a directed propagation tree sequence is constructed , use graph attention network GAT to process the propagation tree, calculate the attention coefficient between nodes , automatically focus on the key propagation path that contributes to rumor discrimination, and obtain the global representation of the propagation cascade in each time window through multi-layer graph attention network GAT aggregation and pooling operation ; S4, constructing a spatio-temporal dynamic fusion module: based on the heterogeneous graph features H of step S2 k and the propagation cascade features P k performing deep fusion on the heterogeneous graph features H of different time windows k and the propagation cascade features P k performing fusion to generate a heterogeneous graph representation and a propagation cascade representation ; specifically comprising the following sub-steps: S4.1, Set historical accumulation feature representation as the fusion feature representation of the previous i time windows, when performing heterogeneous graph feature fusion, let , when propagating cascaded feature fusion, let The cross-attention score is calculated by the following formula: ； wherein , , respectively represent the query matrix, key matrix and value matrix, is the dimension of the key vector, used to scale the dot product, is the normalization function; Q α is the heterogeneous graph feature representation of the kth time window or the propagation cascade representation multiplied by W q to obtain the query matrix Q, and is the historical accumulated feature representation of the k-1th time window multiplied by W k and W v to obtain the key matrix K and value matrix V, respectively, W q , W k , W v represent linear mapping weight matrices; S4.2, the feature vector of the time window k or as Q, the historical accumulated feature vector obtained from the k-1th time window as K and V, identify the key interaction patterns throughout the time window, and obtain the final heterogeneous graph representation and the propagation cascade representation ; S5, constructing a rumor classification model based on a multi-layer perception: the heterogeneous graph representation and the propagation cascade representation are spliced to obtain a final representation, which is input into a rumor classification model based on a multi-layer perception, combined with a cross-entropy loss function, to obtain a prediction result of a rumor or non-rumor.

2. The rumor detection method based on heterogeneous propagation structure and dynamic feature fusion according to claim 1, characterized in that, Step S1 comprises the following sub-steps: S1.1, for the post content in the social network, using a pre-trained BERT model, the static features of the post are obtained by the following formula: ， ； In the formula, represents the static features of the post, represents the original content of the post, represents a pre-trained language model for converting text content into a vector embedding; represents the original attribute information of the user, represents a pre-trained encoding model for extracting high-dimensional features of the user; S1.2, divide the rumor propagation period into k time windows, and calculate the dynamic features of user nodes and post nodes respectively in each time window to obtain user dynamic features and post dynamic features ; S1.3, for each user and post in each time window, the formula Fusion of static and dynamic features, obtain dynamic fusion features of time step, wherein, denotes a vector splicing operation, denotes the dynamic fusion feature of the user or post in the i-th time window, 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. 3.The rumor detection method based on heterogeneous propagation structure and dynamic feature fusion according to claim 1, characterized in that, In step S5, the final representation is obtained by the following formula: ， wherein represents a concatenation operation.

4. The rumor detection method based on heterogeneous propagation structure and dynamic feature fusion according to claim 1, characterized in that, In step S5, the cross-entropy loss function formula is as follows: ， In the formula, denotes a real label, denotes a predicted label, N denotes a sample number, and log() denotes a natural logarithm operation.

5. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 4.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 4.

Images