Rumor detection method based on dual-domain perception structural feature fusion learning

Abstract

The present invention relates to a rumor detection method based on dual-domain perceptual structural feature fusion learning, belonging to the field of natural language processing and information dissemination technology. The method comprises the following steps: obtaining a social media message event dataset, constructing a post propagation network graph and a user social network graph, and extracting post propagation features and user social features; using the propagation density peak and slope partitioning method to identify the key time window for rumor diffusion and extract substructure features; processing the subfeatures through a mutual attention mechanism to obtain fused features; constructing a projection matrix based on the fused features, using the projection matrix to perform weighted summation on the fused features to obtain interaction features, and then generating a first weighted post propagation feature and a first weighted user social feature. After homogeneous interaction information modeling, a final homogeneous interaction comprehensive representation is obtained, which is then spliced with the original tweet features for classification to obtain a rumor detection result. The present invention can improve the accuracy and robustness of rumor detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing and information dissemination, and specifically relates to a rumor detection method based on dual-domain perception structure feature fusion learning. Background Art

[0002] With the rapid development of the internet and social media, the public can access information and exchange opinions more efficiently. Unlike traditional rumor propagation channels, rumors on social media transcend geographical and temporal boundaries and spread rapidly. However, rumors on social media exhibit unique propagation patterns, and effectively leveraging these patterns for detection is crucial for purifying the online environment and maintaining information security. The rapid development of deep learning technology has provided significant support for rumor detection. Researchers extract features from multiple dimensions, such as text content, user characteristics, and propagation paths, to construct high-dimensional event representations and assess information authenticity. For example, recurrent neural networks (RNNs, LSTMs, and GRUs) are well-suited for capturing the temporal dependencies of information propagation due to their excellent temporal modeling capabilities. Convolutional neural networks (CNNs), with their powerful feature extraction capabilities, have been widely used in text analysis and propagation modeling. However, CNNs primarily focus on local features and struggle to effectively model the global structure of information propagation. GNNs, by modeling nodes and their relationships within a graph structure, can effectively capture the complex dynamics of information propagation and show promising application prospects in rumor detection.

[0003] To accurately detect rumors, researchers have conducted in-depth analysis of content features such as text and images, user attributes, and communication structures involved in rumors, and explored methods such as machine learning, tree structure modeling, and graph structure learning. A review of existing research reveals that existing rumor detection methods still face several key issues: First, the mining of propagation features during key time periods is insufficient. Second, rumor propagation is often complex and sudden, and existing methods fail to fully utilize limited time series to mine key features when dealing with dynamic features during intensive propagation phases, making it difficult to accurately capture the core information of rumor propagation. Third, the fusion of multidimensional features is insufficient. Existing methods separate the coupling relationship between user social behavior characteristics and information propagation structures, lack the ability to jointly model the dynamics of user-information interactions, and thus affect detection performance. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a rumor detection method based on dual-domain perception structure feature fusion learning.

[0005] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0006] The present invention provides a rumor detection method based on dual-domain perception structure feature fusion learning, comprising the following steps:

[0007] S1. Obtain a social media message event dataset, construct a post propagation network graph and a user social network graph based on the dataset, and extract post propagation features and user social features respectively through a graph attention network encoder;

[0008] S2. Use the peak and slope partitioning method of propagation density to identify the key time window for rumor spread and extract relevant substructure features from the post propagation network graph and user social network graph;

[0009] S3. Process the post propagation sub-features and user social sub-features through the mutual attention mechanism to obtain the post propagation fusion feature and the user social fusion feature;

[0010] S4. Constructing a projection matrix using the post propagation fusion features and the user social fusion features. Then, using the projection matrix, performing a weighted summation of the post propagation fusion features and the user social fusion features to obtain the post propagation interaction features and the user social interaction features, thereby generating a first weighted post propagation feature and a first weighted user social feature.

[0011] S5. Modeling homogeneous interaction information based on the first weighted post propagation feature and the first weighted user social feature to obtain a final homogeneous interaction comprehensive representation;

[0012] S6. After concatenating the final homogeneous interaction comprehensive representation with the original tweet features, the feature is input into the rumor detection module for classification to obtain the rumor detection results.

[0013] Furthermore, step S1 specifically includes:

[0014] The social media message event dataset ,in, Indicates the events, is the total number of events;

[0015] For each event Constructing a post propagation network graph , the root node of the propagation graph is the event declaration ; It is a set of nodes in the post propagation network graph, where one node represents one tweet; is the edge set of the post propagation network graph, representing the interaction between tweets; the adjacency matrix of the post propagation network graph is ,when When Tweets and There is an interactive relationship between tweets; When Tweets and There is no interactive relationship between tweets;

[0016] For each event Build a user social network graph , is a set of nodes in the user social network graph, where one node represents one user; is the edge set of the user social network graph, representing the social relationship between users; the adjacency matrix of the user social network graph is ,when When User and There is a following relationship between users; when When User and There is no following relationship between users;

[0017] Input the post propagation network graph and the user social network graph into the graph attention network GAT encoder for feature extraction to obtain the post propagation features and user social characteristics ; The graph attention network GAT encoder adopts a two-layer graph attention network GAT encoder. The first layer of the graph attention network GAT encoder performs a preliminary update on the node features, and the second layer of the graph attention network GAT encoder processes the output features of the first layer of the graph attention network GAT encoder.

[0018] Furthermore, step S2 specifically includes:

[0019] For each event Calculating propagation density , the formula is as follows:

[0020] ,

[0021] in, Represents the indicator function, when the time information Belong to time When , the value is 1, otherwise it is 0; Divide by hour; Representation node Belongs to a node collection ;

[0022] According to the event The propagation time and propagation density are calculated to obtain the propagation density map. The sliding window method is used to detect the local maximum and identify the outbreak node. The formula is as follows:

[0023] ,

[0024] in, It is represented as the time point of the local maximum. represents the peak value of fluctuation during the propagation process, Indicates the window size, Indicates the maximum value operation; select the two peaks with the highest propagation density and record them as the first peak time point , the second peak time point ;

[0025] Calculate the slope within each peak neighborhood in the propagation density map. The formula is as follows:

[0026] ,

[0027] in, Represents the propagation density map at time point The density value of Indicates the window length; It represents the slope, which represents the changing trend of the propagation density at that moment;

[0028] From the local maximum point in the propagation density map corresponding to the fluctuation peak during the propagation process To begin, look for slopes of ± The point is taken as the boundary of the outbreak interval, starting from the first peak time point Start by scanning left and right to find the first one that satisfies Time point; determine the boundary of the transmission outbreak interval: starting time: End time: ; Finally, the first peak time point is obtained The corresponding transmission outbreak interval: Similarly, the second peak time point is obtained The corresponding transmission outbreak interval;

[0029] According to the obtained propagation outbreak interval, the nodes and edges that meet the outbreak interval are filtered out from the post propagation network graph to obtain the post propagation subgraph , the post propagation subgraph satisfies the following formula:

[0030] ,

[0031] ,

[0032] in, A set of nodes representing the post propagation subgraph, The edge set representing the post propagation subgraph;

[0033] Through the above process, the first peak time point is extracted from the post propagation network graph Post propagation subgraph and the second peak time point Post propagation subgraph ; Post propagation subgraph and The graph attention network GAT encoder is used to extract features respectively to obtain the corresponding first post propagation sub-features and the second post propagation sub-feature According to the relationship between the post propagation network graph and the user social network graph, the user social subgraph corresponding to each post propagation subgraph is extracted from the user social network graph, and the first user social sub-feature is obtained by using the graph attention network GAT encoder. and the second user's social sub-features .

[0034] Furthermore, step S3 specifically includes:

[0035] Spread the first post sub-feature , Second post propagation sub-feature , first user social sub-features and the second user social sub-feature Perform feature mapping through linear transformation function to obtain the first post propagation mapping sub-feature , Second post propagation mapping sub-feature , first user social mapping sub-features and the second user social mapping sub-feature ;

[0036] The multi-head attention mechanism is used to fuse feature representation. In the multi-head attention mechanism, each attention head learns different attention distributions through independent parameters; for each attention head, the first post propagation mapping sub-feature is calculated and the second post propagates the mapped sub-features The attention weight matrix between , the formula is as follows:

[0037] ,

[0038] in, Indicates the The parameters of the attention head, Softmax represents the normalization function, Represents the scaling factor; similarly, the first user social mapping sub-feature is calculated and the second user social mapping sub-feature The attention weight matrix between ;

[0039] Spread the mapping sub-features for the first post and the second post propagates the mapped sub-features Perform weighted summation to obtain the post propagation fusion sub-features, which are expressed as follows:

[0040] ,

[0041] in, Indicates the The post propagation fusion sub-feature output by the attention head; similarly, the user social fusion sub-feature is obtained In the multi-head attention mechanism, multiple attention heads calculate different attention distributions in parallel, concatenate the outputs of multiple attention heads together, and map them to the final feature space through linear transformation. The formula is as follows:

[0042] ,

[0043] ,

[0044] in, represents the post propagation splicing sub-feature, Represents a splicing operation, represents the first linear transformation matrix, represents the user's social splicing sub-feature, represents the second linear transformation matrix;

[0045] Spread the post and splice sub-features and post propagation characteristics By fusion, we can obtain the post propagation fusion feature, which is expressed as follows:

[0046] ,

[0047] in, Represents the post propagation fusion feature; similarly, we get the user social fusion feature .

[0048] Furthermore, step S4 specifically includes:

[0049] Fusion features propagated through posts and user social integration features Constructing the projection matrix , then use the projection matrix Fusion features of post propagation and user social integration features Perform weighted summation to obtain post propagation interaction features Social interaction features with users , the formula is as follows:

[0050] ,

[0051] ,

[0052] ,

[0053] in, represents the hyperbolic tangent activation function, represents the parameters of the projection matrix, The linear transformation matrix representing the user's social features, The linear transformation matrix representing the propagation characteristics of the post;

[0054] Post propagation interaction features Social interaction features with users Perform attention weight distribution, calculate the importance of each feature dimension through the Softmax function, and perform weighted summation to obtain the first weighted post propagation feature and the first weighted user social features :

[0055] ,

[0056] ,

[0057] in, Represents the Softmax function.

[0058] Furthermore, step S5 specifically includes:

[0059] Propagate features to the first weighted post through the self-attention module and the first weighted user social features Redistribute weights, learn the correlation and dynamic importance of internal features, and obtain the second weighted post propagation feature and the second weighted user social feature , the formula is as follows:

[0060] ,

[0061] ,

[0062] in, Represents the self-attention mechanism;

[0063] Use the gated recurrent unit GRU to and Perform temporal modeling and obtain the final hidden state as , the formula is as follows:

[0064] ,

[0065] in, Represents the gated recurrent unit operation; propagates the first weighted post feature , first weighted user social features and the final hidden state Splice to get the final homogeneous interactive comprehensive representation , the formula is: .

[0066] Furthermore, step S6 specifically includes:

[0067] The final homogeneous interaction is comprehensively represented With event statement Features of original tweets Splice to get the final comprehensive features , the formula is as follows:

[0068] ,

[0069] The predicted probability vector is calculated through the fully connected layer and the Softmax function. The formula is as follows:

[0070] ,

[0071] in, Represents the fully connected layer operation.

[0072] Furthermore, during training, optimization is performed by minimizing the cross entropy loss between the predicted probability and the true label distribution:

[0073] ,

[0074] in, represents the cross entropy loss, is the number of samples in the dataset, represents the number of categories, that is, the total number of categories in the classification task, Indicates that this is Regularization term, representing the model parameters The sum of the squares of is the regularization factor, Representing an event The true label distribution of .

[0075] The advantages of the present invention are:

[0076] The proposed method constructs a post propagation network and a user social network, extracts cross-domain features using a graph attention network (GAT), and uses the slope partitioning method to identify evolving features in key time windows and extract features of high-influence groups. Post propagation features and user social features are fused via a mutual attention mechanism, and combined with heterogeneous and homogeneous interaction modules to capture global dependencies and temporal patterns, achieving accurate rumor detection. Experimental results demonstrate that the proposed method outperforms existing models on multiple real-world social media datasets, particularly after integrating social interaction features, achieving significant improvements in detection accuracy and robustness. Furthermore, the proposed method excels in time complexity and early detection, making it suitable for real-time detection of large-scale social media data. Ablation experiments and hyperparameter sensitivity analysis further validate the effectiveness of the framework, providing optimization guidance for practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention but do not constitute a limitation of the present invention.

[0078] Figure 1 is a flow chart of the steps of the method of the present invention;

[0079] Figure 2 Statistical analysis results for rumors;

[0080] Figure 2 (a) is a statistical analysis chart of rumors on the Twitter15 dataset;

[0081] Figure 2 (b) is a statistical analysis chart of rumors on the Weibo dataset;

[0082] Figure 3 Comparison of early rumor detection results on the Twitter15 dataset;

[0083] Figure 4 Comparison of early rumor detection results on the Twitter16 dataset;

[0084] Figure 5 Comparison of early rumor detection results on the Pheme dataset;

[0085] Figure 6 The ablation experiment results of the method of the present invention are as follows;

[0086] Figure 7 The number of outbreak segments ( ) settings on the impact of the method of the present invention;

[0087] Figure 8 is the ratio of deleted edges ( ) settings on the impact of the method of the present invention;

[0088] Figure 9 is the slope threshold ( ) settings on the impact of the method of the present invention. DETAILED DESCRIPTION

[0089] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0090] Example 1

[0091] In this embodiment, Figure 1 As shown, the present invention provides a rumor detection method based on dual-domain perception structure feature fusion learning, and the specific steps include:

[0092] S1. Obtain a social media message event dataset, build a post propagation network graph and a user social network graph based on the dataset, and extract post propagation features and user social features respectively through a graph attention network encoder.

[0093] Specifically, the social media message event dataset ,in, Indicates the events, is the total number of events;

[0094] For each event Constructing a post propagation network graph , the root node of the propagation graph is the event declaration ; It is a set of nodes in the post propagation network graph, where one node represents one tweet; is the edge set of the post propagation network graph, representing the interaction between tweets; the adjacency matrix of the post propagation network graph is ,when When Tweets and There is an interactive relationship between tweets; When Tweets and There is no interactive relationship between tweets;

[0095] For each event Build a user social network graph , is a set of nodes in the user social network graph, where one node represents one user; is the edge set of the user social network graph, representing the social relationship between users; the adjacency matrix of the user social network graph is ,when When User and There is a following relationship between users; when When User and There is no following relationship between users;

[0096] Input the post propagation network graph and the user social network graph into the graph attention network GAT encoder for feature extraction to obtain the post propagation features and user social characteristics ; The graph attention network GAT encoder adopts a two-layer graph attention network GAT encoder. The first layer of the graph attention network GAT encoder performs a preliminary update on the node features, and the second layer of the graph attention network GAT encoder processes the output features of the first layer of the graph attention network GAT encoder.

[0097] S2. Use the propagation density peak and slope partitioning method to identify the key time window for rumor spread, and extract relevant substructure features from the post propagation network graph and user social network graph.

[0098] Specifically, the spread of rumors is considered as a dynamic process with time series characteristics. In order to identify the outbreak interval in the spread process, it is first necessary to divide the spread behavior of rumor events into time periods and calculate their spread density at different times. ; For events , propagation density Indicates that during this time period, from the propagation node set The number of dissemination behaviors (such as comments and forwarding) that users participated in. The entire dissemination process is divided according to the timestamp attached to each dissemination record in the original dataset;

[0099] For each event Calculating propagation density , the formula is as follows:

[0100] ,

[0101] in, Represents the indicator function, when the time information Belong to time When , the value is 1, otherwise it is 0; Divide by hour; Representation node Belongs to a node collection ;

[0102] According to the event The propagation time and propagation density are calculated to obtain the propagation density map. The sliding window method is used to detect the local maximum and identify the outbreak node. The formula is as follows:

[0103] ,

[0104] in, It is represented as the time point of the local maximum. represents the peak value of fluctuation during the propagation process, Indicates the window size, Indicates the maximum value operation; To ensure the analysis of local transmission outbreaks, this study is based on Select the two peaks with the highest propagation density and record them as the first peak time point , the second peak time point ;

[0105] Calculate the slope within each peak neighborhood in the propagation density map. The formula is as follows:

[0106] ,

[0107] in, Represents the propagation density map at time point The density value of Indicates the window length; It represents the slope, which represents the changing trend of the propagation density at that moment. The slope is used to assist in finding the boundaries of intervals where the density changes dramatically.

[0108] From the local maximum point in the propagation density map corresponding to the fluctuation peak during the propagation process To begin, look for slopes of ± The point is taken as the boundary of the outbreak interval, starting from the first peak time point Start by scanning left and right to find the first one that satisfies Time point; determine the boundary of the transmission outbreak interval: starting time: End time: ; Finally, the first peak time point is obtained The corresponding transmission outbreak interval: ;

[0109] This time period is considered to be the key period when the transmission intensity rises or falls sharply during the event propagation process, which provides a clear time boundary basis for the subsequent transmission substructure feature extraction. Similarly, the second peak time point is obtained. The corresponding transmission outbreak interval;

[0110] According to the obtained propagation outbreak interval, the nodes and edges that meet the outbreak interval are filtered out from the post propagation network graph to obtain the post propagation subgraph , the post propagation subgraph satisfies the following formula:

[0111] ,

[0112] ,

[0113] in, A set of nodes representing the post propagation subgraph, The edge set representing the post propagation subgraph; the propagation subgraph selected through this time window is the substructure of nodes and edges with the highest propagation density and information volume during this time period, which is the basis for subsequent structural feature extraction.

[0114] Through the above process, the first peak time point is extracted from the post propagation network graph Post propagation subgraph and the second peak time point Post propagation subgraph ; Post propagation subgraph and The graph attention network GAT encoder is used to extract features respectively to obtain the corresponding first post propagation sub-features and the second post propagation sub-feature According to the relationship between the post propagation network graph and the user social network graph, the user social subgraph corresponding to each post propagation subgraph is extracted from the user social network graph, and the first user social sub-feature is obtained by using the graph attention network GAT encoder. and the second user's social sub-features .

[0115] S3. The post propagation sub-features and user social sub-features are processed through the mutual attention mechanism to obtain the post propagation fusion feature and user social fusion feature.

[0116] Specifically, the first post propagation sub-feature , Second post propagation sub-feature , first user social sub-features and the second user social sub-feature Perform feature mapping through linear transformation function to obtain the first post propagation mapping sub-feature , Second post propagation mapping sub-feature , first user social mapping sub-features and the second user social mapping sub-feature ; This allows feature vectors from different sources to be mapped into a feature space of the same dimension, facilitating subsequent learning of the mutual attention mechanism. The formula is as follows:

[0117] ,

[0118] ,

[0119] ,

[0120] ,

[0121] in, , They represent the learned post propagation sub-feature mapping weight matrix and user social sub-feature mapping weight matrix respectively, , They represent the bias items of post propagation sub-feature and user social sub-feature respectively; Represents a matrix multiplication operation; ensure that all mapped features , , , The dimensions are consistent and mapped into a unified feature space. This unified feature space provides a feature alignment basis for the subsequent mutual attention mechanism between heterogeneous structures, ensuring that structural features from different sources can be associated modeled and fused in the same semantic space.

[0122] The multi-head attention mechanism is used to fuse feature representation. In the multi-head attention mechanism, each attention head learns different attention distributions through independent parameters; for each attention head, the first post propagation mapping sub-feature is calculated and the second post propagates the mapped sub-features The attention weight matrix between , the formula is as follows:

[0123] ,

[0124] in, Indicates the The parameters of the attention head, Softmax represents the normalization function, Represents the scaling factor; similarly, the first user social mapping sub-feature is calculated and the second user social mapping sub-feature The attention weight matrix between ;

[0125] Spread the mapping sub-features for the first post and the second post propagates the mapped sub-features Perform weighted summation to obtain the post propagation fusion sub-features, which are expressed as follows:

[0126] ,

[0127] in, Indicates the The post propagation fusion sub-feature output by the attention head; similarly, the user social fusion sub-feature is obtained In the multi-head attention mechanism, multiple attention heads calculate different attention distributions in parallel, concatenate the outputs of multiple attention heads together, and map them to the final feature space through linear transformation. The formula is as follows:

[0128] ,

[0129] ,

[0130] in, represents the post propagation splicing sub-feature, Represents a splicing operation, represents the first linear transformation matrix, represents the user's social splicing sub-feature, Represents the second linear transformation matrix; in the above process, and are matrices used to linearly transform the concatenated post propagation sub-features and user social sub-features. Specifically, these linear transformation matrices are obtained by optimizing the backpropagation algorithm during training. In the multi-head attention mechanism, the dimension of the output of the attention head may increase after the concatenation operation, so it is necessary to use the linear transformation matrix and The concatenated features are mapped to the target feature space to ensure consistency of feature dimensions. These matrices are typically initialized with random values and gradually adjusted during training as the model is updated to learn the optimal feature representation. The purpose of the linear transformation is to transform the concatenated feature vectors into a feature space more suitable for subsequent model processing, further improving the model's expressive power.

[0131] Spread the post and splice sub-features and post propagation characteristics By fusion, we can obtain the post propagation fusion feature, which is expressed as follows:

[0132] ,

[0133] in, Represents the post propagation fusion feature; similarly, we get the user social fusion feature .

[0134] S4. Construct a projection matrix through the post propagation fusion features and the user social fusion features, and then use the projection matrix to perform weighted summation on the post propagation fusion features and the user social fusion features to obtain the post propagation interaction features and the user social interaction features, and then generate the first weighted post propagation features and the first weighted user social features.

[0135] Specifically, the fusion features are propagated through posts and user social integration features Constructing the projection matrix , then use the projection matrix Fusion features of post propagation and user social integration features Perform weighted summation to obtain post propagation interaction features Social interaction features with users , the formula is as follows:

[0136] ,

[0137] ,

[0138] ,

[0139] in, Represents the hyperbolic tangent activation function, which is usually used to map the eigenvalues after linear transformation into a range to enhance nonlinear expression capabilities. Parameters representing the projection matrix used to connect post propagation features and user social characteristics , generate the projection matrix . This is a learned parameter matrix. The linear transformation matrix representing the user's social features is used to integrate the user's social fusion features. Perform the transformation, The linear transformation matrix representing the post propagation features is used to fuse the post propagation features. Perform transformations;

[0140] Post propagation interaction features Social interaction features with users Perform attention weight distribution, calculate the importance of each feature dimension through the Softmax function, and perform weighted summation to obtain the first weighted post propagation feature and the first weighted user social features :

[0141] ,

[0142] ,

[0143] in, Represents the Softmax function. It should be noted that the Softmax here is used to distribute attention weights across dimensions. Although similar in form to the self-attention mechanism, it essentially models heterogeneous interactions between structures and differs from the subsequent homogeneous attention mechanism.

[0144] S5. Model the homogeneous interaction information of the first weighted post propagation feature and the first weighted user social feature to obtain the final homogeneous interaction comprehensive representation.

[0145] Specifically, in order to further explore the dynamic evolution characteristics of the post propagation structure and the user social structure, the present invention performs homogeneity modeling on the weighted features obtained in step S4. First, the self-attention mechanism is introduced to capture the importance distribution within the feature, thereby obtaining a more representative second weighted feature; the first weighted post propagation feature is modeled by the self-attention module. and the first weighted user social features Redistribute weights, learn the correlation and dynamic importance of internal features, and obtain the second weighted post propagation feature and the second weighted user social feature , the formula is as follows:

[0146] ,

[0147] ,

[0148] in, Represents the self-attention mechanism;

[0149] Use the gated recurrent unit GRU to and Perform temporal modeling and obtain the final hidden state as , the formula is as follows:

[0150] ,

[0151] in, Represents the gated recurrent unit operation; propagates the first weighted post feature , first weighted user social features and the final hidden state Splice to get the final homogeneous interactive comprehensive representation , the formula is: ;in, The final representation that comprehensively integrates the dynamic evolution features and the original structural features serves as the input for subsequent heterogeneous interactive fusion.

[0152] S6. After concatenating the final homogeneous interaction comprehensive representation with the original tweet features, the feature is input into the rumor detection module for classification to obtain the rumor detection results.

[0153] Specifically, the final homogeneous interaction synthesis representation With event statement Original tweet characteristics Splice to get the final comprehensive features , the formula is as follows:

[0154] ,

[0155] The predicted probability vector is calculated through the fully connected layer and the Softmax function. The formula is as follows:

[0156] ,

[0157] in, Represents the fully connected layer operation.

[0158] Specifically, during training, optimization is performed by minimizing the cross entropy loss between the predicted probability and the true label distribution:

[0159] ,

[0160] in, represents the cross entropy loss, is the number of samples in the dataset, represents the number of categories, that is, the total number of categories in the classification task, Indicates that this is Regularization term, representing the model parameters The sum of the squares of is the regularization factor, Representing an event The true label distribution of .

[0161] Example 2

[0162] This embodiment carries out experimental analysis on the method of the present invention:

[0163] 1. Dataset introduction and statistical analysis

[0164] (1) Dataset Introduction

[0165] In the experiment, three public social media datasets were used to evaluate the performance of the proposed method. The datasets include Twitter15 dataset, Twitter16 dataset, and Pheme dataset. The statistical analysis of the datasets is shown in Table 1.

[0166] Table 1 Statistical analysis of the dataset

[0167]

[0168] (2) Statistical analysis

[0169] This method extracts outbreak segments based on the fluctuating characteristics of rumor propagation, thereby mining dense feature intervals in information dissemination. To verify the feasibility of this method, this paper conducted a statistical analysis on two representative datasets: Twitter15, a typical English social media dataset, and Weibo, a representative Chinese social platform. The two datasets are complementary in terms of language structure and user behavior, and have good experimental representativeness and promotional value. The analysis results are as follows: Figure 2 As shown, Figure 2 (a) is a statistical analysis chart of rumors on the Twitter15 dataset. Figure 2 (b) is a statistical analysis chart of rumors on the Weibo dataset; the horizontal axis represents the time of rumor propagation, and the vertical axis represents the number of comments or reposts, i.e., the propagation density. As can be observed from the figure, the curve of rumor events changing over time exhibits significant volatility, showing characteristic peak fluctuations. These fluctuations not only manifest as rises and falls over time, but also exhibit a certain regularity of rise and fall within different time periods. This pattern is fully verified by the statistical results of both the Twitter15 and Weibo datasets. Specifically, the propagation density rises sharply in certain time periods, forming a short-lived peak, and then gradually decreases over time. This phenomenon reflects the active interaction of users during the information dissemination process and the volatility of information dissemination intensity.

[0170] Fluctuation analysis of information density can identify outbreaks of information dissemination. Peaks in information density correspond to peaks in user interaction, followed by a rapid decline, reflecting the sudden nature of information dissemination. This fluctuation demonstrates that information diffusion does not grow linearly, but rather exhibits periodic bursts influenced by user attention and dissemination mechanisms. Defining outbreak density intervals helps locate key nodes of rumor spread, extract precise temporal features, and further reveal the dynamic patterns of information dissemination. Combining analysis of user interaction behavior and dissemination paths can enhance the ability to detect critical moments, improve detection accuracy and adaptability, and improve response speed and early warning capabilities to outbreaks, especially in dynamic social environments.

[0171] 2. Evaluation indicators

[0172] This paper uses accuracy and F1 score as core evaluation criteria to comprehensively measure the overall performance of the method in rumor detection tasks. The advantage of the F1 score lies in its comprehensive consideration of classification precision and recall, ensuring that the model predicts positive samples with high accuracy while also covering as many positive instances as possible. In rumor detection tasks, it effectively compensates for potential bias when using accuracy alone, enabling the model to maintain good discrimination capabilities even when faced with class imbalance. In experiments, a comprehensive analysis of accuracy and F1 score helps comprehensively measure model performance and provides a reference for algorithm optimization.

[0173] 3. Experimental Setup

[0174] The default optimization configuration of each comparison method was adopted, implemented in the PyTorch framework, and Adam optimization was selected. The present invention uses the processed user social and post interaction dataset and constructs a propagation relationship graph. The specific experimental parameters are as follows: the learning rate is set to 0.001, the batch size is 128, and the dropout rate is 0.2. The text nodes are initialized using TF-IDF (dimension 5000), and the user features use the standard initial vector. The output dimension of the graph attention network (GAT) and the fusion gated unit is 64, and the number of heads of the multi-head attention is 4. Select contains 2 substructures, slope threshold Set to 0.4, 0.4 and 0.5; remove side scale The edge deletion ratio q is used as a regularization and data augmentation strategy during graph representation learning, perturbing the graph structure during training to improve the model's generalization, prevent overfitting, and enhance robustness to noisy edges. This operation constructs a diverse view of the graph structure by randomly deleting edges at a ratio of q, helping the model learn more stable and discriminative representations. An early stopping strategy is used during training, stopping training when the validation set loss shows no decrease for 10 consecutive rounds. Five-fold cross-validation is performed to enhance experimental robustness. Finally, the optimal parameters obtained from the validation set are used to evaluate test set performance.

[0175] 4. Comparison method

[0176] The method of the present invention is compared with several advanced comparative models, including:

[0177] DTC model: A detection method based on a decision tree classifier that relies on manually constructed text features for classification and discrimination.

[0178] RFC model: A detection method based on random forest classifier, which performs detection by fusing user temporal behavior and event correlation features.

[0179] SVM-TK model: A detection method based on linear support vector machines (SVM) that improves discrimination performance by constructing propagation time series features.

[0180] RvNN model: A detection method based on recurrent neural networks (RNNs) that optimizes classification performance by modeling the temporal dependencies during post propagation.

[0181] PLAN model: A rumor propagation tree embedding learning method based on hierarchical self-attention mechanism, which captures the propagation rules by analyzing the cross-level interaction relationships between nodes.

[0182] UPFD model: An end-to-end rumor detection method based on modeling endogenous user preferences and exogenous social context, which achieves detection by integrating individual behavioral characteristics and social environment factors.

[0183] BiGCN model: A detection method based on GCN that enhances the modeling ability of information diffusion characteristics by learning the bidirectional information flow pattern of the rumor propagation tree.

[0184] GCAN model: A detection method based on multi-source heterogeneous information fusion, which achieves detection by integrating text semantics, user profiles and network topology features for graph convolution learning.

[0185] DYNGCN model: A dynamic graph convolutional network detection method based on temporal snapshot partitioning, which performs detection by learning propagation structure evolution in stages.

[0186] DGNF model: A detection method based on the GAT-Transformer hybrid architecture that uses a self-attention mechanism to capture long-range dependencies and local interactions in the propagation sequence for detection.

[0187] DECL model: A rumor detection method based on contrastive learning of time-snapshot dynamic graphs, which captures the evolution of rumors by modeling structural differences across time periods.

[0188] RDMSC model: A joint detection method based on the extraction of highly homogeneous social circle features and the fusion of social interaction information, which optimizes the discrimination performance by integrating multi-source heterogeneous features.

[0189] CoAHRD model: A hybrid detection method based on content, context, user features and co-attention mechanism, achieving fine-grained classification through dynamic weight allocation.

[0190] These comparison models are highly representative, covering different types of rumor detection methods. They can effectively compare the performance with the method of the present invention, and verify the advantages and applicability of the rumor detection task from multiple dimensions.

[0191] 5. Experimental results and analysis

[0192] The present invention will evaluate the performance of the rumor detection framework proposed in the present invention from five aspects, including rumor comparison experiments, early detection capabilities, time complexity analysis, ablation experiments, and hyperparameter sensitivity analysis. First, by comparing with existing rumor detection models, the advantages of the method of the present invention in terms of overall detection performance, early rumor recognition capabilities, and computational efficiency are comprehensively analyzed. Secondly, ablation experiments are used to verify the role of each module in the method of the present invention, and their contribution to the final detection effect is analyzed to prove their necessity. Finally, hyperparameter sensitivity analysis explores the impact of different hyperparameter configurations on detection accuracy and examines the adaptability of the method of the present invention under changes in key parameters.

[0193] (1) Comparative experimental results and analysis

[0194] Tables 2, 3, and 4 show the experimental results of the proposed method on the Twitter15 dataset, Twitter16 dataset, and Pheme dataset, respectively. Comparative analysis shows that the proposed method outperforms existing methods on all three datasets, with accuracy improvements of 1.4%, 1%, and 1.5%, respectively. This demonstrates that integrating user structure information with a joint modeling strategy can effectively improve performance in rumor detection tasks. Furthermore, the experimental results further validate the adaptability and robustness of the proposed method framework in diverse task environments. The detailed analysis is as follows:

[0195] Traditional machine learning methods performed poorly. Experimental results show that the DTC model, RFC model, and SVM-TS model performed poorly, demonstrating the limitations of traditional methods using manual features in identifying rumors. Traditional methods rely on manual features, are susceptible to subjective bias, and struggle to capture the complex characteristics of rumor propagation, resulting in low accuracy.

[0196] Deep learning methods outperform traditional methods. Deep learning methods (RvNN) significantly outperform traditional methods in detection tasks. This is primarily due to their ability to autonomously learn implicit patterns and complex features in the data. Traditional detection methods based on handcrafted features often struggle to fully capture this information, limiting improvements in detection capabilities.

[0197] Advantages of user-text joint detection. Models such as the RDSML model and the CoAHRD model that combine user behavior and text information are better than models that rely solely on text information in terms of detection performance, which shows that integrating user social information and text content can more comprehensively capture the characteristics of rumor propagation. However, these methods fail to fully consider the dynamic volatility in the process of rumor propagation, especially the information-intensive features within the outbreak interval and the influence of noise during the propagation process. In contrast, the method of the present invention can more effectively utilize this key information by focusing on mining the dense features of the outbreak interval, thereby improving the accuracy of rumor detection.

[0198] The advantages of the proposed method framework. The proposed method significantly outperforms all compared models in both the four-class classification task on the Twitter15 and Twitter16 datasets and the two-class classification task on Pheme. It considers the dynamic volatility of rumor propagation, exploits dense features within outbreak intervals, and combines joint features of users and comments to more comprehensively and accurately capture various deceptive signals. Compared with traditional methods, the proposed method is better able to handle the complex temporal features and user interaction characteristics of rumor propagation, thereby improving overall detection performance.

[0199] There are significant differences in the detection performance of the Twitter15 dataset, Twitter16 dataset, and Pheme dataset. From the experimental results, the detection performance of the Pheme dataset is significantly lower than that of the Twitter15 dataset and Twitter16 dataset, which can be mainly attributed to two key factors. First, the topic overlap is high. The Pheme dataset only contains 5 breaking news events with highly overlapping content, which causes the framework to be more inclined to identify topic categories rather than authenticity. In contrast, the statements in the Twitter15 dataset and Twitter16 dataset are more independent and have diverse information sources, which can more effectively distinguish true and false information. Secondly, the amount of available information is limited. The average number of posts per statement in the Pheme dataset is only 26, which is far less than that of the Twitter15 dataset and Twitter16 dataset, resulting in less information available for training, limiting the learning of the propagation pattern and contextual features by the method of the present invention, and affecting the detection effect.

[0200] In summary, the Pheme dataset suffers from topic overlap and limited information, resulting in relatively low detection accuracy and hindering the performance of our method. However, the Twitter15 and Twitter16 datasets, with their more independent topics and richer propagation information, enable our method to fully learn the spread patterns of rumors, ultimately achieving superior detection results. This demonstrates that our method possesses greater adaptability and detection capabilities in diverse and informative data environments.

[0201] Table 2 Rumor detection results of Twitter15 dataset

[0202]

[0203] Table 3. Rumor detection results of Twitter16 dataset

[0204]

[0205] Table 4 Pheme rumor detection results

[0206]

[0207] (2) Early rumor detection results and analysis

[0208] Early detection of rumors is crucial to curbing their spread and mitigating their negative social impact. Therefore, evaluating the performance of the method of the present invention at this stage becomes a key indicator for measuring its effectiveness. In order to evaluate the effectiveness of early detection, the present invention compared different methods based on time. The results are as follows: Figure 3 、 Figure 4 、 Figure 5 As shown. Figure 3 It can be seen that as the time segmentation increases, the detection performance of each method gradually improves. It is worth noting that in the early stage of rumor dissemination, the accuracy of the method of the present invention is significantly better than that of other methods, and its advantages become more prominent over time. This shows that the method of the present invention can effectively capture key features in the early stage of propagation and improve detection accuracy. The main reason is that the method of the present invention fully utilizes the dense features of users and text for joint modeling, and can select representative feature intervals when the information increases, avoiding noise interference, thereby improving accuracy.

[0209] The method of the present invention outperforms other comparative methods in early detection, demonstrating that even when there is little dissemination information, it can still effectively detect rumors by relying on the background knowledge of the source tweet. Furthermore, when there is abundant dissemination information, the method of the present invention can combine background knowledge with dissemination information to further improve detection accuracy. This fully demonstrates the advantages of the method of the present invention in early rumor detection. In summary, the method of the present invention demonstrates significant advantages in the task of early rumor detection, and can dynamically adjust its strategy based on the dissemination environment, significantly improving detection accuracy.

[0210] (3) Time complexity results and analysis

[0211] To evaluate the computational time consumption of our method, we selected several well-performing comparison models (Bi-GCN, DYNGCN, GCAN, DECL, REMSC, and CoAHRD). We recorded the average inference time (in milliseconds) for each model predicting a single sample label in each test set. The experimental results are shown in Table 5.

[0212] Table 5 Comparison of calculation time results of the method of the present invention

[0213]

[0214] Experimental results show that while the method of the present invention improves the accuracy of rumor detection, although the inference time is slightly increased, it still maintains reasonable computational efficiency. The main reason for this is that the method of the present invention enhances the ability to capture complex features of different propagation stages by jointly modeling user behavior and text information and combining a multi-level attention mechanism for feature fusion. However, this efficient information integration also brings additional computational burden. To optimize time complexity, the present invention uses dense features of the outbreak interval for detection during design, rather than modeling the entire propagation process, thereby reducing redundant computational costs. This optimization measure effectively reduces computational complexity, making the inference time of the method of the present invention on large-scale data sets similar to that of other comparison models, maintaining strong application competitiveness.

[0215] In summary, while the inference time of our method on certain datasets is slightly longer than that of other models, this shortcoming is offset by significantly improved detection accuracy. After an optimized design, our method is comparable in computational efficiency to other methods and can provide highly effective rumor detection services in practical applications, demonstrating its strong potential for application.

[0216] (4) Ablation Experiment Results and Analysis

[0217] In order to further analyze the effectiveness of each component of the method of the present invention, a comparison is conducted with different variants. The specific ablation experiment includes the following four parts:

[0218] -U: Remove the user social network graph from method S1 and only use the post propagation network graph as input, while keeping the other parts unchanged. This experiment aims to verify the role of user structure in the framework.

[0219] -A: Remove the global feature P in method S3 and only retain the local features. The purpose is to verify the effectiveness of the global feature in the framework.

[0220] -D: Remove the heterogeneous interaction part in step S4 of the method, and directly perform subsequent operations on the post propagation structure features and user structure features without considering the fine-grained interaction features between the two, aiming to verify the contribution of fine-grained feature fusion.

[0221] -S: Remove the homogeneous interaction part in step S5 of the method, and add the weighted heterogeneous interaction feature posts and user structure Simple splicing ignores the previous and next correlation information and redundant noise in the temporal features, and mainly verifies the contribution of the selective fusion of temporal features to the overall framework.

[0222] The ablation experiment results are as follows Figure 6 The figure shows the performance changes of the proposed method on the Twitter15, Twitter16, and Pheme datasets. The experiment aims to analyze the impact of each module on the detection effect and further verify the rationality of the proposed method design.

[0223] Experimental results for "-U" show that removing this module reduces the accuracy of our method by 5.6%, 5.4%, and 3.6% on the Twitter15, Twitter16, and Pheme datasets, respectively. This demonstrates the crucial importance of user structure for rumor detection. User behavior information provides an effective supplement when sufficient dissemination data is lacking. Our method effectively improves detection performance by jointly learning the dual-channel interaction features of posts and users. Therefore, introducing user structure significantly enhances the robustness and accuracy of our method.

[0224] Experimental results for the "-A" module show that removing this module reduces the accuracy of our method by 1.2%, 1.9%, and 1.5%, respectively, validating the necessity of global feature fusion. Global features provide richer contextual information, which, when combined with local features, can improve detection accuracy. Removing this module results in the loss of some critical contextual information, thus impacting overall performance.

[0225] Experiments with the "-D" module show that removing this module reduces the accuracy of our method by 2.2%, 2.6%, and 2.7%, respectively. This result demonstrates that the heterogeneous interaction module plays a key role in improving detection accuracy. By learning the fine-grained heterogeneous interactions between post propagation structure and user structural features, it effectively captures the complex relationships between the two, thereby enhancing our understanding and identification of rumor propagation patterns. Without this module, detection is performed solely through concatenated information, resulting in decreased performance.

[0226] Regarding the "-S" flag, its removal resulted in a 0.7%, 1.5%, and 0.6% decrease in accuracy for our method on the Twitter15, Twitter16, and Pheme datasets, respectively. The homogeneous interaction module, through a self-attention mechanism, effectively reduces interference from redundant and irrelevant information while optimizing temporal feature extraction. Experiments show that removing this module weakens the model's ability to selectively learn temporal information, thereby impacting overall detection performance.

[0227] In summary, each module of the proposed method plays a key role in improving overall performance. The user structure module provides additional behavioral information, global feature fusion enhances contextual understanding, the heterogeneous interaction module fine-grainedly models the complex relationships between posts and users, and the homogeneous interaction module optimizes temporal feature extraction and suppresses noise. The synergistic effect of these modules enables the proposed method to perform superiorly in rumor detection tasks, with enhanced generalization and stability.

[0228] (5) Experimental parameter results and analysis

[0229] In order to further optimize the performance of the method of the present invention, the present invention conducted a hyperparameter analysis experiment, focusing on evaluating the impact of the following key parameters on the effect of the method of the present invention:

[0230] Number of outbreak segments (b): This method sets b to 1, 2, and 3 to explore the impact of different numbers of outbreak segments on detection effectiveness. The outbreak segment divisions identify key periods in the rumor spread process, and different numbers of segments may affect the method's ability to capture the dynamics of the spread.

[0231] Edge deletion ratio (q): This parameter controls the ratio of edges deleted from the graph, which is used to simulate noise and data incompleteness in information propagation. In the experiment, the robustness under different noise conditions was tested by adjusting the value of q.

[0232] Slope threshold (k): The slope threshold k is used to distinguish different outbreak intervals during the post propagation process. A reasonable slope threshold can more accurately identify the peak period of rumor propagation and thus more accurately identify the key propagation period.

[0233] In this experiment, different settings were made for b, q, and k, and the results were evaluated on multiple datasets. Figure 7 、 Figure 8 、 Figure 9 Based on the experimental results shown in , the present invention can draw the following conclusions:

[0234] Effect of b: When b = 1, the detection accuracy of the method of the present invention is low, which may be because it relies only on a single outbreak segment and fails to fully capture the dynamic characteristics of rumor propagation. As b increases, the detection effect reaches the best when b = 2, indicating that moderate division helps to accurately depict the propagation dynamics. The two outbreak segments can cover most of the key information, thereby improving detection performance. However, when b = 3, although the detection accuracy is slightly improved, the increase is limited, and too many segments may introduce redundant information, increase the computational burden, and affect detection efficiency. Considering the detection accuracy and computational cost, b = 2 is the best choice.

[0235] The impact of q: Appropriate edge removal can remove noise and enhance robustness. Experiments show that q reaches its highest values when it is 0.2, 0.3, and 0.3, respectively, indicating that moderate edge removal helps extract key information. However, excessive edge removal can lead to information loss and reduce detection accuracy. Therefore, choosing a reasonable q is key to improving performance.

[0236] The impact of k: k influences burst segmentation and detection performance. Experimental results show that k values of 0.4, 0.4, and 0.5 optimally identify propagation peaks. However, when k is too high, the segmentation criteria become more stringent, retaining only the steepest propagation phases. While this helps filter weak propagation signals, it can result in the loss of some key propagation modes, thus affecting detection accuracy. Therefore, k = 0.4 and k = 0.5 achieve an optimal balance between propagation mode capture and information integrity, providing optimal performance.

[0237] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A rumor detection method based on dual-domain perception structural feature fusion learning, characterized by: The following steps are involved: S1. Obtain a social media message event dataset, construct a post propagation network graph and a user social network graph based on the dataset, and extract post propagation features and user social features respectively through a graph attention network encoder; S2. Use the peak and slope partitioning method of propagation density to identify the key time window for rumor spread and extract relevant substructure features from the post propagation network graph and user social network graph; S3. Process the post propagation sub-features and user social sub-features through the mutual attention mechanism to obtain the post propagation fusion feature and the user social fusion feature; S4. Constructing a projection matrix using the post propagation fusion features and the user social fusion features. Then, using the projection matrix, performing a weighted summation of the post propagation fusion features and the user social fusion features to obtain the post propagation interaction features and the user social interaction features, thereby generating a first weighted post propagation feature and a first weighted user social feature. S5. Modeling homogeneous interaction information based on the first weighted post propagation feature and the first weighted user social feature to obtain a final homogeneous interaction comprehensive representation; S6. After concatenating the final homogeneous interaction comprehensive representation with the original tweet features, the feature is input into the rumor detection module for classification to obtain the rumor detection results.

2. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 1 is characterized in that: Step S1 specifically includes: The social media message event dataset ,in, Indicates the events, is the total number of events; For each event Constructing a post propagation network graph , the root node of the propagation graph is the event declaration ; It is a set of nodes in the post propagation network graph, where one node represents one tweet; is the edge set of the post propagation network graph, representing the interaction between tweets; the adjacency matrix of the post propagation network graph is ,when When Tweets and There is an interactive relationship between tweets; When Tweets and There is no interactive relationship between tweets; For each event Build a user social network graph , is a set of nodes in the user social network graph, where one node represents one user; is the edge set of the user social network graph, representing the social relationship between users; the adjacency matrix of the user social network graph is ,when When User and There is a following relationship between users; when When User and There is no following relationship between users; Input the post propagation network graph and the user social network graph into the graph attention network GAT encoder for feature extraction to obtain the post propagation features and user social characteristics ; The graph attention network GAT encoder adopts a two-layer graph attention network GAT encoder. The first layer of the graph attention network GAT encoder performs a preliminary update on the node features, and the second layer of the graph attention network GAT encoder processes the output features of the first layer of the graph attention network GAT encoder.

3. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 2 is characterized in that: Step S2 specifically includes: For each event Calculating propagation density , the formula is as follows: ， in, Represents the indicator function, when the time information Belong to time When , the value is 1, otherwise it is 0; Divide by hour; Representation node Belongs to a node collection ; According to the event The propagation time and density are calculated to obtain the propagation density map. The sliding window method is used to detect the local maximum and identify the outbreak node. The formula is as follows: ， in, It is represented as the time point of the local maximum. represents the peak value of fluctuation during the propagation process, Indicates the window size, Indicates the maximum value operation; select the two peaks with the highest propagation density and record them as the first peak time point , the second peak time point ; Calculate the slope within each peak neighborhood in the propagation density map. The formula is as follows: ， in, Represents the propagation density map at time point The density value of Indicates the window length; It represents the slope, which represents the changing trend of the propagation density at that moment; From the local maximum point in the propagation density map corresponding to the fluctuation peak during the propagation process To begin, look for slopes of ± The point is taken as the boundary of the outbreak interval, starting from the first peak time point Start by scanning left and right to find the first one that satisfies Time point; determine the boundary of the transmission outbreak interval: starting time: End time: ; Finally, the first peak time point is obtained The corresponding transmission outbreak interval: Similarly, the second peak time point is obtained The corresponding transmission outbreak interval; According to the obtained propagation outbreak interval, the nodes and edges that meet the outbreak interval are filtered out from the post propagation network graph to obtain the post propagation subgraph , the post propagation subgraph satisfies the following formula: ， ， in, A set of nodes representing the post propagation subgraph, The edge set representing the post propagation subgraph; Through the above process, the first peak time point is extracted from the post propagation network graph Post propagation subgraph and the second peak time point Post propagation subgraph ; Post propagation subgraph and The graph attention network GAT encoder is used to extract features respectively to obtain the corresponding first post propagation sub-features and the second post propagation sub-feature According to the relationship between the post propagation network graph and the user social network graph, the user social subgraph corresponding to each post propagation subgraph is extracted from the user social network graph, and the first user social sub-feature is obtained by using the graph attention network GAT encoder. and the second user's social sub-features .

4. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 3 is characterized in that: Step S3 specifically includes: Spread the first post sub-feature , Second post propagation sub-feature , first user social sub-features and the second user social sub-feature Perform feature mapping through linear transformation function to obtain the first post propagation mapping sub-feature , Second post propagation mapping sub-feature , first user social mapping sub-features and the second user social mapping sub-feature ; The multi-head attention mechanism is used to fuse feature representation. In the multi-head attention mechanism, each attention head learns different attention distributions through independent parameters; for each attention head, the first post propagation mapping sub-feature is calculated and the second post propagates the mapped sub-features The attention weight matrix between , the formula is as follows: ， in, Indicates the The parameters of the attention head, Softmax represents the normalization function, Represents the scaling factor; similarly, the first user social mapping sub-feature is calculated and the second user social mapping sub-feature The attention weight matrix between ; Spread the mapping sub-features for the first post and the second post propagates the mapped sub-features Perform weighted summation to obtain the post propagation fusion sub-features, which are expressed as follows: ， in, Indicates the The post propagation fusion sub-feature output by the attention head; similarly, the user social fusion sub-feature is obtained In the multi-head attention mechanism, multiple attention heads calculate different attention distributions in parallel, concatenate the outputs of multiple attention heads together, and map them to the final feature space through linear transformation. The formula is as follows: ， ， in, represents the post propagation splicing sub-feature, Represents a splicing operation, represents the first linear transformation matrix, represents the user's social splicing sub-feature, represents the second linear transformation matrix; Spread the post and splice sub-features and post propagation characteristics By fusion, we can obtain the post propagation fusion feature, which is expressed as follows: ， in, Represents the post propagation fusion feature; similarly, we get the user social fusion feature .

5. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 4 is characterized in that: Step S4 specifically includes: Fusion features propagated through posts and user social integration features Constructing the projection matrix , then use the projection matrix Fusion features of post propagation and user social integration features Perform weighted summation to obtain post propagation interaction features Social interaction features with users , the formula is as follows: ， ， ， in, represents the hyperbolic tangent activation function, represents the parameters of the projection matrix, The linear transformation matrix representing the user's social features, The linear transformation matrix representing the propagation characteristics of the post; Post propagation interaction features Social interaction features with users Perform attention weight distribution, calculate the importance of each feature dimension through the Softmax function, and perform weighted summation to obtain the first weighted post propagation feature and the first weighted user social features : ， ， in, Represents the Softmax function.

6. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 5 is characterized in that: Step S5 specifically includes: Propagate features to the first weighted post through the self-attention module and the first weighted user social features Redistribute weights, learn the correlation and dynamic importance of internal features, and obtain the second weighted post propagation feature and the second weighted user social feature , the formula is as follows: ， ， in, Represents the self-attention mechanism; Use the gated recurrent unit GRU to and Perform temporal modeling and obtain the final hidden state as , the formula is as follows: ， in, Represents the gated recurrent unit operation; propagates the first weighted post feature , first weighted user social features and the final hidden state Splice to get the final homogeneous interactive comprehensive representation , the formula is: .

7. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 6 is characterized in that: Step S6 specifically includes: The final homogeneous interaction is comprehensively represented With event statement Features of original tweets Splice to get the final comprehensive features , the formula is as follows: ， The predicted probability vector is calculated through the fully connected layer and the Softmax function. The formula is as follows: ， in, Represents the fully connected layer operation.

8. The rumor detection method based on dual-domain perception structure feature fusion learning according to claim 7 is characterized in that: During training, optimization is performed by minimizing the cross entropy loss between the predicted probability and the true label distribution: ， in, represents the cross entropy loss, is the number of samples in the dataset, represents the number of categories, that is, the total number of categories in the classification task, Indicates that this is Regularization term, representing the model parameters The sum of the squares of is the regularization factor, Representing an event The true label distribution of .

Images