Multi-Dimensional Rumor Detection Method Based on Graph Attention Mechanism and Self-Supervised Learning

Abstract

The present application relates to a multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning. The method includes: extracting key features of the original data, and performing preprocessing and encoding representation, wherein the key features include node attribute features, graph attribute features, user features and time features; reconstructing a propagation graph according to the message forwarding list, inputting the key features and the propagation graph into a double-layer graph attention network, and constructing an initial rumor detection classifier; using a graph self-supervised learning model to process the encoded key features, and generating an embedding vector from the propagation graph; constructing an explanation matrix based on the key features and the embedding vector, performing attention weighting based on the explanation matrix to obtain an adjusted feature vector; inputting the adjusted feature vector into the initial rumor detection classifier, optimizing the initial rumor detection classifier, and performing rumor detection based on the optimized initial rumor detection classifier to output a rumor detection result.

Description

Technical Field

[0001] The present application relates to the technical field of social network rumor detection, and in particular to a multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning. Background Art

[0002] In the Internet era, the rise of social media and online platforms has greatly promoted the rapid circulation of information, but at the same time has also brought about the widespread spread of rumors and false information. These rumors not only mislead the public, but may also trigger social panic, disrupt order, and even affect economic stability. Therefore, it has become particularly urgent to develop effective rumor detection technologies.

[0003] Most existing rumor detection methods mainly rely on the semantic information of text. Therefore, in a multilingual or cross-platform environment, the detection effect is often limited. Traditional models usually need to extract features and train separately for different languages, resulting in weak cross-language transfer ability. At the same time, in the initial stage of rumor spread, rumors often spread rapidly in non-text form through users' interaction and forwarding behaviors. Traditional text analysis methods are difficult to capture these behavioral changes in a timely manner, and the response speed is relatively lagging.

[0004] Secondly, many detection methods ignore the significant differences in the time distribution of user forwarding between rumors and true information. Rumors usually spread rapidly in the early stage, accompanied by a large number of users' concentrated forwarding, while the forwarding time distribution of true information is more dispersed and stable. In addition, most existing rumor detection models lack interpretability of the detection results and cannot clearly explain which features play a key role in the classification results. This not only reduces the credibility of the model, but also limits its operability in practical applications.

[0005] Therefore, in the related technologies, there is an urgent need for a method that can effectively identify rumors and improve the speed and accuracy of rumor detection. Summary of the Invention

[0006] Based on this, in view of the above technical problems, it is necessary to provide a multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning that can effectively identify rumors and improve the speed and accuracy of rumor detection.

[0007] In a first aspect, the present application provides a multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning. The method includes:

[0008] Extracting key features of the original data, and performing preprocessing and encoding representation, where the key features include node attribute features, graph attribute features, user features, and time features;

[0009] According to the message forwarding list, reconstruct the propagation graph, input the key features and the propagation graph into a two-layer graph attention network, and construct an initial rumor detection classifier;

[0010] Use a graph self-supervised learning model to process the encoded key features and generate embedding vectors from the propagation graph;

[0011] Construct an explanation matrix based on the key features and the embedding vectors, perform attention weighting based on the explanation matrix, and obtain an adjusted feature vector;

[0012] Input the adjusted feature vector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier, and output the rumor detection result.

[0013] Optionally, in an embodiment of the present application, the preprocessing includes data normalization processing, and its formula is:

[0014]

[0015] where, X scaled represents the result of the normalization processing, X min and X max represent the minimum and maximum values of various features.

[0016] Optionally, in an embodiment of the present application, the calculation formula of the attention mechanism of the two-layer graph attention network is:

[0017]

[0018] where, α ij is the attention weight of node i to node j, h i , h j and h k are the feature vectors of the nodes, is the neighborhood of i, W is the shared weight matrix, a is the learnable attention weight vector, || represents the concatenation of vectors, and LeakyReLU is the activation function.

[0019] Optionally, in an embodiment of the present application, the graph self-supervised learning model uses the DGI model to maximize the mutual information between the global information and the local information in the graph structure and obtain useful embedding vectors of the graph or nodes.

[0020] Optionally, in an embodiment of the present application, the loss function for optimizing and training the DGI model is optimized using sparsity loss, orthogonality loss, and regularization loss.

[0021] Optionally, in an embodiment of the present application, the formula for constructing the explanation matrix is:

[0022]

[0023] Among them, E * represents the interpretation matrix, H * is the optimal embedding vector of the node, represents the outer product of vectors, F is the feature vector used for transformation together with the node embedding, and ||·|| represents the norm of the vector.

[0024] Optionally, in an embodiment of the present application, the attention weighting based on the interpretation matrix to obtain the adjusted feature vector includes:

[0025] Averaging the features of each row and each dimension of the interpretation matrix to obtain the feature average weight;

[0026] Normalizing the feature average weight and applying it to the key features to obtain the adjusted feature vector.

[0027] In a second aspect, the present application also provides a multi-dimensional rumor detection device based on a graph attention mechanism and self-supervised learning. The device includes:

[0028] An original feature processing module, configured to extract the key features of the original data, and perform preprocessing and encoding representation, where the key features include node attribute features, graph attribute features, user features, and time features;

[0029] An initial rumor detection classifier construction module, configured to reconstruct the propagation graph according to the message forwarding list, input the key features and the propagation graph into a double-layer graph attention network, and construct an initial rumor detection classifier;

[0030] An embedding vector generation module, configured to process the encoded key features by using a graph self-supervised learning model, and generate embedding vectors from the propagation graph;

[0031] An interpretation matrix construction module, configured to construct an interpretation matrix based on the key features and the embedding vectors, and perform attention weighting based on the interpretation matrix to obtain the adjusted feature vector;

[0032] A rumor detection result determination module, configured to input the adjusted feature vector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier, and output the rumor detection result.

[0033] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the steps of the methods in the above various embodiments.

[0034] Fourthly, the present application also provides a computer-readable storage medium. On the computer-readable storage medium, a computer program is stored, and when the computer program is executed by a processor, the steps of the methods in the above various embodiments are implemented.

[0035] The above multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning. First, extract the key features of the original data, and perform preprocessing and encoding representation. Among them, the key features include node attribute features, graph attribute features, user features, and time features. Then, according to the message forwarding list, reconstruct the propagation graph, and input the key features and the propagation graph into a two-layer graph attention network to construct an initial rumor detection classifier. Then, use a graph self-supervised learning model to process the encoded key features and generate embedding vectors from the propagation graph. Then, construct an explanation matrix based on the key features and the embedding vectors, and perform attention weighting based on the explanation matrix to obtain an adjusted feature vector. Finally, input the adjusted feature vector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier to output the rumor detection result. That is to say, by fusing multiple features, including node attributes, graph attributes, user features, and time features, the dependence on text features in the past is broken. This multi-dimensional method not only accurately maps the interaction relationship between users and their influence in the network, but also can monitor the propagation trend of rumors over time. By introducing a graph attention network, it can flexibly process complex information flows, accurately capture key nodes and patterns in the propagation chain, thereby enhancing the adaptability in diverse backgrounds. This feature greatly improves the robustness of the model and the detection accuracy between different platforms, ensuring its wide application in different propagation environments. Through self-supervised learning and the explanation matrix mechanism, the accuracy and interpretability of rumor detection are further improved. The explanation matrix can intuitively show the specific contribution of each feature to the final detection result, making the detection process transparent, helping users understand the basis for the model's judgment, and enhancing the credibility of the model's decision-making. At the same time, the present invention has a highly sensitive recognition ability at the early stage of rumor propagation, can quickly lock in potential propagation paths and abnormal propagation behaviors, thereby significantly improving the response speed and accuracy of rumor detection. By quickly intervening at the initial stage of rumor propagation, the model can effectively contain the spread of rumors, providing more reliable technical support and guarantee for response strategies in complex public opinion environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is an application environment diagram of the multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning in an embodiment;

[0037] Figure 2Schematic flow chart of a multi - dimensional rumor detection method based on graph attention mechanism and self - supervised learning in an embodiment;

[0038] Figure 3 Schematic structural diagram of a multi - dimensional rumor detection method model based on graph attention mechanism and self - supervised learning in an embodiment;

[0039] Figure 4 Schematic diagram of the visualization result of the interpretation matrix of the overall news data in an embodiment;

[0040] Figure 5 Schematic diagram of the visualization result of the interpretation matrix of news data of the Politics category in an embodiment;

[0041] Figure 6 Schematic diagram of the visualization result of the interpretation matrix of news data of the Viral Photos / Stories / Urban Legends category;

[0042] Figure 7 Schematic diagram of the visualization result of the interpretation matrix of news data of the War / Terrorism / Shootings category;

[0043] Figure 8 Schematic diagram of the result comparison before and after weighting the original feature vector by the interpretation matrix value;

[0044] Figure 9 Schematic diagram of the comparison of the experimental results of the present invention with those of other methods;

[0045] Figure 10 Schematic diagram of the early detection result of the present invention.

[0046] Figure 11 Schematic block diagram of a multi - dimensional rumor detection device based on graph attention mechanism and self - supervised learning in an embodiment;

[0047] Figure 12 Internal structural diagram of a computer device in an embodiment. Detailed implementation manners

[0048] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] The multi - dimensional rumor detection method based on graph attention mechanism and self - supervised learning provided by the embodiments of the present application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or other network servers. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0050] In one embodiment, as Figure 2 shown, a multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning is provided. Taking the server in Figure 1 as an example, the method includes the following steps:

[0051] S201: Extract the key features of the original data, and perform preprocessing and encoding representation. Among them, the key features include node attribute features, graph attribute features, user features, and time features.

[0052] In the embodiment of the present application, first, key features are extracted from the original data, including node attribute features, graph attribute features, user features, and time features, and these features are preprocessed and encoded. Among them, the node attribute features include the number of connections of the node, clustering coefficient, standard deviation, degree of the average adjacent node, average adjacent clustering coefficient, and centrality. The graph attribute features include the size of the graph, maximum depth, and maximum width. The user features include the average number of user attentions, average participation, average number of fans, and average account age. The time features include the total time of each cascade propagation and the total time of the independent user propagation interval. The preprocessed features are concatenated as the initial feature vector for subsequent training.

[0053] In one embodiment of the present application, the preprocessing includes data normalization processing, and its formula is:

[0054]

[0055] where, X scaled represents the normalization result, X min and X max represent the minimum and maximum values of various features.

[0056] S203: According to the message forwarding list, reconstruct the propagation graph, and input the key features and the propagation graph into a two-layer graph attention network to construct an initial rumor detection classifier.

[0057] In the embodiments of the present application, when a user forwards a certain news or information, this forwarding behavior will form a propagation path in the user network. By analyzing the propagation path, specific cascading events can be identified. Based on this, each news propagation graph is reconstructed. After that, the initial feature vectors extracted from the original data and spliced after encoding and the reconstructed propagation graph are divided into training samples and test samples. The spliced initial feature vectors are used as additional node features of the propagation graph and are input into a two-layer graph attention network for training to construct an initial rumor detection classifier and output preliminary detection results. As Figure 3 shown, the two-layer graph attention network module can improve the feature extraction and rumor recognition capabilities of the model in the propagation network by effectively using the node relationships in the graph structure data and the neighbor information on the graph.

[0058] Specifically, the two-layer graph attention network is an improved version of the Graph Attention Networks (GAT), and GATv2Conv is an enhancement of the original GAT. In the original GAT, the attention mechanism relies on the linear combination of input features and weights, which may lead to insufficient learning of information expressing the graph structure in certain cases. GATv2 proposes a new attention calculation method, using self-attention mechanism, which can dynamically adjust the aggregation of node features according to the importance of neighbor nodes, making the attention mechanism insensitive to the order of input features, that is, solving the symmetry problem of the attention function. This enables GATv2 to flexibly capture the complex dependencies between neighbor nodes.

[0059] In an embodiment of the present application, the calculation formula of the attention mechanism of the two-layer graph attention network is:

[0060]

[0061] where α ij is the attention weight of node i to node j, h i , h j and h k are the feature vectors of the nodes, is the neighborhood of i, W is the shared weight matrix, a is the learnable attention weight vector, || represents the concatenation of vectors, and LeakyReLU is the activation function.

[0062] In an embodiment of the present application, through the attention weight α ij , the feature h i of each node i will aggregate the features of its neighbor nodes and be updated to a new representation h i ′. The specific formula is as follows:

[0063]

[0064] Among them, σ is the activation function. Optionally, the SELU activation function can be adopted.

[0065] The SELU activation function is a self-normalizing activation function that can maintain the zero mean and unit variance of the neuron output in a deep neural network, helping to solve the problem of gradient vanishing or explosion in deep networks. The formula is as follows:

[0066]

[0067] Among them, λ is a scalar constant used to scale the output. α is the scaling factor for the negative part.

[0068] S205: Process the encoded key features using the graph self-supervised learning model and generate embedding vectors from the propagation graph.

[0069] In the embodiments of the present application, the encoded initial feature vectors are processed using the graph self-supervised learning model, and embedding vectors are generated from the propagation graph.

[0070] In one embodiment of the present application, the graph self-supervised learning model adopts the DGI model, which is used to maximize the mutual information between the global information and the local information in the graph structure to obtain useful embedding vectors of the graph or nodes.

[0071] In one embodiment of the present application, as Figure 3 shown, the graph self-supervised learning model adopts the DGI (Deep GraphInfomax) model, which aims to obtain useful embedding vectors of the graph or nodes by maximizing the mutual information between the global information and the local information in the graph structure. Assume that a graph G(X, A) is given, where X is the node feature matrix with a dimension of N×F, where N is the number of nodes and F is the feature dimension of each node, and the graph convolutional network GCN (Graph ConvolutionalNetworks) is used to generate embedding vectors. The basic propagation rule of GCN is:

[0072]

[0073] Among them, is the adjacency matrix with self-loops added. is 's degree matrix. H (l) is the node embedding representation of the l-th layer. Initially, H (0) = X. W (l) is the learnable weight matrix. σ is the non-linear activation function. Here, we adopt the RELU activation function. Then, a READOUT function is used to generate the global graph representation s.

[0074] For contrastive learning, DGI needs to generate negative samples. By randomly shuffling the node features X, a perturbed feature matrix X′ is obtained, while the adjacency matrix A remains unchanged. The node embeddings for generating negative samples are:

[0075] H′ = GCN(X′, A)

[0076] After that, the core objective of DGI is to maximize the mutual information between the local embedding h i of the node and the global graph representation s, while minimizing the correlation between the negative sample node embedding h i ′ and the global representation s. This is achieved through contrastive learning, and the loss function is:

[0077]

[0078] where D(h i , s) is a discriminator used to distinguish positive and negative samples. It usually has the following form:

[0079]

[0080] where W is a learnable weight matrix and σ is the sigmoid function.

[0081] In one embodiment of the present application, the loss function for training the DGI model is optimized using sparsity loss, orthogonality loss, and regularization loss.

[0082] In one embodiment of the present application, in order to further optimize the DGI model and obtain the optimal embedding vector, sparsity loss, orthogonality loss, and regularization loss are added to the loss function.

[0083] Sparsity loss encourages the model to generate sparse representations by controlling the sparsity of the interpretation matrix, thereby improving the generalization ability of the model. Specifically, sparsity loss focuses on making most elements in the interpretation matrix close to zero, while only retaining a few important features. This helps the model focus on the most important graph features, thereby improving its interpretability and generalization ability. The calculation formula for sparsity loss is:

[0084]

[0085] where E is the interpretation matrix, λ sparsity is the weight of the sparsity loss, and ||·||1 is the L1 norm.

[0086] The orthogonality loss avoids overfitting by making the row vectors in the interpretation matrix orthogonal to each other, which helps to increase the discriminative power of the embedding. The orthogonality loss focuses on reducing the correlation between the row vectors of the interpretation matrix, so that each dimension encodes different information as independently as possible. This helps to improve the interpretability of the model because it ensures that each embedding dimension is defined by a different set of graph features. The calculation formula of the orthogonality loss is:

[0087]

[0088] where E T E is the autocorrelation matrix of the interpretation matrix, ||·|| F is the Frobenius norm, and λ orthogonality is the weight of the orthogonality loss.

[0089] The regularization loss is used to limit the size of the embedding, avoid overly large weights or eigenvalues, and improve the stability of the model. The regularization loss focuses on controlling the norm of the embedding vector, preventing the model from relying too much on certain specific features, thereby improving the generalization ability of the model. The calculation formula of the regularization loss is:

[0090]

[0091] where λ regularization is the weight of the regularization loss. H i is the embedding of the i-th node. The final loss function is the weighted sum of multiple loss terms, and the formula is:

[0092]

[0093] By inputting the feature vector into the DGI model and training and optimizing it, the optimal embedding vector can be obtained.

[0094] S207: Construct an interpretation matrix based on the key features and the embedding vector, and perform attention weighting based on the interpretation matrix to obtain an adjusted feature vector.

[0095] In the embodiments of the present application, the interpretation matrix is jointly constructed by combining the embedding vector and the initial feature vector to evaluate the importance of each feature. And attention weighting is performed according to the importance values of the features in each row of the interpretation matrix to adjust the feature vector.

[0096] In an embodiment of the present application, the construction formula of the interpretation matrix is:

[0097]

[0098] where E * represents the interpretation matrix, H * is the optimal embedding vector of the node, Denotes the cross product of vectors, F is the eigenvector used for transformation together with the node embedding, and ||·|| denotes the norm of the vector.

[0099] In one embodiment of the present application, the obtaining of the adjusted eigenvector by performing attention weighting based on the interpretation matrix includes:

[0100] Averaging the features of each row and each dimension of the interpretation matrix to obtain the feature average weight;

[0101] Normalizing the feature average weight and applying it to the key features to obtain the adjusted eigenvector.

[0102] In one embodiment of the present application, first, the features of each row and each dimension of the interpretation matrix are averaged, and the specific formula is as follows:

[0103]

[0104] where w i represents the average weight of feature i, and m represents the number of columns of the interpretation matrix.

[0105] After calculating the average weights of all features, these weights are normalized:

[0106]

[0107] where w norm,i represents the normalized weight of feature i, w min and w max represent the maximum and minimum values in the feature average weights.

[0108] And applying the normalized weights to each feature of the original feature matrix to obtain the adjusted eigenvector. The specific formula is as follows:

[0109] X adjusted = X ⊙ w norm

[0110] where X represents the original feature matrix, and X adjusted represents the adjusted feature matrix.

[0111] S209: Input the adjusted eigenvector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier, and output the rumor detection result.

[0112] In the embodiment of the present application, the adjusted feature matrix is re-input into the constructed initial rumor detection classifier for training to optimize the initial rumor detection classifier, and the final rumor detection result is output.

[0113] It should be noted that in specific applications, starting from the initial moment of rumor spread, the accuracy values of rumor spread within a fixed time interval are detected to obtain early detection results of rumors, so as to achieve rapid identification and response to rumors. For example, every 100 seconds is used as a fixed time interval to monitor the rumor spread within the first 600 seconds.

[0114] In an embodiment of the present application, real spread data is used to verify the usability of the rumor detection model. The real spread data are all verified true and false news reports posted on Twitter from 2006 to 2017 collected in the paper "The spread of true and false news online". Conditional screening is performed on the overall news data and news data of major categories (including Politics, Viral Photos / Stories / Urban Legends, War / Terrorism / Shootings). Finally, 5000 news true-false cascades are screened out from the overall news data, and 1000 news true-false cascades are respectively screened out from the news data of specific categories as samples to calculate the interpretability matrices of their respective feature vectors and embedding vectors, so as to obtain the importance degree of features.

[0115] On the basis of using the DGI model as the embedding model, visual analysis of the interpretation matrix is performed on the overall news data and news data of specific categories (including Politics, Viral Photos / Stories / Urban Legends, War / Terrorism / Shootings), and the results are as Figure 4 、 5 shown in Figures 6 and 7. It can be seen that whether it is the overall or classified rumor data, the ranking of feature importance remains consistent. The important features are, in order, the degree of the average adjacent node of the node, centrality, the size of the graph, the average number of users' followers, the average number of fans, the average account age, the total time of the independent user spread interval, the number of connections of the node, the maximum width of the graph, the total time of each cascade spread, and the maximum depth of the graph. The contribution of other features to rumor spread is almost zero, indicating that they have little impact on the prediction ability of the model.

[0116] In an embodiment of the present application, the interpretation matrices of the overall news data and news data of specific categories (including Politics, Viral Photos / Stories / Urban Legends, War / Terrorism / Shootings) are used to perform weighted processing on the importance of each feature, and the rumor detection results before and after weighting are compared. As Figure 8As shown, in the detection of overall news data, the application of the weighted interpretation matrix significantly improved the model performance. The detection results increased from the original accuracy of 0.8384, precision of 0.9692, recall of 0.6864, and F1-score of 0.7827 to an accuracy of 0.8505, precision of 0.9641, recall of 0.7210, and F1-score of 0.8050.

[0117] For specific types of news data, the weighted interpretation matrix also demonstrated an optimization effect:

[0118] In the news data of the Politics category, the detection results increased from an accuracy of 0.9475, precision of 0.9786, recall of 0.9150, and F1-score of 0.9457 to an accuracy of 0.9651, precision of 0.9947, recall of 0.9350, and F1-score of 0.9639. In the news data of the Viral Photos / Stories / Urban Legends category, the detection results increased from an accuracy of 0.8321, precision of 0.9879, recall of 0.6643, and F1-score of 0.7983 to an accuracy of 0.8429, precision of 0.9983, recall of 0.6857, and F1-score of 0.8136. In the news data of the War / Terrorism / Shootings category, the detection results increased from an accuracy of 0.8679, precision of 0.9893, recall of 0.7357, and F1-score of 0.8477 to an accuracy of 0.9036, precision of 0.9982, recall of 0.8071, and F1-score of 0.8933. It can be seen that by introducing the weighted mechanism of the interpretation matrix, the rumor detection model achieved performance improvement in news data of each category, indicating the important role of the interpretation matrix in improving the prediction accuracy and interpretability of the model.

[0119] In an embodiment of the present application, the DTC (Decision Tree Classifier) model, SVM (Support Vector Machine) model, GraphSAGE model, GCN model, GAE (Graph Auto-Encoder) model, and GRU (Gated Recurrent Unit) model were used as comparative examples for rumor detection under the same data. As Figure 9As shown, when comparing the performance of different models, the DTC model achieved an accuracy of 0.7543, a precision of 0.7292, a recall of 0.7202, and an F1 score of 0.7247; the SVM model achieved an accuracy of 0.7854, a precision of 0.7632, a recall of 0.7378, and an F1 score of 0.7253; the GraphSAGE model achieved an accuracy of 0.8088, a precision of 0.9120, a recall of 0.6356, and an F1 score of 0.7491; the GCN model achieved an accuracy of 0.7939, a precision of 0.8528, a recall of 0.6540, and an F1 score of 0.7403; the GAE model achieved an accuracy of 0.8248, a precision of 0.9642, a recall of 0.6172, and an F1 score of 0.7598; the GRU model achieved an accuracy of 0.8307, a precision of 0.9541, a recall of 0.6421, and an F1 score of 0.7853. In contrast, the multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning achieved excellent performance with an accuracy of 0.8505, a precision of 0.9641, a recall of 0.7210, and an F1 score of 0.8050, outperforming other compared models in all evaluation metrics. This result further verifies the effectiveness of the method of the present invention and its adaptability to the rumor identification task.

[0120] In one embodiment of the present application, as Figure 10As shown, timed monitoring was carried out within the first 600 seconds of rumor spread to evaluate the effectiveness of the method in early rumor detection. Within the first 100 seconds, the accuracy of early detection reached 0.8087, the precision was 0.9598, the recall rate was 0.5968, and the F1 score was 0.7360. When the time reached 200 seconds, the accuracy of early detection increased to 0.8272, the precision was 0.9676, the recall rate was 0.6343, and the F1 score was 0.7663. At 300 seconds, the accuracy of early detection further increased to 0.8380, the precision was 0.9781, the recall rate was 0.6496, and the F1 score was 0.7807. At 400 seconds, the accuracy of early detection slightly increased to 0.8423, the precision was 0.9355, the recall rate was 0.6945, and the F1 score was 0.7972. By 500 seconds, the accuracy of early detection increased to 0.8495, the precision was 0.9893, the recall rate was 0.6698, and the F1 score was 0.7988. At 600 seconds, the accuracy of early detection stabilized at 0.8498, the precision was 0.9609, the recall rate was 0.6917, and the F1 score was 0.8044. It can be seen that the method can quickly identify rumors in a short time, and as time goes by, its detection performance gradually enhances and tends to be stable. This early detection ability is crucial for quickly responding to and controlling the spread of rumors, and helps to take measures before the rumors cause extensive impacts.

[0121] In the above multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning, first, key features of the original data are extracted and preprocessed and encoded. Among them, the key features include node attribute features, graph attribute features, user features, and time features. After that, according to the message forwarding list, the propagation graph is reconstructed, and the key features and the propagation graph are input into a two-layer graph attention network to construct an initial rumor detection classifier. Then, a graph self-supervised learning model is used to process the encoded key features and generate embedding vectors from the propagation graph. Then, an interpretation matrix is constructed based on the key features and the embedding vectors, and attention weighting is performed based on the interpretation matrix to obtain adjusted feature vectors. Finally, the adjusted feature vectors are input into the initial rumor detection classifier to optimize the initial rumor detection classifier, and rumor detection is performed based on the optimized initial rumor detection classifier, and the rumor detection result is output. That is to say, by fusing multiple features, including node attributes, graph attributes, user features, and time features, the dependence on text features in the past is broken. This multi-dimensional method not only accurately maps the interaction relationship between users and their influence in the network, but also can monitor the propagation trend of rumors over time. By introducing a graph attention network, complex information flows can be flexibly processed, key nodes and patterns in the propagation chain can be accurately captured, thereby enhancing the adaptability in diverse backgrounds. This feature greatly improves the robustness of the model and the detection accuracy between different platforms, ensuring its wide application in different propagation environments. Through self-supervised learning and the interpretation matrix mechanism, the accuracy and interpretability of rumor detection are further improved. The interpretation matrix can intuitively show the specific contribution of each feature to the final detection result, making the detection process transparent, helping users understand the judgment basis of the model, and enhancing the credibility of the model's decision-making. At the same time, the present invention has a highly sensitive recognition ability in the early stage of rumor propagation, can quickly lock potential propagation paths and abnormal propagation behaviors, thereby significantly improving the response speed and accuracy of rumor detection. By quickly intervening in the initial stage of rumor propagation, the model can effectively contain the spread of rumors, providing more reliable technical support and guarantee for coping strategies in complex public opinion environments.

[0122] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0123] Based on the same inventive concept, an embodiment of the present application further provides a multi-dimensional rumor detection device based on graph attention mechanism and self-supervised learning for implementing the above-mentioned multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the multi-dimensional rumor detection device based on graph attention mechanism and self-supervised learning provided below can refer to the limitations on the multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning in the above text, and will not be repeated here.

[0124] In one embodiment, as Figure 11 shown, a multi-dimensional rumor detection device 1100 based on graph attention mechanism and self-supervised learning is provided, including: an original feature processing module 1101, an initial rumor detection classifier construction module 1103, an embedding vector generation module 1105, an interpretation matrix construction module 1107, and a rumor detection result determination module 1109, where:

[0125] The original feature processing module 1101 is configured to extract key features of the original data, and perform preprocessing and encoding representation, where the key features include node attribute features, graph attribute features, user features, and time features.

[0126] The initial rumor detection classifier construction module 1103 is configured to reconstruct the propagation graph according to the message forwarding list, and input the key features and the propagation graph into a two-layer graph attention network to construct an initial rumor detection classifier.

[0127] The embedding vector generation module 1105 is configured to process the encoded key features by using a graph self-supervised learning model, and generate embedding vectors from the propagation graph.

[0128] The interpretation matrix construction module 1107 is used to construct an interpretation matrix based on the key features and embedding vectors, perform attention weighting based on the interpretation matrix, and obtain an adjusted feature vector.

[0129] The rumor detection result determination module 1109 is used to input the adjusted feature vector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier, and output a rumor detection result.

[0130] In an embodiment of the present application, the preprocessing includes data normalization processing, and its formula is:

[0131]

[0132] where, X scaled represents the result of normalization processing, X min and X max represent the minimum and maximum values of various features.

[0133] In an embodiment of the present application, the calculation formula of the attention mechanism of the double-layer graph attention network is:

[0134]

[0135] where, α ij is the attention weight of node i to node j, h i , h j and h k are the feature vectors of the nodes, is the domain of i, W is the shared weight matrix, a is the learnable attention weight vector, || represents the concatenation of vectors, and LeakyReLU is the activation function.

[0136] In an embodiment of the present application, the graph self-supervised learning model adopts the DGI model, which is used to maximize the mutual information between the global information and the local information in the graph structure, and obtain useful embedding vectors of the graph or nodes.

[0137] In an embodiment of the present application, the loss function for optimizing and training the DGI model is optimized by using sparsity loss, orthogonality loss, and regularization loss.

[0138] In an embodiment of the present application, the formula for constructing the interpretation matrix is:

[0139]

[0140] where, E * represents the interpretation matrix, H * is the optimal embedding vector of the node, Denotes the outer product of vectors, F is the eigenvector used for transformation together with the node embedding, and ||·|| denotes the norm of the vector.

[0141] In one embodiment of the present application, the obtaining of the adjusted eigenvector by performing attention weighting based on the interpretation matrix includes:

[0142] Averaging the features of each row and each dimension of the interpretation matrix to obtain the feature average weight;

[0143] Normalizing the feature average weight and applying it to the key features to obtain the adjusted eigenvector.

[0144] Each module in the above multi-dimensional rumor detection device based on the graph attention mechanism and self-supervised learning can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0145] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 12 shown. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a multi-dimensional rumor detection method based on the graph attention mechanism and self-supervised learning. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the shell of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0146] Those skilled in the art can understand that Figure 12 the structure shown in

[0147] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented.

[0148] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0149] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties.

[0151] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0152] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0153] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning, characterized in that The method includes: extracting key features of the original data, and performing preprocessing and encoded representation, where the key features include node attribute features, graph attribute features, user features, and time features; reconstructing a propagation graph according to the message forwarding list, inputting the key features and the propagation graph into a double-layer graph attention network, and constructing an initial rumor detection classifier; processing the encoded key features by using a graph self-supervised learning model, and generating an embedding vector from the propagation graph; constructing an interpretation matrix based on the key features and the embedding vector, and performing attention weighting based on the interpretation matrix to obtain an adjusted feature vector; inputting the adjusted feature vector into the initial rumor detection classifier, optimizing the initial rumor detection classifier, and performing rumor detection based on the optimized initial rumor detection classifier to output a rumor detection result; The formula for constructing the interpretation matrix is: Among them, represents the interpretation matrix, is the optimal embedding vector of the node, represents the outer product of vectors, is the eigenvector used for transformation together with the node embedding, represents the norm of the vector; The performing attention weighting based on the interpretation matrix to obtain an adjusted feature vector includes: averaging the features of each row and each dimension of the interpretation matrix to obtain an average feature weight; performing normalization processing on the average feature weight and applying it to the key features to obtain an adjusted feature vector.

2. The multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning according to claim 1, characterized in that, The preprocessing includes data normalization processing, and its formula is: Among them, represents the normalization result, and represent the minimum and maximum values of various features.

3. The multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning according to claim 1, wherein The calculation formula of the attention mechanism of the double-layer graph attention network is: Among them, is the node 's attention weight on the node , , and are the feature vectors of the node, is 's domain, is the shared weight matrix, is the learnable attention weight vector, represents the concatenation of vectors, is the activation function.

4. The multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning according to claim 1, characterized in that The graph self-supervised learning model adopts a DGI model, which is used to maximize the mutual information between the global information and the local information in the graph structure, and obtain useful embedding vectors of the graph or nodes.

5. The multi-dimensional rumor detection method based on graph attention mechanism and self-supervised learning according to claim 4, characterized in that, The loss function of the DGI model is optimized by using sparsity loss, orthogonality loss, and regularization loss.

6. A multi-dimensional rumor detection device based on graph attention mechanism and self-supervised learning, characterized in that, The device includes: an original feature processing module, which is used to extract key features of the original data, and perform preprocessing and encoded representation, where the key features include node attribute features, graph attribute features, user features, and time features; an initial rumor detection classifier construction module, which is used to reconstruct a propagation graph according to the message forwarding list, input the key features and the propagation graph into a double-layer graph attention network, and construct an initial rumor detection classifier; an embedding vector generation module, which is used to process the encoded key features by using a graph self-supervised learning model, and generate an embedding vector from the propagation graph; an interpretation matrix construction module, which is used to construct an interpretation matrix based on the key features and the embedding vector, and perform attention weighting based on the interpretation matrix to obtain an adjusted feature vector; a rumor detection result determination module, which is used to input the adjusted feature vector into the initial rumor detection classifier, optimize the initial rumor detection classifier, and perform rumor detection based on the optimized initial rumor detection classifier to output a rumor detection result; The formula for constructing the interpretation matrix is: Among them, represents the interpretation matrix, is the optimal embedding vector of the node, represents the outer product of vectors, is the eigenvector used for transformation together with the node embedding, represents the modulus of the vector; The performing attention weighting based on the interpretation matrix to obtain an adjusted feature vector includes: averaging the features of each row and each dimension of the interpretation matrix to obtain an average feature weight; performing normalization processing on the average feature weight and applying it to the key features to obtain an adjusted feature vector.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images