A rumor detection method, system and device based on multi-scale and double view

Abstract

The application discloses a rumor detection method, system and device based on multi-scale and double-view, and the method comprises the following steps: acquiring tweet data; based on the tweet data, constructing a propagation graph and generating a training set; constructing a graph neural network model based on a multi-scale framework and a double-view framework to obtain a detection model; training the detection model based on the training set to obtain a trained detection model; and applying the trained detection model to a rumor detection task. The system comprises a data acquisition module, a preprocessing module, a model construction module, a training module and an application module. The device comprises a memory and a processor for executing the rumor detection method based on multi-scale and double-view. By using the application, the accuracy of the rumor detection task can be improved. The application can be widely applied to the field of model application.

Description

Technical Field

[0001] This invention relates to the field of model applications, and in particular to a rumor detection method, system, and apparatus based on multi-scale and dual-view methods. Background Technology

[0002] The widespread dissemination of rumors not only distorts the public's judgment of information but can also exacerbate social conflicts, triggering unnecessary panic and chaos. In some extreme cases, rumors can even lead to mass incidents, disrupting public order and posing a serious threat to social stability. The goal of rumor detection on social media is to assess unverified events on social media, determining whether they are factual information or rumors. The task of rumor detection has received increasing attention in academia.

[0003] Although research on rumor detection on social media has made some breakthroughs and achievements in recent years, the field is still in a developing stage overall. Deep learning-based methods often ignore or fail to effectively utilize the graph structure of rumor propagation. Summary of the Invention

[0004] In view of this, in order to solve the technical problem that existing rumor detection methods cannot effectively utilize the graph structure information of rumor propagation, thus leading to low detection accuracy, firstly, this invention proposes a rumor detection method based on multi-scale and dual-view, the method comprising the following steps:

[0005] Get tweet data;

[0006] Based on the tweet data, a propagation graph is constructed and a training set is generated;

[0007] A graph neural network model based on a multi-scale framework and a dual-view framework is constructed to obtain the detection model;

[0008] The detection model is trained based on the training set to obtain a trained detection model;

[0009] The trained detection model is then applied to a rumor detection task.

[0010] In some embodiments, the step of constructing a propagation graph and generating a training set based on the tweet data specifically includes:

[0011] Based on the tweet data, events and tags are defined to obtain preprocessed data;

[0012] The preprocessed data is encoded to obtain tweet features;

[0013] Based on the tweet characteristics, a propagation graph is constructed according to the propagation process of the event, and a training set is obtained.

[0014] This optimized step completes the preprocessing of tweet data, including encoding and propagation graph construction, and then the training set is divided for subsequent model training.

[0015] In some embodiments, it also includes:

[0016] The tweet data is segmented into words, and stop words and special symbols are removed.

[0017] This optimized step performs preprocessing steps commonly used in text processing on the tweet data, making it easier to organize the text data.

[0018] In some embodiments, the step of training the detection model based on the training set to obtain a trained detection model specifically includes:

[0019] Construct an extended graph from the propagation graph in the training set;

[0020] Based on the dual-view framework, information in the extended graph is captured, node features are updated, and feature view features and topology view features are obtained.

[0021] An attention mechanism is introduced to fuse the feature view features of different scales in the node to obtain the first fused feature;

[0022] Based on the attention mechanism, the topological view features of different scales in the nodes are fused to obtain the second fused feature;

[0023] The InfoNCE loss function value is calculated based on the first fusion feature and the second fusion feature to obtain the first loss value;

[0024] Based on the first fusion feature and the second fusion feature, graph-level vectors are constructed using graph pooling;

[0025] Based on the graph-level vectors, the predicted labels are obtained;

[0026] The second loss value is obtained by calculating the loss based on the predicted label and the true label;

[0027] Calculate the final loss value based on the first loss value and the second loss value;

[0028] The detection model is backpropagated based on the final loss value to update its parameters, resulting in a fully trained detection model.

[0029] This optimization process determines the basic framework and training steps of the model. After training is completed, the model is applied to real-world scenarios. The data processing process in actual application is not significantly different from the training process.

[0030] In some embodiments, the step of capturing information in the extended graph based on the dual-view framework, updating node features, and obtaining feature view features and topology view features specifically includes:

[0031] The dual-view framework includes a graph attention network and a graph isomorphism network;

[0032] The attention coefficients between a node and its neighboring nodes are calculated based on the graph attention network, and the feature view features of the node are output.

[0033] Based on the aforementioned co-directional network, a parameter vector is defined, and the topological view features of the nodes are output.

[0034] In the dual-view framework of this embodiment, Graph Attention Networks (GAT) are used to capture the feature view, while Graph Isomorphism Networks (GIN) are used to capture the structure view.

[0035] In some embodiments, the step of introducing an attention mechanism to fuse the feature view features of different scales in a node to obtain a first fused feature specifically includes:

[0036] Introduce an attention mechanism;

[0037] Calculate the attention scores for nodes at different scales;

[0038] The attention score is normalized to obtain the normalized attention score;

[0039] Based on the normalized attention score, the feature view features of the node at different scales are weighted and summed to obtain the first fused feature.

[0040] This optimization step introduces an attention mechanism to finely evaluate the contribution of nodes with different hop counts to the final representation.

[0041] In some embodiments, the formula for calculating the first loss value is as follows:

[0042]

[0043] in, Functions represent calculations and The similarity, where τ represents the scaling parameter. This represents the first fusion feature of node i. This represents the second fusion feature of node i. This represents the first fusion feature of node j. Let j represent the second fusion feature of node j, and n represent the total number of nodes.

[0044] In this embodiment, the loss function is used to adjust the similarity between the feature view and the structure view.

[0045] In some embodiments, the step of constructing a graph-level vector using graph pooling based on the first fusion feature and the second fusion feature specifically includes:

[0046] The first fusion feature and the second fusion feature are integrated respectively to obtain the first fusion feature matrix and the second fusion feature matrix;

[0047] The first fusion feature matrix and the second fusion feature matrix are subjected to mean and maximum pooling operations, and then concatenated with the vector of the root node of the propagation graph to obtain a graph-level vector.

[0048] Secondly, the present invention also proposes a rumor detection system based on multi-scale and dual-view architecture, the system comprising:

[0049] The data acquisition module is used to acquire tweet data;

[0050] The preprocessing module constructs a propagation graph and generates a training set based on the tweet data;

[0051] The model building module is used to build graph neural network models based on multi-scale and dual-view frames to obtain the detection model;

[0052] The training module trains the detection model based on the training set to obtain a trained detection model.

[0053] The application module is used to apply the trained detection model to the rumor detection task.

[0054] This invention also proposes a rumor detection device based on multi-scale and dual-view, comprising:

[0055] At least one processor;

[0056] At least one memory for storing at least one program;

[0057] When the at least one program is executed by the at least one processor, the at least one processor implements a rumor detection method based on multi-scale and dual-view as described above.

[0058] Based on the above scheme, this invention provides a rumor detection method, system, and device based on multi-scale and dual-view. Addressing the shortcoming of being unable to capture social influences at long distances, this invention proposes a multi-scale framework to capture social influences at different distances, particularly ensuring that long-distance social influences are not ignored. Furthermore, addressing the problem of insufficient information obtained from a single view, this invention proposes a dual-view framework that integrates feature views and structural views, extracting common information from both views while maintaining their independence. This achieves the goal of processing the details of node features while revealing the complex connections and propagation patterns between nodes.

[0059] This invention addresses the need for rumor detection by incorporating a multi-scale framework and a dual-view framework to construct a rumor detection model based on graph neural networks. This model captures comprehensive features of the rumor propagation process, thereby improving the accuracy and reliability of rumor detection in social media. Attached Figure Description

[0060] Figure 1 This is a flowchart of the steps of a rumor detection method based on multi-scale and dual-view according to the present invention;

[0061] Figure 2 This is a schematic diagram of the detection model in a specific embodiment of the present invention;

[0062] Figure 3 This is a propagation diagram of an event according to a specific embodiment of the present invention;

[0063] Figure 4 This is a structural block diagram of a rumor detection system based on multi-scale and dual-view according to the present invention. Detailed Implementation

[0064] To address the technical problem mentioned in the background art that rumor detection methods cannot effectively utilize the graph structure information of rumor propagation, thus leading to low detection accuracy, this invention constructs a propagation graph based on the propagation process of an event and uses it as raw data, enabling the model to learn the graph structure information, thereby overcoming the aforementioned technical problem.

[0065] In addition to the technical issues mentioned above, some methods, when involved in graph structure construction, fail to capture long-distance social influences due to oversmoothing, or often focus only on a single view—either the feature view of a node or the structural view of the graph—without considering both simultaneously. This invention overcomes these problems by capturing social influences at different distances through a multi-scale framework and extracting shared information from two views through a dual-view framework.

[0066] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0067] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0068] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0069] Unless the context explicitly indicates an exception, words such as "a," "an," "a kind," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list; a method or apparatus may also include other steps or elements. An element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0070] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0071] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0072] Reference Figure 1 This is a schematic flowchart of an optional example of the rumor detection method based on multi-scale and dual-view proposed in this invention. The method can be applied to computer devices, and the imaging method proposed in this embodiment may include, but is not limited to, the following steps:

[0073] Step S1: Obtain tweet data;

[0074] The tweet data includes the source tweet corresponding to the event and the responding tweets.

[0075] Step S2: Based on the tweet data, construct a propagation graph and generate a training set;

[0076] All tweet data is preprocessed.

[0077] Step S3: Construct a graph neural network model based on a multi-scale frame and a dual-view frame to obtain the detection model;

[0078] The structure of the detection model is as follows: Figure 2 As shown, the model is composed of a multi-scale framework and a dual-view framework. This model leverages the synergy between multi-scale feature extraction and attention mechanisms, adaptively aggregating features from different scales to ensure that direct and long-distance social influences are appropriately considered during rumor detection, thus improving the ability to identify user interaction patterns. The dual-view framework, consisting of a feature view and a structure view, is used to process the rumor propagation graph. Simultaneously, contrastive learning is employed to quantify the degree of information sharing between views, encouraging diverse learned representations.

[0079] Step S4: Train the detection model based on the training set to obtain the trained detection model;

[0080] Step S5: Apply the trained detection model to rumor detection.

[0081] In some feasible embodiments, step S2 specifically includes:

[0082] S2.1. Based on the tweet data, define events and tags to obtain preprocessed data;

[0083] Prior to step S2.1, the tweet data can be segmented and stop words and special symbols can be removed.

[0084] In this embodiment, the i-th event is defined as Where n is the total number of tweets related to the event. This indicates that the source tweet is the news information. This represents the j-th comment or retweet in event i. Each event has a tag y. i ∈{rumors, not rumors}.

[0085] S2.2 Encode the preprocessed data to obtain tweet features;

[0086] Next, the processed tweets are encoded using a natural language processing library to obtain the tweet's features. These features can be represented as... Indicates the source tweet Feature vectors, other features Tweet The feature vectors have a dimension of d.

[0087] S2.3 Based on the tweet features, construct a propagation graph according to the propagation process of the event, and divide it into training sets.

[0088] Then, a propagation graph of the event is constructed based on the propagation process of the event, denoted as G. i = <V i E i ,X i >,V i Let be the set of nodes representing the tweets of the i-th event, where each node represents a tweet, defined as . node Corresponding to the tweet E i It is an edge set, defined as If the k-th tweet retweets or comments on the j-th tweet, then there exists a directed edge from the k-th tweet to the j-th tweet. point to Right now The adjacency matrix A can be constructed from the edge set. i ∈{0,1} n×n .if So

[0089] An example of a propagation graph for a specific event is as follows: Figure 3 As shown.

[0090] In some feasible embodiments, step S3 specifically includes:

[0091] S3.1 Construct an extended graph for the propagation graph in the training set;

[0092] In this step, given a graph, the undirected adjacency matrix A is first used. UD To construct a series of extended graphs, these graphs correspond to different powers of the adjacency matrix, ranging from 1 to K, representing different hop counts and scales. Formally, this can be represented as a set of graphs {G}. 1 G 2 ,…,G K}, where each graph G k Defined as X represents the node feature matrix.

[0093] S3.2. Based on the dual-view framework, capture the information in the extended graph, update the node features, and obtain the feature view features and the topology view features;

[0094] In the complex network of rumor propagation, the fusion of different views is key to understanding and detecting rumors. To fully utilize the information in the propagation graph, this embodiment proposes a dual-view framework designed to capture diverse information within the graph.

[0095] S3.3 Introduce an attention mechanism to fuse the feature view features of different scales in the node to obtain the first fused feature;

[0096] S3.4. Based on the attention mechanism, the topological view features of different scales in the node are fused to obtain the second fused feature;

[0097] This step utilizes an attention mechanism to fuse the feature view features and structural view features of nodes at various scales.

[0098] S3.5 Calculate the InfoNCE loss function value based on the first fusion feature and the second fusion feature to obtain the first loss value;

[0099] The InfoNCE (Information Noise Contrastive Estimation) loss function is a loss function used in unsupervised learning, commonly found in autoencoders and contrastive learning tasks. By maximizing the mutual information of positive sample pairs, the InfoNCE loss function encourages the model to map similar samples to close embedding vector spaces, thereby improving the performance of contrastive learning tasks. Simultaneously, by minimizing the mutual information of negative sample pairs, it can also enhance the model's ability to distinguish between different samples. Mutual information is a measure of the dependency between two random variables and can be used to measure the model's understanding of the input data.

[0100] S3.6. Based on the first fusion feature and the second fusion feature, construct a graph-level vector using graph pooling;

[0101] S3.7. Based on the graph-level vectors, perform discrimination to obtain the predicted labels;

[0102] S3.8 Calculate the loss based on the predicted label and the true label to obtain the second loss value;

[0103] S3.9 Calculate the final loss value based on the first loss value and the second loss value;

[0104] S3.10. Perform backpropagation on the detection model based on the final loss value, update the parameters of the detection model, and obtain the trained detection model.

[0105] In some feasible embodiments, step S3.2 specifically includes:

[0106] The dual-view framework includes a graph attention network and a graph isomorphism network;

[0107] In this framework, Graph Attention Networks (GAT) are used to capture the feature view, while Graph Isomorphism Networks (GIN) are used to capture the structure view. The feature view focuses on the semantic and sentiment content of user interactions, while the structure view emphasizes the topological structure of the propagation graph. Within this framework, the principles of contrastive learning are used to optimize the representations of these two views, ensuring that shared information is captured while maintaining their uniqueness, thus providing a comprehensive representation of the rumor propagation graph.

[0108] S3.2.1 Calculate the attention coefficients between a node and its neighboring nodes based on the graph attention network, and output the feature view features of the node;

[0109] Graph attention networks employ a self-attention mechanism, which fully considers the importance of its first-order neighbors by assigning weights. For a given input node, the graph attention network uses formula (1) to calculate the attention coefficient between the node and its neighboring nodes to determine the importance of the neighboring nodes to the node. The higher the attention coefficient value, the more important the node is.

[0110]

[0111] Where, x i and x j The feature vectors representing nodes i and j, · T The symbol represents the transpose operation, || is the concatenation operation, φ represents the activation function (such as the LeakyReLU function or the Sigmoid function), and a T is a weight vector, W is a shared weight matrix, and N i It is the set of first-order neighbors of node i in the k-th view.

[0112] The output feature of node i in the k-th view of the graph attention network. As shown in formula (2).

[0113]

[0114] Here, σ(.) is a non-linear activation function.

[0115] S3.2.2 Based on the aforementioned co-directional network, set the parameter vector and output the topological view features of the nodes.

[0116] Graph isomorphic networks have been shown to achieve the same expressive power as the Weisfeiler-Lehman test, thus effectively utilizing the structural features of graphs. Unlike the node update rules of graph attention networks, graph isomorphic networks use a learnable parameter vector ∈ to control the importance of self-node features during node updates. Then, a multilayer perceptron (MLP) is employed as the node update function. Because MLPs can non-linearly combine and map multiple neurons, they can capture richer feature information and complex interaction patterns in node representations, thereby improving the model's expressiveness.

[0117] The graph isomorphic network updates the output feature of node i in the k-th view. As shown in formula (3).

[0118]

[0119] Here, ∈ is a learner parameter.

[0120] Next, the matrix formed by integrating the feature views and topological views of all nodes in the k-th view scale graph is represented as follows: and

[0121] In some feasible embodiments, step S3.3 specifically includes:

[0122] This model introduces an attention mechanism to finely evaluate the contribution of nodes with different hop counts to the final representation. By calculating the attention scores of nodes with different hop counts, normalizing them, multiplying them by their corresponding feature representations, and summing the results across all scales, a comprehensive node representation can be obtained, thus balancing the importance of different scales.

[0123] S3.3.1, Introduce an attention mechanism;

[0124] S3.3.2 Calculate the attention scores of nodes at different scales;

[0125] Taking node i in the feature view as an example, its node at the k-th hop is represented as First, a nonlinear transformation is applied to the node representation to enhance the model's expressive power, resulting in attention coefficients. This process can be represented by the following formula:

[0126]

[0127] Where q is the attention vector, W is the weight matrix, and b is the bias vector.

[0128] S3.3.3 Normalize the attention score to obtain the normalized attention score;

[0129] Subsequently, these K attention scores are normalized using formula (5). The higher the normalized attention coefficient value, the more important the corresponding node feature.

[0130]

[0131] S3.3.4. Based on the normalized attention score, the feature view features of the node at different scales are weighted and summed to obtain the first fused feature.

[0132] The final representation of node i in the feature view is obtained by performing a weighted summation according to formula (6).

[0133]

[0134] Based on the above process, the final node representation of node i in the topological view can be obtained similarly.

[0135] Finally, the output representation Z, consisting of all nodes in the feature view and the topology view, is obtained through integration. F and Z T .

[0136] In some feasible embodiments, step S3.5, the formula for calculating the first loss value is expressed as follows:

[0137]

[0138] in, Functions represent calculations and The similarity is calculated using cosine similarity; τ represents the scaling parameter, which controls the scaling of the similarity score. This represents the first fusion feature of node i. This represents the second fusion feature of node i. This represents the first fusion feature of node j. Let j represent the second fusion feature of node j, and n represent the total number of nodes.

[0139] In this formula, the numerator calculates the similarity between the feature view and the structural view of the same node, while the summation in the denominator sums the similarity scores of all negative sample pairs, ensuring that the model not only brings the views of positive sample pairs closer together, but also pushes the views of all negative sample pairs further apart.

[0140] In this embodiment, the loss function is used to adjust the similarity between the feature view and the structure view. Specifically, the feature view and structure view of each node are considered positive sample pairs, while combinations with views from other nodes are considered negative sample pairs. The InfoNCE loss function encourages the model to learn that the feature view and structure view of the same node are closer in the embedding space, while keeping different nodes of the same type further apart in the embedding space. The learned representations can better distinguish different nodes, improving the model's generalization ability. Finally, this model reduces the weight of the contrastive loss in the final loss, retaining the uniqueness of the remaining information while fully utilizing and integrating the shared information of the two views.

[0141] In some feasible embodiments, step S3.6 specifically includes:

[0142] S3.6.1 Integrate the first fusion feature and the second fusion feature respectively to obtain the first fusion feature matrix and the second fusion feature matrix;

[0143] The output representation Z is obtained by integrating all nodes in the feature view and the topology view. F and Z T .

[0144] S3.6.2 Perform mean and maximum pooling operations on the first fusion feature matrix and the second fusion feature matrix, and concatenate them with the vector of the root node of the propagation graph to obtain a graph-level vector.

[0145] For Z F and Z T These two matrices are pooled using mean and maximum values, and then concatenated with the vector of the root node of the rumor propagation graph to form the final graph-level vector Z. G .

[0146] Z G =MEAN(Z) F )+MAX(Z F )+MEAN(Z T )+MAX(Z T )+X root #(8)

[0147] In some feasible embodiments, step S3.7 specifically includes:

[0148] The graph-level vectors are fed into the MLP layer and then normalized using softmax to obtain the predicted labels. Let C represent the probability of the graph belonging to each category, and let C represent the type of rumor.

[0149]

[0150] In some feasible embodiments, step S3.8 specifically includes:

[0151] A supervised loss function was constructed by calculating the cross-entropy between the model's predicted values ​​and the true label categories in the dataset using formula (10).

[0152]

[0153] In some feasible embodiments, step S3.9 specifically includes:

[0154] The total loss is a weighted sum of the contrast loss and the supervised classification loss, as shown in formula (11).

[0155]

[0156] Where λ is the hyperparameter of the weight.

[0157] In some feasible embodiments, the method further includes dividing a test set and using the test set to validate the model, specifically as follows:

[0158] Obtain the text content corresponding to each test event in the test set, preprocess all the text content to obtain the nodes, adjacency matrix, and node features of the test events; input the adjacency matrix and node features into the trained graph neural network for authenticity judgment to obtain the predicted authenticity label.

[0159] Based on the above scheme, this invention utilizes the synergistic effect between multi-scale feature extraction and attention mechanisms to ensure that the social impact of different distances is appropriately considered during rumor detection. It employs a dual-view approach—a feature view and a structural view—to process the rumor propagation graph. Simultaneously, it utilizes contrastive learning to quantify the degree of information sharing between the two views, encouraging the learned representations to be critical and diverse. Through these two mechanisms, the accuracy of rumor identification can be effectively improved.

[0160] The problems addressed include: constructing nodes at different scales to capture social impact at different distances, thus solving the problem that graph neural networks cannot capture social impact over long distances; and utilizing contrastive learning and dual views to overcome the limitations of previous methods in fully capturing the characteristics of rumor propagation.

[0161] like Figure 4 As shown, a rumor detection system based on multi-scale and dual-view includes:

[0162] The data acquisition module is used to acquire tweet data;

[0163] The preprocessing module constructs a propagation graph and generates a training set based on the tweet data;

[0164] The model building module is used to build graph neural network models based on multi-scale and dual-view frames to obtain the detection model;

[0165] The training module trains the detection model based on the training set to obtain a trained detection model.

[0166] The application module is used to apply the trained detection model to the rumor detection task.

[0167] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0168] A rumor detection device based on multi-scale and dual-view:

[0169] At least one processor;

[0170] At least one memory for storing at least one program;

[0171] When the at least one program is executed by the at least one processor, the at least one processor implements a rumor detection method based on multi-scale and dual-view as described above.

[0172] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0173] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a rumor detection method based on multi-scale and dual-view as described above.

[0174] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0175] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A rumor detection method based on multi-scale and dual-view, characterized in that, Includes the following steps: Get tweet data; Based on the tweet data, a propagation graph is constructed and a training set is generated; A graph neural network model based on a multi-scale framework and a dual-view framework is constructed to obtain the detection model; The detection model is trained based on the training set to obtain a trained detection model; The trained detection model is then applied to a rumor detection task. The step of training the detection model based on the training set to obtain the trained detection model specifically includes: Construct an extended graph from the propagation graph in the training set; Based on the dual-view framework, information in the extended graph is captured, node features are updated, and feature view features and topology view features are obtained. An attention mechanism is introduced to fuse the feature view features of different scales in the node to obtain the first fused feature; Based on the attention mechanism, the topological view features of different scales in the nodes are fused to obtain the second fused feature; The InfoNCE loss function value is calculated based on the first fusion feature and the second fusion feature to obtain the first loss value; Based on the first fusion feature and the second fusion feature, graph-level vectors are constructed using graph pooling; Based on the graph-level vectors, the predicted labels are obtained; The second loss value is obtained by calculating the loss based on the predicted label and the true label; Calculate the final loss value based on the first loss value and the second loss value; The detection model is backpropagated based on the final loss value to update the parameters of the detection model, thus obtaining the trained detection model. The step of capturing information in the extended graph based on the dual-view framework, updating node features, and obtaining feature view features and topology view features specifically includes: The dual-view framework includes a graph attention network and a graph isomorphism network; The attention coefficients between a node and its neighboring nodes are calculated based on the graph attention network, and the feature view features of the node are output. Based on the graph isomorphic network, a parameter vector is set, and the topological view features of the nodes are output. The step of constructing graph-level vectors using graph pooling based on the first fusion feature and the second fusion feature specifically includes: The first fusion feature and the second fusion feature are integrated respectively to obtain the first fusion feature matrix and the second fusion feature matrix; The first fusion feature matrix and the second fusion feature matrix are subjected to mean and maximum pooling operations, and then concatenated with the vector of the root node of the propagation graph to obtain a graph-level vector.

2. The rumor detection method based on multi-scale and dual-view according to claim 1, characterized in that, The step of constructing a propagation graph and generating a training set based on the tweet data specifically includes: Based on the tweet data, events and tags are defined to obtain preprocessed data; The preprocessed data is encoded to obtain tweet features; Based on the tweet characteristics, a propagation graph is constructed according to the propagation process of the event, and a training set is obtained.

3. The rumor detection method based on multi-scale and dual-view according to claim 2, characterized in that, Also includes: The tweet data is segmented into words, and stop words and special symbols are removed.

4. The rumor detection method based on multi-scale and dual-view according to claim 1, characterized in that, The step of introducing an attention mechanism to fuse the feature view features of different scales in a node to obtain a first fused feature specifically includes: Introduce an attention mechanism; Calculate the attention scores for nodes at different scales; The attention score is normalized to obtain the normalized attention score; Based on the normalized attention score, the feature view features of the node at different scales are weighted and summed to obtain the first fused feature.

5. The rumor detection method based on multi-scale and dual-view according to claim 1, characterized in that, The formula for calculating the first loss value is as follows: in, Functions represent calculations and similarity, This represents the scaling parameter. This represents the first fusion feature of node i. This represents the second fusion feature of node i. This represents the first fusion feature of node j. Let j represent the second fusion feature of node j, and n represent the total number of nodes.

6. A rumor detection system based on multi-scale and dual-view architecture, characterized in that, For performing the rumor detection method based on multi-scale and dual-view as described in claim 1, comprising: The data acquisition module is used to acquire tweet data; The preprocessing module constructs a propagation graph and generates a training set based on the tweet data; The model building module is used to build graph neural network models based on multi-scale and dual-view frames to obtain the detection model; The training module trains the detection model based on the training set to obtain a trained detection model. The application module is used to apply the trained detection model to the rumor detection task.

7. A rumor detection device based on multi-scale and dual-view architecture, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a rumor detection method based on multi-scale and dual-view as described in any one of claims 1-5.

Images