Rumor detection method

Abstract

The invention discloses a rumor detection method. The rumor detection method comprises a graph construction enhancement stage, a pre-training stage and a fine adjustment stage. According to the method, the propagation characteristics of rumors are deeply mined by constructing the two-way graph; and constructing a top-down propagation graph based on the original data features and the adjacent matrixes corresponding to the original data features, performing transposition operation on the adjacent matrixes, and generating a bottom-up scatter diagram in combination with the original data features. In a pre-training stage, a graph isomorphic network is adopted to generate potential representation of each graph view, and an objective function is designed according to an information bottleneck theory and a contrast learning normal form so as to reduce negative effects caused by noise. And entering a fine tuning stage, and realizing variational reasoning and supervised comparative learning by utilizing graph embedding obtained in the pre-training stage, so as to enhance the uncertainty robustness of the model to noise, and improve the classification precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of rumor detection, and in particular relates to a rumor detection method. Background Art

[0002] The rapid development of social media has exacerbated the spread of rumors, and the demand for effective rumor detection methods has become more intense. Previous research on rumor detection has mainly adopted machine learning techniques, which focus on identifying rumors based on text content, including source posts and their interactions. Methods such as decision trees and support vector machines have been used for this purpose. However, these methods often ignore the structure of rumor propagation, which has been shown to be crucial for effective rumor identification. In addition, these methods rely heavily on time-consuming and labor-intensive feature engineering.

[0003] Recently, many deep learning methods, such as recurrent neural networks and convolutional neural networks, have been successfully used in the rumor detection problem due to their powerful ability to model high-level representations. Among them, graph neural networks (GNNs) have attracted widespread attention due to their ability to encode features through topological structures and thus simulate the spread between entities. GNN-based methods make full use of the topological relationships between nodes and their neighbors and are particularly effective in learning the potential interaction representations of nodes, which helps capture important information about the behavior patterns of nodes.

[0004] Despite the achievements of GNNs in rumor detection, many studies have failed to address the interpretability of these models, which limits their effectiveness due to the black-box nature of GNNs. First, data from various sources often contain large uncertainties and noise, as data can come from any accessible source. For example, a tweet may attract supportive, opposing, and irrelevant comments. This complex combination of responses makes it more difficult to evaluate the accuracy of the original post, resulting in less robust performance of existing models. Second, unlike large language models (LLMs) trained on extensive and diverse datasets, rumor detection models are often developed and trained on smaller and more limited datasets. These datasets may not account for all forms of noise or irrelevant information present in the original data, which affects the model's generalization and ability to handle interference, resulting in models trained on limited datasets being more easily misled by noise during training. Therefore, a deep understanding of how noise affects message propagation is crucial to improving the interpretability and noise-resistance performance of the model.

[0005] In addition, existing methods have the following problems: (1) In social media, noise in posts and events can significantly disrupt the propagation network. When applying GNN for rumor detection, noisy messages are mixed with normal posts through the GNN operator, which damages the feature representation of normal posts. The disturbed representation reduces the model's ability to generalize effectively. (2) Extracting meaningful and relevant information from the original features is crucial for accurate rumor detection. Although high-dimensional vectors are often used to represent node attributes, not all dimensions have the same effect on detecting rumors. High-dimensional features complicate the detection process because some dimensions may exhibit ambiguous or nonlinear relationships with the features being analyzed, which may introduce noise. (3) Noise has a significant impact on both the training process and the detection results. Although interpretable methods can reduce the impact of noise, the abstraction of the prediction results may generate uncertainty. This means that the noise in the data distribution increases the uncertainty of predictions for different categories. It is very important to quantify and evaluate distribution characteristics to reduce noise in the decision process and improve the confidence of model predictions. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the purpose of an embodiment of the present invention is to provide a rumor detection method to improve the noise resistance and interpretability of rumor detection; effectively filter and manage noise in complex social media data, while ensuring that rumors can be accurately and reliably generalized and predicted.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is a rumor detection method, which first constructs a top-down propagation graph and a bottom-up scatter graph of an event, and performs edge perturbation, subsampling and re-parameterized node mask enhancement on the bidirectional graph; the encoder performs multi-level feature capture and fusion on the enhanced graph to obtain three enhanced representations of the bidirectional graph, and pre-trains the encoder to learn the discriminative representation of rumors to obtain a pre-trained encoder; the pre-trained encoder outputs enhanced graph features with discriminability;

[0008] Then, a classifier model is deployed, and the enhanced graph features with discriminability are passed through the classifier model. Variational inference and supervised contrastive learning are introduced to fine-tune the classifier model to obtain the final objective function; finally, the classifier model is guided by the final objective function to output the prediction result.

[0009] Furthermore, the process of constructing the event propagation graph and scatter graph is specifically as follows:

[0010] make Represents a set of events, where c i represents the i-th event, m is the number of events; among them,

[0011]

[0012] Among them, s i Indicates the original post, represents the jth interactive post in the i-th event, n i is the number of posts in the entire event, Indicates event c i social communication network, among which

[0013]

[0014] The node set Edge Set They represent the set of posts in the i-th event, the edge set from the replied posts to the responded or forwarded posts, represents the edge from the pth node to the qth node in the i-th event, express The adjacency matrix of Refers to the matrix A i The dimension is n i ×n i ;

[0015] Among them, if and only if hour, otherwise is the adjacency matrix A i The element at row p and column q; Represents the set of all post nodes in the i-th event The characteristic matrix of , d is the node vector dimension, Indicates the nth i Posts Features;

[0016] Each event c i Top-down propagation diagram of Bottom-up scatter plot in Represents the adjacency matrix A i The transposed matrix of .

[0017] Furthermore, the specific process of edge disturbance enhancement is as follows:

[0018] A ep =A⊙M insert +(1-A)⊙M drop

[0019] Among them, M insert is the new perturbation edge generated in the graph structure, M drop is the edge in the original moment adjacency matrix A that is deleted; ⊙ is the Hadamard product, A eprepresents the adjacency matrix after edge perturbation; the enhanced subgraph after edge perturbation enhancement is R n×n Refers to the matrix A ep The dimension is n×n, d is the node vector dimension, Represents the set of all post nodes in the i-th event The feature matrix of .

[0020] Furthermore, the specific process of subsampling enhancement is as follows: introduce a subgraph sampling function into the bidirectional graph, gradually sample from the root node through random walks of nodes, spread to different neighbors, and set the iteration time to generate a sampling subgraph as Where n' is the number of sampling nodes, R n×n It means that the dimension of matrix A is n'×n', d is the node vector dimension, and X represents the feature matrix of the set of all post nodes in the event.

[0021] Furthermore, the process of reparameterized node mask enhancement is as follows: for the original graph of any event graph Create a weight matrix W = [w1,w2,…,w n ], where w j (j=1,2,…,n)∈[0,1] represents the transformed sampling probability, j represents the jth post, and n is the number of posts in the entire event; then define a key r for each node j :

[0022] r j =softmax(log(w j )+g j )

[0023] where g j is random noise, softmax(·) is a normalization function, and then the n keys are sorted, and the first K keys are selected as the subgraph generated, and the corresponding K nodes are used as the output of the node reparameterization sampling, and finally the reparameterized node mask enhanced subgraph is obtained Where n" is the sampled node, A is the adjacency matrix, R n″×d The dimension of matrix A is n"×n", d is the node vector dimension, and X represents the feature matrix of all post node sets in the event.

[0024] Furthermore, the process of capturing and fusing multi-level features of the enhanced image through the encoder is as follows: Then a three-layer GIN encoder is used to capture the multi-level potential features of the three enhanced images:

[0025]

[0026] where φ(·) is the LeakyReLU activation function, A is the adjacency matrix, CONCAT(·,·) is the concatenation operation, represents the root node feature of the lth layer, M l+1 represents the root node feature of the l+1th layer, H l and H l+1 Represent the node feature matrices of the lth layer and the l+1th layer respectively, and b l are the weight parameters and bias parameters of the l-th layer encoder network; after encoding, three enhanced representations of the propagation map and the scatter map are obtained.

[0027] Furthermore, the total pre-training objective function of learning the discriminative representation of rumors through pre-training for:

[0028]

[0029] Among them, η and μ are hyperparameters, is the unsupervised alignment loss function:

[0030]

[0031] is the sampler loss function:

[0032]

[0033] is the mutual information loss function:

[0034]

[0035] Among them, β is a hyperparameter, p(H|X) represents the encoded node features, is the KL divergence, represents the expected distribution with parameter θ, p(H) is a random Gaussian distribution, is a set of augmented graphs, α o For the weight of each enhanced graph o, Enhance the root features of graph o, f θ (X) is the embedded representation of the enhanced graph; p(H|f θ (X)) represents the latent features encoded from the original view, represents the latent features encoded from the sampled graph, It means to find the expected value of function f for event i, g θ (·,·) represents the scaled cosine similarity function, Enhanced local representation of view, H j A graph representing all samples; H tIt represents the graph representation of all the remaining samples after removing the target sample i.

[0036] Furthermore, the deployment process of the classifier model is specifically as follows: using the pre-trained weight vector W, according to the calculated key value, sampling the first K nodes from the original node feature X, eliminating the propagation of noise information, and obtaining the denoised graph representation H of the propagation view of the pre-trained encoder P and the denoised graph representation of the scatter view H D ;

[0037] Then, the representations of the two propagation views are concatenated to generate a comprehensive representation of the event graph. And deploy a Bayesian layer to learn the distribution:

[0038]

[0039] Where Bay(·|W b ) represents the Bayesian network, W b is the learnable parameter of the Bayesian network, σ(·) is the PreLU activation function, H b represents the comprehensive representation after encoding; OCNCAT(·) is the concatenation function;

[0040] Use original root features and their graph-level representation as auxiliary evidence for rumor detection:

[0041]

[0042] in, is the comprehensive representation of the original features, X root and Represent the original root feature and its graph-level representation respectively; two Beyasian layers are used to obtain the transformed representation of the original feature:

[0043] Xb +1 =σ(Bay b (X b |W b ))

[0044] Bay b (·) represents the b-th layer Bayesian network, W b represents the weight parameter of the b-th layer Bayesian network, X b and X b+1 Represent the input node features of the b-th layer and the b+1-th layer Bayesian network respectively;

[0045] Finally, the two parts of the high-quality representation are concatenated and a Beyasian layer with normalization is used to achieve the detection task:

[0046]

[0047] in To predict the label, Bay c (·) represents the output layer based on the Bayesian network, W c is the weight parameter of the output layer; c represents the last layer, i.e. the output layer, and b represents the intermediate layer.

[0048] Furthermore, the final objective function is specifically:

[0049]

[0050] where y i is the true label of event i, is the predicted label of event i, λ and κ are the weight parameters of the objective function, m is the number of events, is the objective function for supervised contrastive learning;

[0051] The variational inference is specifically to fit the rescaled prior distribution to the posterior distribution through KL divergence:

[0052]

[0053] Where q(ω|θ,H) is the posterior distribution of the Bayesian network parameters, p(ω) is the scaled prior distribution of the Bayesian network parameters, represents the variational inference loss function, It means to find the KL divergence between distribution p and distribution q

[0054] The objective function of supervised contrastive learning is for:

[0055]

[0056] in, It means to find the expectation of batch B, f k (x t ,θ) represents a classification model with parameter θ, x t represents the potential embedded X and H b The splicing features, x j is a single sample in batch B.

[0057] Compared with the existing technology, the rumor detection method of the present invention has demonstrated excellent capabilities in processing complex and variable data sets. First, the present invention has a powerful noise processing mechanism. By integrating multiple enhancement strategies such as node samplers, it effectively eliminates interference from irrelevant or low-quality information, thereby improving the generalization ability and robustness of the overall model. In addition, the method combines contrastive learning with information bottleneck theory to extract key features from complex social network graphs. These features not only help reduce the negative impact of noise, but also further enhance the transparency and interpretability of the model decision-making process.

[0058] In order to further improve the performance of the model under uncertain conditions, the present invention also introduces a Bayesian network based on variational inference to fine-tune the model. This measure enables the model to show higher stability and reliability when dealing with various uncertainty problems caused by noise. Experimental results show that on two widely used benchmark datasets, Twitter15 and Twitter16, the present invention not only significantly improves the accuracy of rumor detection, but also shows high robustness and stability in combating noise interference, which is significantly better than the existing technical solutions. The present invention provides a new, more reliable and easy-to-understand approach for social media rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 is a flow chart of the rumor detection method in this embodiment;

[0061] Figure 2 is a diagram of ablation experiment results of the method of this embodiment;

[0062] Figure 3 1 is a graph of early rumor detection results of this embodiment; (a) is the result on the Twitter15 dataset, and (b) is the result on the Twitter16 dataset;

[0063] Figure 4 is a visualization diagram of the interpretability of external noise during the pre-training process of this embodiment; (a) is a propagation view, and (b) is a scattering view;

[0064] Figure 5It is a visualization diagram of the interpretability of endogenous noise during pre-training; (a) is a schematic diagram of the embedded feature changes of the first selected node, (b) is a schematic diagram of the embedded feature changes of the second selected node, (c) is a schematic diagram of the embedded feature changes of the third selected node, and (d) is a schematic diagram of the embedded feature changes of the fourth selected node;

[0065] Figure 6 3 is a visualization diagram of the interpretable fine-tuning process of the present embodiment for noise; (a) is a feature view of the 0th iteration of the fine-tuning process, (b) is a feature view of the 10th iteration of the fine-tuning process, (c) is a feature view of the 20th iteration of the fine-tuning process, and (d) is a feature view of the 30th iteration of the fine-tuning process. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] This embodiment provides a rumor detection method, specifically, it uses an interpretable rumor detection model to address the problems existing in the prior art. This embodiment adopts multiple enhancement strategies to generate multiple enhanced graph views, especially using the method of reparameterized sampling to eliminate noise nodes, thereby enhancing the information propagation process in the graph. This embodiment uses a three-layer graph encoder to capture multi-level representation information from the graph. In order to reduce endogenous noise and obtain high-quality representation, this embodiment implements contrastive learning based on the information bottleneck (IB) theory. Unlike traditional self-supervised learning methods, the IB theory retains the basic information of the original data while compressing the noise, thereby improving the interpretability of the potential embedding space. Finally, based on the pre-trained model, the variational inference method is used to fine-tune the Bayesian network to improve the confidence of the model in handling uncertainty.

[0068] First, the rumor detection problem is defined as Represents a set of events, where c i represents the i-th event, and m is the number of events. where s i Indicates the original post, represents the jth interactive post in the i-th event, n i is the number of posts in the entire event. Indicates event c i Specifically, The node set Edge Set denote the set of posts in the ith event and the edge set from the replied posts to the responded or forwarded posts, respectively. represents the edge from the pth node to the qth node in the i-th event, express The adjacency matrix of Specify the matrix A i The dimension is n i ×n i . If and only if hour, otherwise is the matrix A i The element at row p and column q. Represents the set of all post nodes in the i-th event The feature matrix of , d is the node vector dimension. Specifically, if Reply or repost and So and There will be a directed arrow pointing to For example, and A i middle nuclear The value of will be set to 1, Represents the set of all post nodes in the i-th event The node feature matrix of , d is the node vector dimension, where x1 represents the root node feature, x j (2≤j≤n i -1) indicates the jth post For ease of description, this embodiment is described below using

[0069] In this embodiment, each event c i Top-down propagation diagram of and bottom-up scatter plots Contains different key information for rumor detection, including Indicates A i To simplify the representation, the subscript i is omitted in the following text of this embodiment.

[0070] In the implementation, each event is attached with a label y, which belongs to one of four classes {N, F, T, U}. N, F, T, U represent non-rumor, false rumor, true rumor, and unconfirmed rumor, respectively. According to the above description, the goal of the rumor detection problem is to learn a classifier f(·) that combines content embedding with structural information to predict The label of each event in .

[0071] like Figure 1 , the rumor detection method described in this embodiment mainly includes a graph construction stage, a pre-training stage and a fine-tuning stage. Traditional detection methods usually only focus on the one-way rumor propagation structure, but ignore important information from different propagation modes, which may lead to a decrease in the utilization rate of discriminatory interactive information. Propagation graphs and scatter graphs contain different key information for rumor detection, which have been proven to be effective for rumor detection. This embodiment establishes a bidirectional graph (propagation graph and scatter graph) to explore the deep propagation characteristics of rumors. Specifically, the corresponding top-down propagation graph is obtained through the original data features and the corresponding adjacency matrix, and then the adjacency matrix is ​​transposed, and the bottom-up scatter graph is obtained in combination with the original data features.

[0072] In the pre-training stage, this implementation uses a graph isomorphism network to generate a latent representation of each graph view. Then, a detailed objective function is designed based on the IB theory and the contrastive learning paradigm to mitigate the harmful effects of noise. In the fine-tuning stage, this implementation uses the pre-trained embeddings of the graph to implement variational reasoning and supervised contrastive learning to enhance the model's robustness to noise-induced uncertainty and improve classification accuracy.

[0073] In some possible implementations, the present implementation enhances the propagation graph and the scatter graph through three enhancement strategies to generate three types of enhanced graphs, namely edge perturbation, subsampling, and parameterized node masking, to generate various pairs full of attribute and topological information.

[0074] S1. Construction and enhancement of bidirectional graphs

[0075] S101, edge disturbance

[0076] In order to solve the problem that the repeated spread of rumors is more likely to be reinforced by the public as a reliable and realistic information, this implementation method implements edge perturbations on both the propagation graph and the scatter graph to alleviate the problem of repeated message propagation, which is defined as follows:

[0077] A ep =A⊙M insert +(1-A)⊙M drop

[0078] Among them, M insert ~Bernoulli(pedge ), M drop ~Bernoulli(p edge ), indicating that the parameter is p edge A random variable under the Bernoulli distribution, where M insert represents the random insertion of values ​​in the original adjacency matrix, that is, generating new perturbation edges in the graph structure, then M drop Indicates deleting the edges in the original matrix, ⊙ is the Hadamard product, A ep represents the adjacency matrix after edge perturbation. For any event graph (event subscript omitted), the original graph can be expressed as Then its enhancer graph is

[0079] S102, sub-image sampling

[0080] Compared with the entire life cycle of rumors, rumors in the early stages have different characteristics. Due to the lack of factual evidence, people are easily influenced by rumors in the early stages. To address this potential phenomenon, this implementation introduces the subgraph sampling function of the Python-based graph learning library pytorch_geometric, which gradually samples from the root node through random walks of nodes, spreads to different neighbors, and sets a specific iteration time. The sampling probability is p sub For any event graph (omitting event subscripts), the original graph can be expressed as Then its sampling subgraph is Where n' is the number of sampling nodes.

[0081] S103, re-parameterized node sampling

[0082] Most existing studies rely only on random node sampling and fail to filter out noisy nodes in the graph. This implementation solves this problem through parameterized top-K sampling, making the graph representation more discriminative. Specifically:

[0083] For any event graph (event subscripts omitted), the original graph can be expressed as Create a weight matrix W = [w1,w2,…,w n ], where w j (j=1,2,…,n)∈[0,1] represents the transformed sampling probability. Then define a key for each node

[0084] r j =softmax(log(w j )+g j )

[0085] where g j~Gumbel(0,1), indicating random noise that follows the Gumbel distribution, and softmax(·) is a normalization function. Then sort the n keys, select the first K keys as the subgraph, and let the corresponding K nodes be the output of the node reparameterization sampling, and finally get Where n" is the node after sampling, that is, |n"|=K.

[0086] This embodiment defines differentiable discrete distribution sampling by reparameterizing the samples into a deterministic transformation of uncertain noise g and parameter W, so that the sampling process can be learned through back propagation. This embodiment can iteratively learn to avoid noise nodes when sampling, thereby creating a noise-diluted enhanced graph.

[0087] S104, multi-level feature capture and fusion of the enhanced image through the encoder

[0088] GNN can extract the potential features represented by nodes based on the graph topology; this implementation uses GIN as a graph encoder. Compared with traditional GNN (such as GCN), the GIN network can capture and distinguish different subgraphs in the graph structure, more accurately represent the features of nodes and graphs, and has good generalization performance and flexibility.

[0089] For a single view, this embodiment generates an enhanced graph by edge perturbation, subsampling, and parameterized node sampling. and Then a three-layer GIN encoder is used to capture the multi-level latent features of the enhanced graph. Specifically, for the GIN encoder of the lth layer, since the root node retains a lot of basic information, the encoded embedding features are concatenated with the features of the root node, which helps to learn more accurate node representations for the rumor detection problem. The process can be expressed as:

[0090]

[0091]

[0092] where φ(·) is the LeakyReLU activation function, A is the adjacency matrix, CONCAT(·,·) is the concatenation operation, represents the root node feature of the lth layer, M l+1 represents the root node feature of the l+1th layer, H l and H l+1 Represent the node feature matrices of the lth layer and the l+1th layer respectively, and b l are the weight parameters and bias parameters of the lth encoder network. Note that when l is equal to 0, H0 and Represents the original node feature matrix X of the event and the original root node feature of the event That is, H0 = X;

[0093] After encoding, three enhanced representations of propagation and scatter diagrams are obtained.

[0094] S2. Pre-training

[0095] This implementation method uses the pre-training method based on IB theory to learn the discriminative representation of rumors. IB theory understands the deep learning process as a process of encoding and decoding. The training goal of IB theory is to minimize the mutual information between the input and the encoded embedding and maximize the mutual information between the output and the encoded embedding:

[0096]

[0097] Where β is a trade-off parameter, IN, MID, and OUT represent the input, encoded embedded features, and output, respectively. In this way, the output can retain the most important and representative information of the input. It means finding the value of variable H that minimizes the objective function f(H).

[0098] Based on the above formula, in order to get rid of the dependence on the real label, this implementation regards the enhanced graph as the label. The optimization objective function is to maximize the global graph representation H and its enhanced representation The mutual information between the global graph representation H and the local node embedding X is minimized. This embodiment optimizes the above equation by the following formula The upper bound of :

[0099]

[0100] in is the mutual information loss function, β is a hyperparameter, p(H|X) represents the encoded node features, is the KL divergence, represents the expected distribution with parameter θ, p(H) is a random Gaussian distribution, is a set of augmented graphs, α o For the weight of each enhanced graph o, Enhance the root features of graph o, f θ (X) is the embedding representation of the enhanced graph. This approach can preserve the discriminative information between the central node and its propagation graph view and identify invariants in the rumor propagation pattern.

[0101] In order to improve the ability of the reparameterized node sampler to learn node classification information, this embodiment adopts a sampler loss function to eliminate irrelevant nodes that contribute little in the message propagation process. Sampler loss function The definition is as follows:

[0102]

[0103] where p(H|f θ (X)) represents the latent features encoded from the original view, represents the latent features encoded from the sampled graph, represents the sampler loss function.

[0104] In order to further eliminate the noise in different augmented graph views and improve the consistency of augmented samples, this implementation introduces a node-level alignment strategy. The original graph view and its augmented graph view are regarded as positive pairs, and different other samples are used to generate negative pairs. Specifically, for the local representation h of event i, i And its corresponding graph represents H i , unsupervised alignment loss function It can be formulated as follows:

[0105]

[0106] in, It means to find the expected value of function f for event i, g θ (·,·) represents the scaled cosine similarity function, Enhanced local representation of view, H j The graph representation of all samples, H t (t≠i) represents the graph representation of all remaining samples after removing the target sample i. The first term in the formula is to reduce the noise between the original image and its augmentation, while making different samples more distinguishable. The second term is to generate a more exclusive representation of the target local representation relative to other samples.

[0107] Finally, this implementation incorporates the sampler loss function into the pre-training objective, and the total pre-training objective function It can be defined as follows:

[0108]

[0109] where η and μ are hyperparameters.

[0110] S3. Classification-based fine-tuning

[0111] Through pre-training, denoised graph representations of propagation graphs and scatter graphs and pre-trained encoders can be obtained. In order to further improve the detection capability of the model and enhance the confidence in the recognition output, this implementation fine-tunes the downstream classification model, which consists of a Bayesian network and parameter perturbations. The Bayesian network is designed for decision-making problems with uncertainty, and it can provide the model with decision confidence as a valuable guide, making the model more robust and resilient.

[0112] S301. Deployment of classification model

[0113] This implementation deploys a pre-trained model to obtain high-quality representations for downstream tasks; using the pre-trained weight vector W, the first K nodes are sampled from the original node features X according to the calculated key values, thereby eliminating the propagation of noise information. Through masking, the denoised graph representation H of the propagation view encoded by the pre-trained GIN model can be obtained. P and the denoised graph representation of the scatter view H D .

[0114] Then, the representations of the two propagation views are concatenated to generate a comprehensive representation of the event graph. And deploy the Bayesian layer to learn the distribution, the formula is as follows:

[0115]

[0116] Where Bay(·|W b ) represents the Bayesian network, W b is the learnable parameter of the Bayesian network, σ(·) is the PreLU activation function, H b Represents the comprehensive representation after encoding.

[0117] In addition, in order to obtain more usable information, this implementation also uses the original root features and their graph-level representation as auxiliary evidence for rumor detection:

[0118]

[0119] in, is the comprehensive representation of the original features, X root and Represent the original root feature and its graph-level representation respectively. Two Beyasian layers are used to obtain the transformed representation of the original feature:

[0120] X b+1 =σ(Bay b (X b |W b ))

[0121] Bay b (·) represents the b-th layer Bayesian network, W b represents the weight parameter of the b-th layer Bayesian network, X b and X b+1 They represent the input node features of the b-th layer and the b+1-th layer Bayesian network respectively. When b=0, X0 represents the comprehensive representation of the original features, that is,

[0122] Finally, the two parts of the high-quality representation are concatenated and a Beyasian layer with normalization is used to achieve the detection task:

[0123]

[0124] in To predict the label, Bay c (·) represents the output layer based on the Bayesian network, W c is the weight parameter of the output layer; c represents the last layer, i.e. the output layer, and b represents the intermediate layer.

[0125] S4, anti-noise fine-tuning

[0126] In the noise-resistant fine-tuning stage, this implementation uses a pre-trained model to encode the original data to obtain discriminative embedded features, based on which downstream classification tasks are performed. In order to improve the robustness and accuracy of the classification model, a complex objective function consisting of three parts is introduced. First, in order to reduce the uncertainty of model prediction and improve the accuracy of detection, variational inference is introduced, that is, the rescaled prior distribution is fitted to the posterior distribution through KL divergence, and the formula is as follows:

[0127]

[0128] Where q(ω|θ,H) is the posterior distribution of the Bayesian network parameters, p(ω) is the scaled prior distribution of the Bayesian network parameters, represents the variational inference loss function, It means to find the KL divergence between distribution p and distribution q.

[0129] In order to promote the discrimination of positive samples from negative samples, a contrastive learning paradigm with a strong supervisory signal is used to bring samples with the same label closer together. In this way, the noise of the same category in different individuals can be minimized. The supervised contrastive learning objective function is defined as follows:

[0130]

[0131] in It means to find the expectation of batch B, f k (x t ,θ) represents a classification model with parameter θ, x t represents the potential embedded X and H b The splicing features, x j is a single sample in batch B.

[0132] Finally, by adding the cross entropy loss function, the final objective function of the fine-tuning stage is formed:

[0133]

[0134] where y i is the true label of event i, is the predicted label of event i, λ and κ are the weight parameters of the objective function, and m is the number of events.

[0135] The initial input of this implementation is a plurality of original event graphs. The training is divided into two stages: pre-training and fine-tuning. The purpose of pre-training is to filter out the embedding features with discrimination, and the purpose of fine-tuning is to further transform the embedding features according to the needs, so that the model can quickly and accurately mine the potential feature space for downstream tasks. The relationship between the two is that without the pre-training process, the performance of fine-tuning (or downstream tasks) will be greatly reduced. In the pre-training stage, Three enhanced graphs are generated through the enhancement strategy, and the enhanced graphs are encoded using GIN. The designed loss function is used to optimize the parameters of the GIN encoder to generate enhanced graph features with discrimination. The output of the pre-training stage is the trained GIN encoder. The input in the fine-tuning stage is the same as the input in the pre-training stage, which is the original event graph. First, the trained GIN encoder is used to encode the discriminative bidirectional event graph features (note that no enhancement processing is required in the fine-tuning stage. The enhancement processing is only to train discriminative features in the pre-training stage. Here, GIN has been trained and can directly obtain discriminative features). Then the encoded features are passed through the Bayesian network and fine-tuned using the designed loss function. Finally, the prediction results of the rumor category are output.

[0136] This embodiment introduces two real-world datasets to evaluate the performance of the model proposed in this embodiment: Twitter15 and Twitter16, all of which are collected from the Twitter social platform. In these two datasets, nodes represent users and edges represent social interactions, such as forwarding or replying. Each node is initialized with a 5000-dimensional vector. Both Twitter datasets include four categories: non-rumor (N), false rumor (F), true rumor (T), and unconfirmed rumor (U). The label of each event in the two datasets is annotated based on the analysis and identification of websites that debunk rumors. Table 1 lists the details of each dataset.

[0137] Table 1 Dataset information

[0138] Statistic Twitter15 Twitter16 #ofusers 276663 173487 #ofevents 1490 818 #offalserumors 370 205 #oftruerumors 374 205 #ofunverifiedrumors 374 203 #ofnon-rumors 372 205 #ofpostings 331612 204820 Avg.#ofpostings / event 223 251 Max#ofpostings / event 1768 2765 Min # of postings / event 55 81

[0139] In order to prove the effectiveness and advancement of the method of this embodiment in the problem of rumor anomaly detection, this method is compared with several existing models, including:

[0140] DTC is a decision tree-based rumor detection method that uses various comprehensive features to analyze credibility.

[0141] SVM-TS utilizes linear SVM to identify rumors by constructing time series using handcrafted features.

[0142] RvNN is a gated recurrent unit model for detecting fake news through a tree structure.

[0143] MVAE is an end-to-end fake news detection model that is able to learn shared multimodal representations.

[0144] GCN is a graph neural network that deploys convolution operations directly on graphs.

[0145] BiGCN is a fake news detection model that takes into account the propagation and diffusion patterns of fake news in the network.

[0146] GACL is a framework for rumor detection methods that utilizes supervised contrastive learning to discriminate features through adversarial enhancement.

[0147] EBGCN introduces the uncertainty caused by unreliable relations in the rumor detection propagation structure.

[0148] RDEA is a rumor detection model based on contrastive learning that deploys three enhancement strategies by modifying the attribute and structural information in news graphs.

[0149] VAE-GCN is a framework that integrates propagation and structural information for rumor reconstruction.

[0150] BiMGCL is a rumor detector that exploits multi-level contrastive learning on bidirectional graphs.

[0151] For the above methods, we implemented them through the source code provided by their developers, and the proposed models were implemented using PyTorch. All methods were executed in Python 3.9.12 environment, using a single NVIDIA A40 GPU, 40GB RAM and 2.60GHz Xeon(R) Gold 6240 CPU.

[0152] Accuracy (ACC) and F-score (F_1) are used as evaluation indicators. Specifically, ACC represents the proportion of true classification results, and F_1 represents the harmonic mean of the recall and precision indicators of the four categories.

[0153] Table 2 and Table 3 show the experimental results of the above method and the method of the present invention on Twitter15 dataset and Twitter16 dataset respectively.

[0154] Table 2 Detection results on Twitter15 dataset

[0155]

[0156] Table 3 Detection results on Twitter16 dataset

[0157]

[0158]

[0159] The results in the table show that DTC and SVM-TS rely only on handcrafted features to learn shallow representations, which causes them to perform inferior to other baseline methods. In contrast, RvNN and MVAE perform better by capturing latent high-level representations in text content, but they fail to fully utilize important nonlinear structures and relational information, which limits their performance in rumor detection tasks.

[0160] The GCN-based method has achieved remarkable results, and its performance has surpassed all traditional methods. However, under multiple evaluation indicators, it is still inferior to the method proposed in this embodiment. This method not only surpasses the existing methods on the Twitter15 and Twitter16 datasets, but also improves the accuracy of the GCN-based method by at least 4.75% and 2.23% on the Twitter16 dataset.

[0161] This implementation effectively avoids the interference of exogenous noise nodes on the model by adopting a variety of data enhancement strategies and adaptive optimization techniques. At the same time, with the help of the comparative pre-training method guided by the information bottleneck (IB) theory, a highly discriminative representation that can reduce endogenous noise is obtained. In addition, special fine-tuning is performed for classification tasks to reduce the uncertainty caused by noise and enhance the consistency of internal representations of the same category. These comprehensive measures significantly improve the performance of the model, especially in rumor detection applications.

[0162] In order to evaluate the rationality and effectiveness of each component of the method of the present invention, an ablation experiment was conducted on the model of this embodiment. The model has six variants: -IB, -SL, -USP, -VI and -CL, where -IB means removing the IB loss, -SL means removing the sampler loss, -USP means removing the unsupervised alignment loss function, -VI means removing the uncertainty loss, and -CL means removing the supervised contrast loss. The results are shown in Figure 2. Figure 2 shown.

[0163] From the observations of the ablation experiments, the absence of any of the five components leads to a drop in performance, which demonstrates the utility of the components of the model of the present invention. Among all variants, -IB and -CL show relatively large drops, which indicates that the learning paradigm of the present invention helps filter out the noise that interferes with representation learning, and that category-based contrastive learning plays a vital role in rumor detection. In addition, -SL, -USP, and -VI show smaller drops, which also demonstrates the practicality of these modules in deriving discriminative representations for rumor detection.

[0164] Early rumor detection is to evaluate the model's ability to identify rumors before they become ubiquitous. Good performance in early rumor detection is conducive to social stability. In order to test the performance of early rumor detection, this implementation sets a series of deadlines and only uses posts published before the deadline to compare with the baseline. The results are shown in Figure 3 .

[0165] From the observed results, it is clear that the proposed method has higher performance than the selected baseline at each deadline, which shows that the proposed method is not only suitable for long-term rumor detection, but also performs well in early rumor detection. In addition, the proposed method can achieve high recognition accuracy before a relatively short deadline, which also proves the effectiveness of the proposed model.

[0166] This embodiment verifies the interpretability of the method of the present invention from the pre-training and fine-tuning processes. In the pre-training stage, an enhancement strategy is deployed on the propagation graph and the dispersion graph to generate an enhanced view. Specifically, in order to verify the interpretability of the external noise exclusion capability, the distance between the encoding embedding enhanced by the node masking strategy and the encoding original graph of the corresponding propagation view and dispersion view at different iterations is visualized and analyzed. Figure 4 As shown, for both propagation views, as iterations continue, the distribution distance gradually shrinks to a very small threshold range close to zero, which means that the similarity between the original and enhanced images is enhanced. Therefore, these observations strongly demonstrate the ability and effectiveness of modeling external noise.

[0167] Finally, the model’s ability to resist node-intrinsic noise is demonstrated by comparing the features of masked nodes in different iterations. The evolution of embeddings of several selected nodes is visualized in the propagation graph and dispersion graph views, respectively. Figure 5 The first and second nodes are selected from the propagation graph, and the remaining nodes are selected from the scatter graph. Figure 5 (a) and Figure 5 (b) shows that there are significantly more elements equal to zero in the 30th iteration than in the 10th iteration. From a distribution point of view, the distribution of the 30th cycle is more compact, and its mean is closer to zero than that of the 10th cycle. Figure 5 (c) and Figure 5 The same trend can also be seen in (d). Therefore, it can be concluded that with the increase in the number of iterations, more noise nodes are masked while more discriminative and important information is retained.

[0168] During the fine-tuning phase, contrastive learning paradigm and variational inference are used to avoid noise-induced uncertainty and perform accurate classification. To further demonstrate the model’s interpretability to noise, the combined embeddings of some selected batches during training are visualized using the t-SNE method, and the results are shown in Figure 2. Figure 6 shown.

[0169] like Figure 6 As shown in (a)(b), at the 0th iteration, the integrated embedding of the graph is mixed together, and there is no clear boundary between different categories, which means that the noise in the embedding is still uncertain. After dozens of iterations, the model achieves a decrease in the intra-class distance and an increase in the inter-class distance, resulting in a clear boundary between different categories, as shown in Figure 6 (c)(d) Therefore, it can be concluded that the method of the present invention has strong interpretability to internal and external noise and uncertainty, which helps to improve the performance of rumor detection.

[0170] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0171] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A rumor detection method, characterized in that: Firstly, a top-down propagation graph and a bottom-up scattering graph of the event are constructed, and the bidirectional graph is enhanced by edge perturbation, subsampling and re-parameterized node mask; the enhanced graph is subjected to multi-level feature capture and fusion by an encoder to obtain three enhanced representations of the bidirectional graph, and the encoder is made to learn the discriminative representation of rumors by pre-training to obtain a pre-trained encoder; the pre-trained encoder outputs enhanced graph features with discriminability; Then, a classifier model is deployed, and the enhanced graph features with discriminability are passed through the classifier model. Variational inference and supervised contrastive learning are introduced to fine-tune the classifier model to obtain the final objective function; finally, the classifier model is guided by the final objective function to output the prediction result.

2. A rumor detection method according to claim 1, characterized in that: The process of constructing the event propagation graph and scatter graph is specifically as follows: make Represents a set of events, where c i represents the i-th event, m is the number of events; among them, Among them, s i Indicates the original post, represents the jth interactive post in the i-th event, n i is the number of posts in the entire event, Indicates event c i social communication network, among which The node set Edge Set They represent the set of posts in the ith event, the edge set from the replied posts to the responded or forwarded posts, represents the edge from the pth node to the qth node in the i-th event, express The adjacency matrix of Refers to the matrix A i The dimension is n i ×n i ; Among them, if and only if hour, otherwise is the adjacency matrix A i The element at row p and column q; Represents the set of all post nodes in the i-th event The feature matrix of , d is the node vector dimension, Indicates the nth i Posts Features; Each event c i Top-down propagation diagram of Bottom-up scatter plot in Represents the adjacency matrix A i The transposed matrix of .

3. A rumor detection method according to claim 1, characterized in that: The specific process of edge disturbance enhancement is: A ep =A⊙M insert +(1-A)⊙M drop Among them, M insert is the new perturbation edge generated in the graph structure, M drop is the edge in the original moment adjacency matrix A that is deleted; ⊙ is the Hadamard product, A ep represents the adjacency matrix after edge perturbation; the enhanced subgraph after edge perturbation enhancement is R n×n Refers to the matrix A ep The dimension is n×n, d is the node vector dimension, Represents the set of all post nodes in the i-th event The feature matrix of .

4. A rumor detection method according to claim 1, characterized in that: The specific process of subsampling enhancement is as follows: introduce a subgraph sampling function into the bidirectional graph, gradually sample from the root node through random walks of nodes, spread to different neighbors, and set the iteration time to generate a sampling subgraph: Where n′ is the number of sampling nodes, R n×n It means that the dimension of matrix A is n′×n′, d is the node vector dimension, and X represents the feature matrix of all post node sets in the event.

5. A rumor detection method according to claim 1, characterized in that: The process of reparameterized node mask enhancement is as follows: for any original event graph Create a weight matrix W = [w1, w2, ..., w n ], where w j (j=1, 2, ..., n)∈[0, 1] represents the transformed sampling probability, j represents the jth post, and n is the number of posts in the entire event; then a key r is defined for each node j : r j =softmax(log(w j )+g j ) where g j is random noise, softmax(·) is a normalization function, and then the n keys are sorted, and the first K keys are selected as the subgraph generated, and the corresponding K nodes are used as the output of the node reparameterization sampling, and finally the reparameterized node mask enhanced subgraph is obtained Where n″ is the sampled node, A is the adjacency matrix, R n×d The dimension of matrix A is n″×n″, d is the node vector dimension, and X represents the feature matrix of all post node sets in the event.

6. A rumor detection method according to claim 1, characterized in that: The process of capturing and fusing multi-level features of the enhanced image through the encoder is as follows: Then a three-layer GIN encoder is used to capture the multi-level potential features of the three enhanced images: where φ(·) is the LeakyReLU activation function, A is the adjacency matrix, CONCAT(·,·) is the concatenation operation, represents the root node feature of the lth layer, M l+1 represents the root node feature of the l+1th layer, H l and H l+1 Represent the node feature matrices of the lth layer and the l+1th layer respectively, and b l are the weight parameters and bias parameters of the l-th layer encoder network; after encoding, three enhanced representations of the propagation map and the scatter map are obtained.

7. A rumor detection method according to claim 1, characterized in that: The total pre-training objective function described above and learning the discriminative representation of rumors through pre-training for: Among them, η and μ are hyperparameters, is the unsupervised alignment loss function: is the sampler loss function: is the mutual information loss function: Among them, β is a hyperparameter, p(H|X) represents the encoded node features, is the KL divergence, represents the expected distribution with parameter θ, p(H) is a random Gaussian distribution, is a set of augmented graphs, α o For the weight of each augmented graph o, Enhance the root features of graph o, f θ (X) is the embedded representation of the enhanced graph; p(H|f θ (X)) represents the latent features encoded from the original view, represents the latent features encoded from the sampled graph, It means to find the expected value of function f for event i, g θ (·,·) represents the scaled cosine similarity function, Enhanced local representation of view, H j A graph representing all samples; H t It represents the graph representation of all the remaining samples after removing the target sample i.

8. A rumor detection method according to claim 1, characterized in that: The deployment process of the classifier model is as follows: using the pre-trained weight vector W, according to the calculated key value, sampling the first K nodes from the original node feature X, eliminating the propagation of noise information, and obtaining the denoised graph representation H of the propagation view of the pre-trained encoder P and the denoised graph representation of the scatter view H D ; Then, the representations of the two propagation views are concatenated to generate a comprehensive representation of the event graph. And deploy a Bayesian layer to learn the distribution: Where Bay(·|W b ) represents the Bayesian network, W b is the learnable parameter of the Bayesian network, σ(·) is the PreLU activation function, H b represents the comprehensive representation after encoding; CONCAT(·) is the concatenation function; Use original root features and their graph-level representation as auxiliary evidence for rumor detection: in, is the comprehensive representation of the original features, X root and Represent the original root feature and its graph-level representation respectively; two Beyasian layers are used to obtain the transformed representation of the original feature: X b+1 =σ(Bay b (X b |W b )) Bay b (·) represents the b-th layer Bayesian network, W b represents the weight parameter of the b-th layer Bayesian network, X b and X b+1 Represent the input node features of the b-th layer and the b+1-th layer Bayesian network respectively; Finally, the two parts of the high-quality representation are concatenated and a Beyasian layer with normalization is used to achieve the detection task: in To predict the label, Bay c (·) represents the output layer based on the Bayesian network, W c is the weight parameter of the output layer; c represents the last layer, i.e. the output layer, and b represents the intermediate layer.

9. A rumor detection method according to claim 1, characterized in that: The final objective function is specifically: where yi is the true label of event i, is the predicted label of event i, λ and κ are the weight parameters of the objective function, m is the number of events, is the objective function for supervised contrastive learning; The variational inference is specifically to fit the rescaled prior distribution to the posterior distribution through KL divergence: Where q(ω|θ, H) is the posterior distribution of the Bayesian network parameters, p(ω) is the scaled prior distribution of the Bayesian network parameters, represents the variational inference loss function, It means to find the KL divergence between distribution p and distribution q The objective function of supervised contrastive learning is for: in, It means to find the expectation for batch B, f k (x t ,θ) represents a classification model with parameter θ, x t represents the potential embedded X and H b The splicing features, x j is a single sample in batch B.

Images