Graph contrast learning rumor detection method based on large language model and node entanglement
By using the graph comparison learning method of large language model and node entanglement in rumors detection, a rumor dissemination graph and feature enhancement view is constructed, and the problem of relying on a large amount of labeled data and models in the existing technology is solved, achieving more efficient and accurate rumor detection.
Patent Information
- Application Number
- CN202510101649.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-05-16
AI Technical Summary
Existing rumor detection methods rely on a large amount of manual annotation data, are costly and time-consuming, and the model is prone to overfitting, and have weak generalization ability and robustness.
A graph comparison learning method based on large language model and node entanglement is adopted, and the generalization ability and robustness of the model are enhanced by constructing rumors propagation graphs and feature enhancement views, combined with graph neural network classifiers, and the rumor detection model is trained to enhance the generalization ability and robustness of the model.
It realizes more efficient and accurate rumor detection, reduces dependence on labeled data, and improves the generalization ability and robustness of the model.
Smart Images
Figure CN120011567A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of computer application, and in particular relates to graph comparison learning. Background Art
[0002] With the rapid development of the Internet, social media has become the main way for people to obtain information. However, the quality of information on social media varies, and some users deliberately spread rumors, making it difficult to ensure the authenticity of the information. The spread of rumors will not only disrupt social order, but also may harm the interests of the public. Therefore, it is particularly important to develop efficient and accurate rumor detection technology.
[0003] Most existing rumor detection methods usually adopt the framework of supervised learning, but this approach has two major problems. First, it relies on a large amount of manually labeled data, which is not only costly but also very time-consuming. Second, since the model is prone to overfitting, its generalization ability is weak, and the model is also less robust when facing label-related attacks. Summary of the invention
[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a rumor detection method based on graph comparative learning of large language models and node entanglement with strong generalization ability and robustness and accurate detection.
[0005] The technical solution adopted to solve the above technical problems consists of the following steps:
[0006] (1) Divide the data set
[0007] Download twitter15 as the rumor dataset from https: / / www.dropbox.com / s / 7ewzdrbelpmrnxu / rumdetect2017.zip?dl=0 and divide the rumor dataset into training set and test set according to 8:2.
[0008] (2) Constructing a rumor propagation graph
[0009] Construct the rumor propagation graph G as follows i :
[0010] G i =(V i ,E i )
[0011]
[0012] E i ={e1,e2,…,e m}
[0013] Among them, Vi represents the set of tweet nodes, r i represents the source tweet of the ith rumor event, Indicates the nth source tweet of the ith rumor event i -1 retweet or comment on a tweet, n i Indicates the number of nodes, E i represents the edge set, e m represents the mth edge; the source tweet and the forwarded or commented tweets constitute the original tweet content S O .
[0014] (3) Constructing feature-enhanced views
[0015] 1) Build a large language model prompt template
[0016] The large language model prompt template T is composed of the original tweet content S O , whether the tweet constitutes a rumor.
[0017] 2) Construct the final text attributes
[0018] Select the large language model GPT3.5, call the application programming interface of the large language model GPT3.5, input the large language model prompt template T into the large language model GPT3.5, obtain the model output result, which includes the rumor category judgment C of the tweet and the text explanation E of the judgment, save the output result in JSON format, use the json library to parse the JSON format output result, and extract the text explanation E of the judgment as the enhanced attribute S A .
[0019] The append function in Python language is used to extract the enhanced attribute S A Add to original tweet content S O At the end of the text, we get the final text attribute S F .
[0020] 3) Determine node characteristics
[0021] Determine the node feature X as follows:
[0022] X=f Bert (S)
[0023] Among them, S represents the text information to be encoded, S∈{S O ,S F}, f Bert Represents the Bert model.
[0024] 4) Construct feature-enhanced views
[0025] The original tweet content S Oand the final text attribute S F Input the Bert model and get the corresponding node feature X O and node feature X F , X O ∈R n×d , X F ∈R n×d , that is, with node feature X O Original picture of rumor event G k and with node feature X F Feature Enhanced View
[0026] (4) Build a structural enhancement view
[0027] 1) Convert the original graph G into a line graph L G
[0028] Use the line_graph function in the networkx library to convert the original graph G into a line graph L G .
[0029] 2) Identify important nodes
[0030] Determine the important nodes according to formula (1)
[0031]
[0032] L1=D1-A1
[0033] L2=D2-A2
[0034]
[0035] Among them, β is a hyperparameter, β takes a finite positive integer, Z1 and Z2 are intermediate variables, and D1 represents the graph L after deleting the node v in the line graph. G_v The degree matrix of A1 means L G_v The adjacency matrix of D2 means L G The degree matrix of A2 means L G The adjacency matrix of N1 represents the number of eigenvalues of the Laplace matrix L1, N1∈[1,n i -1], N2 represents the number of eigenvalues of the Laplace matrix L2, N2∈[1,n i ],n i is a finite positive integer, represents the i-th eigenvalue of the Laplacian matrix L1, represents the i-th eigenvalue of the Laplacian matrix L2.
[0036] 3) Build structure enhancement view
[0037] According to formula (2), the modified subset of the original edge set E is sampled The probability P is:
[0038]
[0039] Among them, p c is a hyperparameter, p s is the cut-off probability.
[0040] Keep a subset The edges with the largest probability P constitute the modified subset Get the node feature X O Structural Enhanced View
[0041] (5) Building a rumor detection model
[0042] The rumor detection model includes an encoding model and a graph neural network classifier.
[0043] The encoding model is composed of a graph convolutional layer 1 and a graph convolutional layer 2 connected in series.
[0044] The graph neural network classifier is composed of a fully connected layer and a softmax layer in series.
[0045] (6) Training rumor detection model
[0046] 1) Constructing graph-level embedding vectors
[0047] The undirected rumor event original graph G k , Enhanced View Enhanced View Input the encoding model and get the corresponding graph-level embedding vectors, which are h k ,
[0048] 2) Mapping to contrast space
[0049] The graph-level embedding vector is mapped to the contrast space as follows.
[0050] z=g(h)
[0051] Here, g(·) is a nonlinear transformation consisting of two layers of perceptrons, and h represents the graph-level embedding vector.
[0052] Embedding graph level vectors and graph-level embedding vector Mapped to the contrast space, we get the graph-level embedding vector in the contrast space and graph-level embedding vector
[0053] 3) Constructing graph contrast loss function
[0054] According to formula (3), the graph contrast loss function L is constructed gcl :
[0055]
[0056] Where τ is the temperature coefficient, sim(·) represents the similarity metric function, represents a positive sample pair, represents a negative sample pair.
[0057] 4) Constructing classification loss function
[0058] According to formula (4), the classification loss function L is constructed class :
[0059]
[0060] Among them, y i represents the true value of the i-th rumor event, represents the predicted value of the ith rumor event, N represents the number of rumor events, W and b are parameters, σ represents the activation function, and h i is to transform the original graph G of the i-th rumor event k_i Input the embedding vector obtained by the encoding model.
[0061] 5) Constructing the loss function of rumor detection model
[0062] The rumor detection model loss function is constructed according to formula (5):
[0063] L=λL gcl +L class (5)
[0064] Among them, λ represents the weight coefficient;
[0065] 6) Training rumor detection model
[0066] The training set is input into the rumor detection model for training. The training parameters are: initial learning rate is 0.0001, weight decay is 0.00001, number of iterations is 3000, and Adam optimizer is used to optimize model parameters. The training is carried out until the rumor detection model loss function L converges.
[0067] (7) Testing the rumor detection model
[0068] The test set is input into the rumor detection model for testing and the rumor detection results are output.
[0069] In the formula (1) of step (4) of the present invention to enhance the structure space, the β is a hyperparameter, β∈[1,200], N1 represents the number of eigenvalues of L1, N1∈[1,n i -1], N2 represents the number of eigenvalues of L2, N2∈[1,n i ],n i The specific value of is the same as the number of nodes in the rumor dataset.
[0070] In the formula (2) for enhancing the structure space in step (4) of the present invention, the p c is a hyperparameter, p c ∈(0,1), p s is the cut-off probability, p s ∈(0,1).
[0071] In the formula (3) of the rumor detection model trained in step (6) of the present invention, the τ is the temperature coefficient, τ∈(0,1).
[0072] In formula (4) of step (6) of the present invention for training the rumor detection model, N represents the number of samples, and the value of N is at least 1; W is a parameter, and the value of W is (0,1); b is a parameter, and the value of b is [0,1).
[0073] In the formula (5) of the step (6) of the present invention for training the rumor detection model, the λ represents a weight coefficient, and the value of λ is (0, 1).
[0074] Since the present invention adopts feature space enhancement based on a large language model and structural space enhancement based on node entanglement, the coding model can capture the more essential graph structure features of rumors, effectively alleviating the technical problem of existing rumor detection methods relying on a large amount of labeled data, and enhancing the generalization ability and robustness of the rumor detection model. By using the powerful understanding ability of the large language model on sentence attributes, the node features are made richer and more accurate, achieving more accurate rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0075] Figure 1 This is a flow chart of Example 1 of the present invention. DETAILED DESCRIPTION
[0076] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples, but the present invention is not limited to the following embodiments.
[0077] Example 1
[0078] exist Figure 1 In the embodiment, the rumor detection method based on graph contrast learning of large language model and node entanglement is composed of the following steps:
[0079] (1) Divide the data set
[0080] Download twitter15 as the rumor dataset from https: / / www.dropbox.com / s / 7ewzdrbelpmrnxu / rumdetect2017.zip?dl=0 and divide the rumor dataset into training set and test set according to 8:2.
[0081] (2) Constructing a rumor propagation graph
[0082] Construct the rumor propagation graph G as follows i :
[0083] G i =(V i ,E i )
[0084]
[0085] E i ={e1,e2,…,e m}
[0086] Among them, V i represents the set of tweet nodes, r i represents the source tweet of the ith rumor event, Indicates the nth source tweet of the ith rumor event i -1 retweet or comment on a tweet, n i Indicates the number of nodes, E i represents the edge set, e m represents the mth edge; the source tweet and the forwarded or commented tweets constitute the original tweet content S O .
[0087] (3) Constructing feature-enhanced views
[0088] 1) Build a large language model prompt template
[0089] The large language model prompt template T is composed of the original tweet content S O , whether the tweet constitutes a rumor.
[0090] 2) Construct the final text attributes
[0091] Select the large language model GPT3.5, call the application programming interface of the large language model GPT3.5, input the large language model prompt template T into the large language model GPT3.5, obtain the model output result, which includes the rumor category judgment C of the tweet and the text explanation E of the judgment, save the output result in JSON format, use the json library to parse the JSON format output result, and extract the text explanation E of the judgment as the enhanced attribute S A .
[0092] The append function in Python language is used to extract the enhanced attribute S A Add to original tweet content S O At the end of the text, we get the final text attribute S F .
[0093] 3) Determine node characteristics
[0094] Determine the node feature X as follows:
[0095] X=f Bert (S)
[0096] Among them, S represents the text information to be encoded, S∈{S O ,S F}, f Bert Represents the Bert model.
[0097] 4) Construct feature-enhanced views
[0098] The original tweet content S O and the final text attribute S F Input the Bert model and get the corresponding node feature X O and node feature X F , X O ∈R n×d , X F ∈R n×d , that is, with node feature X O Original picture of rumor event G k and with node feature X F Feature Enhanced View
[0099] (4) Build a structural enhancement view
[0100] 1) Convert the original graph G into a line graph L G
[0101] Use the line_graph function in the networkx library to convert the original graph G into a line graph L G .
[0102] 2) Identify important nodes
[0103] Determine the important nodes according to formula (1)
[0104]
[0105] L1=D1-A1
[0106] L2=D2-A2
[0107]
[0108] Where β is a hyperparameter, β∈[1,200], and the value of β in this embodiment is 100. Z1 and Z2 are intermediate variables, and D1 represents the graph L after deleting the node v in the line graph. G_v The degree matrix of A1 means L G_v The adjacency matrix of D2 means L G The degree matrix of A2 means L G The adjacency matrix of N1 represents the number of eigenvalues of the Laplace matrix L1, N1∈[1,n i -1], N2 represents the number of eigenvalues of the Laplace matrix L2, N2∈[1,n i ],n i The specific value of is the same as the number of nodes in the rumor dataset. represents the i-th eigenvalue of the Laplacian matrix L2.
[0109] 3) Build structure enhancement view
[0110] According to formula (2), the modified subset of the original edge set E is sampled The probability P is:
[0111]
[0112] Among them, p c is a hyperparameter, p c ∈(0,1), p in this embodiment c The value is 0.5, p s is the cut-off probability, p s ∈(0,1), p in this embodiment s The value is 0.5.
[0113] Keep a subset The edges with the largest probability P constitute the modified subset Get the node feature X O Structural Enhanced View
[0114] (5) Building a rumor detection model
[0115] The rumor detection model includes an encoding model and a graph neural network classifier.
[0116] The encoding model of this embodiment is composed of a graph convolutional layer 1 and a graph convolutional layer 2 connected in series.
[0117] The graph neural network classifier of this embodiment is composed of a fully connected layer and a softmax layer in series.
[0118] (6) Training rumor detection model
[0119] 1) Constructing graph-level embedding vectors
[0120] The undirected rumor event original graph G k , Enhanced View Enhanced View Input the encoding model and get the corresponding graph-level embedding vectors, which are h k ,
[0121] 2) Mapping to contrast space
[0122] The graph-level embedding vector is mapped to the contrast space as follows:
[0123] z=g(h)
[0124] Among them, g(·) is a nonlinear transformation consisting of two layers of perceptrons, and h represents the graph-level embedding vector;
[0125] Embedding graph level vectors and graph-level embedding vector Mapped to the contrast space, we get the graph-level embedding vector in the contrast space and graph-level embedding vector
[0126] 3) Constructing graph contrast loss function
[0127] According to formula (3), the graph contrast loss function L is constructed gcl :
[0128]
[0129] Wherein, τ is the temperature coefficient, τ∈(0,1), the value of τ in this embodiment is 0.5, sim(·) represents the similarity measurement function, represents a positive sample pair, represents a negative sample pair.
[0130] 4) Constructing classification loss function
[0131] According to formula (4), the classification loss function L is constructed class :
[0132]
[0133] Among them, y i represents the true value of the i-th rumor event, represents the predicted value of the ith rumor event, N represents the number of rumor events, N is at least 1, and in this embodiment, N is 100, W and b are parameters, W is (0, 1), and in this embodiment, W is 0.5, b is [0, 1), and in this embodiment, b is 0.5, σ represents the activation function, h i is to transform the original graph G of the i-th rumor event k_i Input the embedding vector obtained by the encoding model.
[0134] 5) Constructing the loss function of rumor detection model
[0135] The rumor detection model loss function is constructed according to formula (5):
[0136] L=λL gcl +L class (5)
[0137] Here, λ represents a weight coefficient, and the value of λ is (0, 1). In this embodiment, the value of λ is 0.5.
[0138] 6) Training rumor detection model
[0139] The training set is input into the rumor detection model for training. The training parameters are: initial learning rate is 0.0001, weight decay is 0.00001, number of iterations is 3000, and Adam optimizer is used to optimize model parameters. The training is carried out until the rumor detection model loss function L converges.
[0140] (7) Testing the rumor detection model
[0141] The test set is input into the rumor detection model for testing and the rumor detection results are output.
[0142] Graph contrastive learning rumor detection method based on large language model and node entanglement.
[0143] Example 2
[0144] The rumor detection method based on graph contrast learning of large language model and node entanglement in this embodiment consists of the following steps:
[0145] (1) Divide the data set
[0146] This step is the same as in Example 1.
[0147] (2) Constructing a rumor propagation graph
[0148] This step is the same as in Example 1.
[0149] (3) Constructing feature-enhanced views
[0150] This step is the same as in Example 1.
[0151] (4) Build a structural enhancement view
[0152] 1) Convert the original graph G into a line graph L G
[0153] This step is the same as in Example 1.
[0154] 2) Identify important nodes
[0155] Determine the important nodes according to formula (1)
[0156] The expression of formula (1) is the same as that of Example 1.
[0157] In formula (1), β is a hyperparameter, β∈[1,200], and the value of β in this embodiment is 1. The meanings and value ranges of other parameters and variables are the same as those in embodiment 1.
[0158] 3) Build structure enhancement view
[0159] According to formula (2), the modified subset of the original edge set E is sampled The probability P is:
[0160] The expression of formula (2) is the same as that of Example 1.
[0161] In formula (2), p c is a hyperparameter, p c ∈(0,1), p in this embodiment c The value is 0.1, p s is the cut-off probability, p s ∈(0,1), p in this embodiment s The value is 0.1.
[0162] (5) Building a rumor detection model
[0163] This step is the same as in Example 1.
[0164] (6) Training rumor detection model
[0165] 1) Constructing graph-level embedding vectors
[0166] This step is the same as in Example 1.
[0167] 2) Mapping to contrast space
[0168] This step is the same as in Example 1.
[0169] 3) Constructing graph contrast loss function
[0170] According to formula (3), the graph contrast loss function L is constructed gcl :
[0171] The expression of formula (3) is the same as that of Example 1.
[0172] In formula (3), τ is the temperature coefficient, τ∈(0,1), and the value of τ in this embodiment is 0.1. The meanings and value ranges of other parameters and variables are the same as those in embodiment 1.
[0173] 4) Constructing classification loss function
[0174] According to formula (4), the classification loss function L is constructed class :
[0175] The expression of formula (4) is the same as that of Example 1.
[0176] In formula (4), N represents the number of rumor events, and the value of N is at least 1. In this embodiment, the value of N is 1. W and b are parameters, and the value of W is (0, 1). In this embodiment, the value of W is 0.1, and the value of b is [0, 1). In this embodiment, the value of b is 0. The meanings and value ranges of other parameters and variables are the same as those in Example 1.
[0177] 5) Constructing the loss function of rumor detection model
[0178] The rumor detection model loss function is constructed according to formula (5):
[0179] The expression of formula (5) is the same as that of Example 1.
[0180] In formula (5), λ represents a weight coefficient, and the value of λ is (0, 1). In this embodiment, the value of λ is 0.1.
[0181] The other steps are the same as those in Example 1. A graph contrast learning rumor detection method based on a large language model and node entanglement is completed.
[0182] Example 3
[0183] The rumor detection method based on graph contrast learning of large language model and node entanglement in this embodiment consists of the following steps:
[0184] (1) Divide the data set
[0185] This step is the same as in Example 1.
[0186] (2) Constructing a rumor propagation graph
[0187] This step is the same as in Example 1.
[0188] (3) Constructing feature-enhanced views
[0189] This step is the same as in Example 1.
[0190] (5) Build structure enhancement view
[0191] 1) Convert the original graph G into a line graph L G
[0192] This step is the same as in Example 1.
[0193] 2) Identify important nodes
[0194] Determine the important nodes according to formula (1)
[0195] The expression of formula (1) is the same as that of Example 1.
[0196] In formula (1), β is a hyperparameter, β∈[1,200], and the value of β in this embodiment is 200. The meanings and value ranges of other parameters and variables are the same as those in embodiment 1.
[0197] 3) Build structure enhancement view
[0198] According to formula (2), the modified subset of the original edge set E is sampled The probability P is:
[0199] The expression of formula (2) is the same as that of Example 1.
[0200] In formula (2), p c is a hyperparameter, p c ∈(0,1), p in this embodiment c The value is 0.9, p s is the cut-off probability, p s ∈(0,1), p in this embodiment s The value is 0.9.
[0201] (5) Building a rumor detection model
[0202] This step is the same as in Example 1.
[0203] (6) Training rumor detection model
[0204] 1) Constructing graph-level embedding vectors
[0205] This step is the same as in Example 1.
[0206] 2) Mapping to contrast space
[0207] This step is the same as in Example 1.
[0208] 3) Constructing graph contrast loss function
[0209] According to formula (3), the graph contrast loss function L is constructed gcl :
[0210] The expression of formula (3) is the same as that of Example 1.
[0211] In formula (3), τ is the temperature coefficient, τ∈(0,1), and the value of τ in this embodiment is 0.9. The meanings and value ranges of other parameters and variables are the same as those in embodiment 1.
[0212] 4) Constructing classification loss function
[0213] According to formula (4), the classification loss function L is constructed class :
[0214] The expression of formula (4) is the same as that of Example 1.
[0215] In formula (4), N represents the number of rumor events, and the value of N is at least 1. In this embodiment, the value of N is 1490. W and b are parameters, and the value of W is (0, 1). In this embodiment, the value of W is 0.9, and the value of b is [0, 1). In this embodiment, the value of b is 0.9. The meanings and value ranges of other parameters and variables are the same as those in Example 1.
[0216] 5) Constructing the loss function of rumor detection model
[0217] The rumor detection model loss function is constructed according to formula (5):
[0218] The expression of formula (5) is the same as that of Example 1.
[0219] In formula (5), λ represents a weight coefficient, and the value of λ is (0, 1). In this embodiment, the value of λ is 0.9.
[0220] The other steps are the same as those in Example 1. A graph contrast learning rumor detection method based on a large language model and node entanglement is completed.
Claims
1. A graph contrastive learning rumor detection method based on large language model and node entanglement, characterized by It consists of the following steps: (1) Divide the data set Download twitter15 as the rumor dataset from https: / / www.dropbox.com / s / 7ewzdrbelpmrnxu / rumdetect2017.zip?dl=0, and divide the rumor dataset into training set and test set according to 8:2; (2) Constructing a rumor propagation graph Construct the rumor propagation graph G as follows i : G i =(V i ,E i ) THE i }{e1,e2,…,e m } Among them, V i represents the set of tweet nodes, r i represents the source tweet of the ith rumor event, Indicates the nth source tweet of the ith rumor event i -1 retweet or comment on a tweet, n i Indicates the number of nodes, E i represents the edge set, e m represents the mth edge; the source tweet and the forwarded or commented tweets constitute the original tweet content S O ; (3) Constructing feature-enhanced views 1) Build a large language model prompt template The large language model prompt template T is composed of the original tweet content S O , whether the tweet constitutes a rumor; 2) Construct the final text attributes Select the large language model GPT3.5, call the application programming interface of the large language model GPT3.5, input the large language model prompt template T into the large language model GPT3.5, obtain the model output result, which includes the rumor category judgment C of the tweet and the text explanation E of the judgment, save the output result in JSON format, use the json library to parse the JSON format output result, and extract the text explanation E of the judgment as the enhanced attribute S A ; The append function in Python language is used to extract the enhanced attribute S A Add to original tweet content S O At the end of the text, we get the final text attribute S F ; 3) Determine node characteristics Determine the node feature X as follows: X=f Bert (S) Among them, S represents the text information to be encoded, S∈{S O ,S F }, f Bert Represents the Bert model; 4) Construct feature-enhanced views The original tweet content S O and the final text attribute S F Input the Bert model and get the corresponding node feature X O and node feature X F , X O ∈R n×d , X F ∈R n×d , that is, with node feature X O Original picture of rumor event G k and with node feature X F Feature Enhanced View (4) Build a structural enhancement view 1) Convert the original graph G into a line graph L G Use the line_graph function in the networkx library to convert the original graph G into a line graph L G ; 2) Identify important nodes Determine the important nodes according to formula (1) L1=D1-A1 L2=D2-A2 Among them, β is a hyperparameter, β takes a finite positive integer, Z1 and Z2 are intermediate variables, and D1 represents the graph L after deleting the node v in the line graph. G_v The degree matrix of A1 means L G_v The adjacency matrix of D2 means L G The degree matrix of A2 means L G The adjacency matrix of N1 represents the number of eigenvalues of the Laplace matrix L1, N1∈[1,n i -1], N2 represents the number of eigenvalues of the Laplace matrix L2, N2∈[1,n i ],n i is a finite positive integer, represents the i-th eigenvalue of the Laplacian matrix L1, represents the i-th eigenvalue of the Laplacian matrix L2; 3) Build structure enhancement view According to formula (2), the modified subset of the original edge set E is sampled The probability P is: Among them, p c is a hyperparameter, p s is the cut-off probability; Keep a subset The edges with the largest probability P constitute the modified subset Get the node feature X O Structural Enhanced View (5) Building a rumor detection model The rumor detection model includes an encoding model and a graph neural network classifier; The encoding model is composed of a graph convolution layer 1 and a graph convolution layer 2 connected in series; The graph neural network classifier is composed of a fully connected layer and a softmax layer in series; (6) Training rumor detection model 1) Constructing graph-level embedding vectors The undirected rumor event original graph G k , Enhanced View Enhanced View Input the encoding model and get the corresponding graph-level embedding vectors, which are h k , 2) Mapping to contrast space The graph-level embedding vector is mapped to the contrast space as follows: z=g(h) Among them, g(·) is a nonlinear transformation consisting of two layers of perceptrons, and h represents the graph-level embedding vector; Embedding graph level vectors and graph-level embedding vector Mapped to the contrast space, we get the graph-level embedding vector in the contrast space and graph-level embedding vector 3) Constructing graph contrast loss function According to formula (3), the graph contrast loss function L is constructed gcl : Where τ is the temperature coefficient, sim(·) represents the similarity metric function, represents a positive sample pair, represents a negative sample pair; 4) Constructing classification loss function According to formula (4), the classification loss function L is constructed class : Among them, y i represents the true value of the i-th rumor event, represents the predicted value of the ith rumor event, N represents the number of rumor events, W and b are parameters, σ represents the activation function, and h i is to transform the original graph G of the i-th rumor event k_i Input the embedding vector obtained by the encoding model; 5) Constructing the loss function of rumor detection model The rumor detection model loss function is constructed according to formula (5): L=λL gcl +L class (5)where λ represents the weight coefficient; 6) Training rumor detection model The training set is input into the rumor detection model for training. The training parameters are: initial learning rate is 0.0001, weight decay is 0.00001, number of iterations is 3000, and Adam optimizer is used to optimize model parameters. The training is conducted until the rumor detection model loss function L converges. (7) Testing the rumor detection model The test set is input into the rumor detection model for testing and the rumor detection results are output.
2. The rumor detection method based on graph contrastive learning of large language model and node entanglement according to claim 1 is characterized in that: In step (4) of enhancing the structure space, in formula (1), β is a hyperparameter, β∈[1,200], N1 represents the number of eigenvalues of L1, N1∈[1,n i -1], N2 represents the number of eigenvalues of L2, N2∈[1,n i ],n i The specific value of is the same as the number of nodes in the rumor dataset.
3. The rumor detection method based on graph contrastive learning with large language model and node entanglement according to claim 1 is characterized in that: In step (4) of enhancing the structure space, in formula (2), the p c is a hyperparameter, p c ∈(0,1), p s is the cut-off probability, p s ∈(0,1).
4. The rumor detection method based on graph contrastive learning of large language model and node entanglement according to claim 1 is characterized in that: In formula (3) of step (6) for training the rumor detection model, τ is the temperature coefficient, τ∈(0,1).
5. The rumor detection method based on graph contrastive learning of large language model and node entanglement according to claim 1 is characterized in that: In formula (4) of step (6) for training the rumor detection model, N represents the number of samples, and the value of N is at least 1; W is a parameter, and the value of W is (0,1); b is a parameter, and the value of b is [0,1).
6. The rumor detection method based on graph contrastive learning with large language model and node entanglement according to claim 1, characterized in that: In formula (5) of step (6) for training the rumor detection model, the λ represents the weight coefficient, and the value of λ is (0,1).