Rumor detection technology with enhanced embedding consistency in graph contrast learning

Abstract

The invention discloses an embedded consistency enhanced rumor detection technology in graph contrast learning, which comprises the following steps of: for each propagation structure, generating two enhanced propagation structures by considering the importance of edges; a negative sample capturer and a negative sample controller are designed, the negative sample capturer indirectly detects gradient confusion by observing the similarity difference of sample pairs in training iteration, and negative sample pairs affected by EIC are captured; the negative sample controller neglects the negative sample pairs, so that the propagation structure encoder adaptively learns from neglected negative samples, neglected nodes are updated in each iteration, all negative samples are ensured to be used for training, the contrast loss is optimized by supplementing gradient, and even if a large number of negative samples are neglected, the propagation structure encoder is ensured to be used for training. The gradient of the contrast loss is still relatively stable; the enhanced propagation structures are connected together to serve as final representation of the event, and cross entropy loss and contrast loss are optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer science, in particular to rumor detection technology with embedding consistency enhancement in graph contrastive learning. BACKGROUND

[0002] In the prior art, rumors are like undercurrents in information streams, which can have a profound impact on public opinion, public perception, and even group behavior. Therefore, efficient and accurate rumor detection technology has become a hot and focused research topic. Graph contrastive learning (GCL) is an important breakthrough in the field of machine learning, and with its powerful self-supervised ability, it has shown unique advantages and broad application prospects in rumor detection. GCL can cleverly extract key information from complex propagation structures without relying on large amounts of manually labeled data through self-supervised learning, and convert it into simple and effective low-dimensional vector representations. In the process of pursuing propagation structure optimization and applying GCL at multiple levels, the conflict between the contrastive loss based on InfoNCE and the message passing mechanism in graph neural networks (GNN) leads to the inability of the model to clearly distinguish negative samples during the learning process, resulting in the feature representation of negative samples being too close, ultimately making it difficult for the model to accurately identify rumors, and the performance of rumor detection cannot meet the strict requirements of accuracy and efficiency in actual applications.

[0003] Therefore, the present application provides rumor detection technology with embedding consistency enhancement in graph contrastive learning. SUMMARY

[0004] The present application aims to provide rumor detection technology with embedding consistency enhancement in graph contrastive learning to solve at least one of the above-mentioned problems in the prior art.

[0005] The object of the present application can be achieved by the following technical solutions:

[0006] The rumor detection technology with embedding consistency enhancement in graph contrastive learning includes the following steps:

[0007] Establish a BERT model, and for each propagation structure of the rumor corresponding data set, generate two enhanced propagation structures by considering the importance of the edges;

[0008] Design a negative sample capturer and a negative sample controller. The negative sample capturer indirectly detects the confusion of the gradient by observing the difference in similarity of the sample pairs in the training iteration, and captures the negative sample pairs affected by EIC.

[0009] The negative sample controller ignores the negative sample pair, so that the propagation structure encoder adaptively learns from the ignored negative sample, and the ignored node is updated in each iteration to ensure that all negative samples are used for training;

[0010] The contrastive loss is optimized by supplementing the gradient, so that the gradient of the contrastive loss remains relatively stable even when a large number of negative samples are ignored;

[0011] The enhanced propagation structure is connected as the final representation of the event, and the cross-entropy loss and the contrastive loss are optimized.

[0012] The beneficial effects of the present application are:

[0013] 1. The edge importance is evaluated by edge centrality, the edge enhancement operation is implemented in probability, the node text is discarded and the mask is processed, and the propagation structure is enriched; the propagation structure encoder of the two-layer GCN is used to effectively extract the features of the enhanced propagation structure, and the model is helped to understand the complex structure; the negative sample catcher and the controller are designed, the former accurately locates the negative samples affected by the EIC, and the latter sets the ignore set and dynamically adjusts, and reasonably uses the negative samples; rumor prediction is trained jointly, cross-entropy and GCL loss are optimized, weights are dynamically adjusted, negative sample gradient is amplified, and model convergence speed and prediction accuracy are improved. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0015] Figure 1 is the step flow chart of the rumor detection technology embedded with consistency enhancement in graph contrastive learning provided by the embodiment 1 of the present application;

[0016] Figure 2 is the propagation structure schematic diagram of the rumor detection technology embedded with consistency enhancement in graph contrastive learning provided by the embodiment 2 of the present application;

[0017] Figure 3 is the RERG model schematic diagram of the rumor detection technology embedded with consistency enhancement in graph contrastive learning provided by the embodiment 3 of the present application;

[0018] Figure 4 is the ablation experiment comparison schematic diagram of the rumor detection technology embedded with consistency enhancement in graph contrastive learning provided by the embodiment 3 of the present application;

[0019] Figure 5Schematic diagram of parameter analysis of rumor detection technology embedded with consistency enhancement in graph contrast learning provided by Example 3 of the present invention;

[0020] Figure 6 This is a schematic diagram of early rumor detection using a rumor detection technology with consistency enhancement embedded in graph contrast learning provided by Example 3 of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0022] Example 1

[0023] like Figure 1 As shown, the rumor detection technology with enhanced consistency embedded in graph contrast learning provided by the embodiment of the present invention specifically includes the following steps:

[0024] Step 1: Combining the advantages of graph convolutional learning (GCL) and graph neural networks (GNNs), we define a dataset for rumor detection and analyze the propagation structure of events to implement rumor detection. GCL is then used based on graph convolutional learning to eliminate EIC in the rumor detection field.

[0025] In some specific embodiments, the BERT model is defined based on combining graph convolutional learning (GCL) and graph neural network (GNN), and the rumor detection dataset is defined as D = {e1, e2, ..., e N}, where e i represents the i-th rumor, N is the number of rumors, where s i is the source event, each represents the post associated with the jth source event;

[0026] The propagation structure of the i-th rumor in the dataset is defined as G i = <V i ,E i >, where V i Represents a node in the event propagation structure, E i Represents an edge in the event propagation structure;

[0027] Based on the obtained rumor detection data set, analyze the rumor response situation. right If there is a reply, there will be a directed edge like s i has a reply, then the directed edge G i is defined as X i ∈R |V|×d , where d represents the degree of the node in the propagation structure, and the adjacency matrix of G i is defined as A i ∈{0,1} |V|×|V| ;

[0028] Based on the rumor detection dataset, the target of rumor detection is defined as a rumor classifier f: D i →Y i , where Y i ∈{T;F;U;N}, T represents a true rumor, F represents a false rumor, U represents an unverified rumor, and N represents a non-rumor;

[0029] Figure 2 As shown in (a), (b) and (c), based on the combination of graph convolution learning GCL and graph neural network GNN, a three-node propagation structure composed of a source node v1 and two related nodes v2 and v3 is established, based on a single-layer GNN containing one-hop neighbors of each node, under the message passing mechanism of GNN, in the forward propagation process, v1 aggregates features from itself and v2, v3, v2 aggregates features from itself and v1, and v3 aggregates features from itself and v1;

[0030] In the process of graph convolution learning GCL, v2 and v3 form a negative pair with v1; as shown in Figure 2 (b), for the negative pair formed by v1 and v2, the loss function based on InfoNCE embeds features from v3, highlighting the difference between the negative pairs, while ignoring information from v1 and v2, minimizing the similarity of the three nodes;

[0031] Based on the negative pair formed by v1 and v3, the loss function of InfoNCE embeds features from v2, while ignoring information from v1 and v3; therefore, when GNN generates the representation of v1, the loss based on InfoNCE requires GNN to embed and ignore the aggregated features from nodes v2 and v3 at the same time;

[0032] As shown in Figure 3 , based on any one propagation structure, two enhanced propagation structures are generated by considering the importance of the edges; by observing the difference in similarity of sample pairs in training iterations, the confusion of gradients is indirectly detected, and negative sample pairs affected by EIC are captured;

[0033] Based on the obtained negative sample pair affected by the EIC, the propagation structure encoder is adapted to learn from the ignored negative sample, the ignored node is updated in each iteration to ensure that all negative samples are used for training, and the contrast loss is optimized by supplementing the gradient to ensure the relative stability of the gradient of the contrast loss.

[0034] It should be noted that the propagation structure encoder represents a responsible rumor propagation structure data processing and feature extraction, based on the principles of graph convolution learning GCL and graph neural network GNN, the information of the nodes and edges in the propagation structure is aggregated and converted.

[0035] The technical scheme of the embodiment of the application is: combining the advantages of graph convolution learning GCL and graph neural network GNN, the rumor detection model is more comprehensive and efficient in feature extraction and information processing, the accuracy of rumor detection is improved, the related node features are reasonably embedded and ignored in the InfoNCE loss function, the problem that the negative sample cannot effectively move away is avoided, and the performance of the model is improved, when processing the propagation structure, the enhanced propagation structure is generated, the gradient confusion is detected, and the negative sample pair affected by the EIC is ignored, the propagation structure encoder is adapted to learn from the ignored negative sample, and the ignored node is updated in each iteration to ensure that all negative samples can participate in training, the contrast loss is optimized by supplementing the gradient, the gradient of the contrast loss is relatively stable when a large number of negative samples are ignored, and the stability of the training process and the robustness of the model are improved.

[0036] Embodiment 2

[0037] As shown in Figure 1 , the rumor detection technology embedded with consistency enhancement in graph contrast learning provided by the embodiment of the application specifically includes the following steps:

[0038] Step two: based on the propagation structure obtained by combining graph convolution learning GCL and graph neural network GNN, the obtained propagation structure is enhanced, the enhanced propagation structure is feature extracted by the propagation structure encoder, and the negative sample pair of the EIC is filtered by the negative sample controller based on the negative sample capturer; based on the filtering, rumor detection is performed and joint training is performed;

[0039] In some specific embodiments, the propagation structure obtained by combining graph convolution learning GCL and graph neural network GNN is evaluated by edge centrality to evaluate the importance of the edge, which is determined by the centrality of the connected node, and the node centrality deg(·) is calculated by the degree of the node.

[0040] Based on the fact that the tail node has a great influence on the edge, in the directed propagation structure, the centrality of the tail node During training, edge augmentation operations including removal and addition are applied based on the probability P(r), defined by the formula To mitigate the impact of highly connected nodes, the formula

[0041]

[0042] The probability of an edge in the propagation structure being removed is obtained, where, and are the maximum and average values of , respectively, and p e is the overall probability of controlling edge removal, p r <1 represents the truncation probability;

[0043] For the text content of the nodes in the propagation structure, a dropout mask is applied to create noisy text samples with minimal loss of information;

[0044] The BERT model encodes the source post and the comment respectively, and concatenates the source post and the comment in the format "[CLS]Source[SEP]Comment[SEP]", highlighting the importance of the source post, and uses the final hidden state of "[CLS]" as the corresponding node representation;

[0045] Based on the propagation structure G i After augmentation, the augmented propagation structure G Two augmented propagation structures G and are obtained. i Transformed into and i Transformed into and Based on the augmented propagation structure G A two-layer GCN is used to capture the embedding, and the formula is

[0046]

[0047] is obtained, where is the weight matrix, and σ is the ReLU function; is the normalized adjacency matrix, is the degree of the i-th node;

[0048] Based on the effective identification and processing of negative sample pairs affected by EIC in the propagation structure, a negative sample capturer is designed to identify negative sample pairs that cause gradient confusion; The negative sample capturer quantifies the gradient impact introduced by a specific negative sample through the following formula:

[0049]

[0050] where W​e represents the parameters of the propagation structure encoder, based on the negative sample pair that triggers the gradient confusion, the similarity change caused by the negative sample pair is used to approximate its influence on the gradient, and the formula is as follows:

[0051]

[0052] Where ΔS ij is the gradient influence caused by the negative sample pair (r i , r j ), is the cosine similarity of r i and r j in the current iteration, and iter represents the number of iterations;

[0053] The negative sample controller is designed to filter out negative sample pairs greater than the sample threshold to generate an ignore set, which ignores the negative sample pairs that trigger EIC, and the formula is as follows:

[0054] IgS={NP(i,j)∣ΔS ij >ΔS r}

[0055] Where r% is the ignore ratio, and ΔS r is the minimum value in the first r nodes after sorting;

[0056] It should be noted that for the negative sample pairs contained in the ignore set, the negative sample controller temporarily excludes them from self-supervised sampling, and most of the negative sample pairs in the ignore set are significantly affected by EIC; After training, part of the ignored negative sample pairs provide enough information for training, and the corresponding similarity does not decrease, and the corresponding sample pair participates in training again to ensure that all negative samples are used for training while not triggering EIC;

[0057] For each event e i , the formula is used to calculate the representations H' i and H″ i of the two enhanced propagation structures, and H' i and H″ i are spliced through the formula

[0058] m i =concat(H′ i ,H″ i )

[0059] The final representation of e i is obtained, and m i is input into the fully connected layer and the softmax layer, and the formula is as follows:

[0060]

[0061] where, is the predicted probability distribution, FC(m i ) represents the fully connected layer;

[0062] Joint training is performed based on rumor prediction, and the cross-entropy loss of the predicted and the real label y i is minimized, and the formula is as follows:

[0063]

[0064] For the GCL loss, based on the prevention of ignoring part of the node gradient too small, the gradient of the negative sample is amplified by the parameter a, and the positive sample is introduced with the parameter β; finally, the GCL loss definition formula is as follows:

[0065]

[0066] Where κ is the temperature parameter, the final loss is composed of cross-entropy loss and GCL loss, and the definition formula is as follows:

[0067]

[0068] In the training process, the weight ξ is dynamically adjusted to ensure that the two losses are emphasized according to the corresponding current importance, and the data set statistics are as follows:

[0069]

[0070] Table 1 data set statistics

[0071] The technical scheme of the embodiment of the present application is: the importance of the edge is evaluated by using the edge centrality, the edge enhancement operation is performed based on the probability, the noise text sample is created by applying the discard mask to the node text content, the propagation structure is more diverse and representative, the effective feature extraction of the enhanced propagation structure is performed by using two layers of GCN through the propagation structure encoder, the high-quality feature representation is provided, and the understanding and learning ability of the model for the complex propagation structure are improved; the negative sample capturer and the negative sample controller are designed, the problem negative sample pair is accurately positioned by the negative sample capturer through the quantitative gradient influence and the similar change approximate gradient influence; the negative sample controller temporarily excludes the problem negative sample pair by setting the ignore set, dynamically adjusts in the training process, ensures that all negative samples can be reasonably utilized, avoids the negative influence of the EIC on the model performance, fully excavates the value of the negative sample, and enhances the stability and robustness of the model; the joint training of the rumor prediction is performed, the cross entropy loss of the prediction and the real label is minimized, the GCL loss is optimized, the gradient problem of too small is prevented by amplifying the negative sample gradient of the parameter and introducing the positive sample parameter, and the weights of the cross entropy loss and the GCL loss are dynamically adjusted, so that the model can balance the effects of the two losses according to the needs of different stages in the training process, and the convergence speed and the prediction accuracy of the model are improved.

[0072] Embodiment 3

[0073] As shown in Figure 4 , Figure 5 and Figure 6 , the rumor detection technology for embedding consistency enhancement in graph contrast learning provided by the embodiment of the present application specifically includes the following steps:

[0074] In some specific embodiments, the RERG model is evaluated on two real-world data sets, the RERG model is implemented by using PyTorch, the parameters pe, pr, K, r, X, N and P(r) are respectively set as 0.1, 0.7, 0.5, 5, 0.001, 20 and (0.5, 0.5), the learning rate is initialized as 5x10 -4 , and gradually decreases according to the decay rate of 1x10 -4 in the training process;

[0075] The parameters of the RERG model are updated by gradient descent and optimized by using the Adam algorithm; the topology structure of the post graph is constructed by using the users, sources and comments in the two data sets, the text content of each graph node is represented by BERT, and the graph structure is captured by GCN;

[0076] Based on the effectiveness of the RERG model in verifying different modules, the ablation study is performed, and the experimental results are as shown in Figure 4As shown in the figure, the performance of RERG-EDGE on the Twitter15 dataset drops by 1.2%, and the performance on the Twitter16 dataset drops by 0.9%. This indicates that ignoring the importance of edges in the process of propagation structure enhancement leads to the loss of key information, which has an adverse impact on the effectiveness of GCL.

[0077] like Figure 4 As shown, the performance of RERG-EIC on the Twitter15 dataset drops by 1.9%, and the performance on the Twitter16 dataset drops by 1.7%. Based on the comparison between RERG-EIC and RERG, it is shown that without solving the EIC problem, using GCL for rumor detection may produce implicit conflicts, which confuses the model. The conflict stems from the contradiction between the message passing mechanism of GNN and the InfoNCE loss function, which limits the effectiveness of GCL.

[0078] Figure 5 (a) Shows the accuracy of rumor detection on the Twitter15 and Twitter16 datasets under different ignoring ratios r. For the Twitter15 dataset, when r increases from 1% to 5%, the accuracy increases. When r increases further, the accuracy begins to fluctuate and continues to decline. For the Twitter16 dataset, the accuracy improves when r increases from 1% to 4%, but decreases when r increases further. This observation shows that ignoring some negative samples reduces EIC and improves accuracy. Excessively high r prevents the model from obtaining enough negative sample pairs for training, resulting in performance degradation.

[0079] Figure 5 (b) shows the effectiveness of the overall probability pe and the cutoff probability pr. By adjusting the hyperparameters pe from 0.2 to 0.6 and pr from 0.5 to 0.9, the impact of edge importance is evaluated. Based on the increase of pe and pr, the performance of the RERG model is improved. On the Twitter15 and Twitter16 datasets, it reaches the best level when pe = 0.4 and pr = 0.7;

[0080] After exceeding the optimal level, the performance of the RERG model begins to decline. The experimental results are consistent with expectations. There are many unreliable relationships in the event propagation structure, which weaken the robustness of the model and lead to poor performance. Based on the increase of pe and pr, the number of unreliable relationships decreases, which enhances the performance of the RERG model. Based on excessive edge removal, pe and pr are too high, and the event propagation structure loses key structural components, which reduces the performance of the RERG model.

[0081] like Figure 5 As shown, eight different time points were set up to test whether the model can accurately detect rumors based on the limited information available in the early stages;

[0082] As Figure 6 shown, the performance of the RERG model in the early rumor detection task is compared with other methods; the performance of the methods in the three data sets changes roughly the same, at time 0, when the input data only contains source information, due to insufficient training data and lack of comment information, the performance of the model is generally poor; after 10 minutes, the performance of the model is significantly improved, the model based on GCL shows particularly rapid improvement in accuracy, and the input data gradually contains more rich structural features; the RERG model shows superior and consistent performance at all time points, and is always superior to the benchmark;

[0083] The above describes one embodiment of the present application in detail, but the content is only the preferred embodiment of the present application, and cannot be considered to limit the scope of the present application; all equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage of the present application.

Claims

1. Rumor detection technology embedded in graph contrastive learning with consistency enhancement, characterized by: The following steps are involved: Build a BERT model. For each propagation structure in the rumor dataset, generate two enhanced propagation structures by considering the importance of edges. Design a negative sample catcher and a negative sample controller. The negative sample catcher indirectly detects gradient confusion by observing the difference in sample similarity between training iterations and captures negative sample pairs affected by EIC. The negative sample controller ignores negative sample pairs, allowing the propagation structure encoder to adaptively learn from the ignored negative samples. The ignored nodes are updated in each iteration to ensure that all negative samples are used for training; Optimize the contrast loss by supplementing the gradient, ensuring that the gradient of the contrast loss remains relatively stable even when a large number of negative samples are ignored; The augmented propagation structures are concatenated as the final representation of the event, and the cross-entropy loss and contrastive loss are optimized.

2. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific process of obtaining the data set is as follows: Based on the combination of graph convolutional learning GCL and graph neural network GNN, the BERT model is defined, and the dataset for rumor detection is defined as D = {e1, e2, ..., e N }, where e i represents the i-th rumor, N is the number of rumors, where s i is the source event, each represents the post associated with the jth source event; The propagation structure of the i-th rumor in the dataset is defined as G i = <V i ,E i >, where V i Represents a node in the event propagation structure, E i Represents an edge in the event propagation structure; Based on the obtained rumor detection data set, analyze the rumor response situation. right If there is a reply, there will be a directed edge like to s i If there is a reply, there will be an edge G i The characteristic matrix is ​​defined as X i ∈R |V|×d , where d represents the degree of the node in the propagation structure, G i The adjacency matrix of A is defined as i ∈{0,1} |V|×|V| .

3. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific classification process of the rumors is as follows: Based on the rumor detection dataset, the goal of rumor detection is defined as the rumor classifier f:D i →Y i , where Y i ∈{T; F; U; N}, T represents a true rumor, F represents a false rumor, U represents an unconfirmed rumor, and N represents a non-rumor.

4. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that The specific acquisition process of the BERT model is as follows: Based on the combination of graph convolutional learning (GCL) and graph neural network (GNN), a three-node propagation structure consisting of a source node v1 and two related nodes v2 and v3 is established. Based on a single-layer GNN containing one-hop neighbors for each node, under the GNN message passing mechanism, during the forward propagation process, v1 aggregates features from itself and v2 and v3, v2 aggregates features from itself and v1, and v3 aggregates features from itself and v1; In the process of learning GCL based on graph convolution, both v2 and v3 form negative pairs with v1. As shown in Figure 1(b), for the negative pairs formed by v1 and v2, the InfoNCE loss embeds the features from v3, highlighting the differences between the negative pairs while ignoring the information from v1 and v2, minimizing the similarity of the three nodes. Based on the negative pairs formed by v1 and v3, the loss function of InfoNCE embeds the features from v2 while ignoring the information from v1 and v3; Therefore, when GNN generates the representation of v1, the InfoNCE-based loss requires GNN to simultaneously embed and ignore the aggregated features from nodes v2 and v3.

5. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific process of ignoring negative sample pairs is as follows: Based on any propagation structure, two enhanced propagation structures are generated by considering the importance of edges. By observing the difference in similarity between sample pairs in training iterations, gradient confusion is indirectly detected, and negative sample pairs affected by EIC are captured. Based on the obtained negative sample pairs affected by EIC, the negative sample pairs are ignored, so that the propagation structure encoder can adaptively learn from the ignored negative samples; The ignored nodes are updated in each iteration to ensure that all negative samples are used for training; the contrast loss is optimized by supplementing the gradient to ensure that the gradient of the contrast loss is relatively stable.

6. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific analysis process of the propagation structure is as follows: Combining graph convolution learning GCL and graph neural network GNN to obtain the propagation structure, through edge centrality To evaluate the importance of edges, It is determined by the centrality of connected nodes, and the node centrality deg(·) is calculated by the degree of the node; Based on the large influence of the tail node on the edge, in the directed propagation structure, the centrality of the tail node is used. During training, edge enhancement operations, including removal and addition, are applied based on the probability P(r); Define the formula To mitigate the impact of height nodes, we use the formula The probability of the edge of the propagation structure being removed is obtained, where and They are The maximum and average values ​​of p e Controls the overall probability of edge removal, p r <1 indicates the cutoff probability.

7. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific acquisition process of the enhanced propagation structure is as follows: The BERT model encodes the source post and comments separately, concatenating them in the format "[CLS]Source[SEP]Comment[SEP]" to highlight the criticality of the source post, and uses the final hidden state of "[CLS]" as the corresponding node representation; Based on the propagation structure G i After enhancement, the enhanced propagation structure is obtained Obtain two enhanced propagation structures and Adjacency matrix A i Convert to and Feature matrix X i Convert to and Based on enhanced propagation structure A two-layer GCN is used to capture embedding, through the formula Get, among them represents the weight matrix, and σ represents the ReLU function; represents the normalized adjacency matrix, represents the degree of the i-th node.

8. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific acquisition process of the supplementary gradient is as follows: Based on the effective identification and processing of negative sample pairs affected by EIC in the propagation structure, a negative sample catcher is designed to identify negative sample pairs that cause gradient confusion; the negative sample catcher quantifies the gradient impact introduced by specific negative samples through the following formula: Among them, W e Represents the parameters of the propagation structure encoder. Based on the fact that the negative sample pairs that cause gradient confusion are not isolated, the similarity changes caused by the negative sample pairs are used to approximate their impact on the gradient. The formula We get, where ΔS ij is the negative sample pair (r i ,r j ) caused by the gradient effect, is the current iteration r i and r j The cosine similarity of , iter represents the number of iterations; Design a negative sample controller, generate an ignore set by filtering out negative sample pairs larger than the sample threshold, ignore the negative sample pairs that cause EIC, and use the formula IgS={NP(i,j)∣ΔS ij >ΔS r } We get, where r% is the neglected ratio, ΔS r It is the smallest value among the first r nodes after sorting.

9. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific process of obtaining the final representation is as follows: For each event e i , the representation H′ of the two enhanced propagation structures is calculated using the formula i and H″ i , H′ i and H″ i Splicing by formula m i =concat(H′ i ,H″ i ) Get e i The final representation of m i Input the fully connected layer and softmax layer, through the formula Get, among them, is the predicted probability distribution, FC(m i ) represents a fully connected layer.

10. The rumor detection technology with embedded consistency enhancement in graph contrastive learning according to claim 1 is characterized in that: The specific optimization process of the cross entropy loss and contrast loss is: Joint training based on rumor prediction to minimize prediction and the true label y i The cross entropy loss is as follows: For GCL loss, in order to prevent the gradient from being too small when ignoring some nodes, the gradient of negative samples is amplified by parameter α, and parameter β is introduced for positive samples. The final GCL loss is defined as follows: Where κ is the temperature parameter, and the final loss is composed of cross entropy loss and GCL loss, which is defined by the formula get.

Images