A label-propagation fake news detection method based on cross-modal tweet graph

Abstract

The present invention relates to the technical field of social media analysis and false information detection, and in particular to a label-propagation false news detection method based on a cross-modal tweet graph, comprising the following steps: S1, constructing a cross-modal tweet graph, which is used to capture the relevance of tweets; S2, based on the cross-modal tweet graph, using a feature context network to learn the relevance between tweets, and using the learning results to predict the truth or falsehood of tweets; S3, using a label propagation network to assign bidirectional label attention to the prediction results, so as to capture the positive and negative correlations of tweet prediction labels in the prediction results; S4, training the feature context network and the label propagation network to obtain a false news detection model; the cross-modal tweet graph constructed based on CLIP successfully captures the potential structure of true and false news of unknown events; at the same time, the discrimination between true and false news tweets is further enhanced by FCN; and the detection capability of the false news detection model for unseen events is improved by domain generalization loss.

Description

Technical Field

[0001] The present invention relates to the technical field of social media analysis and false information detection, and in particular to a method for detecting false news propagated by tags based on a cross-modal tweet graph. Background Art

[0002] Fake news detection is often treated as a binary classification problem. Early methods used features extracted from text content to train fake news classifiers, which requires expensive preprocessing and tedious feature engineering. Some existing studies have attempted to leverage explicit social context to establish connections between tweets, such as constructing graphs based on tweet retweets and tags and using graph neural networks (GNNs) to model the relationships between them. With the help of GNNs, adjacent tweets can be aggregated to obtain a more comprehensive representation.

[0003] However, existing tweet detection methods have the following flaws: 1. Relying solely on a single tweet to determine the authenticity of the original event is one-sided and unreliable. Many existing methods only use a single tweet for inference, without taking into account the relationships and interactions between tweets of the same type of event, and ignoring the social context information that can help detect fake news; 2. Graphs constructed solely based on social context are relatively sparse, and such sparse connections hinder the capture of possible interactions between tweets; 3. When training models for fake news detection, overfitting the training data is a major challenge, resulting in poor performance on unseen tweets. Summary of the Invention

[0004] The purpose of the present invention is to provide a label-propagation fake news detection method based on cross-modal tweet graphs, which mainly solves the technical problems mentioned in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A label-propagation fake news detection method based on cross-modal tweet graphs includes the following steps:

[0007] S1. Construct a cross-modal tweet graph, where the cross-modal tweet graph is used to capture the relevance of tweets.

[0008] S2. Based on the cross-modal tweet graph, a feature context network is used to learn the correlation between tweets, and the learning results are used to predict the authenticity of tweets;

[0009] S3. Using a label propagation network to assign bidirectional label attention to the prediction results obtained in S2, so as to capture the positive and negative correlation of tweet prediction labels in the prediction results;

[0010] S4. Training the feature context network and the label propagation network to obtain a fake news detection model;

[0011] S5. Use the fake news detection model to perform tweet detection.

[0012] Preferably, in said S1, the construction of said cross-modal tweet graph includes:

[0013] The CLIP model is used to extract tweet features, including features of the image modality and the text modality. The cosine similarity of all modality combinations is calculated, and a similarity threshold is set.

[0014] If the cosine similarity of at least one modality combination in the tweets exceeds the similarity threshold, there is potential correlation between the tweets.

[0015] Preferably, the modal combination specifically includes:

[0016] Image to image;

[0017] Text to text;

[0018] Image to text;

[0019] Text to Image.

[0020] Preferably, in said S2, the learning method of said feature context network includes:

[0021] The feature context network aggregates neighboring tweets for comparison with other tweets.

[0022] Preferably, in S3, the label propagation network is used to integrate the prediction results of adjacent tweets, smooth the labels by stacking bidirectional label attention layers, and analyze the positive and negative correlations of the labels to integrate the predicted labels corresponding to the tweets, wherein;

[0023] Set the final layer context features from the feature context network And the final layer context features The vector is input into the multi-layer perceptron to generate a two-dimensional logits vector, which is then converted into a class probability through the softmax function. in, Represents tweet d i is the true probability, Represents tweet d i The probability of being false;

[0024] In the label propagation network, L' label attention layers are used and the label from the lth label attention layer is given Used to pass category probabilities to subsequent label attention layers for updating and adjustment, where;

[0025] Labels for the second label attention layer To update:

[0026]

[0027]

[0028] in, ⊙ is element-by-element multiplication, is the bidirectional label attention weight between tweet i and tweet j, and the bidirectional label attention weight α ij Based on contextual features, is the input category probability; [·,·] is the connection, tanh is the hyperbolic tangent function, A and V are learnable weight matrices, and A∈R 2×2d , V∈R d×d , is the contextual feature of tweet i in layer l, is the contextual feature of tweet j in layer l.

[0029] Preferably, in S4, the training data set is divided into two disjoint subsets D s and subset D u , the subset D s For training the feature context network and the label propagation network, the subset D u Used to update the feature context network;

[0030] Using the subset D s The feature context network and the label propagation network are trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the feature context network. FCNmin and the minimized cross entropy loss L of the label propagation network LPNmin , the label propagation network from the subset D s The cross-modal tweet graph is used to learn initial classification prediction labels based on tweet features in the dataset. The label propagation network is used to effectively aggregate labels of adjacent tweets to capture the correlation between adjacent tweets. The cross-modal tweet graph is then generalized to unseen tweets, ensuring accurate classification on seen tweets through reasonable generalization. The cross-modal tweet graph is then normalized to maintain a consistent distribution on unseen tweets.

[0031] Set the feature set P of the seen dataset and the feature set Q of the unseen dataset of the tweets, and calculate the maximum mean difference (MMD) (P, Q) of the feature set P of the seen dataset and the feature set Q of the unseen dataset in the feature space. max:

[0032]

[0033] Where p is the sample in the feature set P of the seen dataset, q is the sample in the feature set Q of the unseen dataset, |P| is the number of samples in the feature set P of the seen dataset, and |Q| is the number of samples in the feature set Q of the unseen dataset;

[0034] The maximum mean difference is calculated for the true labels and false labels of the known feature set and the unknown feature set respectively, and the sum of the maximum mean differences is taken as the overall loss L MMD :

[0035]

[0036] in, is a set of known features with true labels, is a set of known features with false labels, is a set of unknown features with true labels, is a set of unknown features with false labels;

[0037] By minimizing the overall loss L MMD , ensuring that the feature set is consistent between known and unknown tweets in the feature space, and then minimizing the cross entropy loss L of the feature context network FCNmin , the minimized cross entropy loss L of the label propagation network LPNmin and the overall loss L MMD Combine them and use the loss function L all The fake news detection model is obtained by training the feature context network and the label propagation network end-to-end:

[0038] L all =L FCN +λL LPN +μL MMD

[0039] Among them, λ and μ are balancing hyperparameters.

[0040] Preferably, using the subset D s The feature context network is trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the feature context network FCNmin , and use the feature context network to minimize the cross entropy loss L FCNmin To add node features Cross-category changes:

[0041]

[0042] in, For tweets i The actual label of k is the real or false category of the label, is the prediction result corresponding to the label.

[0043] Preferably, the label propagation network is trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the label propagation network. LPNmin :

[0044]

[0045] in, For tweets i The actual label of k is the real or false category of the label, is the last L' layer output of the label propagation network, is the prediction result corresponding to the label.

[0046] Preferably, the subset D u Using the feature update method of graph convolutional network, the tweet features in the feature context network are updated. Perform normalization update on the first layer to obtain aggregated features

[0047]

[0048] N i ={d j |(d i ,d j )∈E}

[0049] N j ={d k |(d j ,d k )∈E}

[0050] Among them, tweet features is the activation function, W l is the learnable weight matrix, N i For tweets i The set of connected adjacent tweets, N j For tweets j The set of connected adjacent tweets, d k For tweets j Connected tweets.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The cross-modal tweet graph constructed based on CLIP successfully captures the latent structure of true and false news about unknown events. Furthermore, the discrimination between true and false news tweets is further enhanced through FCN, where the FCN features encode the differences between true and false news tweets, qualitatively confirming that LP can model the positive and negative correlations in FCN features. Furthermore, the domain generalization loss is used to match the feature distribution of unseen tweets with that of seen tweets, improving the fake news detection model's ability to detect unseen events. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 Flowchart of the method for detecting false news by label propagation based on cross-modal tweet graph of the present invention;

[0055] Figure 2 This is a structural block diagram of the label propagation fake news detection method based on cross-modal tweet graph of the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0057] like Figure 1-Figure 2 As shown, the present invention provides a label-propagation false news detection method based on a cross-modal tweet graph, comprising the following steps:

[0058] A label-propagation fake news detection method based on cross-modal tweet graphs includes the following steps:

[0059] Step 1: Construct a cross-modal tweet graph. The cross-modal tweet graph is used to capture the relevance of tweets. The construction of the cross-modal tweet graph includes:

[0060] The CLIP model is used to extract tweet features, which include features of image modality and text modality. The cosine similarity of all modality combinations is calculated, and a similarity threshold is set.

[0061] If the cosine similarity of at least one modality combination in the tweets exceeds the similarity threshold, there is potential correlation between the tweets.

[0062] Specifically, we use cross-modal tweet similarity to build a graph to potentially capture the correlation between tweets; first, we use CLIP to encode images and text into a unified space for similarity calculation, and calculate the cosine similarity of all possible modality combinations. Among them, modal combination C = {image to image, text to text, image to text, text to image}, a similarity threshold τ is defined in the set of all possible modal combinations C, if the similarity corresponding to the modal combination C in the tweet exceeds τ, the corresponding edge is included in the cross-modal tweet graph (E).

[0063] Formally, let D be a set of D tweets, each with d i Contains an image v i and text t i , where tweet d i =(v i ,t i )∈D,d j =(v j ,t j )∈D; and use the undirected graph G=(D,E) to represent the correlation between tweets. If tweet d i and tweets j The similarity between them is very high, then E contains edge e ij , where e ij =(d i ,d j ), and (d i ,d j ) is the CLIP feature extracted from the image modality and text modality, that is, when at least one modality combination c When, e ij ∈E.

[0064] Step 2: Based on the cross-modal tweet graph, a feature context network (FCN) is used to learn the correlation between tweets, and the learning results are used to predict the authenticity of tweets. The learning method of the feature context network includes comparative learning using tweets adjacent to a certain tweet.

[0065] Step 3: Use the Label Propagation Network (LPN) to assign bidirectional label attention to the prediction results obtained in Step 2 to capture the positive and negative correlation of tweet prediction labels in the prediction results;

[0066] Specifically, the label propagation network is used to integrate the prediction results of adjacent tweets, smooth the labels by stacking bidirectional label attention layers, and analyze the positive and negative correlations of the labels to integrate the predicted labels corresponding to the tweets, where;

[0067] Set the final layer context features from the feature context network And the final layer context features The vector is input into the multi-layer perceptron (MLP) to generate a two-dimensional logits vector, which is then converted into class probability through the softmax function. in, Represents tweet d i is the true probability, Represents tweet d i The probability of being false;

[0068] Use L' label attention layers in the label propagation network and give the label from the lth label attention layer (l=1,...,L'-1) Used to pass category probabilities to subsequent label attention layers for updating and adjustment, where;

[0069] Labels for the second label attention layer To update:

[0070]

[0071]

[0072] in, ⊙ is element-by-element multiplication, is the bidirectional label attention weight between tweet i and tweet j, and the bidirectional label attention weight α ij Based on contextual features, is the input category probability; [·,·] is the connection, tanh is the hyperbolic tangent function, A and V are learnable weight matrices, and A∈R 2×2d , V∈R d×d , is the contextual feature of tweet i in layer l, is the contextual feature of tweet j in layer l.

[0073] It should be noted that: in this embodiment, when α ij When it is positive, it means that the tweet d' j and tweets i ' has a positive correlation with the label probability, which means that in a certain layer of the label attention layer, if the tweet d' j The probability of being considered true increases, then the next layer of tweets d i' is also considered to be true. In this case, the label propagation network will be more inclined to influence the label probabilities between adjacent tweets, believing that they are describing similar content. When α ij When it is a negative value, it means that the tweet d' j and tweets i ' has a negative correlation with the label probability, which means that in a certain layer of the label attention layer, if the tweet d' j The probability of being considered true increases, then the next layer of tweets d i 'The probability of being considered true decreases; in this case, the label propagation network will be more cautious in judging the label probabilities between adjacent tweets, believing that they are describing opposite facts or opinions; this consideration of competition and contradictory relationships can help the label propagation network better understand the relationship between tweets and improve the reliability and accuracy of the label propagation network.

[0074] Step 4: Train the feature context network and label propagation network to obtain a fake news detection model. When training a fake news detection model for fake news detection, the training data is prone to overfitting, resulting in poor performance in unseen tweets (especially tweets about new events). In order to achieve generalization of the fake news detection model, it is key to ensure that the feature distributions of seen and unseen training data are consistent, and the training dataset is divided into two disjoint subsets D. s and subset D u , subset D s For training feature context network and label propagation network, subset D u Used to update the feature context network;

[0075] Among them, using subset D s The feature context network and label propagation network are trained by minimizing the cross entropy loss of the predicted label, and the minimized cross entropy loss L of the feature context network is obtained. FCNmin and minimize the cross entropy loss L of the label propagation network LPNmin , the label propagation network from subset D s The cross-modal tweet graph is generalized to unseen tweets, ensuring accurate classification on seen tweets through reasonable generalization, and normalizing the cross-modal tweet graph to maintain a consistent distribution on unseen tweets.

[0076] It should be noted that: in this embodiment, reasonable generalization refers to the ability of the fake news detection model to accurately classify unseen tweets while maintaining the consistent distribution of unseen tweets and seen tweets in the feature space; normalization means minimizing the maximum mean difference (MMD) loss to ensure that the features learned by the feature context network are consistent between seen tweets and unseen tweets, thereby improving the generalization ability of the fake news detection model on unseen tweets.

[0077] Then set the feature set P of the seen dataset and the feature set Q of the unseen dataset of the tweets, and calculate the maximum mean difference (MMD) (P, Q) of the feature set P of the seen dataset and the feature set Q of the unseen dataset in the feature space. max :

[0078]

[0079] Where p is the sample in the feature set P of the seen dataset, q is the sample in the feature set Q of the unseen dataset, |P| is the number of samples in the feature set P of the seen dataset, and |Q| is the number of samples in the feature set Q of the unseen dataset;

[0080] The maximum mean difference is calculated for the true labels and false labels of the known feature set and the unknown feature set respectively, and the sum of the maximum mean differences is taken as the overall loss L MMD :

[0081]

[0082] in, is a set of known features with true labels, is a set of known features with false labels, is a set of unknown features with true labels, is a set of unknown features with false labels.

[0083] By minimizing the overall loss L MMD , ensuring that the feature set is consistent between known and unknown tweets in the feature space, and then minimizing the cross entropy loss L of the feature context network FCNmin , minimize the cross entropy loss L of the label propagation network LPNmin and the overall loss L MMD Combine them and use the loss function L all By training the feature context network and label propagation network end-to-end, we obtain a fake news detection model:

[0084] L all =L FCN +λL LPN +μL MMD

[0085] Among them, λ and μ are balancing hyperparameters.

[0086] Furthermore, using subset D s The feature context network is trained by minimizing the cross entropy loss of the predicted label, and the minimized cross entropy loss L of the feature context network is obtained. FCNmin , and use the minimized cross entropy loss L of the feature context network FCNmin To add node features Cross-category changes:

[0087]

[0088] in, For tweets i The actual label of k is the real or false category of the label, is the prediction result corresponding to the label.

[0089] Furthermore, the label propagation network is trained by minimizing the cross entropy loss of the predicted label, and the minimized cross entropy loss L of the label propagation network is obtained. LPNmin :

[0090]

[0091] in, For tweets i The actual label of k is the real or false category of the label, is the last L' layer output of the label propagation network, is the prediction result corresponding to the label.

[0092] Furthermore, subset D u Using the feature update method of graph convolutional network (GCN), tweet features in the feature context network are updated. Perform normalization update in the first layer (l=1,...,L-1) to obtain aggregated features

[0093] N i ={d j |(d i ,d j )∈E}

[0094] N j ={d k |(d j ,d k )∈E}

[0095] Among them, tweet features is the activation function (such as ReLU), W l is the learnable weight matrix, N i For tweets i The set of connected adjacent tweets, N j For tweets j The set of connected adjacent tweets, d k For tweets j Connected tweets.

[0096] Step 5: Use fake news detection model to detect tweets.

[0097] In this example, to evaluate the proposed fake news detection model, we compared it with several leading detection methods on the Twitter dataset. These include EANN, SpotFake, MVAE, SAFE, MCAN, HMCAN, and CLIP. Following the conventions of fake news detection metrics, we used accuracy, precision, recall, and F1 score to evaluate the performance of the fake news detection model. The comparative experimental results are shown in the following table:

[0098]

[0099]

[0100] Specifically, FCN-LP can be cascaded into any fake news detector to improve its performance. To this end, a detector is first trained on a fake news dataset, and then tweet features are extracted from the layer before the prediction head. The extracted tweet features are used as node features of FCN-LP; for the CLIP baseline, image and text features are directly concatenated through CLIP as node features, and the node features are fixed when training FCN-LP.

[0101] Furthermore, FCN-LP can improve all indicators of existing multimodal fake news detectors, ranging from 0.57% to 8.07%; the interaction between tweet features and the positive and negative correlation based on multimodal similarity graph modeling are conducive to detection, especially the detection based on CLIP features shows excellent performance on the Twitter dataset; because CLIP is pre-trained on a larger multimodal dataset, the CLIP features have better generalization ability when applied to the test set; and, on the one hand, after training on the fake news dataset, tweet features extracted from other methods will be affected by potential overfitting, which can easily lead to limited performance improvement of FCN-LP relative to the CLIP-based baseline; on the other hand, because the pre-training dataset of CLIP mainly comes from the English corpus, it does not show strong performance on the Weibo dataset dominated by Chinese text. MCAN and HMCAN use Chinese BERT to extract text features. By using this fake news detection model, the performance of the Weibo dataset can be improved.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A label-propagation fake news detection method based on cross-modal tweet graph, characterized by: The following steps are involved: S1. Construct a cross-modal tweet graph, where the cross-modal tweet graph is used to capture the relevance of tweets. S2. Based on the cross-modal tweet graph, a feature context network is used to learn the correlation between tweets, and the learning results are used to predict the authenticity of tweets; S3. Using a label propagation network to assign bidirectional label attention to the prediction results obtained in S2, so as to capture the positive and negative correlation of tweet prediction labels in the prediction results; S4. Training the feature context network and the label propagation network to obtain a fake news detection model; The training dataset is divided into two disjoint subsets D s and subset D u , the subset D s For training the feature context network and the label propagation network, the subset D u Used to update the feature context network; Using the subset D s The feature context network and the label propagation network are trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the feature context network. FCNmin and the minimized cross entropy loss L of the label propagation network LPNmin , the label propagation network from the subset D s The cross-modal tweet graph is used to learn initial classification prediction labels based on tweet features in the dataset. The label propagation network is used to effectively aggregate labels of adjacent tweets to capture the correlation between adjacent tweets. The cross-modal tweet graph is then generalized to unseen tweets, ensuring accurate classification on seen tweets through reasonable generalization. The cross-modal tweet graph is then normalized to maintain a consistent distribution on unseen tweets. Set the feature set P of the seen dataset and the feature set Q of the unseen dataset of the tweets, and calculate the maximum mean difference (MMD) (P, Q) of the feature set P of the seen dataset and the feature set Q of the unseen dataset in the feature space. max : Where p is the sample in the feature set P of the seen dataset, q is the sample in the feature set Q of the unseen dataset, |P| is the number of samples in the feature set P of the seen dataset, and |Q| is the number of samples in the feature set Q of the unseen dataset; The maximum mean difference is calculated for the true labels and false labels of the known feature set and the unknown feature set respectively, and the sum of the maximum mean differences is taken as the overall loss L MMD : in, is a set of known features with true labels, is a set of known features with false labels, is a set of unknown features with true labels, is a set of unknown features with false labels; By minimizing the overall loss L MMD , ensuring that the feature set is consistent between known and unknown tweets in the feature space, and then minimizing the cross entropy loss L of the feature context network FCNmin , the minimized cross entropy loss L of the label propagation network LPNmin and the overall loss L MMD Combine them and use the loss function L all The fake news detection model is obtained by training the feature context network and the label propagation network end-to-end: L all =L FCN +λL LPN +μL MMD Among them, λ and μ are balancing hyperparameters; S5. Use the fake news detection model to perform tweet detection.

2. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: In said S1, the construction of said cross-modal tweet graph includes: The CLIP model is used to extract tweet features, including features of the image modality and the text modality. The cosine similarity of all modality combinations is calculated, and a similarity threshold is set. If the cosine similarity of at least one modality combination in the tweets exceeds the similarity threshold, there is potential correlation between the tweets.

3. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 2, characterized in that: The modal combination specifically includes: Image to image; Text to text; Image to text; Text to Image.

4. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: In said S2, the learning method of said feature context network includes: The feature context network aggregates neighboring tweets for comparison with other tweets.

5. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: In S3, the label propagation network is used to integrate the prediction results of adjacent tweets, smooth the labels by stacking bidirectional label attention layers, and analyze the positive and negative correlations of the labels to integrate the predicted labels corresponding to the tweets, wherein; Set the final layer context features from the feature context network And the final layer context features The vector is input into the multi-layer perceptron to generate a two-dimensional logits vector, which is then converted into a class probability through the softmax function. in, Represents tweet d i is the true probability, Represents tweet d i The probability of being false; In the label propagation network, L' label attention layers are used and the label from the lth label attention layer is given Used to pass category probabilities to subsequent label attention layers for updating and adjustment, where; Labels for the second label attention layer To update: in, ⊙ is element-by-element multiplication, is the bidirectional label attention weight between tweet i and tweet j, and the bidirectional label attention weight α ij Based on contextual features, is the input category probability; [·,·] is the connection, tanh is the hyperbolic tangent function, A and V are learnable weight matrices, and A∈R 2×2d , V∈R d×d , is the contextual feature of tweet i in layer l, are the contextual features of tweet j in layer l.

6. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: Using the subset D s The feature context network is trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the feature context network FCNmin , and use the feature context network to minimize the cross entropy loss L FCNmin To add node features Cross-category changes: in, For tweets i The actual label of k is the real or false category of the label, is the prediction result corresponding to the label.

7. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: The label propagation network is trained by minimizing the cross entropy loss of the predicted label to obtain the minimized cross entropy loss L of the label propagation network. LPNmin : in, For tweets i The actual label of k is the real or false category of the label, is the last L' layer output of the label propagation network, is the prediction result corresponding to the label.

8. The method for detecting fake news propagated by hashtags based on a cross-modal tweet graph according to claim 1, characterized in that: The subset D u Using the feature update method of graph convolutional network, the tweet features in the feature context network are updated. Perform normalization update on the first layer to obtain aggregated features N i ={d j |(d i ,d j )∈E} N j ={d k |(d j ,d k )∈E} Among them, tweet features is the activation function, W l is the learnable weight matrix, N i For tweets i The set of connected adjacent tweets, N j For tweets j The set of connected adjacent tweets, d k For tweets j Connected tweets.

Images