A fake news identification method based on heterogeneous graph contrastive learning

Abstract

The present invention relates to a false news identification method based on heterogeneous graph contrast learning, comprising the following steps: 1) constructing a false news heterogeneous graph about news nodes; 2) obtaining text feature information F of news; 3) constructing a subgraph based on two meta-paths of news-user-news and news-topic-news; 4) obtaining a first group of views through identity transformation for the subgraphs, and obtaining a second group of views through random structural transformation; 5) extracting features for the two groups of views through a graph neural network, and obtaining corresponding feature matrices respectively; 6) training the parameters of the graph neural networks GNN1 and GNN2 through contrast learning and back propagation, inputting the initial text features of the news nodes to obtain news node features; 7) using the obtained news node features as input of a multi-layer perceptron and training it to identify false news. Compared with the prior art, the present invention has the advantages of being adaptable to small sample annotation, good robustness, and high practicality.

Description

Technical Field

[0001] The present invention relates to the field of network interactive false news detection, and in particular to a false news identification method based on heterogeneous graph contrast learning. Background Art

[0002] In today's society, online social networks have become an indispensable information exchange platform for people. At the same time, social platforms are also becoming a hotbed for the spread of false news without people noticing. Due to the popularity of social platforms, the spread of these false news has also greatly increased. In summary, false news identification for social platforms is a technology with great practical value.

[0003] Fake news identification has a long history of research in the industry, and many researchers have proposed some targeted solutions to this problem. If the techniques for fake news identification are classified based on methods, they can be divided into text-based methods and methods that integrate social network interaction information. Text-based methods regard the task of fake news identification as a text classification task in natural language processing. Kaliyar and Goswamiz proposed a method to identify fake news using the BERT language model. The disadvantage of this type of method is that a large amount of text corpus is required to train the language model, and self-supervised pre-training may not necessarily improve the performance on supervised downstream tasks. In addition, when news text that does not appear in some corpora appears, this type of method performs poorly. For adversarial text news generated by algorithms, pure text methods may be completely ineffective. To compensate for these problems, the interactive relationship between users and news in social networks can be integrated on the basis of text. Through this interactive information, the deep structural relationship between news and news can be discovered, and then this information can be mined through graph representation learning methods to improve the performance of fake news identification. Dou and Shu considered the propagation paths between news posts on the Twitter platform, constructed a tree-like propagation graph, obtained the community information of the news through representation learning, and then combined the text with the community information for classification. Ren and Wang modeled users, news, and topics together into a heterogeneous graph, and then used the learned news node features for classification. However, most of these methods still require a large amount of labeled data for training, which is inconsistent with the conditions in reality.

[0004] In summary, the existing fake news identification methods have the following shortcomings:

[0005] (1) Algorithms based purely on text lack robustness and are highly dependent on corpora, making them unable to effectively identify new data.

[0006] (2) Both text-based and community information-integrated methods require a large amount of labeled data for training, which is inconsistent with actual conditions and also has certain requirements on the quality of the labeled data.

[0007] (3) Some methods of integrating community information are too complex and require obtaining a large amount of user personal information to generate user features, which increases the complexity of the model. If the quality of user feature data is not high, it will significantly affect the performance of the model. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a fake news identification method based on heterogeneous graph comparative learning.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A fake news identification method based on heterogeneous graph contrast learning includes the following steps:

[0011] 1) Obtain news-related data from social networking platforms, including the original text content of the news The user groups that interact and forward news and the category and subject information of the news are collected, and a heterogeneous graph of fake news about news nodes is constructed;

[0012] 2) According to the original text content of the news, the text feature information F of the news is obtained through the pre-trained language model as the initial text feature of the news node;

[0013] 3) Through the two meta-paths of news-user and news-topic, construct sub-graphs based on the two meta-paths of news-user-news and news-topic-news and

[0014] 4) For subgraphs and The first set of views A are obtained by identity transformation bu With A bo , the second set of views A′ is obtained by random structural transformation bu and A′ no ;

[0015] 5) For the two sets of views, feature extraction is performed through the graph neural network GNN1 and GNN2, and the corresponding feature matrices Z = GNN1 (A nu ,A no ) and Z′=GNN2(A′ nu ,A′ no );

[0016] 6) The parameters of the graph neural networks GNN1 and GNN2 are trained through comparative learning and back propagation, the graph neural network GNN1 of the first set of views is used as the encoder of the news node, and the initial text features of the news node are input to obtain the news node features;

[0017] 7) The obtained news node features are used as input of a multi-layer perceptron and trained to obtain a trained classification model, and false news is identified based on the trained classification model.

[0018] In the step 1), for the user group that interacts with the news and forwards it, only the connection relationship between the news and the user is obtained. ni ; For the category and topic information of news, if there is a relevant topic in the news, it is used directly, otherwise the topic classification of news is obtained through the topic generation method of LDA class, and then the relationship between news and topic categories is constructed. no .

[0019] In the step 2), for short texts, they are directly used as inputs of the language model; for texts with titles and body parts, they are used as inputs of the model separately and then concatenated; for the text containing the interactive comment sequence below, the interactive text information is captured through the LSTM network, and all features are concatenated as the final text feature information.

[0020] In the step 4), the identity transformation specifically does not make any changes to the two subgraphs, and the random structural transformation includes edge sampling, edge addition and Personalized Pagerank.

[0021] The edge sampling is specifically as follows:

[0022] Sample the adjacency matrix of the subgraph, and randomly select some elements in the adjacency matrix and set them to 0;

[0023] The edge increase is specifically:

[0024] Randomly add elements to the adjacency matrix of the subgraph and randomly select a portion of the elements and set them to 1.

[0025] The Personalized Pagerank is specifically:

[0026] The PPR score reflecting the importance of the element is calculated for each element in the adjacency matrix, and the edges with low importance are removed after sorting the importance.

[0027] The step 5) specifically comprises the following steps:

[0028] 51) Input the adjacency matrix of the subgraph into a graph convolutional neural network to obtain the feature vector;

[0029] 52) Substitute the feature vector into a semantic attention network to obtain the feature matrix.

[0030] In the step 52), the calculation formula for each point in the feature matrix Z is:

[0031]

[0032] In the formula, z i is the i-th row in the feature matrix Z, corresponding to the feature of the i-th news node, N represents the set of news nodes, and the semantic attention network consists of a single-layer perceptron and a weight parameter. Parameters a, W, and b are the parameters to be learned by the semantic attention network, where parameters W and b are the learning parameters of the single-layer perceptron, and a is the weight parameter, which is used to adaptively adjust the representation under different views. In feature z i The superscript and subscript k represent the kth sub-image in the view, and k = 1, 2, β k is the representation corresponding to different subgraphs The weight of fusion, and the weight w k After normalization.

[0033] The step 6) specifically comprises the following steps:

[0034] 61) For each news node, select positive and negative samples based on the heterogeneous structure similarity HPS;

[0035] 62) According to the obtained positive and negative samples, the parameters of the graph neural network are trained by minimizing the total loss function, specifically:

[0036] The loss function for the i-th news node in the first set of views after identity transformation is The expression is:

[0037]

[0038] Similarly, the loss function of the i-th news node in the second set of views obtained after random structural transformation is The expression is:

[0039]

[0040] in, represents the positive sample set for the i-th news node, represents the negative sample set for the i-th news node, ξ is an adjustable parameter, z i is the node feature of the i-th news node in the first group of views, z j′ is the node feature of the jth news node in the second group of views, sim(z i ,z′ j ) represents the news node feature z i and z′ j The similarity between

[0041] The total loss function is obtained by summing up the losses of all news nodes and averaging them. The expression is:

[0042]

[0043] Among them, θ is a manually set hyperparameter.

[0044] In the step 61), the expression of the heterogeneous structure similarity HPS of news nodes i and j is:

[0045]

[0046] Where S is the number of meta-paths, and S = 2, λ s To balance the weight parameters between different meta-paths, and are the heterogeneous neighbor sets on the meta-path s respectively.

[0047] In the step 61), for each news node, the first τ points with the largest HPS are selected as positive samples, and the rest are selected as negative samples.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. The fake news identification method provided by the present invention utilizes the contrastive learning technology, which can still achieve good performance under the condition of small sample annotation, which makes it more applicable in real business scenarios.

[0050] 2. Compared with text-based recognition technology, the fake news recognition method provided by the present invention can effectively integrate the interactive attributes of the news itself, making it more robust when facing new events and text counterattacks.

[0051] 3. Compared with other methods that require complex user information, the present invention uses the method of meta-path extraction to bypass the problem of displaying user features, and has better practicality in scenarios where user information is difficult to obtain. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is an explanatory diagram of the principle method of the present invention.

[0053] Figure 2This is a performance comparison between the present invention and the best method HAN other than the present invention when the proportion of labeled samples is 5%, 10%, 20%, 30% and 50% respectively. The horizontal axis represents the proportion of labeled samples to all data, and the vertical axis represents the classification accuracy. HAN represents the HAN method, and FnCRL represents the present invention. FIG. (2a) is a performance comparison on the PolitiFact-old dataset, and FIG. (2b) is a performance comparison on the PolitiFact-new dataset.

[0054] Figure 3 It is the projection of the node features learned by the present invention on the two-dimensional plane, among which, Figure (3a) is the projection of the SentenceTransformers algorithm, Figure (3b) is the projection of the HAN algorithm, Figure (3c) is the projection of the HCN algorithm, and Figure (3d) is the projection of the method of the present invention. DETAILED DESCRIPTION

[0055] The following is a detailed description of the embodiments of the present invention: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. It should be pointed out that for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

[0056] Example

[0057] like Figure 1 As shown, the present invention provides a fake news identification method based on heterogeneous graph contrast learning, and the main steps of the method are as follows:

[0058] The first step is to crawl relevant news content, users participating in the interaction, and the topic of the news itself from the Internet according to the news that needs to be identified. If the news itself does not provide a relevant topic, the LDA algorithm can be used to generate the topic.

[0059] The second step is to obtain the feature information of the news text through the pre-trained language model Sentence Transformer. For short texts, directly use them as the input of the language model. For texts with titles and text parts, use them as the input of the model separately and then splice them together. If there is a long sequence of interactive comments below the text, LSTM can be used to capture the interactive text information between them, and all features can be spliced ​​together as the final text feature content. For the case of inconsistent dimensions after splicing, the PCA method can be used to adjust the dimensions. Regardless of which of the above methods is used, a text feature F with the same dimension is obtained in the end = [F1,…,F |N|}.

[0060] The third step is to construct subgraphs based on the two meta-paths of news-user and news-topic through the two meta-paths of news-user-news and news-topic-news. and

[0061] Step 4: For the two subgraphs and Apply two different sets of transformations. The first set of transformations is the identity transformation, which does not make any changes to the two sub-images, and obtains the first set of views A. nu With A no The second set of transformations is a random structural transformation, which makes some changes to the edge structure of the two subgraphs to enhance the effect of contrastive learning. Common transformations include edge sampling, edge augmentation, and Personalized Pagerank. In this example, the appropriate graph transformation can be selected based on the performance on the actual data set. Through the second set of transformations, the second set of views A′ can be obtained. nu and A′ no .

[0062] In the fifth step, for the two sets of views obtained, the graph neural network is used to learn their representations. Taking the first set of views as an example, the adjacency matrix A of the subgraphs corresponding to the two meta-paths of news-user and news-topic under the first set of views is obtained from the previous step. nu With A no The first layer of the network converts the A under the first set of views nu With A no Input into a graph convolutional neural network to get the feature vector representation and Where F is the text feature obtained in the second step, and for Next, we will transform h 1 With h 2 Substitute into a semantic attention network to obtain the final feature matrix Z. Each point z in the matrix Z i The calculation formula is as follows:

[0063]

[0064] Among them, z i is the i-th row in the feature matrix Z, corresponding to the feature of the i-th point. The semantic attention network consists of a single-layer perceptron and a weight parameter. Parameters a, W, and b are the parameters to be learned by the semantic attention network. Parameters W and b correspond to the parameters of the perceptron, and a is the weight parameter, which can adaptively adjust the representation under different views. In the total feature z iHere, the superscript and subscript k correspond to the kth subgraph in the view (in this algorithm, k = 2), and the subscript i represents the i-th node. k After normalization, we get β k , as the corresponding subgraphs The weight of the fusion.

[0065] Similarly, we can get the feature vector h′ of the second view with the graph transformation added 1 and h′ 2 , and then calculate the corresponding feature matrix Z′;

[0066] The sixth step is to call contrastive learning to perform representation learning on the heterogeneous graph. Specifically, the heterogeneous structure similarity (HPS) is used as the basis for selecting positive and negative samples. The expression of the heterogeneous structure similarity HPS of news nodes i and j is:

[0067]

[0068] Where S is the number of meta-paths, and S = 2, λ s To balance the weight parameters between different meta-paths, and are the heterogeneous neighbor sets on the meta-path s respectively.

[0069] Define the first τ points with the largest HPS as positive samples of a certain point, and the rest as negative samples. Note that under this definition, a node must be its own positive sample, which is consistent with the intuition of positive sample design.

[0070] Step 7: Based on the positive and negative samples obtained in the previous step, the loss function for node i in the first untransformed view is as follows:

[0071]

[0072] Symmetrically, the loss function for node i in the second view after transformation is as follows

[0073]

[0074] In the formula, Represents the positive sample set for the i-th node, represents the negative sample set for the ith node, ξ is an adjustable parameter. i is the node feature of the i-th news node in the first group of views, z′ j is the node feature of the jth news node in the second group of views, sim(z i ,z′ j ) represents node z i and z′ jThe similarity of can generally be taken as the inner product, that is, sim(z i ,z′ j )=z i T z′ j .

[0075] The total loss function can be obtained by summing up the losses of all nodes and averaging them. θ is a manually set hyperparameter with a default value of 0.5.

[0076]

[0077] Finally, the initial text feature F of the news node is input into the graph convolutional neural network and semantic attention model under the first view to obtain the learned news node representation. This representation is input into a multi-layer perceptron and trained to obtain the final classification model. To identify a piece of fake news, you only need to input its text features into the trained multi-layer perceptron to get the final result.

[0078] In order to more clearly illustrate the technical solution in this embodiment, the following specific application implementations are listed for further explanation:

[0079] Suppose a company wants to manage information on the Internet. First, it needs to find out the fake news on the Internet. For example, if the company wants to crack down on fake news on Weibo, it can be expressed as the following mathematical problem: Given the text content of the news itself And the edge relationship between users who interact with news and news nu , the edge relationship between news-related topics and news no , hoping to learn a z for each news node that takes into account both text features and network structure features i , and then use this feature to train the final classifier.

[0080] In order to achieve the above goals, the first step is to crawl the data in the required format. During the crawling process, you can use the data acquisition interface provided by the network platform, or you can write your own script to crawl the data. Taking the Weibo platform as an example, the official provides data acquisition interfaces in various programming languages, which can easily obtain data in the required format.

[0081] In the actual application of this method, the second step may encounter multilingual problems. Taking Weibo as an example, if you encounter Chinese news, the Sentence Transformer model trained on English corpus will not perform well. In order to solve this problem, you can follow the pre-training method of Sentence Transformer to train a Chinese version of the pre-trained model on the Chinese corpus. The Chinese corpus required for the pre-training process is free and open on the Internet. In summary, the multilingual problem can be solved by replacing the pre-trained model, which does not affect the generality of this method.

[0082] Next, the method is implemented according to steps 3 to 7 in the embodiment. It is worth noting that the graph transformation in step 5 needs to be selected based on the performance on the actual data set. The following common graph transformation techniques are defined, including:

[0083] (1) Edge sampling: given the adjacency matrix of a graph right To sample the elements in , randomly select some elements and set them to 0. In order to ensure that edge sampling does not have too much impact on the structure, a parameter λ is set to control the strength of edge sampling and require

[0084] (2) Edge enhancement: Given the adjacency matrix of a graph right The elements in are randomly added, and a part of the elements are randomly selected and set to 1. In order to ensure that edge sampling does not have too much impact on the structure, a parameter λ′ is set to control the strength of edge enhancement and requires

[0085] (3) Personalized Pagerank: The PPR score is calculated for each element in the adjacency matrix, reflecting the importance of this element in the graph structure. After sorting the importance, those edges with low importance are removed.

[0086] Other common graph transformation techniques include Markov diffusion, etc., which are not listed here one by one. In actual application scenarios, users can choose the appropriate graph transformation based on the performance on the validation set during training. In the experimental results, edge sampling and edge enhancement are the two most effective graph transformations. The performance difference between the two is not large, and they can be used as the default choice.

[0087] Finally, we only need to input the learned representation of news nodes into a classifier for training to obtain the classification model for news node classification. Here, a certain amount of news needs to be labeled as true news and false news. In the experimental results, the classification accuracy can reach more than 90% on the experimental data set as long as the proportion of labeled data in the total data reaches 5%.

[0088] The effectiveness of this embodiment can be further illustrated by the following simulation experiment. It should be noted that the parameters used in the experiment do not affect the generality of the present invention.

[0089] 1) Simulation conditions:

[0090] CPU Intel(R)Xeon(R)E5-2630 v4@2.20GHz, RAM 256.00GB, operating system CentOS 7, GPU NVIDIA TITAN RTX, and the simulation program is written in Python.

[0091] 2) Simulation content:

[0092] In order to examine the performance of the present invention on an actual data set, a data set from PolitiFact, a news authenticity verification website, was selected. PolitiFact is an American nonprofit organization that specializes in studying the authenticity of various news on the Internet, especially on social platforms. First, the PolitiFact part of the FakeNewsNet data set, which was open sourced by researchers at Ohio State University, was selected. This data set includes the text content of the news itself and user information that interacts with the news. In addition, the news topics on the original news website were crawled through the original news links. These serve as the first data set PolitiFact-old in the simulation. In order to further verify the effectiveness of the present invention, the latest verified news spread on the Twitter platform was crawled from the PolitiFact website. Like PolitiFact-old, these data include the text content of the news itself, user information that interacts with the news, and news topic information provided by the PolitiFact website.

[0093] In order to verify the effect of the present invention, the following three experiments were mainly conducted. First, the performance of the model in a weakly supervised real-world scenario was compared. Weak supervision means that the labeled training data accounts for 5% of the total data. 20% of the total data was selected as the validation set, and 30% of the total data was selected as the test set. In order to ensure the fairness of the classification problem, the ratio of positive and negative samples in the validation set and the test set is 1:1. The results are shown in Table 1, which compares the performance of the model in a weakly supervised real-world scenario. Secondly, the accuracy of HAN, the best method other than the present invention, and the present invention when the total proportion of labeled data is 5%, 10%, 20%, 30% and 50% is compared. The results are Figure 2 Finally, the classification results of the pure language model, GCN, HAN and the news node features learned by the present invention are visually compared on a two-dimensional plane. Figure 3 ,From the effect, it can be seen that the features learned by the ,presentation are separated best.

[0094] Table 1 Experimental results

[0095] PolitiFact-old PolitiFact-new Accuracy Accuracy Text Model (Semi-Supervised) 0.737 0.857 GCN (weak supervision) 0.717 0.750 DGI (weak supervision) 0.848 0.912 HAN (Weak Supervision) 0.800 0.883 The present invention (weak supervision) 0.900 0.946

[0096] There are several methods for comparison:

[0097] 1. Sentence Transformers: pure language model, without considering network structure information

[0098] 2. GCN: Graph Convolutional Neural Network, published by Kipf and Welling at the ICLR conference in 2017. Since it is a homogeneous graph method, the better result of the two meta-path views is used for comparison.

[0099] 3. DGI: Contrastive Learning Technique on Homogeneous Graphs, by Fedus et al. published this method at the ICLR conference in 2019. Since it is a homogeneous graph method, the better result of the two meta-path views is used for comparison.

[0100] 4. HAN: Heterogeneous Graph Attention Networks, by Wang, Ji et al., published in the 2019 WWW conference.

[0101] It can be seen from the final results that the present invention outperforms other methods on both datasets, especially under the condition of small sample annotation.

[0102] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A fake news identification method based on heterogeneous graph contrast learning, characterized in that: The following steps are involved: 1) Obtain news-related data from social networking platforms, including the original text content of the news The user groups that interact and forward news and the category and subject information of the news are collected, and a heterogeneous graph of fake news about news nodes is constructed; 2) According to the original text content of the news, the text feature information F of the news is obtained through the pre-trained language model as the initial text feature of the news node; 3) Through the two meta-paths of news-user and news-topic, construct sub-graphs based on the two meta-paths of news-user-news and news-topic-news and 4) For subgraphs and The first set of views A are obtained by identity transformation nu With A no , the second set of views A is obtained by random structural transformation ′ nu With A ′ no ; 5) For the two sets of views, feature extraction is performed through the graph neural network GNN1 and GNN2, and the corresponding feature matrices Z = GNN1 (A nu ,A no ) and Z′=GNN2(A′ nu ,A′ no ); 6) The parameters of the graph neural networks GNN1 and GNN2 are trained through comparative learning and back propagation, the graph neural network GNN1 of the first set of views is used as the encoder of the news node, and the initial text features of the news node are input to obtain the news node features; 7) The obtained news node features are used as the input of the multi-layer perceptron and trained to obtain a trained classification model, and false news identification is performed based on the trained classification model; in the step 1), for the user group that interacts with the news and forwards it, only the connection relationship between the news and the user is obtained. nu ; For the category and topic information of news, if there is a relevant topic in the news, it is used directly, otherwise the topic classification of news is obtained through the topic generation method of LDA class, and then the relationship between news and topic categories is constructed. no .

2. According to claim 1, a method for identifying fake news based on heterogeneous graph contrastive learning is characterized in that: In the step 2), for short texts, they are directly used as inputs of the language model; for texts with titles and body parts, they are used as inputs of the model separately and then concatenated; for the text containing the interactive comment sequence below, the interactive text information is captured through the LSTM network, and all features are concatenated as the final text feature information.

3. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 1, characterized in that: In the step 4), the identity transformation specifically does not make any changes to the two subgraphs, and the random structural transformation includes edge sampling, edge addition and Personalized Pagerank.

4. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 3, characterized in that: The edge sampling is specifically: Sample the adjacency matrix of the subgraph, and randomly select some elements in the adjacency matrix and set them to 0; The edge increase is specifically: Randomly add elements to the adjacency matrix of the subgraph, and randomly select some elements and set them to 1; The Personalized Pagerank is specifically: The PPR score reflecting the importance of the element is calculated for each element in the adjacency matrix, and the edges with low importance are removed after sorting the importance.

5. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 1, characterized in that: The step 5) specifically comprises the following steps: 51) Input the adjacency matrix of the subgraph into a graph convolutional neural network to obtain the feature vector; 52) Substitute the feature vector into a semantic attention network to obtain the feature matrix.

6. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 5, characterized in that: In the step 52), the calculation formula for each point in the feature matrix Z is: In the formula, z i is the i-th row in the feature matrix Z, corresponding to the feature of the i-th news node, N represents the set of news nodes, and the semantic attention network consists of a single-layer perceptron and a weight parameter. Parameters a, W, and b are the parameters to be learned by the semantic attention network, where parameters W and b are the learning parameters of the single-layer perceptron, and a is the weight parameter, which is used to adaptively adjust the representation under different views. In feature z i The superscript and subscript k represent the kth sub-image in the view, and k = 1, 2, β k is the representation corresponding to different subgraphs The weight of fusion, and the weight w k After normalization.

7. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 6, characterized in that: The step 6) specifically comprises the following steps: 61) For each news node, select positive and negative samples based on the heterogeneous structure similarity HPS; 62) According to the obtained positive and negative samples, the parameters of the graph neural network are trained by minimizing the total loss function, specifically: The loss function for the i-th news node in the first set of views after identity transformation is The expression is: Similarly, the loss function of the i-th news node in the second set of views obtained after random structural transformation is The expression is: in, represents the positive sample set for the i-th news node, represents the negative sample set for the i-th news node, ξ is an adjustable parameter, z i is the node feature of the i-th news node in the first group of views, z j ′ is the node feature of the jth news node in the second group of views, sim(z i ,z j ′ ) represents the news node feature z i With z j ′ The similarity between The total loss function is obtained by summing up and averaging the losses of all news nodes. The expression is: Among them, θ is a manually set hyperparameter.

8. The method for identifying fake news based on heterogeneous graph contrastive learning according to claim 7, characterized in that: In the step 61), the expression of the heterogeneous structure similarity HPS of news nodes i and j is: Where S is the number of meta-paths, and S = 2, λ s To balance the weight parameters between different meta-paths, and are the heterogeneous neighbor sets on the meta-path s respectively.

9. The method for identifying fake news based on heterogeneous graph contrast learning according to claim 8, characterized in that: In the step 61), for each news node, the first τ points with the largest HPS are selected as positive samples, and the rest are selected as negative samples.

Images