Method for training a social media fake news detection model

Abstract

The application provides a training method of a social media fake news detection model, which comprises the following steps: obtaining a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a propagation path of news; obtaining a plurality of news propagation graphs of a second field in the training set; the ratio of the number of the news propagation graphs of the second field to the number of the news propagation graphs of the first field is less than a threshold value; training the social media fake news detection model according to the plurality of news propagation graphs of the first field, the plurality of news propagation graphs of the second field and a first target loss function, obtaining the social media fake news detection model trained based on the training set; the social media fake news detection model is used for detecting the authenticity of news; and the first target loss function is determined by a classification loss, a global contrast loss and a local contrast loss. The method of the application realizes the accuracy of the detection of fake news of the second field with a smaller data volume and a shorter propagation time.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to the training of a social media fake news detection model. Background Technology

[0002] With the rapid development of the internet, social media has become an important platform for people to obtain information, express opinions, and communicate daily. However, as the number of users on social media platforms grows, if fake news becomes a hot topic and is discussed and spread by a large number of users, it will affect social stability and bring potential economic losses. In order to determine the authenticity of news, the social media fake news detection task has been proposed, aiming to determine the authenticity of news on social media.

[0003] In related technologies, automatic fake news detection models trained in high-resource domains with abundant data can accurately detect fake news in high-resource domains; however, for emerging fields involving breaking news, the accuracy of automatic fake news detection is low due to insufficient data. Therefore, how to accurately detect the authenticity of news in low-resource domains when high-resource data is abundant and low-resource data is scarce is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the problems in the prior art, embodiments of the present invention provide a training method for a social media fake news detection model.

[0005] Specifically, the embodiments of the present invention provide the following technical solutions:

[0006] In a first aspect, embodiments of the present invention provide a method for training a social media fake news detection model, comprising:

[0007] Obtain multiple news dissemination graphs from the first domain of the training set; news dissemination graphs are used to represent the dissemination path of news.

[0008] Obtain multiple news dissemination graphs from the second domain in the training set; the ratio of the number of news dissemination graphs from the second domain to the number of news dissemination graphs from the first domain is less than a threshold.

[0009] Based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and the first objective loss function, a social media fake news detection model is trained to obtain a social media fake news detection model trained on the training set. The social media fake news detection model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. Global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type augmented graph of the news dissemination graph, and features of the news dissemination graph. Local contrastive loss represents the degree of correlation between node features in the second type augmented graph of the news dissemination graph and node features in the third type augmented graph of the news dissemination graph. Classification loss represents the accuracy of the classification result.

[0010] Furthermore, the social media fake news detection model includes at least one of the following:

[0011] Feature extraction module; The feature extraction module is used to extract features from news dissemination graphs;

[0012] Classification module; The classification module is used to predict the authenticity of social media news corresponding to a news dissemination graph based on its features;

[0013] The self-supervised learning module is used to determine the global contrast loss based on the node features in the news dissemination graph, the node features in the first type of augmented graph of the news dissemination graph, and the features of the news dissemination graph; and to determine the local contrast loss based on the node features in the second type of augmented graph of the news dissemination graph and the node features in the third type of augmented graph of the news dissemination graph.

[0014] Furthermore, the social media fake news detection model is trained in the following way:

[0015] Multiple news dissemination graphs from the first domain and multiple news dissemination graphs from the second domain in the training set are input into the social media fake news detection model, and the model outputs the authenticity detection results of the social media news corresponding to the news dissemination graphs. Based on the authenticity detection results of the social media news and the tag information of the news, the classification loss of the social media fake news detection model is obtained. The tag information is used to label the authenticity of the news.

[0016] The first type of augmented graph, the second type of augmented graph, and the third type of augmented graph in the first domain of the training set are input into the social media fake news detection model to obtain the global contrast loss and local contrast loss of multiple news items in the first domain.

[0017] The training set contains multiple news dissemination graphs in the second domain, the first type augmented graphs corresponding to multiple news dissemination graphs in the second domain, the second type augmented graphs corresponding to multiple news dissemination graphs in the second domain, and the third type augmented graphs corresponding to multiple news dissemination graphs in the second domain. These are then input into the social media fake news detection model to obtain the global contrast loss and local contrast loss for multiple news items in the second domain.

[0018] The weighted sum of the classification loss of the social media fake news detection model, the global and local contrast loss of multiple news items in the first domain, and the global and local contrast loss of multiple news items in the second domain is used as the value of the first objective loss function.

[0019] Based on the value of the first objective loss function, the social media fake news detection model is trained to obtain the social media fake news detection model trained on the training set.

[0020] Furthermore, social media fake news detection models also include:

[0021] The data adaptive constraint module is used to determine the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set.

[0022] Furthermore, based on the value of the first objective loss function, the social media fake news detection model is trained. After obtaining the social media fake news detection model trained on the training set, the following steps are also included:

[0023] The first type of augmented graph, the second type of augmented graph, and the third type of augmented graph in the second domain of the test set are input into the social media fake news detection model trained on the training set to obtain the global contrast loss and local contrast loss of multiple news items in the second domain of the test set.

[0024] The value of the second objective loss function is determined by the difference between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set, and by the weighted sum of the global contrast loss and the local contrast loss of multiple news items in the second domain of the test set.

[0025] Based on the value of the second objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model based on the latent features of the test set.

[0026] Secondly, embodiments of the present invention also provide a method for detecting fake news on social media, comprising:

[0027] Obtain the news dissemination map of the second domain to be detected;

[0028] The news dissemination graph of the second domain to be detected is input into the social media fake news detection model to obtain the authenticity detection result of the social media news corresponding to the news dissemination graph of the second domain; the social media fake news detection model is trained based on the training method of the social media fake news detection model in the first aspect.

[0029] Thirdly, embodiments of the present invention also provide a social media fake news detection device, comprising:

[0030] Obtain the news dissemination map of the second domain to be detected;

[0031] The news dissemination graph of the second domain to be detected is input into the social media fake news detection model to obtain the authenticity detection result of the social media news corresponding to the news dissemination graph of the second domain; the social media fake news detection model is trained based on the training method of the social media fake news detection model as described in the first aspect.

[0032] Fourthly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the training method for the social media fake news detection model as described in the first aspect or the social media fake news detection method as described in the second aspect.

[0033] Fifthly, embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the training method for the social media fake news detection model as described in the first aspect or the social media fake news detection method as described in the second aspect.

[0034] In a sixth aspect, embodiments of the present invention also provide a computer program product, including a computer program that, when executed by a processor, implements the training method for the social media fake news detection model as described in the first aspect or the social media fake news detection method as described in the second aspect.

[0035] The training method for the social media fake news detection model provided in this invention trains the model based on multiple news dissemination graphs in a first domain, multiple news dissemination graphs in a second domain, and a first objective loss function. This allows the trained model to accurately extract feature information from news in the first domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. Furthermore, it can accurately extract feature information from news in the second domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. This enables the trained model to accurately and comprehensively extract feature information from fake news in the second domain, where data volume is limited and dissemination time is short. This effectively promotes and improves the accuracy of fake news detection, thus enabling the model to accurately detect the authenticity of news in the second domain, where data volume is limited and dissemination time is short, thereby improving the accuracy of fake news detection in this area. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0037] Figure 1 This is one of the flowcharts illustrating the training method of the social media fake news detection model provided in this embodiment of the invention;

[0038] Figure 2 This is a second schematic flowchart of the training method for the social media fake news detection model provided in this embodiment of the invention;

[0039] Figure 3 This is the third flowchart illustrating the training method of the social media fake news detection model provided in this embodiment of the invention;

[0040] Figure 4 This is the fourth flowchart illustrating the training method of the social media fake news detection model provided in this embodiment of the invention;

[0041] Figure 5 This is a schematic diagram of the structure of the training device for the social media fake news detection model provided in an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0044] The method of this invention can be applied to fake news detection scenarios, improving the accuracy of fake news detection in the second domain with less data and shorter dissemination time.

[0045] In related technologies, automatic fake news detection models trained in high-resource domains with abundant data can accurately detect fake news in high-resource domains; however, for emerging fields involving breaking news, the accuracy of automatic fake news detection is low due to insufficient data. Therefore, how to accurately detect the authenticity of news in low-resource domains when high-resource data is abundant and low-resource data is scarce is a problem that urgently needs to be solved by those skilled in the art.

[0046] The training method of the social media fake news detection model in this invention trains the model based on multiple news dissemination graphs in a first domain, multiple news dissemination graphs in a second domain, and a first objective loss function. This allows the trained model to accurately extract feature information of news in the first domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. Furthermore, it can accurately extract feature information of news in the second domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. This enables the trained model to accurately and comprehensively extract feature information of fake news in the second domain, where data volume is limited and dissemination time is short. This effectively promotes and improves the accuracy of fake news detection, thus enabling the model to accurately detect the authenticity of news in the second domain, where data volume is limited and dissemination time is short, thereby improving the accuracy of fake news detection in this area.

[0047] The following is combined with Figures 1-6 The technical solution of the present invention will be described in detail with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0048] Figure 1This is a flowchart illustrating an embodiment of the training method for a social media fake news detection model provided by this invention. Figure 1 As shown, the method provided in this embodiment includes:

[0049] Step 101: Obtain multiple news dissemination graphs from the first domain in the training set; news dissemination graphs are used to represent the dissemination path of news.

[0050] Specifically, existing technologies that rely on human subjective judgment to determine and verify the authenticity of news are inefficient. For example, ... Figure 2 As shown in the embodiment of this application, the domain with a larger amount of data is the source domain, and the domain with a smaller amount of data is the target domain. The model trained by the source domain and the target domain can accurately identify fake news in the first domain (with only one error), while the detection result for fake news in the second domain will have a large number of errors (6 errors).

[0051] To address the aforementioned issues, this embodiment first obtains multiple news dissemination graphs in the first domain of the training set. These news dissemination graphs represent the dissemination path of news. Optionally, the dissemination of news on social media primarily relies on user behavior. Users' reposting, commenting, or liking of news on social media constitutes a social media news dissemination graph. In other words, users' forwarding, commenting, or liking of news are used as nodes in the news dissemination graph. Based on the nodes and time information in the news dissemination graph, a news dissemination graph can be constructed, thereby accurately representing the dissemination path of news.

[0052] Step 102: Obtain multiple news dissemination graphs in the second domain of the training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold.

[0053] Specifically, when a sudden event occurs, that is, when a new domain emerges, how to accurately detect the authenticity of news in this newly emerging domain when data is scarce, and thus take timely measures to prevent the further spread of fake news, is a problem that urgently needs to be solved by those skilled in the art. In the embodiments of this application, after obtaining multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain are also obtained in the training set. The number of news dissemination graphs in the second domain is much smaller than the number of news dissemination graphs in the first domain; for example, the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold; for example, the number of news dissemination graphs in the second domain is 10% of the number of news dissemination graphs in the first domain. That is, in the process of training the social media fake news detection model, this application not only obtains news from the first domain with a large amount of data and a long dissemination time as training samples, but also obtains news from the second domain with a small amount of data and a short dissemination time as training samples. This allows the trained social media fake news detection model to accurately detect the authenticity of news with a large amount of data and a long dissemination time, as well as the authenticity of social media news that has just occurred with a small amount of data and a short dissemination time.

[0054] Step 103: Based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and the first objective loss function, train the social media fake news detection model to obtain a social media fake news detection model trained on the training set. The social media fake news detection model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. Global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type augmented graph of the news dissemination graph, and features of the news dissemination graph. Local contrastive loss represents the degree of correlation between node features in the second type augmented graph of the news dissemination graph and node features in the third type augmented graph of the news dissemination graph. Classification loss represents the accuracy of the classification result.

[0055] Specifically, in this embodiment, after acquiring multiple news dissemination graphs in a first domain with a large amount of data and a long dissemination time, and news dissemination graphs in a second domain with a small amount of data and a short dissemination time, the social media fake news detection model is trained based on the multiple news dissemination graphs in the first domain, the multiple news dissemination graphs in the second domain, and a first objective loss function, resulting in a social media fake news detection model trained on the training set. The social media fake news detection model is used to detect the authenticity of news. Optionally, the first objective loss function is determined by classification loss, global contrast loss, and local contrast loss. The global contrast loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type of augmented graph of the news dissemination graph, and features of the news dissemination graph. The local contrast loss represents the degree of correlation between node features in the second type of augmented graph of the news dissemination graph and node features in the third type of augmented graph of the news dissemination graph. In other words, the global contrast loss represents the degree of correlation between nodes and the news dissemination graph, and the local contrast loss represents the degree of correlation between nodes. Optionally, the global contrastive loss can be learned by comparing the original news dissemination graph and the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph, to obtain the relationship features between each node in the original news dissemination graph and the whole graph. The local contrastive loss can be learned by comparing the two types of augmented graphs obtained by selectively discarding edges of the original news dissemination graph and selectively masking the node features of the original graph, to obtain the relationship features between each node in the graph. That is, by adding global and local contrastive losses in this embodiment, the feature extractor in the social media fake news detection model is optimized, so that the optimized feature extractor and the social media fake news detection model can extract features in the news dissemination graph more accurately and comprehensively, extract fake news features more accurately and completely, and thus extract more important information for news classification, thereby achieving better and more accurate classification of the authenticity of news. On the other hand, the training method of the social media fake news detection model in this embodiment of the invention performs local and global contrastive learning on the second domain fake news data and its augmented data, learns the local and global information of the second domain fake news data, and improves the model's generalization ability on the second domain fake news data.

[0056] In other words, after training the social media fake news detection model based on the first objective loss function, the trained social media fake news detection model can not only accurately extract the feature information of news in the first domain and the global contrast loss between nodes in the news dissemination graph and the news dissemination graph, and the local contrast loss between augmented graph nodes in the news dissemination graph, but also accurately extract the feature information of news in the second domain and the global contrast loss between nodes in the news dissemination graph and the news dissemination graph, and the local contrast loss between augmented graph nodes in the news dissemination graph. This enables the social media fake news detection model to accurately identify and extract the feature information of fake news, effectively promoting and improving the accuracy of fake news detection. Consequently, the trained social media fake news detection model can accurately detect the authenticity of news in the second domain with less data and shorter dissemination time, improving the accuracy of fake news detection in the second domain with less data and shorter dissemination time.

[0057] The method described in the above embodiments trains a social media fake news detection model based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and a first objective loss function. This allows the trained model to accurately extract feature information from news in the first domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. Furthermore, it can accurately extract feature information from news in the second domain, as well as the global contrast loss between nodes in the news dissemination graph and the news dissemination graph itself, and the local contrast loss between nodes in the augmented graph. This enables the trained model to accurately and comprehensively extract feature information from fake news in the second domain, where data volume is limited and dissemination time is short. This effectively promotes and improves the accuracy of fake news detection, thus enabling the model to accurately detect the authenticity of news in the second domain, where data volume is limited and dissemination time is short, thereby improving the accuracy of fake news detection in this area.

[0058] In one embodiment, the social media fake news detection model includes at least one of the following:

[0059] Feature extraction module; The feature extraction module is used to extract features from news dissemination graphs;

[0060] Classification module; The classification module is used to predict the authenticity of social media news corresponding to a news dissemination graph based on its features;

[0061] The self-supervised learning module is used to determine the global contrast loss based on the node features in the news dissemination graph, the node features in the first type of augmented graph of the news dissemination graph, and the features of the news dissemination graph; and to determine the local contrast loss based on the node features in the second type of augmented graph of the news dissemination graph and the node features in the third type of augmented graph of the news dissemination graph.

[0062] Specifically, in this embodiment, the social media fake news detection model includes a feature extraction module, a classification module, and a self-supervised learning module. The feature extraction module extracts features from the news dissemination graph. Optionally, the feature extraction module is based on a Graph Convolutional Network (GCN). The classification module predicts the authenticity of the social media news corresponding to the news dissemination graph based on its features. The self-supervised learning module determines the global contrast loss based on node features in the news dissemination graph, node features in the first type of augmented graph of the news dissemination graph, and the news dissemination graph features. It also determines the local contrast loss based on node features in the second type of augmented graph of the news dissemination graph and node features in the third type of augmented graph of the news dissemination graph. This allows the social media fake news detection model to accurately extract not only the feature information of news in the first domain but also the global contrast loss between nodes in the news dissemination graph and the news dissemination graph. The contrast loss, the local contrast loss between augmented graph nodes in the news dissemination graph, and the global contrast loss between nodes in the news dissemination graph and the news dissemination graph, as well as the local contrast loss between augmented graph nodes in the news dissemination graph, can accurately extract the feature information of news in the second domain. Optionally, the global contrast loss and local contrast loss corresponding to fake news are different from those corresponding to real news. The global contrast loss and local contrast loss corresponding to fake news in the first domain are also different from those corresponding to fake news in the second domain. This allows the trained social media fake news detection model to accurately identify and extract the feature information of fake news, realize the detection of the authenticity of news in the second domain with less data and shorter dissemination time, and improve the accuracy of fake news detection in the second domain with less data and shorter dissemination time.

[0063] In one embodiment, the social media fake news detection model is trained in the following manner:

[0064] Multiple news dissemination graphs from the first domain and multiple news dissemination graphs from the second domain in the training set are input into the social media fake news detection model, and the model outputs the authenticity detection results of the social media news corresponding to the news dissemination graphs. Based on the authenticity detection results of the social media news and the tag information of the news, the classification loss of the social media fake news detection model is obtained. The tag information is used to label the authenticity of the news.

[0065] The first type of augmented graph, the second type of augmented graph, and the third type of augmented graph in the first domain of the training set are input into the social media fake news detection model to obtain the global contrast loss and local contrast loss of multiple news items in the first domain.

[0066] The training set contains multiple news dissemination graphs in the second domain, the first type augmented graphs corresponding to multiple news dissemination graphs in the second domain, the second type augmented graphs corresponding to multiple news dissemination graphs in the second domain, and the third type augmented graphs corresponding to multiple news dissemination graphs in the second domain. These are then input into the social media fake news detection model to obtain the global contrast loss and local contrast loss for multiple news items in the second domain.

[0067] The weighted sum of the classification loss of the social media fake news detection model, the global and local contrast loss of multiple news items in the first domain, and the global and local contrast loss of multiple news items in the second domain is used as the value of the first objective loss function.

[0068] Based on the value of the first objective loss function, the social media fake news detection model is trained to obtain the social media fake news detection model trained on the training set.

[0069] Specifically, in the training process of the social media fake news detection model, this application first inputs multiple news dissemination graphs from the first domain and multiple news dissemination graphs from the second domain into the social media fake news detection model. The classification module of the social media fake news detection model outputs the authenticity detection results of the social media news corresponding to the news dissemination graphs. Then, based on the authenticity detection results of the social media news and the label information of the news, the classification loss (detection accuracy) of the social media fake news detection model can be obtained.

[0070] Furthermore, in this embodiment, multiple news dissemination graphs in the first domain, the first type of augmented graphs corresponding to the multiple news dissemination graphs in the first domain, the second type of augmented graphs corresponding to the multiple news dissemination graphs in the first domain, and the third type of augmented graphs corresponding to the multiple news dissemination graphs in the first domain are input into the social media fake news detection model to obtain the global contrast loss and local contrast loss of multiple news items in the first domain. That is, the self-supervised learning module in the social media fake news detection model obtains the local contrast loss and global contrast loss of the nodes in the graph based on the features of the original news dissemination graph and its three types of augmented graphs. Among them, the global contrast loss is obtained by performing comparative learning on the original news dissemination graph and the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph to obtain the relationship features between each node in the original news dissemination graph and the entire graph. The local contrast loss is obtained by performing comparative learning on the two types of augmented graphs obtained by selectively discarding the edges of the original news dissemination graph and selectively covering the node features of the original graph to obtain the relationship features between each node in the graph.

[0071] For example, the global contrast loss between each node in the original news dissemination graph and the entire graph can be modeled using the following formula:

[0072] D(Z si ,s)=Sigmoid(Z si *s)

[0073] Among them, Z si represents the node feature representation of the news dissemination graph; s represents the feature representation of the news dissemination graph; * represents the inner product; D represents the discriminator, which calculates the relevance scores of positive and negative samples respectively; Sigmoid is the activation function.

[0074] The global contrast loss between each node and the entire graph in the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph is determined based on the following formula:

[0075]

[0076] Where N represents the number of nodes in the input graph; Z 0i Z represents a node in the original news dissemination graph; 1i This represents the nodes in the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph.

[0077] Alternatively, the local contrast loss can be modeled using the following formula:

[0078]

[0079] Z2 and Z3 represent two types of augmented graphs of news data; optionally, they can be two types of augmented graphs obtained by selectively discarding the edges of the original news dissemination graph and selectively masking the node features of the original graph.

[0080] (Z 2i Z 3j (i,j∈{1,...,N},i≠j), where N represents the number of nodes in the graph, cos() represents the cosine similarity function, τ is a hyperparameter, and g() is used to further enhance the representational power of the model. The nodes calculated by g() are represented as Z2' and Z3';

[0081]

[0082] Where I represents the identity matrix, the final definition of the local contrast loss of a node is as follows:

[0083]

[0084] Furthermore, the multiple news dissemination graphs in the second domain of the training set, the first type augmented graphs corresponding to the multiple news dissemination graphs in the second domain, the second type augmented graphs corresponding to the multiple news dissemination graphs in the second domain, and the third type augmented graphs corresponding to the multiple news dissemination graphs in the second domain are input into the social media fake news detection model to obtain the global contrast loss and local contrast loss of multiple news items in the second domain. Optionally, the global contrast loss and local contrast loss of multiple news items in the second domain are determined in the same way as the global contrast loss and local contrast loss of the first domain, which will not be repeated in the embodiments of this application.

[0085] Optionally, after determining the classification loss of the social media fake news detection model, the global and local contrast losses of multiple news items in the first domain, and the global and local contrast losses of multiple news items in the second domain, in this embodiment, the weighted sum of the classification loss of the social media fake news detection model, the global and local contrast losses of multiple news items in the first domain, and the global and local contrast losses of multiple news items in the second domain is used as the value of the first objective loss function. That is, in the process of training the social media fake news detection model, this embodiment not only considers the classification loss of the social media fake news detection model, but also considers whether the social media fake news detection model can accurately and comprehensively extract the feature information of fake news in the second domain with less data and shorter dissemination time. Thus, after accurately extracting the feature information of fake news in the second domain with less data and shorter dissemination time, the accuracy of fake news identification can be effectively improved. This fundamentally solves the problem of not being able to accurately and comprehensively extract the feature information of fake news in the second domain with less data and shorter dissemination time, and effectively improves the accuracy of fake news detection in the second domain with less data and shorter dissemination time.

[0086] The method described in the above embodiments, during the training of the social media fake news detection model, not only considers the classification loss of the social media fake news detection model, but also considers whether the social media fake news detection model can accurately and comprehensively extract the feature information of fake news in the second domain with less data and shorter dissemination time. Therefore, after accurately extracting the feature information of fake news in the second domain with less data and shorter dissemination time, it can effectively improve the accuracy of fake news identification, fundamentally solving the problem of not being able to accurately and comprehensively extract the feature information of fake news in the second domain with less data and shorter dissemination time, and effectively improving the accuracy of fake news detection in the second domain with less data and shorter dissemination time.

[0087] In one embodiment, the social media fake news detection model further includes:

[0088] The data adaptive constraint module is used to determine the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set.

[0089] Specifically, in this embodiment, the social media fake news detection model, in addition to the feature extraction module, classification module, and self-supervised learning module, also adds a data adaptive constraint module. This data adaptive constraint module is used to determine the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set. That is, after training the social media fake news detection model based on the training set, this embodiment uses the data adaptive constraint module to determine the differences between the news dissemination graph features in the training set and the test set, thereby fine-tuning the social media fake news detection model. This ensures that the performance of the feature extractor on the test set is similar to that on the training set, optimizing the feature extractor and preventing overfitting. Therefore, by adding the data adaptive constraint module to constrain the differences between the news dissemination graph features in the training set and the test set, the social media fake news detection model can more accurately identify fake news in the second domain to be detected.

[0090] For example, the differences between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set can be determined based on the following method:

[0091]

[0092] Where h is the feature matrix of the training set data, μ represents the feature mean, and Σ represents the covariance matrix. The feature mean μ of the test set is obtained in the same way. t , and covariance matrix Σ t The differences between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set are as follows:

[0093]

[0094] In other words, the data adaptive constraint module in this embodiment calculates the difference between two statistical values ​​of fake news data in the training set and fake news data in the test set, ensuring that the social media fake news detection model does not overfit to the fake news data and improving the model's generalization ability.

[0095] The method described in the above embodiment allows the social media fake news detection model to fine-tune the model by adding a data adaptive constraint module to constrain the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set. This makes the performance of the feature extractor on the test set similar to that on the training set, thus optimizing the feature extractor and preventing overfitting. Consequently, the social media fake news detection model, after constraining the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set by adding the data adaptive constraint module, can more accurately identify fake news in the second domain to be detected.

[0096] In one embodiment, after training the social media fake news detection model based on the value of the first objective loss function to obtain the social media fake news detection model trained on the training set, the method further includes:

[0097] The first type of augmented graph, the second type of augmented graph, and the third type of augmented graph in the second domain of the test set are input into the social media fake news detection model trained on the training set to obtain the global contrast loss and local contrast loss of multiple news items in the second domain of the test set.

[0098] The value of the second objective loss function is determined by the difference between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set, and by the weighted sum of the global contrast loss and the local contrast loss of multiple news items in the second domain of the test set.

[0099] Based on the value of the second objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model based on the latent features of the test set.

[0100] Specifically, in this embodiment, multiple news propagation graphs in the second domain of the test set, the first type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, the second type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, and the third type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set are input into a social media fake news detection model trained on the training set to obtain the global contrast loss and local contrast loss of multiple news items in the second domain of the test set. Optionally, the method steps for determining the global contrast loss and local contrast loss of multiple news items in the second domain of the test set are similar to the method steps for determining the global contrast loss and local contrast loss of multiple news items in the training set, and will not be repeated in this embodiment.

[0101] Furthermore, after obtaining the global and local contrastive losses for multiple news items in the second domain of the test set, the differences between the news dissemination graph features in the second domain of the training set and the news dissemination graph features in the second domain of the test set can be determined based on the data adaptive constraint module in the social media fake news detection model. Then, the difference between the news dissemination graph features in the second domain of the training set and the news dissemination graph features in the second domain of the test set, and the weighted sum of the global and local contrastive losses for multiple news items in the second domain of the test set can be used as the value of the second objective loss function. Based on the value of the second objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model based on the latent features of the test set. In other words, by adding a constraint on the difference between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set to the loss function of the training set, the social media fake news detection model is fine-tuned. This makes the performance of the feature extractor on the test set similar to that on the training set, optimizing the feature extractor and preventing overfitting. Therefore, by adding the constraint on the difference between the news dissemination graph features in the training set and the test set, the social media fake news detection model can more accurately identify fake news in the second domain. Alternatively, if the constraint on the difference between the news dissemination graph features of the second domain in the training set and the test set is not added to the loss function, overfitting will occur, making the social media fake news detection model trained on the training set unable to accurately detect the authenticity of news in the second domain, and thus unable to accurately and effectively identify fake news in the second domain.

[0102] The method described in the above embodiment adds a constraint on the difference between the news dissemination graph features in the second domain of the training set and the news dissemination graph features in the second domain of the test set to the loss function. This enables fine-tuning of the social media fake news detection model and optimization of the feature extractor, making the performance of the feature extractor on the test set similar to that on the training set. This effectively prevents overfitting and allows the social media fake news detection model with the added constraint on the difference between the news dissemination graph features in the training set and the news dissemination graph features in the test set to more accurately identify fake news in the second domain to be detected.

[0103] This application provides a method for detecting fake news on social media, including:

[0104] Obtain the news dissemination map of the second domain to be detected;

[0105] The news dissemination graph of the second domain to be detected is input into the social media fake news detection model to obtain the authenticity detection result of the social media news corresponding to the news dissemination graph of the second domain; the social media fake news detection model is trained based on the training method of the social media fake news detection model mentioned above.

[0106] Specifically, after training a social media fake news detection model based on training and testing sets, the trained model can accurately detect fake news in the second domain. Optionally, a news dissemination graph of the second domain to be detected can be obtained first, and then input into the trained model. This yields the authenticity detection results of the social media news corresponding to the news dissemination graph of the second domain, thus enabling the detection of authenticity for second-domain news with limited data and short dissemination time, and improving the accuracy of fake news detection in this area.

[0107] For example, the specific process of training the social media fake news detection model in this application embodiment is as follows: Figure 3 and Figure 4 As shown, the details are as follows:

[0108] 1. Training a social media fake news detection model based on news from the first domain (high-resource) in the training set.

[0109] Three different types of augmentation operations—edge dropping, feature shuffling, and feature masking—are applied to the first-domain news data to generate three different types of augmented maps. A feature extractor is used to obtain high-dimensional feature matrices of the first-domain news data and the three types of augmented maps. The main task and the auxiliary task share a single feature extractor. The first-domain fake news data is input into the main task, while the first-domain news data and the three types of augmented maps are input into the auxiliary task. Local comparative learning is performed on the augmented maps with edge dropping and feature shuffling operations, and global comparative learning is performed on the first-domain news data and the augmented map with feature shuffling operations.

[0110] 1.1 Global Comparative Learning Based on Graph Convolutional Neural Networks

[0111] The purpose of global contrastive learning is to help node representations in graph data obtain global information about the entire graph. Given an input event... Different graphs can be generated through various types of data augmentation. For global contrastive learning, two views were used: an original view (View0) and an augmented view (View1), where the node attributes of all nodes in the graph were randomly assigned. With these two graphs, two corresponding node representations, Z0 and Z1, can be obtained through a shared feature extractor. Subsequently, a global graph representation s can be summarized from the node representation matrix Z0 of the original view (View0) using a multilayer perceptron.

[0112] In global contrastive learning, positive samples consist of node-graph representation pairs, where both the node and graph representations originate from the original view, View0. Negative samples also consist of node-graph representation pairs, where the node representation comes from View1 and the graph representation comes from View0. The discriminator D calculates scores for both positive and negative samples; higher scores are awarded for positive samples, and lower scores for negative samples. The representation of D is defined as follows:

[0113] D(Z si ,s)=Sigmoid(Z si *s)

[0114] Z si The node representations of the graph views are shown, and * denotes the inner product. The loss function for global contrastive learning is as follows:

[0115]

[0116] Where N represents the number of nodes in the input graph.

[0117] 1.2 Local Contrast Learning Based on Graph Convolutional Neural Networks

[0118] Through local contrastive learning, the model can learn richer node feature representations. By comparing the feature differences between a node and its neighbors, the model can learn more discriminative node representations, reducing the impact of node location, noise, and missing values ​​on graph data and improving the robustness and interpretability of the graph data. Local contrastive learning also allows the model to better capture the similarities and differences between a node and its neighbors, thus leading to a better understanding of the node's contextual information.

[0119] Given an input event Two distinct augmented graphs are obtained using graph augmentation strategies (edge ​​removal and node attribute masking), which serve as inputs to the feature extractor. The output of the feature extractor is two node representation matrices, Z2 and Z3. The fundamental goal of local contrastive learning is to distinguish whether two nodes from the augmented views are the same node; therefore, (Z... 2i Z 3i (i∈{1,...,N}) represents a pair, where N is the number of nodes, (Z2i Z 3j ) and (Z 2i Z 2j (i,j∈{1,...,N},i≠j) represents negative pairs within a class and negative pairs between classes. The objective function for positive pairs is defined as follows:

[0120]

[0121]

[0122] Where cos() represents the cosine similarity function, τ is a hyperparameter, and g() is a two-layer MLP. The nodes processed by the MLP are represented as Z2' and Z3', and a regularization modifier is used on the redefined node representation, as follows:

[0123]

[0124] The final loss function for local contrast is defined as follows:

[0125]

[0126] The training method based on high-resource fake news data trained during test time has the following final loss function for the auxiliary task:

[0127] L s =L g +αL l

[0128] The training method based on high-resource fake news data trained during test time has the following loss function:

[0129] L = L m +γL s

[0130] Where Lm is the loss of the main task.

[0131] 2. Training a social media fake news detection model based on news from the second domain (low-resource) in the training set.

[0132] Three different types of augmentation operations are performed on the second-domain news data: edge dropping, feature shuffling, and feature masking, generating three different types of augmented maps. A feature extractor is used to obtain the high-dimensional feature matrix of the second-domain news data and the three types of augmented maps. The feature extractor here is the same as the feature extractor mentioned in the first part. The second-domain news data and the three types of augmented maps are input into the auxiliary task. The augmented maps with edge dropping and feature shuffling operations are used for local comparative learning, and the second-domain fake news data and the augmented maps with feature shuffling operations are used for global comparative learning. The social media fake news detection model is trained based on the second-domain news in the training set. It also adopts a self-supervised learning framework that includes global comparative learning and local comparative learning. The model is the same as the auxiliary task model in the method of training the social media fake news detection model based on the first-domain news in the training set, so it will not be described again here.

[0133] 3. Adaptive constraint method based on training data during testing.

[0134] The READOUT function is used to obtain the feature matrix of the news data. Two statistical parameters of the news data feature matrix are then calculated: the feature mean and the covariance matrix. The two statistical parameters, eigenvalues ​​and covariance matrix, are calculated for each piece of fake news data in the second domain, and a loss function is constructed.

[0135] definition Use the READOUT function (the READOUT function is a feature extractor).

[0136]

[0137] To obtain the feature matrix of the news data, calculate two statistical data points from the feature matrix: the feature mean and the covariance matrix. The definitions of the feature mean and covariance matrix are as follows:

[0138]

[0139]

[0140] The feature matrix of the second-domain low-resource fake news data is calculated using the READOUT function. Similarly, its two statistical data, the feature mean and covariance matrix, are calculated, and the loss function is constructed as follows:

[0141]

[0142] Where μ t ,∑ t represents the feature mean and covariance matrix of low-resource fake news data, respectively.

[0143] The method described in the above embodiments determines whether news data is fake news by acquiring news data to be detected and inputting it into a fake news detection model. The model training method of this invention integrates fake news data features from two different domains into the model representation, thereby achieving better modeling of the target data domain. Furthermore, the model is retrained on the target data during the testing phase, enhancing its ability to represent the target data. This allows for more accurate and efficient detection of fake news, improving the accuracy of cross-domain fake news detection.

[0144] The training apparatus for the social media fake news detection model provided by the present invention will be described below. The training apparatus for the social media fake news detection model described below can be referred to in correspondence with the training method for the social media fake news detection model described above.

[0145] Figure 5 This is a schematic diagram of the training device for the social media fake news detection model provided by the present invention. The training device for the social media fake news detection model provided in this embodiment includes:

[0146] The first acquisition module 710 is used to acquire multiple news dissemination graphs in the first domain of the training set; the news dissemination graph is used to represent the dissemination path of news;

[0147] The second acquisition module 720 is used to acquire multiple news dissemination graphs in the second domain of the training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold.

[0148] Training module 730 is used to train a social media fake news detection model based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and a first objective loss function, resulting in a social media fake news detection model trained on the training set. The social media fake news detection model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. Global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type augmented graph of the news dissemination graph, and features of the news dissemination graph. Local contrastive loss represents the degree of correlation between node features in the second type augmented graph of the news dissemination graph and node features in the third type augmented graph of the news dissemination graph. Classification loss represents the accuracy of the classification result.

[0149] Optionally, the social media fake news detection model includes at least one of the following:

[0150] Feature extraction module; The feature extraction module is used to extract features from news dissemination graphs;

[0151] Classification module; The classification module is used to predict the authenticity of social media news corresponding to a news dissemination graph based on its features;

[0152] The self-supervised learning module is used to determine the global contrast loss based on the node features in the news dissemination graph, the node features in the first type of augmented graph of the news dissemination graph, and the features of the news dissemination graph; and to determine the local contrast loss based on the node features in the second type of augmented graph of the news dissemination graph and the node features in the third type of augmented graph of the news dissemination graph.

[0153] Optionally, the social media fake news detection model is trained in the following way:

[0154] Multiple news dissemination graphs from the first domain and multiple news dissemination graphs from the second domain in the training set are input into the social media fake news detection model, and the model outputs the authenticity detection results of the social media news corresponding to the news dissemination graphs. Based on the authenticity detection results of the social media news and the tag information of the news, the classification loss of the social media fake news detection model is obtained. The tag information is used to label the authenticity of the news.

[0155] The first type of augmented graph, the second type of augmented graph, and the third type of augmented graph in the first domain of the training set are input into the social media fake news detection model to obtain the global contrast loss and local contrast loss of multiple news items in the first domain.

[0156] The training set contains multiple news dissemination graphs in the second domain, the first type augmented graphs corresponding to multiple news dissemination graphs in the second domain, the second type augmented graphs corresponding to multiple news dissemination graphs in the second domain, and the third type augmented graphs corresponding to multiple news dissemination graphs in the second domain. These are then input into the social media fake news detection model to obtain the global contrast loss and local contrast loss for multiple news items in the second domain.

[0157] The weighted sum of the classification loss of the social media fake news detection model, the global and local contrast loss of multiple news items in the first domain, and the global and local contrast loss of multiple news items in the second domain is used as the value of the first objective loss function.

[0158] Based on the value of the first objective loss function, the social media fake news detection model is trained to obtain the social media fake news detection model trained on the training set.

[0159] Optionally, the social media fake news detection model also includes:

[0160] The data adaptive constraint module is used to determine the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set.

[0161] Optionally, the training module 730 is further configured to input multiple news propagation graphs in the second domain of the test set, the first type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, the second type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, and the third type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set into the social media fake news detection model trained on the training set, so as to obtain the global contrast loss and local contrast loss of multiple news items in the second domain of the test set;

[0162] The value of the second objective loss function is determined by the difference between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set, and by the weighted sum of the global contrast loss and the local contrast loss of multiple news items in the second domain of the test set.

[0163] Based on the value of the second objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model based on the latent features of the test set.

[0164] The apparatus of this invention is used to execute the method in any of the foregoing method embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

[0165] Figure 6The example illustrates the physical structure of an electronic device, which may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logic instructions in the memory 830 to execute a training method for a social media fake news detection model. This method includes: acquiring multiple news dissemination graphs in a first domain of the training set; the news dissemination graphs represent the dissemination path of news; acquiring multiple news dissemination graphs in a second domain of the training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold; training the social media fake news detection model based on the multiple news dissemination graphs in the first domain, the multiple news dissemination graphs in the second domain, and a first objective loss function to obtain a social media fake news detection model trained on the training set; the social media fake news detection model is used to detect the authenticity of news; the first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss; the global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type augmented graph of the news dissemination graph, and features of the news dissemination graph; the local contrastive loss represents the degree of correlation between node features in the second type augmented graph of the news dissemination graph and node features in the third type augmented graph of the news dissemination graph; and the classification loss represents the accuracy of the classification result.

[0166] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is capable of executing the training method for the social media fake news detection model provided by the above methods, the method comprising: acquiring multiple news dissemination graphs in a first domain of a training set; the news dissemination graphs being used to represent the dissemination path of news; acquiring multiple news dissemination graphs in a second domain of a training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain being less than a threshold; and, based on the multiple news dissemination graphs in the first domain and the multiple news dissemination graphs in the second domain... A social media fake news detection model is trained using a news dissemination graph and a first objective loss function, resulting in a model trained on the training set. This model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. Global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type augmented graph of the news dissemination graph, and features in the news dissemination graph itself. Local contrastive loss represents the degree of correlation between node features in the second type augmented graph of the news dissemination graph and node features in the third type augmented graph of the news dissemination graph. The classification loss represents the accuracy of the classification result.

[0168] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the training methods for the social media fake news detection models provided above.

[0169] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0170] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a social media fake news detection model, characterized in that, include: Obtain multiple news dissemination graphs from the first domain of the training set; The news dissemination diagram is used to represent the dissemination path of news; Obtain multiple news dissemination graphs in the second domain of the training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold. Based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and a first objective loss function, a social media fake news detection model is trained to obtain a social media fake news detection model trained on the training set. This model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. The global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type of augmented graph of the news dissemination graph, and features of the news dissemination graph itself. The local contrastive loss represents the degree of correlation between node features in the second type of augmented graph of the news dissemination graph and node features in the third type of augmented graph of the news dissemination graph. The classification loss represents the accuracy of the classification result. The global comparison loss includes: and ； Z represents the global contrast loss between each node in the original news dissemination graph and the entire graph; si represents the node feature representation of the news dissemination graph; s represents the feature representation of the news dissemination graph; * represents the inner product; D represents the discriminator, which calculates the relevance scores of positive and negative samples respectively; Sigmoid is the activation function; Z represents the global contrast loss between each node in the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph and the entire graph; N represents the number of nodes in the input graph; Z represents the global contrast loss between each node in the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph and the entire graph. 0i Z represents a node in the original news dissemination graph; 1i This represents the nodes in the augmented graph obtained by randomly shuffling the nodes of the original news dissemination graph. Local contrast loss includes: ； ； in, Two types of augmented graphs representing news data; N represents the number of nodes in the graph; cos() represents the cosine similarity function; τ represents a hyperparameter; and g() is used to enhance the representational power of the model.

2. The training method for the social media fake news detection model according to claim 1, characterized in that, The social media fake news detection model includes at least one of the following: Feature extraction module; The feature extraction module is used to extract features from news dissemination graphs; A classification module; the classification module is used to predict the authenticity of social media news corresponding to the news dissemination graph based on the features of the news dissemination graph; The self-supervised learning module is used to determine the global contrast loss based on the node features in the news dissemination graph, the node features in the first type augmented graph of the news dissemination graph, and the features of the news dissemination graph. The local contrast loss is determined based on the node features in the second type augmented map of the news dissemination graph and the node features in the third type augmented map of the news dissemination graph.

3. The training method for the social media fake news detection model according to claim 2, characterized in that, The social media fake news detection model is trained in the following way: Multiple news dissemination graphs from the first domain and multiple news dissemination graphs from the second domain in the training set are input into the social media fake news detection model, and the authenticity detection results of the social media news corresponding to the news dissemination graphs are output; based on the authenticity detection results of the social media news and the tag information of the news, the classification loss of the social media fake news detection model is obtained. The tag information is used to indicate the authenticity of the news; The training set contains multiple news propagation graphs in the first domain, the first type of augmented graphs corresponding to the multiple news propagation graphs in the first domain, the second type of augmented graphs corresponding to the multiple news propagation graphs in the first domain, and the third type of augmented graphs corresponding to the multiple news propagation graphs in the first domain. These are then input into the social media fake news detection model to obtain the global contrast loss and local contrast loss for multiple news items in the first domain. The training set contains multiple news propagation graphs in the second domain, the first type augmented graphs corresponding to the multiple news propagation graphs in the second domain, the second type augmented graphs corresponding to the multiple news propagation graphs in the second domain, and the third type augmented graphs corresponding to the multiple news propagation graphs in the second domain. These are then input into the social media fake news detection model to obtain the global contrast loss and local contrast loss for multiple news items in the second domain. The weighted sum of the classification loss of the social media fake news detection model, the global and local contrast loss of multiple news items in the first domain, and the global and local contrast loss of multiple news items in the second domain is used as the value of the first objective loss function. Based on the value of the first objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model trained on the training set.

4. The training method for the social media fake news detection model according to claim 3, characterized in that, The social media fake news detection model also includes: The data adaptive constraint module is used to determine the differences between the news dissemination graph features in the training set and the news dissemination graph features in the test set.

5. The training method for the social media fake news detection model according to claim 4, characterized in that, After training the social media fake news detection model based on the value of the first objective loss function to obtain the social media fake news detection model trained on the training set, the method further includes: The multiple news propagation graphs in the second domain of the test set, the first type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, the second type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set, and the third type augmented graphs corresponding to the multiple news propagation graphs in the second domain of the test set are input into the social media fake news detection model trained on the training set to obtain the global contrast loss and local contrast loss of multiple news items in the second domain of the test set. The value of the second objective loss function is determined by the difference between the news dissemination graph features of the second domain in the training set and the news dissemination graph features of the second domain in the test set, and by the weighted sum of the global contrast loss and the local contrast loss of multiple news items in the second domain of the test set. Based on the value of the second objective loss function, the social media fake news detection model is trained to obtain a social media fake news detection model based on the latent features of the test set.

6. A method for detecting fake news on social media, characterized in that, include: Obtain the news dissemination map of the second domain to be detected; The news dissemination graph of the second domain to be detected is input into the social media fake news detection model to obtain the authenticity detection result of the social media news corresponding to the news dissemination graph of the second domain; the social media fake news detection model is trained based on the method described in any one of claims 1-5.

7. A training apparatus for a social media fake news detection model, used to implement the training method for the social media fake news detection model as described in any one of claims 1 to 5 or the social media fake news detection method as described in claim 6, characterized in that, include: The first acquisition module is used to acquire multiple news dissemination graphs in the first domain of the training set; The news dissemination diagram is used to represent the dissemination path of news; The second acquisition module is used to acquire multiple news dissemination graphs in the second domain of the training set; the ratio of the number of news dissemination graphs in the second domain to the number of news dissemination graphs in the first domain is less than a threshold. The training module is used to train a social media fake news detection model based on multiple news dissemination graphs in the first domain, multiple news dissemination graphs in the second domain, and a first objective loss function, to obtain a social media fake news detection model trained on a training set. The social media fake news detection model is used to detect the authenticity of news. The first objective loss function is determined by classification loss, global contrastive loss, and local contrastive loss. The global contrastive loss represents the degree of correlation between node features in the news dissemination graph, node features in the first type of augmented graph of the news dissemination graph, and features of the news dissemination graph. The local contrastive loss represents the degree of correlation between node features in the second type of augmented graph of the news dissemination graph and node features in the third type of augmented graph of the news dissemination graph. The classification loss represents the accuracy of the classification result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the training method of the social media fake news detection model as described in any one of claims 1 to 5 or the social media fake news detection method as described in claim 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the training method for the social media fake news detection model as described in any one of claims 1 to 5 or the social media fake news detection method as described in claim 6.

10. A computer program product having executable instructions stored thereon, characterized in that, When executed by the processor, this instruction causes the processor to implement the training method of the social media fake news detection model as described in any one of claims 1 to 5 or the social media fake news detection method as described in claim 6.

Images