A fake news detection method based on enhanced prompt learning and graph diffusion strategy

Abstract

The application discloses a false news detection method based on enhanced prompt learning and graph diffusion strategy, comprising: false news preliminary classification based on enhanced prompt learning: an enhanced prompt learning model trained by a small number of labeled samples is used to preliminarily predict the true or false label of a large number of input false news texts and user forwarding records thereof; a true or false propagation consistency network is constructed: a true or false propagation consistency network is constructed based on the user forwarding records, news is taken as a node, the connection between news sharing the same social user is taken as an edge, and an undirected graph is formed; news final classification by applying a graph feature diffusion strategy: a graph feature diffusion strategy and a high-order neighbor perception diffusion equation are applied on the network, the features of the network news nodes are diffused until the network is stable, the true or false features of the nodes are finally integrated and updated, the news is finally predicted to be true or false, and therefore, more accurate false news detection results are output, and the accuracy and robustness of the detection are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of internet fake news detection technology, specifically involving a fake news detection method based on enhanced cueing learning and graph diffusion strategy. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, the field of fake news detection technology has benefited from this rapid advancement and is currently experiencing rapid growth, playing an increasingly important role in social production and daily life. Furthermore, the widespread use of social media, news websites, and various online platforms has significantly increased the speed and impact of fake news dissemination, making the performance requirements for fake news detection technology increasingly urgent. Comprehensive and intelligent detection methods will become the future trend of international development.

[0003] Despite the numerous advantages of current mainstream fake news detection methods, some difficulties and challenges remain in practical applications:

[0004] 1. Limitations of Feature Extraction and Representation: Current mainstream fake news detection methods typically rely on traditional text feature extraction techniques, such as TF-IDF and word embedding. These methods have certain limitations in processing text semantics and contextual information, especially in capturing complex syntactic structures and implicit semantics, thus limiting detection accuracy.

[0005] 2. Lack of contextual information: The spread of fake news is often accompanied by complex contextual information, making it difficult to obtain accurate results by relying solely on the content of a single article. Existing methods perform poorly when handling cross-document contextual relationships and cannot fully utilize user behavior data and propagation paths to improve detection effectiveness;

[0006] 3. Insufficient generalization ability of the model: In practical applications, fake news takes many forms and has diverse content. Traditional detection models often lack sufficient generalization ability when faced with new and diverse fake news. The model is prone to overfitting the training data, resulting in unstable detection performance in real-world applications.

[0007] 4. Scarcity of labeled data: Fake news detection tasks require a large amount of labeled data for model training, but obtaining high-quality labeled data is often difficult and time-consuming. This leads to a significant imbalance in training samples, especially for a few categories of fake news, where existing models perform poorly. Summary of the Invention

[0008] The purpose of this invention is to provide a fake news detection method based on augmented cue learning and graph diffusion strategy. This method combines an augmented cue learning model with a graph diffusion strategy, uses user forwarding records to construct a consistency network for the propagation of truth and falsehood, and applies a high-order neighbor-aware diffusion equation to achieve efficient and accurate detection of fake news.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The present invention discloses a method for detecting fake news based on enhanced cueing learning and graph diffusion strategy, which is aimed at the field of fake news detection technology. The main steps of the method include:

[0011] (1) Preliminary Classification of Fake News Based on Augmented Cue Learning: A large set of fake news texts to be detected is used as input, and an augmented cue learning model trained with a small number of labeled samples is used for preliminary classification. Specific steps include: training based on a Mask Language Model (such as BERT) through pseudo-cue fine-tuning. The pseudo-cue fine-tuning technique uses a fixed cue template, such as "This is a [MASK] news article: [News content]". During training, the news content is inserted into the cue template to fine-tune the BERT model, enabling it to make preliminary true / false label predictions for all fake news texts to be detected. Adaptive learning rate and gradient truncation techniques are used during training to ensure stable model convergence.

[0012] (2) Constructing a Consistent Network for Authenticity and Falsehood Propagation: Based on user forwarding records, a consistent network for authenticity and falsehood propagation is constructed. This network uses a large number of news items as nodes, and the connections between news items shared by the same social users as edges, forming an undirected graph. Specific steps include: extracting news forwarding data from user forwarding records, using news items as nodes, and the connections between news items forwarded by the same user as edges to form an undirected graph. To ensure the simplicity and efficiency of the graph, only active users and trending news items with forwarding frequencies exceeding a set threshold are retained, and duplicate graph connections are deleted. Finally, the filtered nodes and edges are integrated into a graph database such as Neo4j for storage and management.

[0013] (3) Final News Classification Using Graph Feature Diffusion Strategy: On the aforementioned consistent network for propagating truth and falsehood, a higher-order neighbor-aware diffusion equation is applied to diffuse the features of adjacent network news nodes to the current node until the network stabilizes. Specific steps include: initializing the features of the initially classified news nodes, where the node features are represented as X = {x1, x2, ..., xN}, where xi represents the feature vector of the i-th news node. The node features are then propagated and updated layer by layer using the higher-order neighbor-aware diffusion equation until the network converges.

[0014] During the diffusion process, the initial features of a node and the features of its neighbors are combined, and a fidelity term is used to ensure the stability and accuracy of feature updates. The diffusion equation is expressed as:

[0015]

[0016] Where xi(t) represents the feature of the i-th node at time t, α and β are diffusion coefficients, N(i) represents the set of neighboring nodes of the i-th node, Aij represents the element in the adjacency matrix, and di represents the degree of the i-th node. The feature update formula is:

[0017]

[0018] Where γ and η are the update weights, and wij is the weight between node i and node j. The criterion for feature convergence is:

[0019] ||x i (t+1)-x i (t)||<∈

[0020] Where ∈ is the convergence threshold. Finally, the diffused features are used to predict the veracity of the news, outputting more accurate fake news detection results. The prediction result is calculated using the following formula:

[0021] y i =softmax(Wx i +b)

[0022] Where yi is the predicted label of the i-th news node, and W and b are model parameters.

[0023] Through the above process, the task of detecting fake news based on the method of augmented cue learning and graph diffusion strategy is completed.

[0024] Preferably, step one of the method execution flow is the process of preliminary classification of fake news based on augmented cue learning. The augmented cue learning model is based on the Mask Language Model (such as BERT) and is trained through fine-tuning with pseudo-cues. Specific steps include: taking the news text as input, using the encoder of the BERT model to generate a contextual representation of each word, and predicting the word at the [MASK] position based on the probability distribution output by the model.

[0025] Given input text S = {si}, where i = 1 to N, and si represents the i-th word, construct a prompt template containing the [MASK] marker, T = "This is a [MASK] news item: " + S. During training, the model is optimized by maximizing the probability of words at the [MASK] position, enabling it to recognize and process input in this format. The specific formula is as follows:

[0026]

[0027] Where yi represents the true label, and pi represents the probability of the i-th word predicted by the BERT model at the [MASK] position.

[0028] Preferably, step two in the method execution flow: the process of constructing a true / false propagation consistency network, specifically includes the following steps:

[0029] 1. Extract news forwarding data from user forwarding records, using news items as nodes to form a preliminary node set. Specifically, extract all news forwarding data from user forwarding records. For each forwarding record, record the user ID and the corresponding news ID, forming a preliminary node set and edge set.

[0030] 2. Determine the edges of each node, that is, the connections between news forwarded by the same user, forming the edge set of an undirected graph.

[0031] 3. Filter users based on forwarding frequency, retaining only active users whose forwarding frequency exceeds a set threshold. Specific steps include: counting the number of forwards for each user, retaining only active users whose forwarding frequency exceeds the set threshold, storing these users' forwarding records in the dataset, and deleting the forwarding records of other users.

[0032] 4. Filter news based on forwarding frequency, retaining trending news and removing news with a forwarding frequency below a set threshold. Specific steps include: counting the number of times each news item is forwarded, retaining only trending news with a forwarding frequency above the set threshold, storing the forwarding records of these news items in the dataset, and deleting the forwarding records of other news items.

[0033] 5. Remove duplicate connections in the graph to ensure its simplicity and efficiency. This includes: traversing the edge set; if multiple edges exist between a pair of nodes, keep only one and delete the other duplicate edges.

[0034] 6. Integrate the filtered nodes and edges into the Neo4j graph database for storage and management. This database supports efficient storage and retrieval of nodes and edges, facilitating subsequent application of graph feature diffusion strategies and authenticity detection.

[0035] Preferably, in step three of the method execution flow: during the application of the graph feature diffusion strategy, the graph feature diffusion strategy includes the following steps:

[0036] 1. Feature Initialization: The features of the news nodes after initial classification are initialized, including embedding the true / false labels output by the augmentation prompt learning model into the node features. Specifically, the node features are represented as X = {x1, x2, ..., xN}, where xi represents the feature vector of the i-th news node.

[0037] 2. Applying a higher-order neighbor-aware diffusion equation, the features of each node are propagated layer by layer to its neighboring nodes, and the influence of neighboring nodes on the features of the current node is calculated. The diffusion equation is expressed as:

[0038]

[0039] Where xi(t) represents the characteristic of the i-th node at time t, α and β are diffusion coefficients, N(i) represents the set of neighboring nodes of the i-th node, Aij represents the element in the adjacency matrix, and di represents the degree of the i-th node.

[0040] 3. Feature Update: During each layer of propagation, the initial features of a node are combined with the feature update results from neighboring nodes. A fidelity term ensures the stability and accuracy of the feature update. The feature update formula is:

[0041]

[0042] Where γ and η are the update weights, and wij is the weight between node i and node j.

[0043] 4. Network Convergence: Continue propagating and updating node features until the network features converge and the node features no longer change significantly. The criteria for feature convergence are:

[0044] ||x i (t+1)-x i (t)||<∈

[0045] Here, ∈ is the convergence threshold. When the change magnitude of all node features is less than the threshold ∈, further feature propagation stops.

[0046] 5. Final Prediction: Utilizing the features after diffusion, a final prediction of the news's veracity is made, outputting more accurate fake news detection results. The final prediction result is calculated using the following formula:

[0047] y i =softmax(Wx i +b)

[0048] Where yi is the predicted label of the i-th news node, and W and b are model parameters.

[0049] Preferably, in step three of the method execution process, during the application of the graph feature diffusion strategy, the graph feature diffusion strategy includes a high-order neighbor-aware diffusion equation, the implementation process of which is as follows:

[0050] 1. The initialization and updating of node features depend on the feature differences between the node and its higher-order neighbors. Specifically, the initial node features are represented as X = {x1, x2, ..., xN}, where xi represents the feature vector of the i-th news node. The initial values ​​of the node features are based on the output of the augmented cue learning model.

[0051] 2. A dynamic weighting strategy is adopted, assigning different weights to the features of neighboring nodes at different levels to ensure the balance of feature propagation. Specifically, at each time step t, the feature update formula for node i is:

[0052]

[0053] Where α is the fidelity weight of the initial feature, wij represents the dynamic weight between node i and node j, and di represents the degree of node i.

[0054] 3. During feature update, a fidelity term is calculated using the similarity between the node's initial features and its neighbors' features to ensure the accuracy of feature updates. Specifically, the formula for calculating the fidelity term is:

[0055]

[0056] 4. At the end of each propagation layer, check the magnitude of change in node features. If the magnitude of change is less than a preset threshold, the features are considered to have converged, and further propagation stops. Specifically, the convergence criterion is as follows:

[0057] ||x i (t+1)-x i (t)||<∈

[0058] Here, ∈ is the convergence threshold. When the change magnitude of all node features is less than the threshold ∈, further feature propagation stops.

[0059] The present invention has at least the following beneficial technical effects:

[0060] 1. Comprehensive Contextual Information Capture: An enhanced prompt learning model is introduced, utilizing Mask Language Models (such as BERT) for fine-tuning through pseudo-prompts to construct fixed prompt templates, thereby improving the ability to capture textual contextual information. Compared with traditional feature extraction methods, it can better handle complex syntactic structures and implicit semantics, improving the accuracy of fake news detection;

[0061] 2. Efficient Utilization of Annotated Data: During the augmented reality learning process, the model is trained by combining annotated true and false news with unannotated news collected from the internet, using adaptive learning rate and gradient truncation techniques. This effectively solves the problems of scarce annotated data and imbalanced training samples. The model can generate high-confidence true / false label prediction results based on the input news content, improving the ability to identify a few categories of fake news.

[0062] 3. Enhanced Model Generalization Ability: A true / false propagation consistency network is constructed, utilizing a high-order neighbor-aware diffusion equation for feature propagation and updating, enabling the model to have stronger generalization ability when facing new and diverse types of fake news. The model no longer relies solely on the content of a single article, but makes comprehensive judgments based on user behavior data and propagation paths, improving the stability of detection results;

[0063] 4. Efficient Feature Update and Propagation: By combining a graph diffusion strategy with dynamic weighting and fidelity term calculation, the stability and accuracy of the feature update process are ensured. A high-order neighbor-aware diffusion equation is used to propagate node features layer by layer to neighbor nodes until the network features converge, greatly improving the efficiency and effectiveness of feature propagation. The formulaic feature update and propagation steps guarantee efficient model operation on large-scale data.

[0064] 5. Fusion of Global and Local Information: This invention, by combining an enhanced cueing learning model and a graph diffusion strategy, fully utilizes both global and local information in the text, improving the overall performance of fake news detection. When processing massive amounts of news data, the model can accurately extract and fuse information from multiple sources, providing more comprehensive detection results. Attached Figure Description

[0065] Figure 1 This is a flowchart of a fake news detection method based on augmented cue learning and graph diffusion strategy.

[0066] Figure 2 This is a schematic diagram of an embodiment of a fake news detection method based on enhanced cue learning and graph diffusion strategy. Detailed Implementation

[0067] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the embodiments described below are only some embodiments of the present invention, and not embodiments compatible with all application scenarios. Referring to the embodiments demonstrated in this invention, all other embodiments obtained by those skilled in the art without any inventive step are also within the protection scope of this invention.

[0068] This invention provides a method for detecting fake news based on enhanced cueing learning and graph diffusion strategies, comprising the following steps:

[0069] First, refer to the appendix Figure 1 Step 1 in the process, "Preliminary Classification of Fake News Based on Enhanced Cueing Learning," shows that the implementation process includes the following steps:

[0070] 1. Data Preparation: Collect a large set of fake news texts to be detected and extract relevant user forwarding records. These user forwarding records are used to subsequently build a consistency network for verifying the propagation of genuine and fake news.

[0071] 2. Training the Enhanced Prompt Learning Model: First, a Mask Language Model (such as BERT) is used as the base model, and fine-tuned using a small number of labeled samples for pseudo-prompt training. Specific steps include: First, constructing a prompt template, such as "This is a [MASK] news article: [News content]", and inserting the news content into the template. Then, the model is optimized by maximizing the prediction probability of the word at the [MASK] position. During training, adaptive learning rate and gradient cutoff techniques are used to prevent overfitting and gradient explosion, ensuring stable model convergence. After training, the model can generate preliminary true / false label predictions based on the input news content.

[0072] 3. Using the finely tuned detection model described above, the fake news to be detected is initially classified, and the classification results are obtained. Specifically, all the fake news texts to be detected are input into the trained reinforcement prompting learning model, and preliminary true / false labels are generated for each news item.

[0073] As a specific embodiment of the present invention, please refer to the appendix. Figure 2 In step one, we collected a large number of fake news items to be detected and used a BERT-based pre-trained language model as the Mask Language Model task, and used it as the backbone network of the prediction model for enhanced cue learning.

[0074] BERT (Bidirectional Encoder Representations from Transformers) is a deep learning model that excels in various linguistic tasks within the field of natural language processing. Proposed by Google AI, BERT is a pre-trained model specifically designed for NLP tasks. It utilizes a bidirectional Transformer architecture to simultaneously consider the contextual information of a word, thereby better understanding the semantics and contextual relationships of words. BERT's pre-training includes a Masked Language Model (MLM) and Next Sentence Prediction (NSP), enabling the model to learn rich contextual information and sentence-level associations.

[0075] Subsequently, using a small number of labeled samples of the collected fake news to be detected, the enhanced prompting learning prediction model was fine-tuned for training. Specifically, based on the constructed prompt template: "This is a [MASK] news: [News content]", the "UFO news" was inserted into the prompt template, resulting in the following prompt sentence: "This is a [MASK] news: A giant unidentified flying object was discovered in a certain place, and scientists say that aliens may have visited Earth." Then, the enhanced prompting learning prediction model was optimized by maximizing the prediction probability of the word at the [MASK] position. In this way, given any news to be detected, its authenticity can be preliminarily judged. For example, for the "UFO news", the detection model may classify the description of the news as a fake label, or it may incorrectly classify it as a real label. In this embodiment, the probability distribution of the initial label for the "UFO news" in the initial classification is: xUFO(0)=[0.8,0.2]. Although it can correctly distinguish between true and false, this preliminary classification method based on the enhanced prompting learning prediction model has low accuracy and confidence, and further optimization is required in subsequent steps.

[0076] Further, refer to the appendix. Figure 1 Step two in the process, "Constructing a Consistent Network for True / False Propagation," includes the following implementation steps:

[0077] 1. Extracting User Forwarding Records: This step is mainly responsible for collecting the elements (nodes, edges) for building the consistency network, specifically extracting news forwarding data from user forwarding records. Specific steps include: recording the user ID and corresponding news ID for each forwarding record from the social news forwarding context of the fake news to be detected, forming a preliminary set of nodes and edges.

[0078] 2. Constructing an undirected graph: First, using news articles as nodes, connections between news articles shared by the same user are used as edges to form an undirected graph. Second, to ensure the simplicity and efficiency of the graph, only active users with a forwarding frequency exceeding a set threshold and trending news articles are retained, and duplicate graph connections are removed.

[0079] 3. Data storage: The filtered nodes and edges are integrated into a graph database such as Neo4j for storage and management, which facilitates the application of graph feature diffusion strategies and authenticity detection.

[0080] As a specific embodiment of the present invention, please refer to the appendix. Figure 2 Regarding the rumor news, "A giant unidentified flying object was discovered in a certain location, and scientists say aliens may have visited Earth," as shown in the diagram, users A, B, and C forwarded this "UFO news." For example, user A also forwarded another news item: "The weather suddenly changed in a certain location, and the meteorological bureau issued a warning." A direct edge connection is then formed between the "UFO news" and the "weather warning news." Subsequently, an undirected graph network is constructed, using all forwarded news items from this user group within the forwarding time of the "UFO news" as network nodes, and connections between news items forwarded by the same user as edges, forming a consistent truth-based propagation network. Active users and trending news items with forwarding frequencies exceeding a set threshold are retained, while duplicate connections are deleted. Finally, the filtered nodes and edges are integrated into the Neo4j graph database for storage and management, and for real-time invocation of subsequent graph feature diffusion strategies.

[0081] Further, refer to the appendix. Figure 1 Step 3: The final news classification using the graph feature diffusion strategy is shown in the following implementation process:

[0082] 1. Feature Initialization: The features of the news nodes after initial classification are initialized, including embedding the true / false labels output by the augmentation prompt learning model into the node features. Specifically, the node features are represented as X = {x1, x2, ..., xN}, where xi represents the feature vector of the i-th news node.

[0083] 2. Applying a higher-order neighbor-aware diffusion equation, the features of each node are propagated layer by layer to its neighboring nodes, and the influence of neighboring nodes on the features of the current node is calculated. The diffusion equation is expressed as:

[0084]

[0085] Where xi(t) represents the characteristic of the i-th node at time t, α and β are diffusion coefficients, N(i) represents the set of neighboring nodes of the i-th node, Aij represents the element in the adjacency matrix, and di represents the degree of the i-th node.

[0086] 3. Feature Update: During each layer of propagation, the initial features of a node are combined with the feature update results from neighboring nodes. A fidelity term ensures the stability and accuracy of the feature update. The feature update formula is:

[0087]

[0088] Where γ and η are the update weights, and wij is the weight between node i and node j.

[0089] 4. Network Convergence: Continue propagating and updating node features until the network features converge and the node features no longer change significantly. The criteria for feature convergence are:

[0090] ||x i (t+1)-x i (t)||<∈

[0091] Here, ∈ is the convergence threshold. When the change magnitude of all node features is less than the threshold ∈, further feature propagation stops.

[0092] 1. Final Prediction: Utilizing the features after diffusion, a final prediction of the news's veracity is made, outputting more accurate fake news detection results. The final prediction result is calculated using the following formula:

[0093] y i =softmax(Wx i +b)

[0094] Where yi is the predicted label of the i-th news node, and W and b are model parameters.

[0095] As a specific embodiment of the present invention, please refer to the appendix. Figure 2 After repeatedly updating the propagation and node features using a graph feature diffusion strategy (specifically at time t), for the rumor news "A giant unidentified flying object was discovered in a certain place, and scientists say aliens may have visited Earth," the initial features of the corresponding news node "UFO News" are updated from xUFO(0) = [0.8, 0.2] to xUFO(t) = [0.95, 0.05], indicating that the probability of the "false" label increases to 0.95. Therefore, through network feature diffusion, the final veracity prediction of this example news can be made, meeting higher accuracy requirements.

[0096] Through the above steps, a fake news detection method based on augmented cue learning and graph diffusion strategy was implemented, which can significantly improve the accuracy and robustness of fake news detection.

[0097] To reiterate, the above embodiments are merely some examples of the present invention, and not embodiments compatible with all application scenarios. All other embodiments obtained by those skilled in the art without any inventive step, referring to the embodiments demonstrated in this invention, are also within the scope of protection of this invention.

Claims

1. A method for detecting fake news based on augmented cue learning and graph diffusion strategy, characterized in that, This method takes as input the text of the fake news to be detected and user news forwarding records to construct a consistency network for verifying the spread of fake news, and outputs a label indicating whether the fake news text to be detected is true or false. It includes the following steps: Step 1: Preliminary classification of fake news based on augmented prompt learning: A large set of fake news texts to be detected is used as input, and an augmented prompt learning model is trained with a small number of labeled samples. This includes: training based on the Mask Language Model and fine-tuning through pseudo-prompts; during the training process, news content is inserted into the prompt template to fine-tune the model so that it can make preliminary predictions of the true and false labels of all fake news texts to be detected. Step 2: Construct a consistent propagation network for authenticity verification: Based on user forwarding records, construct a consistent propagation network for authenticity verification. This network uses a large number of news items as nodes, and the connections between news items shared by the same social media user as edges, forming an undirected graph. This includes: extracting news forwarding data from user forwarding records, using news items as nodes, and the connections between news items shared by the same user as edges to form an undirected graph; retaining active users and trending news items with forwarding frequencies exceeding a set threshold, and deleting duplicate graph connections; finally, using a graph database such as Neo4j to store and manage the network data. Step 3: Final News Classification Using Graph Feature Diffusion Strategy: On the constructed consistent network for truth / falsehood propagation, a high-order neighbor-aware diffusion equation is applied to diffuse the features of adjacent network news nodes to the current node until the network stabilizes. This includes: initializing the features of the initially classified news nodes; propagating and updating node features layer by layer through the high-order neighbor-aware diffusion equation until the network converges; during the diffusion process, combining the initial node features and neighbor features, a fidelity term is used to ensure the stability and accuracy of feature updates; finally, the truth / falsehood features of the nodes are integrated, updated, and fused to output the final prediction result for the truth / falsehood of fake news. The specific implementation process of the high-order neighbor-aware diffusion equation is as follows: The initialization and updating of node features depend on the feature differences between the node and its higher-order neighbors; specifically, the initial node features are represented as follows: ={ 1, 2,..., },in, Indicates the first The feature vectors of each news node; the initial values ​​of the node features are based on the output of the augmented cue learning model; A dynamic weighting strategy is adopted, assigning different weights to the features of neighbor nodes at different levels to ensure the balance of feature propagation; specifically, at each time step... ,node The feature update formula is: in, These are the fidelity weights of the initial features. Represents a node and nodes Dynamic weights between them Represents a node The degree; During feature update, a fidelity term is calculated using the similarity between the node's initial features and its neighbors' features to ensure the accuracy of feature updates. Specifically, the formula for calculating the fidelity term is: At each time step The fidelity term is used to adjust the weights in node feature updates to ensure the stability and accuracy of feature updates; At the end of each propagation layer, the magnitude of change in node features is checked. If the magnitude of change is less than a preset threshold, the features are considered to have converged, and further propagation is stopped.

2. The method for detecting fake news based on augmented cueing learning and graph diffusion strategy according to claim 1, characterized in that, Step 1: Preliminary classification of fake news based on augmented cue learning. The augmented cue learning model is based on the Mask Language Model and is trained through fine-tuning with pseudo-cues. This includes: taking the news text as input, using the encoder of the BERT model to generate a contextual representation of each word, and predicting the word at the [MASK] position based on the probability distribution output by the model; given the input text... ={ } =1~ ,in, Indicates the first For each word, a prompt template containing the [MASK] tag is constructed: T = "This is a [MASK] news article:" + S [news content]. During training, the news content is inserted into the prompt template, and the BERT model is fine-tuned by maximizing the probability of words at the [MASK] position to enable it to recognize and process input in this format.

3. The method for detecting fake news based on augmented cue learning and graph diffusion strategy according to claim 2, characterized in that, In step one, the news text is used as input, and the encoder of the BERT model generates a contextual representation of each word. The probability distribution output by the model is then used to predict the word at the [MASK] position. The model is optimized by maximizing the probability of the correct word at the [MASK] position, thereby improving the accuracy of identifying fake news. The specific formula is as follows: in, Indicates the true label, This indicates the BERT model's prediction at position [MASK] for the [MASK]th time. The probability of each word.

4. The method for detecting fake news based on augmented cue learning and graph diffusion strategy according to claim 3, characterized in that, In step one, during the training process, adaptive learning rate and gradient cutoff techniques are used to prevent overfitting and gradient explosion, ensuring stable convergence of the model. After training, the model can generate high-confidence true / false label prediction results based on the input news content.

5. The method for detecting fake news based on augmented cueing learning and graph diffusion strategy according to claim 1, characterized in that, Step two: Constructing a truth-or-false propagation consistency network, which includes the following steps: 201) Extract news forwarding data from user forwarding records, and form a preliminary node set with news as nodes; specifically, extract all news forwarding data from user forwarding records, and record the user ID and the corresponding news ID for each forwarding record to form a preliminary node set and edges; 202) Determine the edges of each node, that is, the connections between news forwarded by the same user, forming the edge set of an undirected graph; 203) Filter users based on forwarding frequency and retain active users whose forwarding frequency is higher than a set threshold, including: counting the number of forwards for each user, retaining active users whose forwarding frequency is higher than a set threshold, keeping the forwarding records of these users in the dataset, and deleting the forwarding records of other users; 204) Filter news based on forwarding frequency, retain hot news, and remove news with forwarding frequency below a set threshold. This includes: counting the number of times each news item is forwarded, retaining hot news with forwarding frequency above a set threshold, keeping the forwarding records of these news items in the dataset, and deleting the forwarding records of other news items. 205) Remove duplicate connections in the graph, including: traversing the edge set, if there are multiple edges between a pair of nodes, keep one and delete the other duplicate edges; 206) Integrate the filtered nodes and edges into the Neo4j graph database for storage and management.

6. The method for detecting fake news based on augmented cue learning and graph diffusion strategy according to claim 1, characterized in that, In step three: the application of the graph feature diffusion strategy includes the following steps: 301) Feature Initialization: Perform feature initialization on the news nodes after preliminary classification, including embedding the true / false labels output by the enhancement prompt learning model into the node features; specifically, the node features are represented as follows: ={ 1, 2,..., },in Indicates the first Feature vectors of each news node; 302) Applying a higher-order neighbor-aware diffusion equation, the features of each node are propagated layer by layer to its neighboring nodes, and the influence of neighboring nodes on the features of the current node is calculated; the diffusion equation is expressed as: in, ( ) indicates the first Each node in time Features and It is the diffusion coefficient. ( ) indicates the first The set of neighboring nodes of a node. Represents the elements in the adjacency matrix Indicates the first The degree of each node; 303) Feature Update: During each layer of propagation, the initial features of a node and the feature update results from neighboring nodes are combined, and a fidelity term is used to ensure the stability and accuracy of the feature update; the feature update formula is: in, and It's about updating the weights. It is a node and nodes The weights between them; 304) Network Convergence: Continue propagating and updating node features until the network features converge and the node features no longer change significantly; the criteria for feature convergence are: in, It is the convergence threshold; when the change in the features of all nodes is less than the threshold. At that time, further feature propagation ceases; 305) Final prediction: Utilize the features after diffusion to make a final prediction of the authenticity of the news, and output more accurate fake news detection results.

7. The method for detecting fake news based on augmented cueing learning and graph diffusion strategy according to claim 6, characterized in that, The prediction result is calculated using the following formula: in, It is the first Predicted tags for each news node, and These are model parameters.

Images