False news detection method and system based on knowledge graph and cognitive disorder

Abstract

The invention provides a false news detection method and system based on a knowledge graph and cognitive disorder, and relates to the technical field of false news detection. The invention aims to solve the technical problem that the accuracy of identifying false news with confusion is not high due to the fact that complex structure information of news transmission is not fully utilized and the psychological cognition process of a user facing contradictory information is difficult to capture and quantify. The method comprises the following steps: constructing an event graph based on obtained social media data, and extracting a context representation vector of each node in the event graph; according to the method, a multi-dimensional cognitive disorder gated attention network is constructed, psychological conflicts experienced by each news propagation participant are finely quantified from two orthogonal dimensions of social contact-content and behavior-content, true and false news classification is realized, the urgent requirements of early discovery and rapid disposal of false news in practical application are met, and the method is suitable for popularization and application. And the method has extremely high engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of fake news detection technology, and in particular relates to a fake news detection method and system based on knowledge graphs and cognitive dissonance. Background Technology

[0002] The widespread adoption of social media has fundamentally changed the way information is produced and disseminated. However, while bringing convenience, it has also provided fertile ground for the proliferation of fake news. Fake news, with its inflammatory, deceptive, and viral spread, poses a serious threat to social stability, public safety, and even individual well-being.

[0003] Existing fake news detection methods mainly fall into two categories. The first is content-based analysis, which typically utilizes Natural Language Processing (NLP) techniques to determine the authenticity of news texts by analyzing their linguistic features, writing style, and sentiment. For example, deep learning models (such as CNN, LSTM, and BERT) are used to extract deep semantic features from the text and train classifiers for recognition. However, these methods have significant limitations: First, they are highly dependent on the text content itself, and their ability to detect "sophisticated" fake news that imitates the style of authoritative news sources and has no obvious linguistic flaws drops significantly. Second, forgers can evade detection by constantly adjusting the text content, requiring frequent model updates and resulting in poor adaptability.

[0004] The second type is the analysis method based on news dissemination patterns. These methods focus on auxiliary information such as the dissemination path of news on social networks and user interaction behavior. They typically model the news dissemination process as a cascaded graph or time-series graph, analyzing its topological structure, dissemination depth, user engagement, and other characteristics. For example, they use graph neural networks (GNNs) to learn representations of the dissemination structure. Although these methods introduce valuable structural information, they also face key bottlenecks: First, most existing methods homogenize the dissemination network, ignoring the heterogeneous relationships between different roles such as users, posts, and entities, thus losing a large amount of valuable semantic information. Second, they mostly analyze dissemination patterns from a macro-statistical perspective, making it difficult to delve into the individual level. They cannot effectively explain "why" certain users become key nodes in the dissemination of fake news, nor can they capture and utilize subtle changes in user cognition during the dissemination process.

[0005] Therefore, designing a system that can accurately identify highly deceptive and inflammatory fake news is a technical challenge that urgently needs to be addressed in the field of automated content moderation and information security. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention aims to solve the technical problems in existing fake news detection methods, such as insufficient utilization of the complex structural information of news dissemination, difficulty in capturing and quantifying users' psychological cognitive processes when faced with contradictory information, and consequently, low accuracy in identifying carefully fabricated and deceptive fake news.

[0007] In a first aspect, the present invention provides a method for detecting fake news based on knowledge graphs and cognitive dissonance, comprising the following steps:

[0008] Acquire multi-source heterogeneous social media data, construct an event graph based on the acquired social media data, and extract the initial feature vector of each node in the event graph;

[0009] Based on the initial feature vector of each node in the event graph, a relational graph convolutional network is used to extract the context representation vector of each node in the event graph;

[0010] A multidimensional cognitive dissonance gating attention network is constructed, including a dissonance gating unit and a multilayer perceptron. The dissonance gating unit is used to generate a graph-level representation vector of the event graph based on the context representation vector of each node in the event graph, and the multilayer perceptron is used to classify the graph-level representation vector into true and false news.

[0011] A labeled dataset was constructed, and a relation graph convolutional network and a multidimensional cognitive dissonance gated attention network were jointly trained and optimized.

[0012] The trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models were deployed as online services for real-time detection of emerging news events.

[0013] Furthermore, the multi-source heterogeneous social media data includes the target news events to be detected and related data, including: participating user data, post data, social relationship data, and dissemination relationship data;

[0014] User participation data includes information on users who participate in posting, forwarding, and commenting on the target news event; post data includes all posts published by users related to the target news event, including original content, forwards, and comments; social relationship data includes the following relationships between users; and dissemination relationship data includes the forwarding or commenting relationships between posts.

[0015] An event graph generation module is built, which is used to construct event graphs based on the acquired social media data. ,in, For a set of nodes, The set of edges and the set of nodes include various types of nodes, including news nodes. Content entity nodes Participating user nodes Post Nodes News Nodes This represents the core news content of the target news event; each event graph contains a unique news node; participating user nodes. This refers to users who participate in posting, forwarding, and commenting on the target news event, and these are considered participating user nodes. It includes basic features and category features. Basic features include account registration days, number of followers, number of following accounts, total number of Weibo posts, bio length, and number of custom tags; category features include authentication type and gender; post nodes. This refers to all posts published by participating users that are related to the target news event, including original content, reposts, and comments; content entity nodes. Represents named entities of predefined categories extracted from target news events and posts, including people's names, place names, organization names, and product names;

[0016] edge set It includes various types of edges, including news-entity edges, post-entity edges, user-post edges, post-post edges, and user-user edges; news-entity edges connect news nodes. and content entity nodes Edges; Post-entity edges connect post nodes. and content entity nodes The edge; the user-post edge connects the participating user nodes. and the post node created by the participating user The edges are: post-post edges are edges connecting two post nodes that have a forwarding or commenting relationship; user-user edges are edges connecting two participating user nodes that have a following relationship. If the current participating user follows another participating user, then the user-user edge between the current participating user and the other participating user is a directed edge from the current participating user to the other participating user.

[0017] Furthermore, the initial feature vector of each node in the event graph is extracted, specifically as follows:

[0018] An initial feature vector extraction module is constructed to initialize feature vectors for each node in the event graph. For news nodes and post nodes, a pre-trained Chinese language model is used for encoding and initial feature vector extraction to obtain the news node's feature vector. initial feature vector and post nodes initial feature vector ;

[0019] For content entity nodes Using a pre-trained large model as the word vector model, the content entity nodes are converted based on the word vector model. Dimensional content entity node initial feature vector ;

[0020] For participating user nodes Initial feature vectors are extracted based on the basic and category features of participating users. The basic features are then Z-score normalized, and the category features are one-hot encoded. These encoded category features are then concatenated with the basic features to obtain the initial feature vectors of the participating user nodes. .

[0021] Furthermore, the context representation vector of each node in the event graph is extracted based on the relational graph convolutional network. The specific method is as follows:

[0022] event diagram enter Layer-wise relational graph convolutional network for extracting event graphs any node in Context representation vector Event diagram any node In the relational graph convolutional network RGCN, the first Layer context representation vector As shown in the formula below:

[0023]

[0024] in, The set of edge types; For the first Types of nodes The set of neighboring nodes, For the set of neighbor nodes The Middle One neighboring node; For the first Type 1 Layer-specific learnable weight matrices; For nodes The transformation matrix; This is the normalization constant; To modify the activation function of the linear unit; For the first The neighboring node is at the _th ... The feature vector of the layer; For nodes In the The feature vector of the layer.

[0025] Furthermore, a graph-level representation vector of the event graph is generated based on the context representation vector of each node in the event graph. The specific method is as follows:

[0026] For event graphs Each participating user node Calculate the participating user nodes respectively Social-content imbalance and behavioral-content dissonance component Generate participating user nodes Multidimensional cognitive dissonance vector ;

[0027] Multidimensional misalignment vector A gating signal with the same dimension as the participating user nodes is generated by transforming the signal through a neural network. ;

[0028] Using gating signals Context representation vector of participating user nodes Perform element-wise multiplication to obtain the gated augmented representation of the participating user nodes. ;

[0029] Multidimensional cognitive dissonance vector The module length serves as the participating user node The overall intensity of the imbalance is calculated using a linear transformation and the Softmax function for participating user nodes. Attention weights;

[0030] Using attention weights and gating enhancement representation of participating user nodes The graph-level representation vector is obtained by weighted summation of all participating user nodes. ;

[0031] The specific method for classifying true and false news using graph-level representation vectors is as follows:

[0032] Representing vectors at the graph level Inputting a two-layer multilayer perceptron (MLP) into a binary classification system for true and false news, the system obtains the predicted graph-level representation vector. This represents the probability distribution of fake news.

[0033] Furthermore, the labeled dataset includes several event graphs and corresponding real and fake labels; the labeled dataset is divided into training set, validation set and test set according to proportions;

[0034] Joint training and optimization of a relation graph convolutional network and a multidimensional cognitive dissonance gated attention network were performed, including:

[0035] Set training parameters, including initial learning rate, weight decay, batch size, and AdamW optimizer;

[0036] Construct a cross-entropy loss function with label smoothing and set the label smoothing parameters. ;

[0037] Before training begins, all learnable parameters of the relational graph convolutional network RGCN and the multidimensional cognitive dissonance gated attention network are randomly initialized using a Xavier uniform distribution; fixed weights of the pre-trained Chinese language model and word vector model are loaded.

[0038] Start the training loop and set the maximum number of rounds; in each round, the model traverses the entire training set in batches. For each batch of data, perform a complete forward propagation and calculate the prediction result corresponding to each batch of data.

[0039] Calculate the loss value of the current batch of data based on the labeled smoothed cross-entropy loss function; execute the backpropagation algorithm to calculate the gradients of all learnable parameters in the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network;

[0040] The AdamW optimizer is invoked to update all parameters in the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network based on the calculated gradients. The AdamW optimizer then completes a training epoch by iterating through all batches of data according to the preset learning rate and weight decay strategy.

[0041] After each training round, based on the relational graph convolutional network and multidimensional cognitive dissonance gated attention network obtained in the current training round, a complete prediction is performed on an independent validation set. The key performance indicators of the current training round on the validation set are calculated and recorded, including accuracy, precision, recall and F1 score.

[0042] Based on the early stopping mechanism, an upper limit is set for the patience counter, which records the highest historical F1 score on the validation set in each training epoch. If the F1 score on the validation set in the current training epoch does not exceed the highest historical F1 score, the patience counter is incremented by one; otherwise, the patience counter is reset to 0. When the patience counter reaches the upper limit, it is determined that the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network have converged, and training is terminated early.

[0043] When the F1 score on the validation set exceeds the highest historical F1 score, update the highest historical F1 score and save all parameter states of the current relational graph convolutional network and multidimensional cognitive dissonance gated attention network as the best model file. When the maximum number of training rounds is reached or the early stopping mechanism is triggered to end the training, the best model file obtained is used as the relational graph convolutional network and multidimensional cognitive dissonance gated attention network model that performed best on the validation set during the entire training process.

[0044] Furthermore, the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models are deployed as online services, specifically as follows:

[0045] Package the trained best model file, event graph generation module, initial feature vector extraction module, relation graph convolutional network, and multidimensional cognitive dissonance gated attention network; use the FastAPI framework to deploy the trained relation graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services;

[0046] When a new news event needs to be detected, a real-time detection process is triggered in response to an external public opinion monitoring system, including:

[0047] When the public opinion monitoring system detects a news event that spreads faster than the threshold of new reposts and comments per minute within a preset time, it obtains the core URL of the news event as a unique identifier.

[0048] The public opinion monitoring system immediately initiates an API call to the load balancer address of the deployed detection service cluster. After receiving the request, the detection service cluster obtains multi-source heterogeneous social media data based on the core URL within a fixed time window, and specifically crawls all posts, user information, forwarding chains and comment data related to the news to be detected. Based on the event graph generation module, it constructs an event graph of the news event to be detected.

[0049] After the event graph is constructed, the service process loads the best model file and model weights that have been trained, and sequentially executes the forward propagation of the relation graph convolutional network and the forward propagation of the multidimensional cognitive dissonance gated attention network to calculate the probability value of the news event to be detected as fake news, obtain the detection result and return it to the public opinion monitoring system.

[0050] Secondly, the present invention provides a fake news detection system based on knowledge graphs and cognitive dissonance, comprising a data layer, a representation layer, a decision layer, and an application layer;

[0051] The data layer is used to acquire multi-source heterogeneous social media data, construct an event graph based on the acquired social media data, and extract the feature vector of each node in the event graph.

[0052] The representation layer is used to extract the context representation vector of each node in the event graph;

[0053] The decision layer is used to generate graph-level representation vectors of the event graph and to classify the graph-level representation vectors as true or false news.

[0054] The application layer is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network model as an online service for real-time detection of emerging news events.

[0055] The data layer includes a data acquisition module, an event graph generation module, and an initial feature vector extraction module; the data acquisition module is used to acquire multi-source heterogeneous social media data; the event graph generation module is used to construct an event graph based on the acquired data; the initial feature vector extraction module is used to extract and encode an initial feature vector for each node in the event graph.

[0056] The representation layer includes a relational graph convolutional network module, which comprises a multi-layer relational graph convolutional network. The first layer of the relational graph convolutional network includes an independent projection matrix, which is used to uniformly project the initial features of nodes of different dimensions to a unified dimension. The other layers, except for the first layer, keep the dimensions unchanged and are used to aggregate the neighbor information of different types of nodes in the event graph, and generate a context representation vector for each node in the event graph that integrates multi-level neighborhood structure and semantic information.

[0057] The decision layer includes an imbalance gating unit and a multilayer perceptron; the imbalance gating unit is used to generate a graph-level representation vector of the event graph; the multilayer perceptron is used to classify the event graph as true or false news, inputting the graph-level representation vector of the event graph into a multilayer perceptron with several layers, and outputting the probability distribution of the news event corresponding to the event graph as false news;

[0058] The application layer includes a service encapsulation module, a containerized deployment module, and a real-time detection and flow control module; the service encapsulation module is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services.

[0059] The containerized deployment module is used to deploy online services to the detection service cluster; the containerized deployment module is also used to dynamically adjust the number of online service instances;

[0060] The real-time detection flow control module is used to coordinate the entire detection task; the real-time detection flow control module includes a triggering unit, a scheduling unit and a feedback unit; the triggering unit is used to connect to an external public opinion monitoring system, extract the core URL of the news event to be detected and initiate a detection request;

[0061] The scheduling unit is used to acquire social media data of the news event to be detected, construct an event graph of the news event to be detected, load the model weights of the best model file, calculate the probability value of the news event to be detected as fake news, and obtain the detection result.

[0062] The feedback unit is used to return the detection results to an external public opinion monitoring system.

[0063] Thirdly, this application proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to perform the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0064] Fourthly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0065] Fifthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0066] The beneficial effects of adopting the above technical solution are as follows: This invention provides a fake news detection method based on knowledge graphs and cognitive dissonance, proposing the concept of a multidimensional cognitive dissonance vector. This vector can quantify the psychological conflict experienced by each participant in news dissemination in a refined manner from two orthogonal dimensions: social-content and behavioral-content. This method transforms abstract psychological theories into calculable, structured features, solving the fundamental problem that traditional methods rely solely on content or dissemination patterns and cannot capture deep cognitive contradictions. This enables the model to identify fake news that appears reasonable in content but causes widespread cognitive dissonance during dissemination.

[0067] This invention designs an imbalance gating unit and an attention mechanism. This mechanism utilizes a multidimensional cognitive imbalance vector to dynamically reshape the user's feature representation through gating operations, simulating the impact of cognitive conflict on individual information processing. Simultaneously, it calculates attention weights, enabling the model to focus on the most conflicted and contradictory user groups. This dual-action mechanism makes the model's decision-making process closer to human judgment logic, significantly improving detection accuracy and interpretability, and achieving a 3-5% improvement in F1 score compared to traditional attention mechanisms.

[0068] This invention constructs an end-to-end, integrated fake news detection system, encompassing data acquisition, graph construction, multi-level representation learning, and cognitive decision-making. The system comprehensively characterizes the multimodal and multi-dimensional information of news events through a heterogeneous knowledge graph and combines it with RGCN for deep structured information extraction. Finally, it makes decisions through an innovative cognitive dissonance network, forming a complete and efficient technical chain. This framework is deployed as a scalable, real-time online service, meeting the urgent need for early detection and rapid handling of fake news in practical applications, and possesses extremely high engineering application value. Attached Figure Description

[0069] Figure 1 Flowchart of the fake news detection method based on knowledge graph and cognitive dissonance provided in Embodiment 1 of the present invention;

[0070] Figure 2 A schematic diagram of the multidimensional cognitive dissonance vector generation mechanism provided in Embodiment 1 of the present invention;

[0071] Figure 3 A schematic diagram of a multidimensional cognitive dissonance gating attention network provided in Embodiment 1 of this invention;

[0072] Figure 4 A schematic diagram of the structure of the fake news detection system based on knowledge graphs and cognitive dissonance provided in Embodiment 2 of the present invention;

[0073] Figure 5 A schematic diagram of the deployment and invocation process of the real-time detection service provided in Embodiment 2 of the present invention. Detailed Implementation

[0074] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0075] Example 1:

[0076] Existing fake news detection methods often overlook a core phenomenon in social psychology: cognitive dissonance. This refers to the psychological tension individuals experience when exposed to information that conflicts with their existing beliefs, the mainstream views of their social circle, or their own behavior. The spread of fake news is essentially a process of continuously creating and exploiting cognitive dissonance within social groups. A carefully crafted piece of fake news can often cleverly trigger strong cognitive conflict and emotional resonance among a specific group, thereby driving its viral spread. Current technologies fail to effectively integrate content semantics with dissemination structure, completely lacking the ability to model and quantify this crucial psychological driver.

[0077] To address the technical problems of existing fake news detection methods, such as insufficient utilization of the complex structural information of news dissemination and difficulty in capturing and quantifying users' psychological cognitive processes when faced with contradictory information, resulting in low accuracy in identifying carefully fabricated and deceptive fake news, this embodiment provides a fake news detection method based on knowledge graphs and cognitive dissonance attention networks. This method deeply integrates news content, dissemination structure, and psychological cognitive dissonance theory. By constructing a heterogeneous knowledge graph to capture the full picture of news dissemination, it proposes a multidimensional cognitive dissonance-gated attention network to simulate and quantify the psychological conflicts users experience during information dissemination, thereby accurately identifying fake news.

[0078] In this embodiment, a fake news detection method based on knowledge graphs and cognitive dissonance, such as... Figure 1As shown, it includes the following steps:

[0079] S1. Obtain multi-source heterogeneous social media data, construct an event graph generation module and an initial feature vector extraction module, use the event graph generation module to construct an event graph based on the acquired social media data, and use the initial feature vector extraction module to initialize a feature vector for each node in the event graph;

[0080] Multi-source heterogeneous social media data includes the target news event to be detected and related data, including: participating user data, post data, social relationship data, and dissemination relationship data;

[0081] User participation data includes information about users who participate in posting, forwarding, and commenting on the target news event; post data includes all posts published by users related to the target news event, including original content, forwards, and comments; social relationship data includes the following relationships between users; and dissemination relationship data includes the forwarding or commenting relationships between posts.

[0082] Constructing an event graph based on acquired social media data ,in, For a set of nodes, The set of edges and the set of nodes include various types of nodes, including news nodes. Content entity nodes Participating user nodes Post Nodes ;

[0083] News Node This represents the core news content of the target news event, with each event graph containing a unique news node.

[0084] Participating user nodes This refers to users who participate in posting, forwarding, and commenting on the target news event;

[0085] Post Node This refers to all posts published by participating users that are related to the target news event, including original content, reposts, and comments;

[0086] Content entity nodes Represents named entities of predefined categories extracted from target news events and posts, including people's names, place names, organization names, and product names;

[0087] edge set It includes various types of edges, including news-entity edges, post-entity edges, user-post edges, post-post edges, and user-user edges;

[0088] News - Entity edges connect news nodes and content entity nodes The edge;

[0089] Posts - Entity edges connect post nodes and content entity nodes The edge;

[0090] The user-post edge connects the participating user nodes. and the post node created by the participating user The edge;

[0091] A post edge is an edge that connects two post nodes that have a forwarding or commenting relationship;

[0092] A user-user edge is an edge that connects two participating user nodes that have a follow relationship. If the current participating user follows another participating user, then the user-user edge between the current participating user and the other participating user is a directed edge from the current participating user to the other participating user.

[0093] Extracting event graphs The initialization feature vector for each node is performed as follows:

[0094] For news nodes and post nodes, a pre-trained Chinese language model is used for encoding and initial feature vector extraction; in this embodiment, news nodes... and post nodes The corresponding text content is input into the pre-trained Chinese RoBERTa-wwm-ext-large model, and the output vector corresponding to the [CLS] flag is used as the news node. and post nodes initial feature vector and ;

[0095] For content entity nodes Using a pre-trained large model as the word vector model, the content entity nodes are converted based on the word vector model. The initial feature vector of dimension; in this embodiment, the FastText model pre-trained on a large Chinese encyclopedia and news corpus is used as the word vector model to convert content entity nodes into... Initial eigenvectors of dimension ;

[0096] For participating user nodes An initial feature vector is extracted based on the basic and category features of participating users. The basic features include the number of days the account has been registered, the number of followers, the number of accounts followed, the total number of Weibo posts, the length of the profile, and the number of custom tags. The category features include the authentication type and gender. The authentication type values ​​are: 0 for none, 1 for individual, and 2 for enterprise. The gender values ​​are: 0 for unknown, 1 for male, and 2 for female.

[0097] Z-score normalization is performed on the basic features, and the first... Basic characteristics Z-score normalized value As shown in the formula below:

[0098]

[0099] in, For the first The original values ​​of the basic features, For the first The mean of the basic characteristics among all users For the first The standard deviation of a basic feature across all users;

[0100] One-hot encoding is performed on the categorical features, and then concatenated with the encoded basic features to obtain the initial feature vector of the participating user nodes. ;

[0101] S2. Extract the context representation vector of each node in the event graph based on the relational graph convolutional network;

[0102] event diagram enter The event graph is processed by the layer relation graph convolutional network RGCN to extract the event graph. any node in Context representation vector Event diagram any node In the relational graph convolutional network RGCN, the first Layer context representation vector As shown in the formula below:

[0103]

[0104] in, The set of edge types; For the first Types of nodes The set of neighboring nodes, For the set of neighbor nodes The Middle One neighboring node; For the first Type 1 Layer-specific learnable weight matrices; For nodes The transformation matrix; Set the normalization constant. ; To modify the activation function of the linear unit; For the first The neighboring node is at the _th ... The feature vector of the layer; For nodes In the The feature vector of the layer;

[0105] In this embodiment, the number of layers of the graph convolutional network RGCN is set. In order to handle input features of different dimensions, such as the 768-dimensional news nodes in this embodiment. Initial feature vector and post nodes of 300-dimensional content entities initial feature vector And 20-dimensional participating user nodes initial feature vector In the relational graph convolutional network RGCN, the 1st Within each layer, a separate projection matrix is ​​set for each node type, projecting the initial feature vectors of all nodes uniformly onto the same matrix. Dimension. The first dimension of the relational graph convolutional network RGCN. The input and output dimensions of the layer are both kept to be .go through The iterative update of the layer relation graph convolutional network RGCN yields the event graph. Each node Context representation vector The context representation vector obtained in this embodiment This embodiment integrates the complex structure and semantic information within the third-order neighborhood of each node. It aggregates multi-order, heterogeneous neighbor node information through a graph neural network to form an event graph. Each node learns a context-aware representation vector that deeply integrates its own content with the graph topology.

[0106] S3. Construct a multidimensional cognitive dissonance gating attention network, including a dissonance gating unit and a multilayer perceptron; the dissonance gating unit is used to generate graph-level representation vectors of the event graph, and the multilayer perceptron is used to classify the graph-level representation vectors as true or false news.

[0107] S31. For event diagrams Each participating user node Quantify participating user nodes from two dimensions: social-content and behavioral-content. The degree of cognitive dissonance was calculated for each participating user node. Social-content imbalance and behavioral-content dissonance component Generate participating user nodes Multidimensional cognitive dissonance vector ;

[0108] The social-content dissonance component measures the conflict between the news content a user encounters and the mainstream views within their social circle. The behavioral-content dissonance component measures the conflict between the user's current dissemination behavior and the emotional / stance expressed in the news content itself.

[0109] The specific method for calculating the social-content dissonance component of participating user nodes is as follows:

[0110] For participating user nodes Neighbors participating in user nodes Context representation vector An attention-weighted average is performed, where the participating user nodes in the neighboring user nodes are the participating user nodes in the sum of the participating user node vectors, to obtain a social consistency representation representing the viewpoints of the social circle. As shown in the formula below:

[0111]

[0112] in, For participating user nodes The neighbors participate in the user node set. For neighboring user node sets The Middle The attention weights of each neighbor participating in a user node are shown in the following formula:

[0113]

[0114] in, For participating user nodes The context representation vector, This is a learnable attention weight matrix;

[0115] Based on the user-post edge on the participating user node, obtain the participating user node. The nodes of the posts being spread Context representation vector And calculate the participating user nodes Information exposure representation As shown in the formula below:

[0116]

[0117] in, Indicates participating user nodes The set of all post nodes associated with the target news event is used as the basis for calculating the semantic information of all posts associated with the user by employing an average pooling strategy.

[0118] Based on computational participation user nodes Social-content imbalance As shown in the formula below:

[0119]

[0120] The specific method for calculating the behavior-content misalignment component of participating user nodes is as follows:

[0121] Based on the user-post edges on the participating user nodes, the participating user nodes In the event diagram All posts below The context representation vector is average-pooled to calculate the behavioral stance representation representing the comprehensive behavior of participating user nodes. As shown in the formula below:

[0122]

[0123] in, A collection of posts published by participating users. For set The first in Each post node For the first The context representation vector of each post node;

[0124] Based on news nodes Context representation vector Calculate the participating user nodes Behavioral-content dissonance component As shown in the formula below:

[0125]

[0126] Participating user nodes Social-content imbalance and behavioral-content dissonance component splicing, such as Figure 2 As shown, the participating user nodes are obtained. Multidimensional cognitive dissonance vector As shown in the formula below:

[0127]

[0128] in, For splicing operations;

[0129] S32. Calculate attention weights based on the imbalance gating unit (DGU) and refine the participating user nodes;

[0130] Multidimensional misalignment vector A gating signal with the same dimension as the participating user nodes is generated by transforming the signal through a neural network. As shown in the formula below:

[0131]

[0132] in, For the weights of the gating network, As a bias for the gated network, the Sigmoid function constrains the gate signal between (0,1) as a gate;

[0133] Using gating signals Context representation vector of participating user nodes Element-wise multiplication is performed to dynamically amplify or reduce information in its feature dimensions, simulating the psychological process of selective attention or ignoring of information caused by cognitive dissonance, resulting in gated augmented representations of participating user nodes. As shown in the formula below:

[0134]

[0135] in, This is an element-wise multiplication operation;

[0136] Multidimensional cognitive dissonance vector The module length serves as the participating user node The overall intensity of the imbalance is calculated using a linear transformation and the Softmax function for participating user nodes. The attention weight is calculated using the following formula:

[0137]

[0138] in, For event diagrams The set of all participating user nodes. For set The Middle One participating user node; and For learnable weights and biases; It is an L2 norm;

[0139] S33, Utilizing Attention Weights and gating enhancement representation of participating user nodes By performing a weighted summation on all participating user nodes, a graph-level representation vector representing complex psychological conflict information is obtained. The calculation formula is as follows:

[0140]

[0141] Representing vectors at the graph level The input is a two-layer multilayer perceptron (MLP) for final binary classification of true / fake news. The calculation formula is as follows:

[0142]

[0143] in, and These are the parameters of the first layer of a multilayer perceptron (MLP). and These are the parameters of the second layer of a multilayer perceptron (MLP). To predict graph-level representation vectors The probability distribution of fake news;

[0144] This embodiment upgrades cognitive dissonance into a multidimensional cognitive dissonance vector and designs a gating mechanism to dynamically adjust the representation of user information, such as... Figure 3 As shown, this allows for a more refined and profound simulation of the impact of users' psychological conflicts on information judgment.

[0145] S4. Construct a labeled dataset and jointly train and optimize the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network through end-to-end supervised learning.

[0146] S41. Construct a labeled dataset, which includes several event graphs and their corresponding true and false labels; divide the labeled dataset into training set, validation set and test set according to the proportions.

[0147] This embodiment uses a combination of the Weibo-16 and PHEME public datasets to construct an labeled dataset, containing 8500 news events with both true and false information. The labeled dataset is strictly divided into training, validation, and test sets in a 7:1.5:1.5 ratio.

[0148] Configure training parameters, including using the AdamW optimizer to handle complex network structures and potential gradient vanishing / exploding problems, and set the initial learning rate. And configured the weight decay coefficient as follows: Set the batch size to 16 to accommodate the memory usage of larger maps;

[0149] S42. Construct a cross-entropy loss function with label smoothing to prevent overfitting and overconfidence in the model, and set the label smoothing parameters. The cross-entropy loss function with label smoothing is 0.1. As shown in the formula below:

[0150]

[0151] in, For batch size, In this embodiment, the number of categories is... , , To predict the first Each sample is a category The probability, The smoothed label is shown in the following formula:

[0152]

[0153] in, For the first Each sample is a category The probability, Set the smoothing parameter to 0.1;

[0154] S43. Jointly train the relation graph convolutional network and the multidimensional cognitive dissonance gated attention network;

[0155] In this embodiment, joint training follows a standard and rigorous iterative optimization process to ensure that the model achieves optimal performance and avoids overfitting, including the following steps:

[0156] (1) Initialization: Before training begins, all learnable parameters of the relation graph convolutional network RGCN and the multidimensional cognitive dissonance gated attention network are randomly initialized using the Xavier uniform distribution; the fixed weights of the pre-trained Chinese language model and word vector model in S1 are loaded.

[0157] (2) Iterative training: Start the training loop and set the maximum number of epochs to 200. In each epoch, the model traverses the entire training set in batches. For each batch of data (16 news event graphs), perform a complete forward propagation, i.e., execute S2 to S3, and calculate the prediction results corresponding to each batch of data. ;

[0158] (3) Loss calculation and backpropagation: based on the loss function defined in S42 Calculate the loss value of the current batch of data; execute the backpropagation algorithm to calculate the gradients of all learnable parameters in the graph convolutional network and the multidimensional cognitive dissonance gated attention network;

[0159] (4) Parameter update: Call the AdamW optimizer to update all parameters in the relation graph convolutional network and the multidimensional cognitive dissonance gated attention network according to the calculated gradient; at the same time, the AdamW optimizer completes the traversal of all batches of data according to the preset learning rate and weight decay strategy, and a training round ends.

[0160] (5) Periodic validation and performance monitoring: After each training round, switch to evaluation mode (deactivating dropout layers etc.), and perform a complete prediction on an independent validation set based on the relational graph convolutional network and multidimensional cognitive dissonance gated attention network obtained in the current training round. Calculate and record the key performance indicators of the current training round on the validation set, including accuracy, precision, recall and F1 score. Use the F1 score as the core monitoring indicator.

[0161] (6) Convergence Judgment and Early Stopping Mechanism: The Early Stopping monitor is activated, and the patience counter upper limit is set to 20. The Early Stopping monitor tracks and records the highest historical F1 score (best_f1) on the validation set in each training round. If the F1 score on the validation set does not exceed the highest historical F1 score best_f1 in the current training round, the patience counter is incremented by one; otherwise, the patience counter is reset to 0. When the upper limit of the patience counter reaches 20, it indicates that the performance of the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network has not improved for 20 consecutive training rounds. It can be considered that the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network have converged, and the training process is terminated early.

[0162] (7) Best model saving: In the periodic validation in step (5), whenever the F1 score of the validation set exceeds the historical highest F1 score best_f1, not only is the value of the historical highest F1 score best_f1 updated, but all parameter states (state_dict) of the current relational graph convolutional network and multidimensional cognitive dissonance gated attention network are immediately saved as the best model file (best_model.pth); when the maximum number of training rounds is reached or the early stopping mechanism is triggered to end the training, the best model file best_model.pth obtained is the relational graph convolutional network and multidimensional cognitive dissonance gated attention network model that performed best on the validation set during the entire training process;

[0163] S5. Deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services for real-time detection of emerging news events;

[0164] S51. Package the trained best model file best_model.pth, the event graph generation module, the initial feature vector extraction module, the relational graph convolutional network, and the multidimensional cognitive dissonance gated attention network; use the FastAPI framework to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models as an online service;

[0165] In this embodiment, an asynchronous, high-performance HTTP API service is built based on the FastAPI framework. The HTTP API service exposes a POST endpoint named ` / detect`. The HTTP API service and all its Python dependencies (PyTorch) are packaged into a Docker image. The Docker image is deployed to a Kubernetes cluster, and horizontal Pod Autoscaling (HPA) is configured to dynamically adjust the number of service instances based on request load, ensuring low-latency response under high concurrency.

[0166] S52. When it is necessary to detect a newly emerging news event, in response to an external public opinion monitoring system, a real-time detection process is triggered, including:

[0167] When the public opinion monitoring system detects a news event whose spread speed exceeds the threshold (more than 50 new reposts and comments per minute) within 10 minutes, it triggers the real and fake news detection and obtains the core URL of the news event to be detected as a unique identifier.

[0168] The public opinion monitoring system immediately initiated an API call to the load balancer address of the deployed detection service cluster: sending an HTTP POST request with the target endpoint being / detect, and the request body being a JSON object in the format {"url": "http: / / news.example.com / article / 123"};

[0169] After receiving a request, a Pod in the detection service cluster retrieves multi-source heterogeneous social media data within a fixed 15-minute time window based on the passed core URL. It then specifically crawls all posts, user information, forwarding chains, and comment data related to the news to be detected and constructs an event graph of the news event to be detected based on the event graph generation module.

[0170] After the event graph is constructed, the service process loads the model weights from the best-trained model file (best_model.pth) into GPU memory. Then, it sequentially executes the forward propagation of the Relationship Graph Convolutional Network (RGCN) and the forward propagation of the Multidimensional Cognitive Dissonance Gated Attention Network (MCA), ultimately calculating the probability that the news event to be detected is fake news.

[0171] The detection service generates a JSON object containing defined fields and returns it as an HTTP response to the public opinion monitoring system. In this embodiment, the JSON object format is {"request_id": "...", "fake_probability": 0.98, "confidence": "high", "suspicious_users": [{"user_id": "...", "dissonance_score": 0.85}, ...], "timestamp": "..."}, where request_id is the ID of the news event to be detected, fake_probability is the predicted probability of fake news, confidence is the predicted confidence level, and suspicious_users is a list of suspicious users, including attention weights. The top 5 users and their dissonance intensity, where user_id is the ID of the participating user, dissonance_score is the user's cognitive dissonance intensity, and timestamp is the timestamp.

[0172] Example 2:

[0173] A fake news detection system based on knowledge graphs and cognitive dissonance, such as Figure 4 As shown, it includes a data layer, a presentation layer, a decision layer, and an application layer.

[0174] The data layer is used to acquire multi-source heterogeneous social media data, construct an event graph based on the acquired social media data, and extract the feature vector of each node in the event graph.

[0175] The data layer includes a data acquisition module, an event graph generation module, and an initial feature vector extraction module;

[0176] The data acquisition module is used to acquire multi-source heterogeneous social media data. The data acquisition module integrates a crawler unit based on the Scrapy framework. When it receives a detection request for the target news, it will crawl all post content, user information, forwarding chain structure and comment data related to the target news within a preset time window according to the input target news URL.

[0177] The event graph generation module is used to construct an event graph based on the collected data. The event graph generation module instantiates social media data into four types of nodes: news nodes, content entity nodes, participating user nodes, and post nodes, and establishes five types of edge relationships based on the association between data: news-entity edge, post-entity edge, user-post edge, post-post edge, and user-user edge.

[0178] The initial feature vector extraction module is used to extract and encode the initial feature vector for each node in the event graph. The initial feature vector extraction module uses a pre-trained Chinese language model to process news nodes and post nodes, extracts the initial feature vectors of news nodes and post nodes, uses a word vector model to convert content entity nodes into initial feature vectors of content entity nodes, and performs initial feature extraction on participating user nodes to obtain the initial feature vectors of participating user nodes.

[0179] The representation layer is used to extract the context representation vector of each node in the event graph, and the representation layer includes a relational graph convolutional network module;

[0180] The relational graph convolutional network module is used to extract the context representation vector of each node in the event graph. The relational graph convolutional network module includes a multi-layer relational graph convolutional network, where the first layer includes an independent projection matrix to project the initial features of nodes of different dimensions to a unified dimension. The other layers keep the dimensions unchanged. By aggregating the neighbor information of different types of nodes in the event graph and updating it through multiple layers, a context representation vector that integrates multi-level neighborhood structure and semantic information is generated for each node in the event graph.

[0181] The decision layer is used to generate graph-level representation vectors of the event graph and to classify the graph-level representation vectors as true or false news.

[0182] The decision-making layer includes a disorientation gating unit and a multilayer perceptron;

[0183] The dissonance gating unit is used to generate graph-level representation vectors of the event graph. First, the dissonance gating unit calculates the social-content dissonance component by performing attention-weighted averaging on the context representation vectors of the user's followed neighbor nodes. Second, it calculates the behavioral-content dissonance component by average pooling the context representation vectors of all relevant posts published by the user. Finally, the two components are concatenated to generate a multidimensional cognitive dissonance vector of the participating user nodes. The gating unit (DGU) generates a gating signal based on the multidimensional dissonance vector, and refines the context representation vectors of the participating user nodes using element-wise multiplication. Subsequently, based on the magnitude of the dissonance vector, the attention weight of each participating user node is calculated through linear transformation and the Softmax function. The calculated attention weights are then weighted and summed on the gating-enhanced user representation to generate a graph-level representation vector.

[0184] Multilayer perceptrons are used to classify news events as true or false. The graph-level representation vector of the event graph is input into a multilayer perceptron (MLP) with several layers, and the output is the probability distribution of whether the news event corresponding to the event graph is true or false news.

[0185] The application layer is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network model as an online service for real-time detection of emerging news events.

[0186] The application layer includes a service encapsulation module, a containerized deployment module, and a real-time detection and flow control module;

[0187] The service encapsulation module is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services using the FastAPI framework. In this embodiment, the service encapsulation module provides an online inference interface, builds an asynchronous HTTP API service based on the FastAPI framework, packages the trained best model (best_model.pth) with the processing logic of the data layer, representation layer and decision layer, and exposes a POST endpoint named " / detect".

[0188] The containerized deployment module is used to achieve high availability and elastic scaling of online services. The containerized deployment module packages API services and their dependencies into Docker images and deploys the images to the Kubernetes cluster. The containerized deployment module is also configured with a horizontal Pod autoscaling (HPA) policy, which can dynamically adjust the number of online service instances according to the request load.

[0189] The real-time detection flow control module coordinates the entire detection process. It comprises a triggering unit, a scheduling unit, and a feedback unit. The triggering unit interfaces with an external public opinion monitoring system. When the Apache Flink-based stream processing task detects that the speed of news dissemination exceeds a set threshold, it extracts the core URL and initiates a detection request. The scheduling unit receives a JSON request body containing the URL, calls the data layer for real-time data crawling and graph construction, and loads model weights into GPU memory for inference. The feedback unit encapsulates the detection results into a JSON object, which contains the predicted probability (fake_probability), confidence score, and a list of suspicious users ranked highest by attention weight (suspicious_users).

[0190] The system, through the collaborative work of the aforementioned layers, achieves a complete process from multi-source data acquisition to cognitive dissonance modeling, and then to real-time online detection, such as... Figure 5 As shown.

[0191] Example 3:

[0192] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0193] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the fake news detection method based on knowledge graphs and cognitive dissonance as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0194] The processor is used to execute all or part of the steps in the knowledge graph-based and cognitive dissonance-based fake news detection method described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0195] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the fake news detection method based on knowledge graph and cognitive dissonance described in the above embodiments.

[0196] Example 4:

[0197] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0198] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the fake news detection method based on knowledge graphs and cognitive dissonance described in the various embodiments of this application.

[0199] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) app stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0200] Example 5:

[0201] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned fake news detection method based on knowledge graphs and cognitive dissonance.

[0202] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0203] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0204] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A method for detecting fake news based on knowledge graphs and cognitive dissonance, characterized in that, Includes the following steps: Acquire multi-source heterogeneous social media data, construct an event graph based on the acquired social media data, and extract the initial feature vector of each node in the event graph; Based on the initial feature vector of each node in the event graph, a relational graph convolutional network is used to extract the context representation vector of each node in the event graph; A multidimensional cognitive dissonance gating attention network is constructed, including a dissonance gating unit and a multilayer perceptron. The dissonance gating unit is used to generate a graph-level representation vector of the event graph based on the context representation vector of each node in the event graph, and the multilayer perceptron is used to classify the graph-level representation vector into true and false news. A labeled dataset was constructed, and a relation graph convolutional network and a multidimensional cognitive dissonance gated attention network were jointly trained and optimized. The trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models were deployed as online services for real-time detection of emerging news events.

2. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 1, characterized in that, Multi-source heterogeneous social media data includes the target news event to be detected and related data, including: participating user data, post data, social relationship data, and dissemination relationship data; User participation data includes information on users who participate in posting, forwarding, and commenting on the target news event; post data includes all posts published by users related to the target news event, including original content, forwards, and comments; social relationship data includes the following relationships between users; and dissemination relationship data includes the forwarding or commenting relationships between posts. An event graph generation module is built, which is used to construct event graphs based on the acquired social media data. ,in, For a set of nodes, The set of edges and the set of nodes include various types of nodes, including news nodes. Content entity nodes Participating user nodes Post Nodes News Nodes This represents the core news content of the target news event; each event graph contains a unique news node; participating user nodes. This refers to users who participate in posting, forwarding, and commenting on the target news event, and these are considered participating user nodes. It includes basic features and category features. Basic features include account registration days, number of followers, number of following accounts, total number of Weibo posts, bio length, and number of custom tags; category features include authentication type and gender; post nodes. This refers to all posts published by participating users that are related to the target news event, including original content, reposts, and comments; content entity nodes. Represents named entities of predefined categories extracted from target news events and posts, including people's names, place names, organization names, and product names; edge set It includes various types of edges, including news-entity edges, post-entity edges, user-post edges, post-post edges, and user-user edges; news-entity edges connect news nodes. and content entity nodes Edges; Post-entity edges connect post nodes. and content entity nodes The edge; the user-post edge connects the participating user nodes. and the post node created by the participating user The edges are: post-post edges are edges connecting two post nodes that have a forwarding or commenting relationship; user-user edges are edges connecting two participating user nodes that have a following relationship. If the current participating user follows another participating user, then the user-user edge between the current participating user and the other participating user is a directed edge from the current participating user to the other participating user.

3. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 2, characterized in that, The initial feature vector of each node in the event graph is extracted using the following method: An initial feature vector extraction module is constructed to initialize feature vectors for each node in the event graph. For news nodes and post nodes, a pre-trained Chinese language model is used for encoding and initial feature vector extraction to obtain the news node's feature vector. initial feature vector and post nodes initial feature vector ; For content entity nodes Using a pre-trained large model as the word vector model, the content entity nodes are converted based on the word vector model. Dimensional content entity node initial feature vector ; For participating user nodes Initial feature vectors are extracted based on the basic and category features of participating users. The basic features are then Z-score normalized, and the category features are one-hot encoded. These encoded category features are then concatenated with the basic features to obtain the initial feature vectors of the participating user nodes. .

4. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 3, characterized in that, The context representation vector of each node in the event graph is extracted based on a relational graph convolutional network. The specific method is as follows: event diagram enter Layer-wise relational graph convolutional network for extracting event graphs any node in Context representation vector Event diagram any node In the relational graph convolutional network RGCN, the first Layer context representation vector As shown in the formula below: in, The set of edge types; For the first Types of nodes The set of neighboring nodes, For the set of neighbor nodes The Middle One neighboring node; For the first Type 1 Layer-specific learnable weight matrices; For nodes The transformation matrix; This is the normalization constant; To modify the activation function of the linear unit; For the first The neighboring node is at the _th ... The feature vector of the layer; For nodes In the The feature vector of the layer.

5. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 4, characterized in that, The graph-level representation vector of the event graph is generated based on the context representation vector of each node in the event graph. The specific method is as follows: For event graphs Each participating user node Calculate the participating user nodes respectively Social-content imbalance and behavioral-content dissonance component Generate participating user nodes Multidimensional cognitive dissonance vector ; Multidimensional misalignment vector A gating signal with the same dimension as the participating user nodes is generated by transforming the signal through a neural network. ; Using gating signals Context representation vector of participating user nodes Perform element-wise multiplication to obtain the gated augmented representation of the participating user nodes. ; Multidimensional cognitive dissonance vector The module length serves as the participating user node The overall intensity of the imbalance is calculated using a linear transformation and the Softmax function for participating user nodes. Attention weights; Using attention weights and gating enhancement representation of participating user nodes The graph-level representation vector is obtained by weighted summation of all participating user nodes. ; The specific method for classifying true and false news using graph-level representation vectors is as follows: Representing vectors at the graph level Inputting a two-layer multilayer perceptron (MLP) into a binary classification system for true and false news, the system obtains the predicted graph-level representation vector. This represents the probability distribution of fake news.

6. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 5, characterized in that, The labeled dataset includes several event graphs and corresponding real and fake labels; the labeled dataset is divided into training set, validation set and test set according to proportions; Joint training and optimization of a relation graph convolutional network and a multidimensional cognitive dissonance gated attention network were performed, including: Set training parameters, including initial learning rate, weight decay, batch size, and AdamW optimizer; Construct a cross-entropy loss function with label smoothing and set the label smoothing parameters. ; Before training begins, all learnable parameters of the relational graph convolutional network RGCN and the multidimensional cognitive dissonance gated attention network are randomly initialized using a Xavier uniform distribution; fixed weights of the pre-trained Chinese language model and word vector model are loaded. Start the training loop and set the maximum number of rounds; in each round, the model traverses the entire training set in batches. For each batch of data, perform a complete forward propagation and calculate the prediction result corresponding to each batch of data. Calculate the loss value of the current batch of data based on the labeled smoothed cross-entropy loss function; execute the backpropagation algorithm to calculate the gradients of all learnable parameters in the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network; The AdamW optimizer is invoked to update all parameters in the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network based on the calculated gradients. The AdamW optimizer then completes a training epoch by iterating through all batches of data according to the preset learning rate and weight decay strategy. After each training round, based on the relational graph convolutional network and multidimensional cognitive dissonance gated attention network obtained in the current training round, a complete prediction is performed on an independent validation set. The key performance indicators of the current training round on the validation set are calculated and recorded, including accuracy, precision, recall and F1 score. Based on the early stopping mechanism, an upper limit is set for the patience counter, which records the highest historical F1 score on the validation set in each training round. If the F1 score on the validation set does not exceed the highest historical F1 score in the current training round, the patience counter is incremented by one; otherwise, the patience counter is reset to 0. When the patience counter reaches the upper limit, it is determined that the relational graph convolutional network and the multidimensional cognitive dissonance gated attention network have converged, and training is terminated early. When the F1 score on the validation set exceeds the highest historical F1 score, update the highest historical F1 score and save all parameter states of the current relational graph convolutional network and multidimensional cognitive dissonance gated attention network as the best model file. When the maximum number of training rounds is reached or the early stopping mechanism is triggered to end the training, the best model file obtained is used as the relational graph convolutional network and multidimensional cognitive dissonance gated attention network model that performed best on the validation set during the entire training process.

7. The method for detecting fake news based on knowledge graphs and cognitive dissonance according to claim 6, characterized in that, The trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models are deployed as online services, specifically as follows: Package the trained best model file, event graph generation module, initial feature vector extraction module, relation graph convolutional network, and multidimensional cognitive dissonance gated attention network; use the FastAPI framework to deploy the trained relation graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services; When a new news event needs to be detected, a real-time detection process is triggered in response to an external public opinion monitoring system, including: When the public opinion monitoring system detects a news event that spreads faster than the threshold of new reposts and comments per minute within a preset time, it obtains the core URL of the news event as a unique identifier. The public opinion monitoring system immediately initiates an API call to the load balancer address of the deployed detection service cluster. After receiving the request, the detection service cluster obtains multi-source heterogeneous social media data based on the core URL within a fixed time window, and specifically crawls all posts, user information, forwarding chains and comment data related to the news to be detected. Based on the event graph generation module, it constructs an event graph of the news event to be detected. After the event graph is constructed, the service process loads the best model file and model weights that have been trained, and sequentially executes the forward propagation of the relation graph convolutional network and the forward propagation of the multidimensional cognitive dissonance gated attention network to calculate the probability value of the news event to be detected as fake news, obtain the detection result and return it to the public opinion monitoring system.

8. A fake news detection system based on knowledge graphs and cognitive dissonance, which performs fake news detection based on the method described in claim 1, characterized in that, It includes the data layer, presentation layer, decision layer, and application layer; The data layer is used to acquire multi-source heterogeneous social media data, construct an event graph based on the acquired social media data, and extract the feature vector of each node in the event graph. The representation layer is used to extract the context representation vector of each node in the event graph; The decision layer is used to generate graph-level representation vectors of the event graph and to classify the graph-level representation vectors as true or false news. The application layer is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network model as an online service for real-time detection of emerging news events.

9. A fake news detection system based on knowledge graphs and cognitive dissonance according to claim 8, characterized in that, The data layer includes a data acquisition module, an event graph generation module, and an initial feature vector extraction module; the data acquisition module is used to acquire multi-source heterogeneous social media data; the event graph generation module is used to construct an event graph based on the acquired data; the initial feature vector extraction module is used to extract and encode an initial feature vector for each node in the event graph. The representation layer includes a relational graph convolutional network module, which comprises a multi-layer relational graph convolutional network. The first layer of the relational graph convolutional network includes an independent projection matrix, which is used to uniformly project the initial features of nodes of different dimensions to a unified dimension. The other layers, except for the first layer, keep the dimensions unchanged and are used to aggregate the neighbor information of different types of nodes in the event graph, and generate a context representation vector for each node in the event graph that integrates multi-level neighborhood structure and semantic information. The decision layer includes an imbalance gating unit and a multilayer perceptron; the imbalance gating unit is used to generate a graph-level representation vector of the event graph; the multilayer perceptron is used to classify the event graph as true or false news, inputting the graph-level representation vector of the event graph into a multilayer perceptron with several layers, and outputting the probability distribution of the news event corresponding to the event graph as false news; The application layer includes a service encapsulation module, a containerized deployment module, and a real-time detection and flow control module; the service encapsulation module is used to deploy the trained relational graph convolutional network and multidimensional cognitive dissonance gated attention network models as online services. The containerized deployment module is used to deploy online services to the detection service cluster; the containerized deployment module is also used to dynamically adjust the number of online service instances; The real-time detection flow control module is used to coordinate the entire detection task; the real-time detection flow control module includes a triggering unit, a scheduling unit and a feedback unit; the triggering unit is used to connect to an external public opinion monitoring system, extract the core URL of the news event to be detected and initiate a detection request; The scheduling unit is used to acquire social media data of the news event to be detected, construct an event graph of the news event to be detected, load the model weights of the best model file, calculate the probability value of the news event to be detected as fake news, and obtain the detection result. The feedback unit is used to return the detection results to an external public opinion monitoring system.

Images