A multi-domain fake news intelligent detection method
By building a dual-branch domain adversarial network model and real-time monitoring system, the problems of time-consuming and labor-intensive training of single-field models and low manual detection efficiency are solved, and the efficiency and comprehensiveness of intelligent detection of fake news in multiple fields are achieved, and the network ecological environment is enhanced Governance capability.
Patent Information
- Application Number
- CN202411298499.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-09-18
AI Technical Summary
Among the existing intelligent detection methods of fake news, single-field model training requires a large amount of labeled data in professional fields, which is time-consuming and labor-intensive, and relies on manual detection efficiency, making it difficult to cover the comprehensive coverage.
Using intelligent detection methods of multi-domain fake news, data preprocessing and building a dual-branch domain adversarial network model by obtaining historical news data, including domain independent text feature extraction, domain shared text feature extraction and feature fusion, combined with backpropagation algorithm training and use real-time monitoring and early warning systems to identify potential fake news.
It solves the problem of data sparseness in single fields, improves detection performance, reduces manual intervention, improves the efficiency of fake news detection, and enhances the ability to manage the network ecological environment.
Smart Images

Figure CN118820957B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fake news detection, and in particular relates to a multi-field fake news intelligent detection method. Background Art
[0002] When the news industry entered the Internet era, fake news also began to have strong characteristics of the Internet era, seriously damaging the credibility of news media. At present, the news data and public comments circulating on social network media platforms in various fields (such as medical, financial, management, business, etc.) are increasing day by day, and fake news information is difficult to identify. When faced with huge news data and the rapid spread of fake news, it is difficult to timely and effectively rectify the network ecological environment by relying solely on regular manual inspections by industry regulatory departments. Therefore, it is necessary to conduct intelligent detection of fake news, real-time warning, and statistical analysis of fake news details.
[0003] The intelligent detection method for fake news is the key to managing fake news and the prerequisite for realizing intelligent monitoring of fake news. There are two problems with the current intelligent detection method for fake news: 1) In the design of the intelligent detection algorithm for fake news, the training of a single-domain model requires a large amount of news data labeled with authenticity, and the authenticity of news usually requires experts in a professional field, which is time-consuming and laborious. At the same time, the amount of true news and fake news data in various fields varies greatly. 2) The application of the intelligent detection method for fake news is still a manual detection method, which requires manual input of news content in the background, and then obtains the result of whether it is abnormal. In this way, a large amount of manpower is still consumed, and the efficiency is low, and it is difficult to cover the whole. For this reason, the present invention proposes a multi-domain intelligent detection method for fake news. Summary of the invention
[0004] The purpose of the present invention is to provide a multi-domain fake news intelligent detection method, aiming to solve the problems raised in the above background technology.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A multi-domain fake news intelligent detection method, comprising the following steps:
[0007] Step 1: Obtain historical news data and split it into training data, verification data and test data;
[0008] Step 2: Data preprocessing: The original text is converted into a digital sequence that can be processed by the BERT model;
[0009] Step 3, constructing a dual-branch domain adversarial network model, including: step 31, domain-independent text feature extraction; step 32, domain-shared text feature extraction; step 33, feature fusion;
[0010] Step 4, constructing the overall objective function of the model, including: step 41, calculating the domain classification loss; step 42, calculating the fake news classification task loss; step 43, calculating the auxiliary classification constraint loss; step 44, calculating the overall model loss;
[0011] Step 5: Model training: Use the back propagation algorithm to train the parameters of the neural network, and calculate the gradient of the overall loss function of the model with respect to the network parameters through step 44 to adjust the network weights;
[0012] Step 6: Intelligently identify fake news to be identified, and conduct real-time monitoring and early warning.
[0013] Furthermore, the specific steps of step 2 are as follows:
[0014] Step 21, initialize the word segmenter: select a suitable BERT pre-trained model to initialize the word segmenter;
[0015] Step 22: Input text and segment it into words: Input the text to be processed into the word segmenter, which splits the text into a series of subwords or word blocks, namely tokens. Tokens are the best word units learned by the BERT model during pre-training.
[0016] Step 23: Convert to BERT input format: Convert tokens to numeric word embedding indices and add special markers to mark the beginning and end of sentences, as well as the separation between different sentences.
[0017] Step 24: Output the word segmentation results: Finally, output the word segmentation results to obtain a token sequence.
[0018] Furthermore, the specific steps of step 3 are as follows:
[0019] Step 31, domain-independent text feature extraction: The pre-trained BERT encoder is combined with the MLP-based mapping function as a domain-independent encoder, wherein the parameters of the BERT encoder do not participate in the update process of the network parameters, and the function of the mapping function is updated in real time; the obtained text token sequence is input into the BERT encoder, and the output result of the BERT encoder is input into the mapping function to obtain the domain-independent coding features;
[0020] ;
[0021] in Represents domain-independent features, represents the pre-trained BERT encoder, represents the frozen parameters of the BERT encoder, represents the mapping function, Represents the network parameters that need to be updated in the mapping function, A token sequence representing news text;
[0022] Step 32, domain-shared text feature extraction: the pre-trained BERT encoder is used as the domain-shared encoder, and the parameters of the BERT encoder participate in the network parameter update process; the obtained text token sequence is input into the domain-shared encoder to obtain domain-independent coding features;
[0023] ;
[0024] in Represents domain-shared features, represents a domain-shared encoder, represents the network parameters that need to be updated in the BERT encoder, A token sequence representing news text;
[0025] Step 33: Feature fusion: Fuse domain-shared features with domain-specific features, and finally fuse features The calculation method is as follows:
[0026] ;
[0027] ;
[0028] in represents element-wise multiplication, represents the addition of the principal elements, represents the weight of domain-independent features, represents the weight of domain-shared features, Indicates the network parameters that need to be updated by the domain sharing module. represents the activation function, The output of the function lies in the interval (0, 1).
[0029] Furthermore, the specific steps of step 4 are as follows:
[0030] Step 41. Calculate domain classification loss: Construct a multi-domain classifier. For D domains, the discriminator is a D classifier and outputs a D-dimensional vector ;vector Each dimension in Corresponding to the predicted probability that the input news text belongs to the kth domain, the domain shared features Domain-independent features Input multiple domain classifiers respectively to obtain prediction results; use classification cross entropy loss to calculate the loss of prediction results of domain shared features, prediction results of domain independent features and true domain labels; the loss calculation method is as follows:
[0031] ;
[0032] ;
[0033] in represents the real domain label, represents the domain classification loss, represents the activation function, represents a domain classifier;
[0034] Step 42. Calculate the loss of fake news classification task: Build a fake news task classifier , input the fusion features into the fake news task classifier to obtain the fake news prediction probability ; The classifier is composed of an MLP network, and the results of the MLP network are The function is transformed into a real number between 0 and 1, which represents the predicted probability of fake news;
[0035] The binary classification cross entropy loss function is used to calculate the loss between the predicted label and the true label. The calculation paradigm is as follows:
[0036] ;
[0037] ;
[0038] in represents the parameters to be updated in the fake news task classifier MLP network, y represents the authenticity label of the real news, represents the loss of fake news classification task;
[0039] Step 43, calculate the auxiliary classification constraint loss: for the feature and Design task-oriented classification constraints, where each branch is treated as a single-task network, where the encoder is coupled to the fake news task classifier. Interaction; Features and Directly input into the fake news task classifier to obtain the predicted probability of fake news in the current task and , used for fake news detection; and The auxiliary constraints are expressed as follows:
[0040] ;
[0041] ;
[0042] in, Auxiliary classification constraints that represent domain-independent features; Auxiliary classification constraints that represent domain-shared features;
[0043] Step 44, calculate the overall loss of the model: The overall loss of the model includes the fake news classification task loss, the auxiliary classification constraints for domain-independent features and domain-shared features, and the domain classification loss, and is calculated as follows:
[0044] ;
[0045] in, Domain classification loss to represent domain-independent features Domain classification loss that represents domain-shared features.
[0046] Furthermore, the specific process of step 5 is as follows:
[0047] For the domain-shared encoder, a domain-adversarial strategy is adopted to update the domain-shared encoder. The gradient reversal network layer is inserted between the domain-shared encoder and the domain classifier to achieve the process of gradient sign reversal during the back-propagation process. Convert to , where λ represents a hyperparameter.
[0048] Furthermore, the specific steps of step 6 are as follows:
[0049] Step 61: Design real-time monitoring indicators for fake news, including real-time quantity, dissemination dynamics, and real-time traffic;
[0050] Step 62: Real-time monitoring of the database, including:
[0051] Step 621: Use change data capture technology to capture updates to the news database in real time, and set database triggers to automatically collect relevant indicator data when news is released;
[0052] Step 622: Use a time series database or a search engine to store and index real-time data;
[0053] Step 623: Use stream processing technology to process and analyze data streams in real time, and analyze data in real time based on a network model to identify potential fake news;
[0054] Step 63: early warning notification; including:
[0055] Step 631: When potential fake news is detected, the monitoring indicators are calculated in real time;
[0056] Step 632: Set warning rules according to the monitoring indicators, and trigger an alarm when the indicators exceed the normal range;
[0057] Step 633: When potential fake news is detected, relevant personnel are notified via email, SMS or application push.
[0058] Compared with the prior art, the present invention has the following beneficial effects:
[0059] The present invention is actually applied to the detection of fake news content on network platforms in various fields and has practical significance, providing technical support for strengthening the governance of the network ecological environment. The method provided by the present invention solves the problem that the model cannot be trained due to sparse data in a single field, and at the same time improves the detection performance of the model. Under the action of the monitoring system, the efficiency of fake news detection can be improved, manual pressure can be relieved, information errors can be avoided, and help can be provided for further governance of fake news. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a flow chart of the present invention.
[0061] Figure 2 It is a structural diagram of the network model in the present invention. DETAILED DESCRIPTION
[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0063] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.
[0064] like Figure 1 As shown, a multi-domain fake news intelligent detection method provided by an embodiment of the present invention includes the following steps:
[0065] Step 1: Obtain historical news data and split it into training data, verification data and test data;
[0066] Step 2: Data preprocessing: The original text is converted into a digital sequence that can be processed by the BERT model. The specific steps of step 2 are as follows:
[0067] Step 21. Initialize the word segmenter: Select a suitable BERT pre-trained model to initialize the word segmenter. For Chinese text, a BERT model pre-trained for Chinese corpus is usually used.
[0068] Step 22: Input text and tokenize: Input the text to be processed into the tokenizer, which splits the text into a series of subwords or tokens. Tokens are the best word units learned by the BERT model during pre-training.
[0069] Step 23: Convert to BERT input format: The subwords after tokenization need to be converted into a format that the BERT model can understand. This includes converting tokens into digital word embedding indexes and adding special markers to mark the beginning and end of sentences, as well as the separation between different sentences.
[0070] Step 24: Output the word segmentation results: Finally, the word segmentation results are output for use in subsequent natural language processing tasks to obtain a token sequence.
[0071] Step 3: Construct a dual-branch domain adversarial network model. The overall structure of the network model is as follows: Figure 2 As shown, the specific steps of step 3 are as follows:
[0072] Step 31, domain-independent text feature extraction: The pre-trained BERT encoder is combined with the MLP-based mapping function as a domain-independent encoder, wherein the parameters of the BERT encoder do not participate in the update process of the network parameters, and the function of the mapping function is updated in real time; the obtained text token sequence is input into the BERT encoder, and the output result of the BERT encoder is input into the mapping function to obtain the domain-independent coding features;
[0073] ;
[0074] in Represents domain-independent features, represents the pre-trained BERT encoder, represents the frozen parameters of the BERT encoder, represents the mapping function, Represents the network parameters that need to be updated in the mapping function, A token sequence representing news text;
[0075] Step 32, domain-shared text feature extraction: the pre-trained BERT encoder is used as the domain-shared encoder, and the parameters of the BERT encoder participate in the network parameter update process; the obtained text token sequence is input into the domain-shared encoder to obtain domain-independent coding features;
[0076] ;
[0077] in Represents domain-shared features, represents a domain-shared encoder, represents the network parameters that need to be updated in the BERT encoder, A token sequence representing news text;
[0078] Step 33, feature fusion: In multi-domain fake news classification, domain-shared features and domain-independent features may have different contributions. Simple fusion operations, such as element-by-element addition or concatenation, cannot effectively emphasize key information. In order to properly adjust the importance of each feature group, it is necessary to assign different weights to each feature group according to their importance. To this end, the present invention fuses domain-shared features with domain-specific features, allowing adaptive adjustment of weights between the two branches. This method takes advantage of the complementary advantages of the two feature types and better adapts to the fake news classification task. The final fused feature The calculation method is as follows:
[0079] ;
[0080] ;
[0081] in Represents element-by-element multiplication. In element-by-element multiplication, the elements at corresponding positions of two matrices or vectors of the same shape are multiplied to obtain a new matrix or vector. It represents principal element addition. In element-by-element addition, the elements at corresponding positions of two matrices or vectors of the same shape are added to obtain a new matrix or vector. represents the weight of domain-independent features, represents the weight of domain-shared features, Indicates the network parameters that need to be updated by the domain sharing module. represents the activation function, The output of the function is in the interval (0, 1). and The sum of the corresponding dimension elements is 1, indicating the complementarity of domain-independent features and domain-shared features.
[0082] Step 4: construct the overall objective function of the model. The specific steps of step 4 are as follows:
[0083] Step 41. Calculate domain classification loss: Construct a multi-domain classifier. For D domains, the discriminator is a D classifier and outputs a D-dimensional vector ;vector Each dimension in Corresponding to the predicted probability that the input news text belongs to the kth domain, the domain shared features Domain-independent features Input multiple domain classifiers respectively to obtain prediction results; use classification cross entropy loss to calculate the loss of prediction results of domain shared features, prediction results of domain independent features and true domain labels; the loss calculation method is as follows:
[0084] ;
[0085] ;
[0086] in represents the real domain label, represents the domain classification loss, represents the activation function, represents a domain classifier;
[0087] Step 42. Calculate the loss of fake news classification task: Build a fake news task classifier , input the fusion features into the fake news task classifier to obtain the fake news prediction probability ; The classifier is composed of an MLP network, and the results of the MLP network are The function is transformed into a real number between 0 and 1, which represents the predicted probability of fake news;
[0088] The binary classification cross entropy loss function is used to calculate the loss between the predicted label and the true label. The calculation paradigm is as follows:
[0089] ;
[0090] ;
[0091] in represents the parameters to be updated in the fake news task classifier MLP network, y represents the authenticity label of the real news, represents the loss of fake news classification task;
[0092] Step 43. Calculate auxiliary classification constraint loss: The ultimate goal of this method is to classify the authenticity of news. Although the domain-shared encoder effectively captures the features of the shared domain using the domain adversarial strategy, and the domain-independent encoder extracts domain-independent features, these two features do not necessarily decompose these features into features specific to the fake news classification task. Specifically, the multi-domain alignment achieved through the domain adversarial mechanism does not explicitly target optimizing downstream task classification, which may result in suboptimal trained domain classifiers. Therefore, in order to ensure that the two encoders can effectively identify task-discriminative features, the features and Design task-oriented classification constraints, where each branch is treated as a single-task network, where the encoder is coupled to the fake news task classifier. Interaction; Features and Directly input into the fake news task classifier to obtain the predicted probability of fake news in the current task and , used for fake news detection; and The auxiliary constraints are expressed as follows:
[0093] ;
[0094] ;
[0095] in, Auxiliary classification constraints that represent domain-independent features; Auxiliary classification constraints that represent domain-shared features;
[0096] Step 44, calculate the overall loss of the model: The overall loss of the model includes the fake news classification task loss, the auxiliary classification constraints for domain-independent features and domain-shared features, and the domain classification loss, and is calculated as follows:
[0097] ;
[0098] in, Domain classification loss to represent domain-independent features Domain classification loss that represents domain-shared features.
[0099] Step 5, model training: Use the back-propagation algorithm to train the parameters of the neural network, and adjust the network weights by calculating the gradient of the model's overall loss function with respect to the network parameters through step 44. For each weight in the network, calculate the partial derivative of the loss function with respect to the weight. This requires repeated application of the chain rule, tracing back from the output layer all the way to the input layer. Once the gradient of each weight is calculated, use gradient descent or other optimization algorithms to update the network weights. The goal of weight updating is to reduce the value of the loss function and thus improve the performance of the network. For the domain-shared encoder, a domain adversarial strategy update is adopted. The process of gradient sign reversal during the back-propagation process is achieved by inserting a gradient reversal network layer between the domain-shared encoder and the domain classifier, that is, the original gradient Convert to , where λ represents a hyperparameter.
[0100] Step 6: Intelligently identify fake news to be identified, monitor in real time, and issue early warnings; timely identify and warn potential fake news to reduce its spread and impact. The specific steps of step 6 are as follows:
[0101] Step 61: Design real-time monitoring indicators for fake news, including:
[0102] 1) Real-time quantity: the number of fake news;
[0103] 2) Spread dynamics: Track the speed and breadth of news sharing in real time through social media APIs;
[0104] 3) Real-time traffic analysis: Use website analysis tools to monitor in real time and count user visits and dwell time.
[0105] Step 62: Real-time monitoring of the database, including:
[0106] 1) Real-time data capture: Use change data capture (CDC) technology to capture updates to the news database in real time. Set up database triggers to automatically collect relevant indicator data when news is published;
[0107] 2) Data storage and indexing: Use a time-series database or search engine (such as Elasticsearch) to store and index real-time data, ensuring that the data storage solution can quickly retrieve and analyze large amounts of data;
[0108] 3) Use stream processing technology (Apache Kafka) to process and analyze data streams in real time. Implement real-time analysis based on fake news intelligent detection models to identify potential fake news.
[0109] Step 63: early warning notification; including:
[0110] 1) When potential fake news is detected, monitoring indicators are calculated in real time;
[0111] 2) Warning rule setting: Set warning rules based on monitoring indicators, and trigger alarms when indicators exceed the normal range. Allow dynamic and continuous adjustment of warning rules;
[0112] 3) Design a notification system to notify relevant personnel via email, SMS or app push when potential fake news is detected.
[0113] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.
Claims
1. A multi-domain fake news intelligent detection method, characterized in that: The following steps are involved: Step 1: Obtain historical news data and split it into training data, verification data and test data; Step 2: Data preprocessing: The original text is converted into a digital sequence that can be processed by the BERT model; Step 3, constructing a dual-branch domain adversarial network model, including: step 31, domain-independent text feature extraction; step 32, domain-shared text feature extraction; step 33, feature fusion; Step 4, constructing the overall objective function of the model, including: step 41, calculating the domain classification loss; step 42, calculating the fake news classification task loss; step 43, calculating the auxiliary classification constraint loss; step 44, calculating the overall model loss; Step 5: Model training: Use the back propagation algorithm to train the parameters of the neural network, and calculate the gradient of the overall loss function of the model with respect to the network parameters through step 44 to adjust the network weights; Step 6: Intelligently identify fake news to be identified, and conduct real-time monitoring and early warning.
2. The multi-domain fake news intelligent detection method according to claim 1 is characterized in that: The specific steps of step 2 are as follows: Step 21, initialize the word segmenter: select a suitable BERT pre-trained model to initialize the word segmenter; Step 22: Input text and segment it into words: Input the text to be processed into the word segmenter, which splits the text into a series of subwords or word blocks, namely tokens. Tokens are the best word units learned by the BERT model during pre-training. Step 23: Convert to BERT input format: Convert tokens to numeric word embedding indices and add special markers to mark the beginning and end of sentences, as well as the separation between different sentences. Step 24: Output the word segmentation results: Finally, output the word segmentation results to obtain a token sequence.
3. The multi-domain fake news intelligent detection method according to claim 2 is characterized in that: The specific steps of step 3 are as follows: Step 31, domain-independent text feature extraction: The pre-trained BERT encoder is combined with the MLP-based mapping function as a domain-independent encoder, wherein the parameters of the BERT encoder do not participate in the update process of the network parameters, and the function of the mapping function is updated in real time; the obtained text token sequence is input into the BERT encoder, and the output result of the BERT encoder is input into the mapping function to obtain the domain-independent features; ; in Represents domain-independent features, represents the pre-trained BERT encoder, represents the frozen parameters of the BERT encoder, represents the mapping function, Represents the network parameters that need to be updated in the mapping function, A token sequence representing news text; Step 32, domain-shared text feature extraction: the pre-trained BERT encoder is used as the domain-shared encoder, and the parameters of the BERT encoder participate in the network parameter update process; the obtained text token sequence is input into the domain-shared encoder to obtain the domain-shared features; ; in Represents domain-shared features, represents a domain-shared encoder, represents the network parameters that need to be updated in the BERT encoder, A token sequence representing news text; Step 33: Feature fusion: Fuse domain-shared features with domain-independent features, and finally fuse features The calculation method is as follows: ; ; in represents element-wise multiplication, represents the addition of the principal elements, represents the weight of domain-independent features, represents the weight of domain-shared features, Indicates the network parameters that need to be updated by the domain sharing module. represents the activation function, The output of the function lies in the interval (0, 1).
4. The multi-domain fake news intelligent detection method according to claim 3 is characterized in that: The specific steps of step 4 are as follows: Step 41, calculate the domain classification loss: build a multi-domain classifier for D domain, the discriminator is a D The classifier outputs a D Dimension vector ;vector Each dimension in Corresponding to the predicted probability that the input news text belongs to the kth domain, the domain shared features Domain-independent features Input multiple domain classifiers respectively to obtain prediction results; use classification cross entropy loss to calculate the loss of prediction results of domain shared features, prediction results of domain independent features and true domain labels; the loss calculation method is as follows: ; ; in represents the real domain label, represents the domain classification loss, represents the activation function, represents a domain classifier; Step 42. Calculate the loss of fake news classification task: Build a fake news task classifier , input the fusion features into the fake news task classifier to obtain the fake news prediction probability ; The classifier is composed of an MLP network, and the results of the MLP network are The function is transformed into a real number between 0 and 1, which represents the predicted probability of fake news; The binary classification cross entropy loss function is used to calculate the loss between the predicted label and the true label. The calculation paradigm is as follows: ; ; in represents the parameters that need to be updated in the fake news task classifier MLP network, y Indicates the authenticity label of real news, represents the loss of fake news classification task; Step 43, calculate the auxiliary classification constraint loss: for the feature and Design task-oriented classification constraints, where each branch is treated as a single-task network, where the encoder is coupled to the fake news task classifier. Interaction; Features and Directly input into the fake news task classifier to obtain the predicted probability of fake news in the current task and , used for fake news detection; and The auxiliary constraints are expressed as follows: ; ; in, Auxiliary classification constraints that represent domain-independent features; Auxiliary classification constraints that represent domain-shared features; Step 44, calculate the overall model loss: The overall model loss includes the fake news classification task loss, the auxiliary classification constraints for domain-independent features and domain-shared features, and the domain classification loss, and is calculated as follows: ; in, Domain classification loss that represents domain-independent features, Domain classification loss that represents domain-shared features.
5. The multi-domain fake news intelligent detection method according to claim 4 is characterized in that: The specific process of step 5 is as follows: For the domain-shared encoder, a domain-adversarial strategy is adopted to update the domain-shared encoder. The gradient reversal network layer is inserted between the domain-shared encoder and the domain classifier to achieve the process of gradient sign reversal during the back-propagation process. Convert to ,in λ Represents a hyperparameter.
6. The multi-domain fake news intelligent detection method according to claim 1 is characterized in that: The specific steps of step 6 are as follows: Step 61: Design real-time monitoring indicators for fake news, including real-time quantity, dissemination dynamics, and real-time traffic; Step 62: Real-time database monitoring, including: Step 621: Use change data capture technology to capture updates to the news database in real time, and set database triggers to automatically collect relevant indicator data when news is released; Step 622: Use a time series database or a search engine to store and index real-time data; Step 623: Use stream processing technology to process and analyze data streams in real time, and analyze data in real time based on a network model to identify potential fake news; Step 63: early warning notification; including: Step 631: When potential fake news is detected, the monitoring indicators are calculated in real time; Step 632: Set warning rules according to the monitoring indicators, and trigger an alarm when the indicators exceed the normal range; Step 633: When potential fake news is detected, relevant personnel are notified via email, SMS or application push.
