A Deep Learning-Based Intelligent Recognition Method for the Credibility of News Information

Abstract

This invention relates to the field of artificial intelligence technology and discloses a deep learning-based intelligent recognition method for news information credibility. The method collects news data from multiple channels, performs text cleaning and structuring, and constructs a multimodal feature model that integrates text semantics, media and account performance, propagation graph structure, and the temporal evolution of user interactions. Credibility inference is then achieved through a deep fusion model. During training, an expectation-constrained performance optimization strategy is introduced, combining fixed random projection and Worst-M memory mechanisms to achieve efficient search and stable convergence in the high-dimensional parameter space. The system classifies risk levels based on credibility scores and links with the content management system to execute prompts or interceptions, forming a closed-loop review process. This invention significantly improves the accuracy, stability, and generalization ability of news credibility recognition.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for intelligent recognition of the credibility of news information based on deep learning. Background Technology

[0002] With the rapid development of news aggregation platforms, social media platforms, and content distribution systems, online news content sources have diversified, dissemination chains have become multi-layered, and interactive behaviors have become more complex. Fake news and misleading information are exhibiting more covert, structured, and contextualized characteristics. Existing intelligent news review technologies are mostly based on deep learning classification models using single-modal text features or fraud detection methods based on knowledge graphs and statistical behavior models. They primarily rely on text semantic similarity judgment, keyword conflict detection, social dissemination feature extraction, or credibility inference based on historical behavior modeling. However, these technologies generally suffer from the following shortcomings: First, existing models struggle to simultaneously integrate multi-dimensional signals such as text semantics, media historical behavior, dissemination graph structure, and user interaction dynamics, lacking a unified representation space, resulting in insufficient robustness and generalization ability for credibility identification in complex news scenarios. Second, existing intelligent models often employ fixed parameter configurations and traditional hyperparameter tuning methods, making efficient exploration and convergence in high-dimensional parameter spaces difficult. The training process is prone to getting stuck in local optima, leading to significant performance fluctuations. Third, existing systems generally lack credibility level interpretation mechanisms and business closed-loop linkage capabilities, making it difficult to support platform-level content security governance needs. Therefore, there is an urgent need for a deep learning-based news credibility recognition method oriented towards multimodal feature fusion. This method can enhance the recognition capability in complex scenarios by modeling the propagation structure and temporal evolution features, and improve the model training efficiency and robustness by combining expected constrained performance optimization strategies, thereby achieving accurate recognition of news credibility and intelligent risk management. Summary of the Invention

[0003] This invention proposes a deep learning-based intelligent recognition method for news information credibility. Addressing the limitations of existing intelligent review technologies in multimodal feature fusion, handling of complex propagation structures, and weak model parameter optimization capabilities, it constructs a unified credibility recognition framework that integrates textual semantics, media credibility, propagation graph structure, and the temporal evolution of interactive behavior. This method jointly models the news propagation path and user behavior dynamics using graph neural networks and temporal convolutional networks, achieving deep expression and correlation representation of complex propagation patterns and multidimensional credibility signals. Through a multimodal feature fusion sub-network, it achieves high consistency alignment of semantic, structural, and behavioral features and credibility inference, significantly improving the robustness and generalization performance of recognition in cross-platform and cross-event news scenarios. Furthermore, this invention introduces an expectation-constrained performance optimization strategy for the first time in the field of news credibility recognition. Combining fixed random projection, Worst-m memory mechanism, and adaptive receptive region adjustment mechanism, it achieves efficient search and stable convergence in the high-dimensional model parameter space, effectively avoiding the performance oscillations and local optimum stagnation problems that traditional deep learning models encounter in complex optimization spaces. This invention constructs an interpretable risk level classification and review linkage control mechanism at the credibility output end, enabling credibility identification results to directly drive the execution of content review strategies, forming an automated risk handling closed loop, and significantly improving the accuracy, real-time performance, and business availability of Internet news content governance.

[0004] This invention provides a deep learning-based intelligent method for recognizing the credibility of news information. The method is executed by a credibility recognition server deployed in a data center. The credibility recognition server includes a processor, memory, and a network interface, and is connected to the Internet, user terminals, and a content management system server via the network interface. The method includes the following steps:

[0005] Step S1: Obtain original news data from multiple sources;

[0006] Step S2: The processor retrieves the preprocessing program from memory to preprocess the multi-source raw news data and obtain structured preprocessed news samples;

[0007] Step S3: The processor of the credibility recognition server takes the structured preprocessed news sample as input, extracts text semantic features, media and account credibility features, propagation structure features and temporal evolution features respectively, and performs unified encoding and dimension alignment to obtain multimodal feature data;

[0008] Step S4: Using multimodal feature data as input, train and validate the credibility recognition model stored in memory. During the training process, the expected constrained performance optimization strategy is used to iteratively optimize the parameter configuration of the credibility recognition model to determine the optimal combination of model parameters, thus obtaining the trained credibility recognition model. The credibility score of news information is output through the trained credibility recognition model. The credibility recognition model includes a text encoding subnetwork, a media and account feature subnetwork, a propagation graph embedding subnetwork, a temporal convolution subnetwork, a feature fusion subnetwork, and a classification subnetwork. The construction process of the expected constrained performance optimization strategy is as follows: Based on the ECP strategy, a fixed random projection mapping, Worst-m memory mechanism, and an adaptive receptive region adjustment mechanism are introduced to optimize the search efficiency, global exploration capability, and convergence stability of the ECP strategy, thus constructing the expected constrained performance optimization strategy.

[0009] Step S5: A set of preset thresholds can be used. Based on the credibility score of the news information and the threshold set, the processor divides the news information into multiple credibility levels, including high credibility, suspicious credibility, and low credibility. When the credibility level is low credibility, a risk interception command is sent to the content management system server to prevent the news from being automatically published and pushes the news to the reviewer's work terminal for manual review. When the credibility level is suspicious, a "risk warning" label is added when the news is displayed on the user's terminal, and it is added to the key monitoring queue. When the credibility level is high credibility, the content management system server is allowed to directly publish the news, and the credibility result is archived in the database for subsequent training and traceability.

[0010] Furthermore, the process of extracting text semantic features, media and account credibility features, dissemination structure features, and temporal evolution features respectively includes the following steps:

[0011] Step S31: Text semantic feature extraction: The processor calls the pre-trained Transformer text encoding model stored in memory to segment and encode the news headlines and body text in the structured preprocessed news samples, generating word-level and sentence-level semantic embedding vectors to form text semantic features;

[0012] Step S32: Media and Account Credibility Feature Extraction: Based on the number of times news from the same media source was manually labeled as fake news, the number of times it was deleted by the platform, the account registration duration and authentication information in the structured preprocessed news samples, the processor performs statistical summarization and scale unification processing to generate media and account credibility features.

[0013] Step S33: Propagation structure feature extraction: The processor constructs a news propagation graph based on the social media comments, reposts and citations in the structured preprocessed news samples. User accounts and topic tags are used as graph nodes, and reposts, comments and @ interactions are used as graph edges. The graph neural network is run on the credibility recognition server to perform multi-layer message passing and aggregation operations on the news propagation graph to extract the propagation structure features.

[0014] Step S34: Temporal Evolution Feature Extraction: The processor constructs a time series input based on the changes in views, reposts, comments, and likes in different time windows of the structured preprocessed news samples, and performs time dynamic modeling through a temporal convolutional network to generate temporal evolution features.

[0015] Furthermore, the process of iteratively optimizing the parameter configuration of the credibility recognition model using an expected constrained performance optimization strategy to determine the optimal combination of model parameters specifically includes the following steps:

[0016] Step E1: Select the learning rate as the performance evaluation index, construct a black-box performance evaluation function to measure the quality of the model; and uniformly represent the learning rate, the number of layers and the size of hidden units in the graph neural network, the kernel size and stride of the temporal convolutional network, and the fusion weights of each modality feature in the feature fusion subnetwork as a high-dimensional parameter vector to form the parameter search space of the desired constrained performance optimization strategy.

[0017] Step E2: Randomly sample initial candidate parameters in the parameter search space, perform model training and validation on the initial candidate parameters, obtain the corresponding performance evaluation results, and construct a historical evaluation sample set; Based on the difference between the maximum and minimum performance values ​​of the historical samples in the historical evaluation sample set and the diameter of the parameter search space, initialize the acceptance region control parameters and adaptive lower bound parameters in the expected constrained performance optimization strategy.

[0018] Step E3: In each iteration, the processor of the credibility recognition server generates new candidate parameters in the parameter search space and calls a fixed random projection mapping to project the high-dimensional parameter vector to the low-dimensional subspace to reduce the complexity of parameter distance calculation and improve global search efficiency. Subsequently, the processor uses the Worst-m memory mechanism to select the worst-performing parameter configuration set in the historical evaluation sample set as the constraint sample set. Based on the acceptance region control parameters, adaptive lower bound parameters, and the distance relationship between the new candidate parameters and the constraint sample set in the low-dimensional projection subspace, the processor calculates the theoretically achievable upper bound of the new candidate parameters and determines whether to pass the acceptance judgment.

[0019] Step E4: When a new candidate parameter passes the acceptance decision, the processor calls the credibility recognition model training and validation process to train and evaluate the new candidate parameter on the training dataset and validation dataset, obtain the corresponding performance evaluation results, add it to the historical evaluation sample set, update the adaptive lower bound parameter and the acceptance region control parameter, and dynamically update the worst-performing parameter configuration set based on the latest performance ranking, so as to realize the continuous adaptive adjustment of the acceptance region in the expected constrained performance optimization strategy and the compressed storage of invalid historical samples.

[0020] Step E5: Iterate from step E2 to step E4. When training terminates, the processor selects the best-performing model parameter combination from the historical evaluation sample set. The learning rate, feature fusion weights, graph neural network structure parameters, and temporal convolutional network structure parameters contained in this parameter combination are written into the configuration file and runtime parameter register of the credibility recognition model. This serves as the target configuration for the credibility recognition model in actual deployment, which is used to improve the recognition accuracy, stability, and generalization ability of news information credibility recognition under different media sources, different event types, and different user dissemination behavior patterns.

[0021] Furthermore, the trustworthiness identification server is a multi-core x86 processor with a vector acceleration instruction set, and its memory includes 32GB of RAM and a GPU graphics card. The GPU graphics card is used to accelerate the inference operations of the Transformer text encoding model and the graph neural network model. The trustworthiness identification server is connected to the news aggregation crawler server, the content management system server, the social media interface server, the relational database server, and the distributed file storage server via 10 Gigabit Ethernet.

[0022] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0023] First, this invention constructs a multimodal deep learning credibility recognition framework that integrates textual semantics, media and account credibility, propagation graph structure, and the temporal evolution of interactive behavior, achieving a three-dimensional modeling and global feature expression of news information credibility. Compared to existing intelligent review models that rely solely on single semantics or static propagation indicators, this invention can accurately capture the true diffusion pattern and signal anomaly patterns of news content among user groups in complex propagation scenarios, enabling precise identification of false propagation chains, concentrated activities of fake accounts, and abnormal growth behaviors. Through a unified feature alignment and cross-modal fusion mechanism, this invention significantly improves the accuracy and stability of credibility recognition for complex news samples, ensuring high robustness of the model across platforms, multiple event types, and multiple propagation environments, effectively solving the problem of insufficient capabilities of traditional intelligent review models in content cross-domain migration and scenario generalization.

[0024] Secondly, the expected-constrained performance optimization strategy introduced in this invention, combining fixed random projection mapping, Worst-M memory mechanism, and adaptive receptive region adjustment mechanism, achieves efficient search and stable convergence in the high-dimensional parameter space of the model. This solves the problems of convergence difficulty, blind search, and easy getting trapped in local optima in the parameter tuning process of traditional deep learning. This strategy can significantly shorten the model training time, improve parameter configuration efficiency, and reduce the waste of computing resources caused by inefficient iteration. It enables the credibility recognition model to adaptively perform dynamic tuning for different news dissemination structures, different content patterns, and different data distributions, enhancing the generalization ability and operational reliability of this invention under real and complex data conditions, and laying the foundation for providing real-time credibility judgment in large-scale content review scenarios.

[0025] Finally, this invention constructs a credibility-level-driven intelligent review and risk management mechanism, achieving closed-loop collaboration between credibility identification and platform content management processes. Through risk grading based on credibility scores, this invention can automatically intercept and manually review low-credibility news, display risk warnings for suspected content, and directly publish high-credibility content, significantly improving content review efficiency and reducing the risk of misinformation and public opinion spread. Simultaneously, the credibility identification results and model judgment criteria support interpretable displays, helping reviewers understand the rationale behind their decisions and improving review transparency and the persuasiveness of the results. This mechanism ensures the real-time and accurate governance of content platforms, significantly enhancing the health of the internet news ecosystem and the ability to protect public information security. Attached Figure Description

[0026] Figure 1 A flowchart illustrating a deep learning-based intelligent recognition method for news information credibility provided by this invention.

[0027] Figure 2 This is a schematic diagram of the graph neural network structure proposed in Example 2.

[0028] Figure 2 In the middle, on the left: the graph structure input layer, where X1, X2, X3, and X4 are nodes, and the lines between nodes represent edges; in the middle: the hidden layers of the graph neural network with multiple rounds of message passing; on the right: the output layer, where Z1, Z2, Z3, and Z4 are the final graph embeddings of the nodes; Y1 and Y4 are the task outputs; on the far right: a visualization of the node embedding space, showing the distribution graph after dimensionality reduction represented by the graph neural network output. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] Example 1, according to Figure 1 This invention provides a deep learning-based intelligent method for recognizing the credibility of news information. Its key feature is that it is applied to internet news aggregation platforms and social media content review platforms, used for recognizing the credibility and issuing risk warnings for news information to be published or already published. The method is executed by a credibility recognition server deployed in a data center. The credibility recognition server includes a processor, memory, and a network interface, and is connected to the internet, user terminals, and content management system servers via the network interface. The method includes the following steps:

[0031] Step S1: News Information Collection and Scene Access: The credibility recognition server connects with the news aggregation crawler server, content management system server, and social media open interface through the network interface to collect news information to be analyzed from news websites, client submissions, third-party content distribution platforms, and social media, and obtain multi-source raw news data including title, body text, cover image link, publishing account, publishing channel, publishing time, reprint source, location, number of reposts, number of comments, and number of likes;

[0032] Step S2: Data Preprocessing and Text Standardization: The processor calls the preprocessing program from memory to preprocess the multi-source raw news data, including: removing HTML tags and script code, performing language detection, Chinese word segmentation and stop word filtering, abnormal symbol cleaning, and time and place name standardization; the cleaned text is used to construct a text sequence in fixed-length segments, and missing value filling, normalization, and format unification are performed on metadata fields including publishing account, publishing channel, reprint source, publishing time, number of views, number of reposts, number of comments, and number of likes to obtain a structured preprocessed news sample;

[0033] Step S3: Multimodal feature construction: The processor of the credibility recognition server takes the structured preprocessed news sample as input, extracts text semantic features, media and account credibility features, propagation structure features and temporal evolution features respectively, and performs unified encoding and dimension alignment to obtain multimodal feature data;

[0034] Step S4: Deep Learning Model Inference: Using multimodal feature data as input, the credibility recognition model stored in memory is trained and validated. During the training process, the expected constrained performance optimization strategy is used to iteratively optimize the parameter configuration of the credibility recognition model to determine the optimal combination of model parameters, thus obtaining the trained credibility recognition model. The credibility recognition model outputs the credibility score of news information. The credibility recognition model includes a text encoding subnetwork, a media and account feature subnetwork, a propagation graph embedding subnetwork, a temporal convolution subnetwork, a feature fusion subnetwork, and a classification subnetwork. The construction process of the expected constrained performance optimization strategy is as follows: Based on the ECP strategy, a fixed random projection mapping, Worst-m memory mechanism, and an adaptive receptive region adjustment mechanism are introduced to optimize the search efficiency, global exploration capability, and convergence stability of the ECP strategy, thus constructing the expected constrained performance optimization strategy.

[0035] Step S5: Credibility Level Classification and Contextualized Handling: Based on a preset threshold set, the processor classifies news information into multiple credibility levels—high credibility, suspicious, and low credibility—based on the credibility score and the threshold set. When the credibility level is low credibility, a risk interception command is sent to the content management system server to prevent the news from being automatically published, and the news is pushed to the reviewer's work terminal for manual review. When the credibility level is suspicious, a "risk warning" label is added when the news is displayed on the user's terminal, and it is added to the key monitoring queue. When the credibility level is high credibility, the content management system server is allowed to directly publish the news, and the credibility result is archived in the database for subsequent training and traceability.

[0036] User terminals include editing workstations with news editing clients installed, reviewer workstations with review management clients installed, and mobile terminals with news reading clients installed.

[0037] The review management client is used to receive low-credibility news push information from the credibility identification server, display the credibility score and a summary of the model's judgment basis, and support reviewers to manually confirm, modify, or return the news.

[0038] The relational database server in the database is used to store basic news fields, account and media attribute fields, and credibility recognition results; the distributed file storage server is used to store the original HTML pages, images, videos, and model training datasets.

[0039] Example 2, according to Figure 2 This embodiment is based on Embodiment 1. In this embodiment, the process of extracting text semantic features, media and account credibility features, propagation structure features and temporal evolution features respectively includes the following steps:

[0040] Step S31: Text semantic feature extraction: The processor calls the pre-trained Transformer text encoding model stored in memory to segment and encode the news headlines and body text in the structured preprocessed news samples, generating word-level and sentence-level semantic embedding vectors to form text semantic features;

[0041] Step S32: Media and Account Credibility Feature Extraction: Based on the number of times news from the same media source was manually labeled as fake news, the number of times it was deleted by the platform, the account registration duration and authentication information in the structured preprocessed news samples, the processor performs statistical summarization and scale unification processing to generate media and account credibility features.

[0042] Step S33: Propagation structure feature extraction: The processor constructs a news propagation graph based on the social media comments, reposts and citations in the structured preprocessed news samples. User accounts and topic tags are used as graph nodes, and reposts, comments and @ interactions are used as graph edges. The graph neural network is run on the credibility recognition server to perform multi-layer message passing and aggregation operations on the news propagation graph to extract the propagation structure features.

[0043] Step S34: Temporal Evolution Feature Extraction: The processor constructs a time series input based on the changes in views, reposts, comments, and likes in different time windows of the structured preprocessed news samples, and performs time dynamic modeling through a temporal convolutional network to generate temporal evolution features.

[0044] Example 3, based on Example 2, describes an iterative optimization of the reliability identification model parameter configuration using a desired constrained performance optimization strategy to determine the optimal combination of model parameters. The specific steps include:

[0045] Step E1: Select the learning rate as the performance evaluation index, construct a black-box performance evaluation function to measure the quality of the model; and uniformly represent the learning rate, the number of layers and the size of hidden units in the graph neural network, the kernel size and stride of the temporal convolutional network, and the fusion weights of each modality feature in the feature fusion subnetwork as a high-dimensional parameter vector to form the parameter search space of the desired constrained performance optimization strategy.

[0046] Step E2: Randomly sample initial candidate parameters in the parameter search space, perform model training and validation on the initial candidate parameters, obtain the corresponding performance evaluation results, and construct a historical evaluation sample set; Based on the difference between the maximum and minimum performance values ​​of historical samples in the historical evaluation sample set and the diameter of the parameter search space, initialize the acceptance region control parameters and adaptive lower bound parameters in the expected constrained performance optimization strategy to establish an initial acceptance region that matches the current recognition task scale.

[0047] The adaptive lower bound parameter is defined as follows:

[0048] ;

[0049] in, Indicates the first Adaptive lower bound parameters during round iteration; and Indicates the historical parameter index. , Indicates the first , A vector of historical parameters; This represents the performance evaluation function, i.e., the learning rate; Indicates the first Performance values ​​of historical parameter configurations; Indicates the first Performance values ​​of historical parameter configurations; Represents the parameter search space. Indicates the diameter of the parameter search space;

[0050] Adaptive lower bound parameter This reflects the proportional relationship between the magnitude of changes in model performance values ​​in the current historical evaluation samples and the scale of the parameter search space. It serves as a theoretical lower bound constraint to ensure that the acceptance region is not excessively shrunk; in other words, This means that in order for candidate parameters to theoretically have the ability to exceed the current optimal parameter combination, the minimum slope requirement of the model performance change trend in the parameter space is used to avoid a large number of candidate parameters being rejected continuously in the early stages of global optimization due to the small acceptance region, which would cause the model training to stagnate, get stuck in local optima, or be unable to explore better parameter combinations.

[0051] In the news information credibility recognition application scenario of this invention, there are significant differences between different news sample sources, different user interaction behavior patterns, and different event propagation structures. The training and inference performance exhibits nonlinearity and high uncertainty in the parameter space. Therefore, through… The established minimum slope constraint mechanism ensures that the parameter search process has sufficient exploratory capabilities, effectively supporting the model to achieve higher recognition stability and generalization performance in complex media environments.

[0052] Step E3: In each iteration, the processor of the credibility recognition server generates new candidate parameters in the parameter search space and calls a fixed random projection mapping to project the high-dimensional parameter vector to a low-dimensional subspace to reduce the complexity of parameter distance calculation and improve global search efficiency; subsequently, the processor uses the Worst-m memory mechanism to select the worst-performing sample from the historical evaluation sample set. The parameter configuration set is used as the constraint sample set. Based on the acceptance region control parameters, adaptive lower bound parameters, and the distance relationship between the new candidate parameters and the constraint sample set in the low-dimensional projection subspace, the theoretically achievable upper bound of the performance of the new candidate parameters is calculated, and it is determined whether to pass the acceptance decision.

[0053] The Worst-m memory mechanism used in this step addresses the problems of excessive computational complexity and insufficient parameter exploration capability caused by excessive shrinkage of the acceptance region in traditional ECP strategies under conditions of large parameter size and continuous accumulation of historical evaluation samples. Based on the Worst-m idea, this invention selects only the m worst-performing parameter configurations from the historical evaluation samples as the constraint sample set, and compares them only with the parameters in this set during the acceptance decision process, without comparing them with all historical parameters. This significantly reduces the computational load and memory overhead, enabling ECPv2 to be applied to complex model structures containing high-dimensional parameter spaces.

[0054] The worst performing The set of indices of each sample The definition is as follows:

[0055] ;

[0056] in, This indicates that the assessment has already been completed. Among the historical parameters, the one with the worst performance. A set of subscripts for each parameter; Indicates from all historical sample sets The size selected in is any subset of; Represents a set Performance values ​​corresponding to all indexes Perform cumulative calculation;

[0057] The formula used to determine whether an application passes the acceptance test is:

[0058] , ;

[0059] in, This represents the candidate parameter vector generated in the current iteration. Represents a fixed random projection matrix. Candidate parameters Representation in low-dimensional space (projection vector). Represents the history parameter vector Representation in low-dimensional space; This represents the squared distance between the candidate parameter vector and the historical parameter vector in the projected subspace; Indicates the parameters controlled by the region. This represents the upper bound of the distance distortion introduced by random projection, that is, the maximum controllable distortion caused by a fixed random projection matrix to the distance in the parameter space; Indicates the adaptive acceptance factor; Indicates all the worst Among the group parameters, take the minimum value among these optimistic upper bound estimates; This represents the best historical performance value to date.

[0060] Specifically, when the upper bound of the performance is less than the maximum performance value in the current historical samples, the candidate parameter is directly judged as a poor configuration and rejected without model training, thereby reducing the invalid evaluation of obviously suboptimal parameter combinations; when the number of consecutive rejections exceeds the preset threshold, the processor expands the acceptance region by amplifying the acceptance region control parameters and updating the adaptive lower bound parameters to avoid the acceptance region approaching an empty set in the long term.

[0061] Step E4: When a new candidate parameter passes the acceptance test, the processor invokes the credibility recognition model training and validation process. It trains and evaluates the new candidate parameter on both the training and validation datasets, obtains the corresponding performance evaluation results, adds them to the historical evaluation sample set, updates the adaptive lower bound parameters and the acceptance region control parameters, and dynamically updates the worst-performing sample based on the latest performance ranking. A set of parameter configurations is provided to enable continuous adaptive adjustment of the acceptance region in the desired constrained performance optimization strategy and compressed storage of invalid historical samples.

[0062] Step E5: Iterate from step E2 to step E4. When training terminates, the processor selects the best-performing model parameter combination from the historical evaluation sample set. The learning rate, feature fusion weights, graph neural network structure parameters, and temporal convolutional network structure parameters contained in this parameter combination are written into the configuration file and runtime parameter register of the credibility recognition model. This serves as the target configuration for the credibility recognition model in actual deployment, which is used to improve the recognition accuracy, stability, and generalization ability of news information credibility recognition under different media sources, different event types, and different user dissemination behavior patterns.

[0063] Example 4, based on Example 3, describes a trustworthiness identification server whose processor is a multi-core x86 processor with a vector acceleration instruction set, and whose memory includes 32GB of RAM and a GPU. The GPU is used to accelerate the inference operations of the Transformer text encoding model and the graph neural network model. The trustworthiness identification server is connected to a news aggregation crawler server, a content management system server, a social media interface server, a relational database server, and a distributed file storage server via 10 Gigabit Ethernet.

[0064] Example 5, this example is based on Example 4, in this example,

[0065] Step S4: Deep learning model inference: Using multimodal feature data as input, the credibility recognition model set in the memory is trained and validated. During the training process of the credibility recognition model, the expected constrained performance optimization strategy is used to iteratively optimize the parameter configuration of the credibility recognition model, determine the optimal combination of model parameters, and obtain the trained credibility recognition model. Through the trained credibility recognition model, the credibility score of the news information is output.

[0066] Step S5: Credibility Level Classification and Contextualized Handling: Based on a preset threshold set, the processor classifies news information into multiple credibility levels—high credibility, suspicious, and low credibility—based on the credibility score and the threshold set. When the credibility level is low credibility, a risk interception command is sent to the content management system server to prevent the news from being automatically published, and the news is pushed to the reviewer's work terminal for manual review. When the credibility level is suspicious, a "risk warning" label is added when the news is displayed on the user's terminal, and it is added to the key monitoring queue. When the credibility level is high credibility, the content management system server is allowed to directly publish the news, and the credibility result is archived in the database for subsequent training and traceability.

[0067] In this embodiment, the credibility recognition server collects a news sample of a social event to be published from social media platform A. The core field data after its structured preprocessing is shown in Table 1 below:

[0068] Table 1

[0069] ;

[0070] After inputting the above data into the credibility recognition model for joint inference, the credibility score of the news information obtained by the credibility recognition server is 0.28.

[0071] In this embodiment, the set of preset confidence level thresholds is shown in Table 2:

[0072] Table 2

[0073] ;

[0074] Based on the aforementioned threshold set, the news is determined to be of low credibility level.

[0075] The trustworthiness assessment server executes the following coordinated processing procedure:

[0076] 1. Send a risk blocking command to the content management system server: prevent the news from being automatically published on the platform and avoid its widespread dissemination without verification;

[0077] 2. Push the news to the reviewer's work terminal: The reviewer's terminal displays a credibility score of 0.28, an abnormal dissemination structure indicator, and a summary of the triggering criteria, prompting manual review;

[0078] 3. Record the recognition results and store them in the database: The data includes text embedding features, propagation graph summaries, and final review status, which are used for subsequent model retraining and for tracing the basis for judgment.

[0079] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent identification method for news information credibility, characterized in that, The method is executed by a trustworthiness identification server deployed in a data center, the server including a processor and memory; the method includes the following steps: Step S1: Obtain original news data from multiple sources; Step S2: The processor retrieves the preprocessing program from memory to preprocess the multi-source raw news data and obtain structured preprocessed news samples; Step S3: The processor takes the structured preprocessed news sample as input, extracts text semantic features, media and account credibility features, dissemination structure features and temporal evolution features respectively, and performs unified encoding and dimension alignment to obtain multimodal feature data; Step S4: Using multimodal feature data as input, train and validate the credibility recognition model set in the memory. During the training process of the credibility recognition model, adopt the expected constrained performance optimization strategy to iteratively optimize the parameter configuration of the credibility recognition model, determine the optimal combination of model parameters, and obtain the trained credibility recognition model. Through the trained credibility recognition model, output the credibility score of the news information. Step S5: Preset threshold set. Based on the credibility score of news information and the threshold set, divide news information into three credibility levels: high credibility, suspicious, and low credibility. 2.The deep learning-based news information credibility intelligent identification method according to claim 1, characterized in that: The credibility recognition model includes a feature fusion subnetwork and a classification subnetwork. 3.The deep learning-based news information credibility intelligent identification method according to claim 1, characterized in that: The process of constructing the expected constrained performance optimization strategy is as follows: Based on the ECP strategy, a fixed random projection mapping, Worst-m memory mechanism, and adaptive receptive region adjustment mechanism are introduced to optimize the ECP strategy in terms of search efficiency, global exploration capability, and convergence stability, thereby constructing the expected constrained performance optimization strategy. 4.The deep learning-based news information credibility intelligent identification method of claim 1, wherein: The process of extracting text semantic features, media and account credibility features, dissemination structure features, and temporal evolution features respectively includes the following steps: Step S31: Call the pre-trained Transformer text encoding model in the memory to segment and encode the news headlines and body text in the structured preprocessed news samples to form text semantic features; Step S32: Based on the number of times news from the same media source in the structured preprocessed news sample was manually labeled as fake news, the number of times it was deleted by the platform, the account registration duration and authentication information, perform statistical summarization and scale unification processing to generate media and account credibility features; Step S33: Construct a news dissemination graph based on the social media comments, reposts, and citations in the structured preprocessed news samples. Use user accounts and topic tags as graph nodes and reposts, comments, and @ interactions as graph edges. Run a graph neural network on the credibility recognition server to perform multi-layer message passing and aggregation operations on the news dissemination graph and extract the dissemination structure features. Step S34: Construct a time series input based on the changes in views, reposts, comments, and likes in different time windows of the structured preprocessed news samples, and perform time dynamic modeling through a temporal convolutional network to generate temporal evolution features.

5. The method of claim 3, wherein the method further comprises: The process of iteratively optimizing the parameter configuration of the credibility identification model using an expected constrained performance optimization strategy to determine the optimal combination of model parameters includes the following steps: Step E1: Select the learning rate as the performance evaluation index; unify the learning rate, the number of layers and the size of hidden units in the graph neural network, the kernel size and stride of the temporal convolutional network, and the fusion weights of each modality feature in the feature fusion subnetwork into a high-dimensional parameter vector to form the parameter search space of the desired constrained performance optimization strategy. Step E2: Randomly sample initial candidate parameters in the parameter search space, perform model training and validation on the initial candidate parameters, obtain the corresponding performance evaluation results, and construct a historical evaluation sample set; Based on the difference between the maximum and minimum performance values ​​of the historical samples in the historical evaluation sample set and the diameter of the parameter search space, initialize the acceptance region control parameters and adaptive lower bound parameters in the expected constrained performance optimization strategy. Step E3: In each iteration, new candidate parameters are generated in the parameter search space and projected onto the low-dimensional subspace using a fixed random projection mapping. The Worst-m memory mechanism is used to select the worst-performing parameter configuration set in the historical evaluation sample set as the constraint sample set. Based on the acceptance region control parameters, adaptive lower bound parameters, and the distance relationship between the new candidate parameters and the constraint sample set in the low-dimensional projection subspace, the performance upper bound of the new candidate parameters is calculated, and it is determined whether they pass the acceptance decision. Step E4: When a new candidate parameter passes the acceptance decision, train and evaluate the new candidate parameter to obtain the corresponding performance evaluation result, add it to the historical evaluation sample set, update the adaptive lower bound parameter and the acceptance region control parameter, and update the parameter configuration set with the worst performance. Step E5: Iterate from step E2 to step E4. When training terminates, the processor selects the best-performing combination of model parameters from the historical evaluation sample set.

6. The deep learning-based news information credibility intelligent identification method according to claim 1, characterized in that: The trustworthiness identification server is a multi-core x86 processor with a vector acceleration instruction set, and its memory includes 32GB of RAM and a GPU graphics card. The GPU graphics card is used to accelerate the inference operations of the Transformer text encoding model and the graph neural network.

Images