Early research and judgment method and device for online social network unreal public opinion information

Abstract

The invention discloses an online social network unreal public opinion information early study and judgment method and device, belongs to the technical field of unreal public opinion information spreading, and particularly relates to spreading modeling of online social network information. The problem that according to an existing judgment method based on public opinion information content, the response time of public opinion managers to public opinion events in an online social network cannot be shortened is solved. The method comprises the following steps of: constructing a detection framework of an online social network unreal public opinion information early study and judgment model; and obtaining a final research and judgment result by obtaining the configured early research and judgment model of the unreal public opinion information of the online social network according to the steps. The online social network unreal public opinion information early research and judgment method and device are suitable for carrying out early research and judgment on online social network unreal public opinion information.

Description

Technical Field

[0001] The present invention relates to the technical field of false public opinion information dissemination, and in particular to the dissemination modeling of online social network information. Background Art

[0002] The popularity of smart mobile devices and the rapid development of Internet technology have made online social networks an indispensable means of communication for people to receive news and real-time information. Online social networks have the characteristics of open registration, freedom of information release, and heterogeneity of users. Online social network users can receive and send news, consultations, and popular science in any network in real time. However, not all online social network information is verified, which makes users vulnerable to false public opinion information when facing public emergencies.

[0003] The online social network space is anonymous, interactive and communicative, which makes it easier for users to obtain information in this network environment, but at the same time it is difficult to guarantee the reliability and authenticity of the content in the online social network. The emergence of this phenomenon has brought huge challenges to the information security of online social networks and the maintenance of a clear network space.

[0004] Specifically, the uncertainty and anonymity of information in online social networks make the virtual crowd in cyberspace unconstrained. In the face of public emergencies, some false public opinion information may appear in online social networks. The rapid spread of false public opinion information in social networks can easily produce huge negative effects.

[0005] It can be seen that the early analysis and control of false public opinion information has become a major problem that needs to be solved urgently in this field. Exploring the early analysis methods of false public opinion information in online social networks has very important theoretical value and practical application value.

[0006] Traditional methods for judging false public opinion information on online social networks mostly rely on reports from Internet users, and the platform conducts manual screening based on the reports. However, this manual method can only judge the information itself, and is time-consuming and laborious. When there is too much false public opinion information, due to the high timeliness of the information, such information spreads rapidly on social networks, and it is difficult to identify it early manually. Therefore, scholars have begun to conduct a lot of research on the automatic identification of false public opinion information on online social networks. Relying on the development of machine learning and the replacement of natural language processing technology and information retrieval technology, some accurate and efficient automatic identification methods have gradually emerged in recent years. By quickly identifying the authenticity of public opinion information in online social networks, it controls and intervenes in the process of its dissemination, further reducing its harm to the country and society. Specifically, existing automatic identification research is mainly divided into two directions: rapid judgment based on the content of public opinion information and identification of information sources based on the dissemination process. The former quickly detects public opinion information, opinions or views posted by online social network users to determine whether they are false public opinion information, while the latter judges the authenticity of the information source based on the network topology structure and network node status generated by the propagation process of public opinion on online social networks.

[0007] In recent years, although the rapid judgment method based on the content of public opinion information has become relatively mature, its specific application in the handling of false public opinion still has certain limitations. Since information spreads at an exponential rate in the network, it is necessary to carry out effective control in the shortest possible time to prevent it from spreading rapidly to the entire online social network. The rapid judgment method based on the content of public opinion information is more targeted at scattered network users, and the types of public opinion are relatively broad, making it difficult to form targeted control targets and effective control measures in a short time. Therefore, compared with the rapid judgment method based on information content, the research on information source identification based on the propagation process is of great significance.

[0008] In summary, the existing judgment methods based on the content of public opinion information cannot shorten the response time to public opinion events in online social networks. Summary of the invention

[0009] The present invention proposes an early analysis method and device for false public opinion information in online social networks, which solves the problem that the existing judgment method based on the content of public opinion information cannot shorten the response time to public opinion events in online social networks.

[0010] The method for early analysis of false public opinion information on online social networks described in the present invention has the following technical solutions:

[0011] The method comprises the following steps:

[0012] Steps for building a detection framework for early detection models of false public opinion information on online social networks;

[0013] The detection framework includes an agent module and a system environment module, wherein:

[0014] The system environment module is used to generate a new source text or comment of public opinion information in each interaction process; it is also used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback;

[0015] The intelligent agent module is used to compile the input public opinion information source text and comments in the online social network into a propagation topology graph;

[0016] The intelligent agent module is used to transfer the compiled propagation topology map to the strategy network, and classify the data of the propagation topology map until a final judgment is made;

[0017] Steps for constructing the state space to support the reinforcement learning process;

[0018] Steps for configuring the action space of the model for early identification of false public opinion information on online social networks;

[0019] Steps for configuring system environment modules;

[0020] Steps for configuring a policy network for an early identification model of false public opinion information on online social networks;

[0021] A step for obtaining a configured online social network false public opinion information early analysis model according to the above steps, and obtaining a final analysis result.

[0022] Further, a preferred embodiment is provided, wherein the steps for constructing a state space to support a reinforcement learning process include the following steps:

[0023] It is used to build a data preprocessing and graph construction pipeline, and compile a propagation topology graph from an online social network public opinion information text dataset;

[0024] The data preprocessing and graph construction pipeline includes an input module, a preprocessing module, a topology graph construction module, and an output module;

[0025] The input module is used to input public opinion information source text and comments;

[0026] The preprocessing module is used to preprocess the source text and comments of public opinion information to obtain the node features of the propagation topology graph;

[0027] The topology map construction module is used to construct a propagation topology map according to the node characteristics of the propagation topology map;

[0028] The output module is used to output the propagation topology map as a state input to the policy network.

[0029] Further, a preferred embodiment is provided, wherein the preprocessing module comprises the following submodules:

[0030] Feature extraction submodule: used to clean the text and comments of public opinion information sources and delete mixed special symbols, expressions and punctuation marks;

[0031] Feature encoding submodule: used to separate the meaningful words in the source text and comments of the cleaned public opinion information. Each separated source text or comment text will be transferred to an ordered word list;

[0032] Feature embedding submodule: used to embed all separated words into a vector space according to the pre-trained dictionary and generate a feature matrix, in which each column is the feature vector of each word;

[0033] Feature averaging submodule: used to reduce data dimension, use feature averaging processing to compress the feature matrix into a feature vector with an average function, and obtain the node features of the propagation topology diagram of public opinion information.

[0034] Furthermore, a preferred implementation is provided, wherein the steps for configuring the action space of the online social network false public opinion information early analysis model include the following steps:

[0035] Set up three discrete actions: true public opinion information sources, false public opinion information sources, and unverified public opinion information sources;

[0036] Get the propagation topology of the public opinion information currently obtained from the state space as the state input, and determine its action type:

[0037] If it is identified as an unverified public opinion information source, the propagation process of the public opinion event will not be terminated, but will be rolled to the propagation topology state of the next timestamp and continue to be iterated and identified by the reinforcement learning algorithm;

[0038] As the propagation time progresses, when it is judged to be a true public opinion information source or a false public opinion information source, the input iteration of the propagation topology map state will stop; the learning process of the early analysis model will be reset and restarted by another public opinion event to be judged.

[0039] Further, a preferred embodiment is provided, wherein the step for configuring the system environment module comprises the following steps:

[0040] Steps for configuring run rules:

[0041] The reinforcement learning model randomly selects real public opinion events or false public opinion events from the training data set;

[0042] When the selected event propagation ends or the agent makes a final judgment, the analysis model will terminate all judgment processes of the current event and start the next public opinion event to be judged;

[0043] The starting time of online social network public opinion events is determined randomly, and a threshold is set to limit the maximum size of the propagation topology graph;

[0044] Steps for configuring the reward equation:

[0045] The reward R is composed of two parts: decision reward r decision and decision time delay reward r time ;

[0046] The reward R is expressed as:

[0047] R=r decision +α·r time

[0048]

[0049] In the formula, α is the unified r decision and r time The parameter set by the unit; Δl is the maximum number of nodes in the current public opinion event propagation topology graph;

[0050] Using r time To punish each additional decision step in the agent's judgment process, this value is relatively small in the early judgment model. Let r time =0.01.

[0051] Further, a preferred implementation is provided, the steps for configuring a policy network for an early analysis model of false public opinion information on online social networks;

[0052] The policy network consists of two modules: a feature extractor and an actor-evaluator architecture;

[0053] The feature extractor consists of a graph neural network module and a multi-layer fully connected neural network module;

[0054] The graph neural network module is used to extract the main features of the public opinion event propagation topology graph, which is implemented by two graph convolutional neural network layers;

[0055] The multi-layer fully connected neural network module contains two fully connected layers, which are used to further process the features extracted by the graph neural network module;

[0056] The actor-evaluator framework consists of two parts: the Actor part and the Critic part, where:

[0057] The Actor part is used to predict the action distribution in the early judgment model of false public opinion information in online social networks, and intelligent

[0058] The body uses the policy gradient method to learn the strategy and then output the corresponding action;

[0059] The Critic part is used to estimate the value function and is responsible for evaluating the action strategy.

[0060] The present invention also proposes an early analysis and judgment device for false public opinion information on online social networks, and its technical solution is as follows:

[0061] The device comprises the following modules:

[0062] A module for building a detection framework for early analysis of false public opinion information on online social networks;

[0063] The detection framework includes an agent module and a system environment module, wherein:

[0064] The system environment module is used to generate a new source text or comment of public opinion information in each interaction process; it is also used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback;

[0065] The intelligent agent module is used to compile the input public opinion information source text and comments in the online social network into a propagation topology graph;

[0066] The intelligent agent module is used to transfer the compiled propagation topology map to the strategy network, and classify the data of the propagation topology map until a final judgment is made;

[0067] Modules for constructing state spaces to support the reinforcement learning process;

[0068] A module for configuring the action space of the early detection model of false public opinion information on online social networks;

[0069] Modules for configuring system environment modules;

[0070] A module for configuring the policy network of the early identification model of false public opinion information on online social networks;

[0071] A module used to obtain the configured online social network false public opinion information early analysis model according to the above steps and obtain the final analysis result.

[0072] The present invention also proposes a computer device, and its technical solution is as follows:

[0073] A computer device includes: a processor and a memory, wherein the memory is used to store executable instructions of the processor, and the processor is configured to execute the above-mentioned method for early analysis of false public opinion information in online social networks by executing the executable instructions.

[0074] The present invention also proposes a computer storage medium, and its technical solution is as follows:

[0075] A computer storage medium stores a computer program, which, when running, executes the above-mentioned method for early analysis of false public opinion information in online social networks.

[0076] The present invention also proposes a computer program product, and its technical solution is as follows:

[0077] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned method for early analysis of false public opinion information in online social networks.

[0078] The present invention has the following beneficial effects:

[0079] 1. The method for early analysis of false public opinion information in online social networks described in the present invention is based on reinforcement learning, takes public opinion information as the research object, introduces the information propagation process into the identification of public opinion information sources, and introduces the actor-evaluator reinforcement learning framework around the two core issues of determining and identifying the source of public opinion information. The public opinion information propagation process is represented as a dynamic graph structure and a graph convolutional neural network is used to extract spatiotemporal features; through a fully connected layer network, public opinion information is analyzed at an early stage; thereby shortening the response time to public opinion events in online social networks.

[0080] 2. The method for early identification of false public opinion information in online social networks described in the present invention combines the proximal policy optimization (PPO) algorithm with the actor--I-Critic method (AC) to identify the authenticity of the public opinion information source as early as possible based on the forwarded information; this method is superior to the neural network learned by the supervised learning method in terms of accuracy and judgment speed, and further optimizes the real-time and accuracy of the identification of false public opinion information.

[0081] The method and device for early analysis of false public opinion information on online social networks described in the present invention are suitable for early analysis of false public opinion information on online social networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0083] Figure 1 A schematic diagram of the structure of a graph neural network in one embodiment of the present invention;

[0084] Figure 2 A schematic diagram of the structure of a graph convolutional neural network in one embodiment of the present invention;

[0085] Figure 3 A schematic diagram of the basic principle of reinforcement learning in one embodiment of the present invention;

[0086] Figure 4 A schematic diagram of the Actor-Critic algorithm training process in one embodiment of the present invention;

[0087] Figure 5 A schematic diagram of a PPO algorithm framework diagram in one embodiment of the present invention;

[0088] Figure 6 A schematic diagram of the overall architecture of a method for early analysis of false public opinion information in online social networks (based on reinforcement learning) in one embodiment of the present invention;

[0089] Figure 7 A schematic diagram of a state space architecture in one embodiment of the present invention;

[0090] Figure 8 A schematic diagram of an action space architecture in one embodiment of the present invention;

[0091] Fig. 9 A schematic diagram of a policy network architecture in one embodiment of the present invention;

[0092] Fig.10 A schematic diagram of a cumulative reward curve of an early online social network false public opinion analysis model in one embodiment of the present invention;

[0093] Fig.11 A schematic diagram of the accuracy stability analysis of an early online social network false public opinion analysis model in one embodiment of the present invention;

[0094] Fig.12 A schematic diagram of the effect of the number of comments on a public opinion information source on the accuracy rate in one embodiment of the present invention;

[0095] Fig.13A schematic diagram of determining the effect of the number of observations on the accuracy rate in one embodiment of the present invention;

[0096] Fig.14 A schematic diagram of the effect of the number of discriminant observations on precision, recall and F1-Score in one embodiment of the present invention;

[0097] Fig.15 The figure is a schematic diagram of a rapid response mechanism for sudden public opinion events in one embodiment of the present invention. DETAILED DESCRIPTION

[0098] In order to make the technical solutions and advantages of the present invention more clearly described, the specific implementation methods of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various implementation methods described below are only part of the preferred solutions of the present invention, rather than all implementation methods; the various implementation methods described below are intended to explain the present invention and cannot be understood as limitations on the present invention; the reasonable combination of technical features defined in the various implementation methods of the present invention, and all other implementation methods obtained by ordinary technicians in the field without creative work based on the implementation methods of the present invention, are within the scope of protection of the present invention.

[0099] In one implementation, a method for early analysis of false public opinion information on online social networks is provided:

[0100] The method (or the model for early identification of false public opinion information in online social networks) is based on reinforcement learning; the overall framework of the method (model) is as follows: Figure 6 shown.

[0101] The method comprises the following steps:

[0102] Steps for building a detection framework for early detection models of false public opinion information on online social networks;

[0103] The detection framework includes an agent module and a system environment module, wherein:

[0104] The system environment module is used to generate a new source text or comment of public opinion information in each interaction process; it is also used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback;

[0105] The intelligent agent module is used to compile the input public opinion information source text and comments in the online social network into a propagation topology graph;

[0106] The agent module is used to transfer the compiled propagation topology map to the strategy network, and classify the data of the propagation topology map until a final judgment (judgment result) is made;

[0107] Steps for constructing the state space to support the reinforcement learning process;

[0108] Steps for configuring the action space of the model for early identification of false public opinion information on online social networks;

[0109] Steps for configuring system environment modules;

[0110] Steps for configuring a policy network for an early identification model of false public opinion information on online social networks;

[0111] A step for obtaining a configured online social network false public opinion information early analysis model according to the above steps, and obtaining a final analysis result.

[0112] In order to give the classification strategy the ability to judge false public opinion information at an early stage, a detection framework (an early judgment model of false public opinion information in online social networks) based on reinforcement learning is designed (constructed). The detection framework includes an agent module and a system environment module, in which:

[0113] The system environment module is a simulator of the public opinion information dissemination process in an online social network. Its tasks are divided into:

[0114] Similar to traditional social media, the system environment module is used to generate a new public opinion information source text or comment (i.e., information data) in each interaction process;

[0115] The system environment module is used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback;

[0116] At the same time, the main tasks of the agent module are:

[0117] The agent module is used to compile the input data, i.e., the text and comments of the public opinion information source in the online social network (i.e., information data), into a propagation topology graph (or information propagation graph);

[0118] The intelligent agent module is used to transfer the compiled propagation topology map to the strategy network (i.e., the information analysis strategy network), and classify the data of the propagation topology map until a final judgment (final analysis result) is made.

[0119] The following is a detailed introduction to the details of the early identification method of false public opinion information on online social networks based on the typical architecture of the reinforcement learning algorithm:

[0120] State Space Design:

[0121] Describes how to construct a state space to support the reinforcement learning process.

[0122] In order to compile a propagation topology graph from the online social network public opinion information text dataset, a data preprocessing and graph construction pipeline is designed (constructed). Its structure is as follows Figure 7 shown.

[0123] The data preprocessing and graph construction pipeline includes an input module, a preprocessing module, a topology graph construction module, and an output module;

[0124] The input module is used to input public opinion information source text and comments;

[0125] The preprocessing module is used to preprocess the source text and comments of public opinion information to obtain the node features of the propagation topology graph;

[0126] The topology map construction module is used to construct a propagation topology map according to the node characteristics of the propagation topology map;

[0127] The output module is used to output the propagation topology map as a state input to the policy network.

[0128] Text Input:

[0129] Collecting public opinion information source text and comments from online social networks; the public opinion information source text and comments include information source ID, information source comment ID, release time and forwarding comment master-slave relationship;

[0130] And according to the authenticity of the information source, the text and comments of the public opinion information source are divided into true public opinion information source and false public opinion information source;

[0131] The public opinion information source text and comments are organized into raw data and input into the data preprocessing and graph construction pipeline.

[0132] Preprocessing:

[0133] Text content is one of the important features in the identification of false public opinion information sources, and its length is usually less than 140 words. However, the text content in the public opinion information of most online social networking platforms is written very casually and is often mixed with a variety of symbols or emoticons. Therefore, before inputting text data into the training model, preprocessing is usually required.

[0134] The preprocessing of public opinion information text data follows the conventional processing flow of natural language processing, with only minor adjustments to a small part of it. A total of four sub-modules are set up: feature extraction, feature encoding, feature embedding, and feature averaging.

[0135] First, the information source and comment text are cleaned to remove mixed special symbols, expressions and punctuation marks (i.e., feature extraction).

[0136] Secondly, the meaningful words in the cleaned information source and comment texts are separated, and each separated source text or comment text will be transferred to an ordered word list (i.e., feature encoding).

[0137] Again, according to the pre-trained dictionary, all separated words are embedded into a vector space and a feature matrix is ​​generated, in which each column is the feature vector (i.e., feature embedding) of each word.

[0138] Finally, in order to reduce the data dimension, this preprocessing module uses feature averaging to compress the feature matrix into a feature vector with an average function, and obtains the node features of the public opinion information propagation topology map (i.e., feature averaging).

[0139] Topology map construction and output:

[0140] After data preprocessing to generate node features of the propagation graph (i.e., propagation topology graph), all node features will be organized into a graph structure;

[0141] Since public opinion information sources, comments and forwarding have temporal correlation, this correlation is converted into spatial correlation of graph structure, that is, the edges of the communication topology graph describe the temporal relationship of comments and forwarding;

[0142] As public opinion information spreads further over time, new information text nodes will be generated in sequence according to their timestamps and connected to the parent text as replies or comments;

[0143] This incremental construction method enables the application of process feature vectors to generate a complete online social network public opinion information propagation topology map, which will serve as the state input of the reinforcement learning algorithm (i.e., the policy network).

[0144] Action Space Design:

[0145] This paper introduces the configuration of action space in the early analysis method of false public opinion information in online social networks. Its architecture is as follows: Figure 8 shown.

[0146] The most important task of this model is to detect false public opinion events as early as possible. To accomplish this task, the strategy learned by the model (i.e., the strategy network) should first avoid making hasty decisions due to lack of information in the early stages of public opinion information dissemination. Later, when the model has enough information for decision-making, the learned strategy should be changed to determine the authenticity of the source of public opinion information on online social networks as early as possible.

[0147] To meet the above strategy learning requirements of the judgment model:

[0148] Set up three discrete actions: true public opinion information sources, false public opinion information sources, and unverified public opinion information sources;

[0149] The current public opinion information propagation topology map obtained from the state space is used as the state input. If the information source is identified as an unverified public opinion information source, the propagation process of the public opinion event will not be terminated, but will be rolled to the propagation topology map state of the next timestamp to continue to be iterated and identified by the reinforcement learning algorithm;

[0150] As the propagation time progresses, when the information source is judged to be a true public opinion information source or a false public opinion information source, the input iteration of the propagation topology map state will stop; the learning process of the early analysis model will be reset and restarted by another public opinion event to be judged.

[0151] System environment design:

[0152] This paper introduces the overall operating rules of the early identification method of false public opinion information on online social networks and the reward equation of the model for the intelligent agent, as follows:

[0153] Operation rules:

[0154] The reinforcement learning model randomly selects real public opinion events or false public opinion events from the training data set;

[0155] When the selected event propagation ends or the agent makes a final judgment, the analysis model will terminate all judgment processes of the current event and start the next public opinion event to be judged;

[0156] In addition, public opinion events on online social networks are accidental, contagious, and delayed.

[0157] In the process of simulating the spread of public opinion events, the proposed early analysis model chooses to determine the start time of the event in a random manner and stipulates a threshold to limit the maximum scale of the information dissemination topology graph;

[0158] While effectively restoring the sudden spread of false public opinion events and the time lag in judgment, it also further improves the judgment efficiency of the model.

[0159] Reward equation:

[0160] The reward (item) R of the early identification method of false public opinion information in online social networks based on reinforcement learning consists of two parts: decision reward r decision and decision time delay reward r time ;

[0161] Specifically, the reward item is set to enable the agent to effectively guide the Actor network to make similar decisions and avoid choosing the opposite wrong decision during the network training process after making the correct decision.

[0162] At the same time, every time the agent interacts with the environment, the reward obtained at that time step will include an r time , so that the intelligent agent can learn that long-term hesitation is not conducive to the early completion of false public opinion information analysis.

[0163] The reward R can be expressed as:

[0164] R=r decision +α·r time

[0165]

[0166] In the formula, α is the unified r decision and r time Parameters set by the unit; Δl is the maximum number of nodes in the current public opinion event propagation topology diagram.

[0167] This setting can normalize the maximum sum of rewards for the early analysis method of false public opinion information in online social networks based on reinforcement learning. No matter how the scale of the public opinion information propagation topology graph changes, the maximum reward value will not exceed 1.

[0168] In addition, in order to enable the agent to quickly determine the authenticity of the information source, r time To punish each additional decision step in the agent's judgment process, this value is relatively small in the early judgment model. Let r time =0.01.

[0169] Strategy Design:

[0170] Introducing the structure of the strategy network for early identification of false public opinion information on online social networks. Its overall architecture meets Fig. 9 .

[0171] The policy network architecture mainly consists of two modules: feature extractor and actor-evaluator architecture;

[0172] The feature extractor consists of a graph neural network module and a multi-layer fully connected neural network (MLP, Multi-Layer Perceptron) module.

[0173] Specifically, the graph neural network module is used to extract the main features of the public opinion event propagation topology graph, which is implemented by two graph convolutional neural network layers; while the multi-layer fully connected neural network module contains two fully connected layers, which is mainly responsible for further processing the features extracted by the graph neural network module;

[0174] The Actor-Critic framework consists of two parts: the Actor part and the Critic part;

[0175] Among them, the Actor part focuses on predicting the action distribution in the early analysis model of false public opinion in online social networks. The intelligent agent uses the policy gradient method to learn the strategy and then output the corresponding action;

[0176] The Critic part focuses on the estimation of the value function and is responsible for the evaluation of the action strategy.

[0177] In this implementation, the process of public opinion information dissemination is converted into a dynamic graph structure, and a reinforcement learning-based method for early analysis of false public opinion information in online social networks is proposed, which is conducive to shortening the response time of sudden public opinion events.

[0178] In this embodiment, in order to shorten the response time to public opinion events in online social networks, an early analysis method for false public opinion information in online social networks (based on reinforcement learning) is proposed, which combines the proximal policy optimization (PPO) algorithm with the actor--I-Critic method (AC) to judge the authenticity of the public opinion information source as early as possible based on the forwarded information; this analysis method is superior to the neural network learned by the supervised learning method in terms of accuracy and judgment speed, and further optimizes the real-time and accuracy of false public opinion information analysis.

[0179] It should be noted that the early analysis method for false public opinion information on online social network platforms proposed in this embodiment provides direction and strategy for the management and control of false public opinion information on online social network platforms. The rapid popularization of smart mobile terminals has led to a sharp increase in the dissemination channels of online social networks, and the scale of public opinion triggered by them has also gradually expanded. At present, online social network platforms have gradually evolved into the main position and battlefield of public opinion in cyberspace. Taking into account the impact of factors such as user traffic and attention competition between different online social network platforms on the process of public opinion management and control, the early analysis method for false public opinion information on online social network platforms and the corresponding management and control strategy proposed in this embodiment have important practical significance for maintaining corporate competitiveness.

[0180] It should be noted that the early analysis and judgment method for false public opinion information on online social network platforms proposed in this implementation method effectively improves the ability to intervene and guide online social network public opinion. In the new media environment, the emergence of online social network public opinion has had a profound impact on the public's views, attitudes and behaviors, which also poses a great challenge to how to conduct precise control. Through the exploration of the three communication perspectives of information, users and platforms, it is proposed to provide scientific decision-making suggestions and management opinions while further reducing the response time to public opinion, so as to better solve the existing control strategy basis, the systematic and scientific nature of the decision-making process and other issues, which has important practical significance for the realization of a "clear cyberspace".

[0181] It should be noted that the premise of formulating a control strategy for false public opinion information is to identify the large amount of false information existing in online social networks. Therefore, the early identification method of false public opinion information in online social networks starts from the information subject in the process of false public opinion information dissemination, and conducts in-depth research on the method of identifying the authenticity of public opinion information sources.

[0182] It should be noted that the spread of information and the spread of infectious diseases are similar to some extent. In the field of infectious disease transmission, the source of transmission refers to the biological individual that initially carries the virus or bacteria. Similar to the source of transmission, the false public opinion information content in online social networks is initially released by several node users, and then forwarded or commented by followers and spread. Among them, the information released by the initial several node accounts is the source of the public opinion. In addition, although the content of false public opinion information may exist in the form of pictures, the release of such pictures is only to increase the negative impact of text descriptions. Therefore, without loss of generality, it is assumed that the content of false public opinion information in online social networks is all carried by text.

[0183] It should be noted that, usually, there are three main ways to deal with infectious diseases: controlling the source of transmission, cutting off the route of transmission, and protecting susceptible populations. Drawing on the source of transmission control in infectious disease control methods, if we can accurately judge the false public opinion incident in the early stage of its occurrence, and accurately locate the source of false public opinion information on online social networks, then we can shorten the response time for handling the false public opinion incident. Furthermore, we can effectively utilize the "golden time" for handling false public opinion incidents to reduce the scale of the impact of the incident. Therefore, how to build an accurate and efficient algorithm to locate and identify the source of false public opinion information as early as possible has become a problem that needs to be solved urgently.

[0184] It should be noted that in recent years, graph neural networks have become a popular method applied in the field of information detection after recurrent neural networks and convolutional neural networks. Taking online social networks as an example, the network can be converted into a graph structure, in which nodes usually represent online social network accounts, and edges represent the connection between accounts (follow, forward or comment, etc.). Nodes and edges can convert online social network structure data into a topological graph, and then compile the structural information into the relationship between network entities. Each node represents a different account in the online social network, and the connection between nodes represents the forwarding and commenting relationship of the network account on a certain information. Therefore, the topological structure can be constructed to represent the propagation process of online social network accounts and information in the network, and it can be used as the input data graph of the information detection graph neural network.

[0185] There are four main types of graph neural network algorithms: graph neural network based on convolution, graph neural network based on recursion, graph neural network based on spatiotemporal data, and graph neural network based on autoencoder. Identifying the source of public opinion information based on the propagation network belongs to the graph classification task. Considering the amount of online social network information data and model efficiency, this implementation method selects a graph neural network based on convolution.

[0186] It should be noted that in the field of information detection, a graph convolutional neural network (GCN) is a network structure that can apply convolution operations to graph data, and its neural network structure is mainly composed of convolution operators and pooling operators. In the public opinion information source analysis task described in this embodiment, the convolution operator refers to a calculation function that can characterize the local data structure of online social network information nodes and reduce noise; the pooling operator is used to reduce the parameter level of the model, while simplifying the computational complexity of the convolutional network and maintaining a certain invariance of the features. From the above-mentioned online social network graph structure, it can be seen that the nodes are online social network accounts, and the edges are the information forwarding relationships between accounts. In the information propagation graph, GCN mainly aggregates the neighbor information of the target node and then learns the representation of the target node, while using the connecting edges to enrich the forwarding relationship information, so that the information source detection task has more non-sequential information.

[0187] The basic network structure of GCN: input a network structure of public opinion information dissemination into the GCN, extract its network features multiple times, so that the information dissemination network can be represented with fewer dimensions and parameters; then apply the nonlinear activation function ReLu to perform nonlinear transformation on the extracted features to increase the expression ability of the model, and obtain the output network structure after multiple training; among them, the hierarchical propagation rule of GCN is defined as follows:

[0188]

[0189] Where: σ——nonlinear activation function;

[0190] H——Feature information of each layer, i.e. hidden state matrix;

[0191] W——parameters to be learned;

[0192] A——adjacency matrix;

[0193] I——identity matrix;

[0194] —— The degree matrix of .

[0195] Since GCN has a high efficiency in the task of identifying the information propagation structure of social networks, this implementation method chooses to use GCN to locate and identify the source of false public opinion information in online social networks.

[0196] It should be noted that the theoretical basis for early analysis of false public opinion information on online social networks is reinforcement learning.

[0197] The theoretical knowledge of reinforcement learning is briefly explained here, and the principles of the actor-evaluator method and proximal strategy optimization algorithm are introduced in detail.

[0198] Basic principles of reinforcement learning:

[0199] The basic idea of ​​reinforcement learning (RL) originates from the study of the excitation and inhibition mechanism of human dopamine and the decision-making process of low-level biological behavior. It is one of the common methods of machine learning. Compared with deep learning, reinforcement learning is more inclined to find the optimal strategy to solve the problem. It is mainly composed of two parts: environment and agent. Among them, the environment represents the external environment. The environment in this embodiment is an online social network where a public opinion event continues to spread; the agent represents a specific algorithm. The agent in this embodiment is a module for identifying false public opinion information sources. The essence of reinforcement learning is to solve the problem of how the agent can obtain the maximum reward of the environment through learning strategies during the interaction between the environment and the agent. Based on discrete time, the environment interacts with the agent. The agent outputs a behavior to the environment through a set of strategies, and the environment will feedback the reward value to the agent, and the environment will transfer to the next state. When the state of the environment changes, the reward obtained by the current agent will also change, and the final strategy of the agent is generated from the collection of behaviors.

[0200] Basic principles of reinforcement learning: First, when the current system environment state is s t When the intelligent experience generates action a according to the current environment state and its own behavior strategy t , which will further react to the state of the system environment; secondly, the system environment is based on the action a of the intelligent agent. t p(s t+1 |s t , a t ) is iteratively updated to s t+1 state; at the same time, the system environment will give feedback to the agent reward r t , the agent will refer to the reward r t Further update and optimize its behavior strategy; finally, the agent continuously interacts with the environment, and the algorithm learns which behavior strategy to choose for different system environments in the process of interaction to obtain higher rewards.

[0201] Reinforcement learning is the most common method for solving optimization problems. Its optimal solution process can be summarized as: solving the optimal strategy - solving the optimal value function of a smaller-scale Markov decision process - optimizing the Bellman equation of a large-scale Markov decision process. As can be seen from the above, the reinforcement learning problem is ultimately transformed into the optimization of the Bellman equation. The specific methods that can be used are: Monte Carlo Method, Dynamic Programming, and Temporal Difference Method.

[0202] Actor-Critic (AC)

[0203] From the above overview of reinforcement learning principles, it can be seen that strategy, state, action and reward are the main components that reinforcement learning methods need to pay attention to. Among them, policy-based reinforcement learning methods are the most common. Strategy is defined as the probability that the agent chooses all potential actions based on the state of the system environment. The policy-based reinforcement learning method is an optimization method that aims to maximize the cumulative reward, parameterizes the strategy itself, and then finds the optimal parameters. Policy Gradient (PG, PolicyGradient) is currently the most commonly used method for policy-based reinforcement learning, that is, the agent will update and iterate according to the gradient of the total reward expected by its strategy with respect to the policy parameters, and finally gradually converge to the optimal strategy in the policy space.

[0204] The AC method is a reinforcement learning method that combines PG with value estimation. The algorithm training process is as follows: Figure 4 As shown in the figure, Actor represents the behavior strategy model, and the agent uses the PG method to learn the strategy and then output the corresponding action; Critic represents the value function model, which is responsible for evaluating the behavior strategy and approximating the value function, and further optimizing the Actor network. The TD-error is the difference between the updated iterations of the calculated value function, and its principle should satisfy the following formula:

[0205] TD Error=r(s t , a t )+v(s t+1 )-v(s t )

[0206] In the formula, r(s t , a t )——the reward value at time t; v(s t )——the value function at time t.

[0207] During the model iteration process, the Actor model outputs action a based on the current system environment state t , and then the system environment will act a according to the output of the agent t Give the corresponding reward r t And finally update and iterate to the next state. When a certain amount of data is obtained during training, the TD-error is calculated using the data and the system is fed back to update the corresponding parameters of the Actor model and the Critic model.

[0208] Proximal Policy Optimization (PPO) algorithm:

[0209] The wide applicability of the AC framework means that most system environment reinforcement learning algorithms with continuous or discrete actions can be optimized using this. However, the framework has limitations such as being sensitive to the optimization step size (convergence rate coefficient) and the large difference between the behavior strategy after the update iteration and the old behavior strategy, which makes it difficult for the algorithm under this framework to ensure monotonic convergence. The Trust Region Policy Optimization (TRPO) algorithm provides a monotonic policy improvement method. Based on the TRPO algorithm, the PPO algorithm further trims and restricts the policy function, simplifying the TRPO algorithm. The PPO algorithm solves the problem that the policy gradient algorithm in the AC framework is difficult to select a suitable step size and the difference between the new and old behavior strategies is large. At the same time, the PPO algorithm strikes a balance between ease of implementation, sample complexity, and ease of tuning, attempting to calculate an update in each step to minimize the cost function while ensuring that the deviation from the previous strategy is relatively small. The PPO algorithm framework is as follows: Figure 5 shown.

[0210] The PPO algorithm uses two networks: the evaluation network (Critic) and the policy network (Actor). Among them, the Actor network represents the strategy learned by the agent. Its input is the current system environment state, and the output result is the action probability distribution. When the PG algorithm is trained, the agent selects the corresponding action from the output action probability distribution of the Actor network and executes it, then interacts with the system environment to collect data and update the Actor network parameters. The Actor network loss function L PG (θ) satisfies the following formula:

[0211]

[0212] In the formula, π θ (a t |s t )——strategy function;

[0213] a t——Agent action at time t;

[0214] s t ——System environment status at time t;

[0215] ——Advantage function at time t.

[0216] The PPO algorithm uses the clip function to truncate the ratio of the new and old strategies, limiting the step size while controlling the distribution difference between the new and old strategies. The Actor network loss function L of the improved PPO algorithm is CLIP (θ) is as follows:

[0217]

[0218] In the formula, r t (θ)——the ratio of the new and old strategies at time t;

[0219] ∈——a hyperparameter, usually ∈=0.2.

[0220] The Critic network in the PPO algorithm is responsible for calculating the advantage function A t and the estimated state value function V(S t ). Its input is the current system environment state, and its output is the state value estimate. During training, the Critic network updates its network parameters by minimizing the loss function. The Critic network loss function L VF (θ) is as follows:

[0221]

[0222] In the formula,

[0223] V θ (S t ) — the state value estimated by the Critic network;

[0224] ——The value function of the target at time t.

[0225] In the above equation, The generalized advantage estimation algorithm (GAE) is used for calculation. This method can balance the bias and variance, and accelerate the convergence of the model while enhancing the stability of model training. The calculation formula is as follows:

[0226]

[0227] δ l-1 =r l-1 +γV(Sl )-V(S l-1 )

[0228] In the formula, δ l-1 ——Time series difference error; γ, λ——Adjustment factors of deviation and variance.

[0229] The pseudo code of the PPO algorithm is shown in the following table:

[0230]

[0231] It should be noted that the definition of false public opinion information analysis is:

[0232] Let the public opinion information dataset P = {p1,p2,···,p m}, where p i represents the i-th data in the data set, and m is the total amount of data in the data set. Any data in the data set can be represented by Indicates that i Represents p i The original text in the data, Indicates i The jth forwarded content, k i For p i The number of public opinion information contained in the data, and G i =(V i ,E i ) is composed of s i , ···, The topological map of public opinion information forwarding in online social networks.

[0233] The main task of solving the problem of false public opinion information is to i That is, to identify the source of public opinion information. According to the coarse-grained information definition method, public opinion information is divided into true public opinion information (T) and false public opinion information (F). In summary, the task of identifying the source of public opinion information is defined as learning a function f:f(p i )→y i , where the classification label value y i ∈{0,1}.

[0234] It should be noted that in recent years, with the increasing popularity of the Internet and mobile smart terminals, the way people exchange information has gradually changed from offline face-to-face interaction to online virtual social interaction, thus forming an online social network with the Internet as the medium. Among them, each node represents a registered user in the network, and the edge represents the virtual connection between nodes. The emergence of online social networks has made the connection between people in the social network closer, and the speed of information dissemination has been greatly improved.

[0235] Based on the context of online social networks, this implementation method defines online social network false public opinion information as "false, harmful, erroneous or even reactionary information, opinions or views related to a public event that is widely concerned by the general public and that are published and disseminated through online social networks, and these comments often have certain tendencies and influences."

[0236] To further refine the above definition, the public events that are widely concerned by the public and that are the focus of this implementation method refer to public safety events that occur suddenly, cause or may cause major casualties, property losses and major social impacts, and thus have a significant impact on social security and stability, such as public health events, emergency disaster safety events, etc. In addition, the "false" in the definition of false public opinion information on online social networks is mainly determined by identifying the authenticity of topics related to public events through the content of published clarification information and various types of erroneous information released in real time by the network platform management center.

[0237] Based on the above definition, the main characteristics of false public opinion information on online social networks can be summarized as follows: 1) Deliberately highlighting controversial elements. Controversial elements are most likely to attract the attention of users in online social networks. Therefore, some users and even platforms tend to highlight certain controversial elements when pushing messages to gain popularity; 2) Taking public information out of context. Online social network users have different prior knowledge of a public opinion event, which causes users with relatively less prior knowledge to have an inadequate understanding of the public information, and then distort it and publish opinions or views that are taken out of context; 3) Fabricating unconfirmed rumors. Some users will use sensational headlines to guide public sentiment based on the public opinion points that the general public is concerned about.

[0238] It should be noted that the purpose of online social network false public opinion research is to use rich information such as text content, comments and forwarding patterns on social media to identify false information in public opinion through a series of methods. At present, technicians in this field focus on the research of public opinion identification, trying to establish a priority and automatic false public opinion information identification method, thereby alleviating the pressure brought to society by public opinion events. Specifically, most of the current false public opinion information identification methods can be summarized into three categories, namely traditional methods, classic deep learning methods and graph neural network methods.

[0239] (1) Traditional method

[0240] Traditional methods for identifying false public opinion information mainly focus on manually extracted text features and classification.

[0241] Based on the preprocessing of tweets, technicians in this field create and fuse two dictionaries to generate feature vectors, and use support vector machines (SVM), K-nearest neighbors (K-NN) and their variants to evaluate the method. For example, Chu et al. used the security threat prediction of Internet public opinion (PSTIR) and the sociological false information propagation (SFIBS) to linearly analyze the proportion of trustworthy fans of a social platform that is popular in the early and future. At the same time, credibility networks, belief propagation (BP) and bounded trust models can also be used as traditional methods for the identification of false public opinion information.

[0242] (2) Methods based on convolutional neural network (CNN).

[0243] CNN is a feedforward neural network with a deep structure, consisting of one or more convolutional layers and a fully connected layer at the top. It is one of the representative algorithms of deep learning. CNN and its improved models are widely used in research fields such as natural language processing and computer vision.

[0244] Currently, most technicians in this field use CNN to extract hidden features of text content and visual information.

[0245] Yu et al. first introduced CNN in the field of false public opinion information identification, and extracted key features from the text content of related posts for identification. On this basis, previous research mainly focused on using CNN to identify text content. Based on the data encoding model of word-level data, Chen et al. proposed an attention residual network, which uses CNN to learn the text content in the information, and then obtain important information about local and global content features. At the same time, CNN based on gated recurrent unit network, CNN model with spatial attention mechanism, CNN based on deterministic factors and other discrimination methods have also been used in the task of public opinion event classification.

[0246] In the latest research, CNN is more used for public opinion information discrimination of pictures and short videos. For example, Zhou et al. used fake news text and visual content for embedding, and then used convolutional neural networks to extract them as feature maps and calculate cross-modal similarity to judge the inconsistency between text and image descriptions. Azri et al. proposed a deepMONI TOR model based on deep neural networks, which uses the text, images and emotions of posts to construct CNN for public opinion information discrimination. Many studies have demonstrated the effectiveness of CNN in extracting public opinion information features, and it can obtain local key features well. However, whether it is text data or propagation sequence, time information can play an important role in the discrimination of false public opinion information, and CNN just ignores this problem.

[0247] (3) Methods based on recurrent neural network (RNN).

[0248] RNN is a type of artificial neural network in which the connections between nodes form a loop and the output of some nodes affects the subsequent input of the same node. The biggest difference between it and convolutional neural network is that it retains the last state during the recursive process and retains time information. Therefore, most existing studies are keen on using recurrent neural networks for information text processing. Ma et al. first applied RNN to the task of identifying false public opinion information, modeled words based on TF-IDF, a common weighted technology for information retrieval and text mining, and used RNN to learn the potential content of information on a social platform. Since then, recurrent neural networks and their variants have been more widely used in the identification of false public opinion information, many of which are the latest research results.

[0249] At present, most studies use RNN to learn text features, image features and multimodal fusion features. Considering the user's attention to information, Singh et al. constructed a long short-term memory network (LSTM) and used text data in 13 different languages ​​and user characteristics to distinguish true from false information. Khattar et al. used multiple stacked bidirectional LSTMs to learn text features and integrate text and social background features. Based on this study, RNN and multi-layer attention models were applied to explore the contextual embedding of attention, and the word-level expression, sentence-level expression and user comment expression of news content were obtained, thereby extracting deep context information from limited text.

[0250] On this basis, some technicians in this field used the time-series feature of RNN to capture the time-series features in the spread of false public opinion information. Ma et al. first restored the information propagation structure to a tree structure and used a gated recurrent unit (GRU) to calculate each branch of the tree sequence. Subsequently, Wu et al. converted the news propagation tree structure in this study into time and used LSTM to represent and classify the propagation path of the message, which was used to solve the problem of malicious communicators disguising fake news as real news.

[0251] In addition, the combined architecture of CNN, RNN and LSTM is also frequently used by scholars. Kumar et al. proposed an LSTM tree based on convolutional units to predict the stance and authenticity of public opinion events in social media conversations. The source blog post is used as the root node and the response is used as the child node. Each node uses LSTM to learn the location of the text or comment. At the same time, more novel and excellent methods such as CNN-RNN, CNN-LSTM, CNN-BiLSTM, etc. are gradually being discovered by scholars. Although RNN can extract time information, the input of RNN is generally a sequence, which causes the structural information of the transmission to be destroyed, and the effect of extracting the structural features of the transmission is not good.

[0252] (4) Methods based on Graph Neural Network (GNN).

[0253] In recent years, GNN has received widespread attention. Social network is a common graph data type that represents the social relations between different individuals or organizations. Compared with CNN and RNN, GNN can better capture the propagation pattern and structure of false public opinion information. Xu et al. first used the GNN method to perform information discrimination tasks on a social platform. Subsequently, a large number of studies on GNN in information discrimination began to emerge.

[0254] At present, the research on GNN mainly focuses on extracting the structural features of information dissemination or the structural features of user interaction in social networks. Bian et al. simulated the propagation of public opinion and the propagation of diffusion direction through a two-layer bidirectional GNN to obtain structural semantics. On this basis, Wei et al. proposed an edge-enhanced graph convolutional neural network (GCN, Graph Convolutional Network) to further strengthen the structural and contextual features through the uncertainty of propagation. Monti et al. used stacked graph convolutional neural network layers to extract propagation structural features from heterogeneous graphs composed of multiple data such as users, user comments, and news dissemination, and automatically identified false public opinion information. In response to the problem of limited data, a semi-supervised fake news identification method has attracted the attention of scholars.

[0255] Some studies have considered propagation features, semantic information, and content features at the same time. For example, Yu et al. created a GCN model that simultaneously considered static features such as user information and text content, and dynamic features such as public opinion diffusion. Serveh Lotfi et al. extracted reply trees and user graphs from conversations in social networks and modeled them, thereby obtaining the structural information of the graph, the content of the conversation tweets, and the interaction information of users in the conversation.

[0256] At the same time, in order to integrate the graph structure and content of topics, posts and comments, methods such as Topic-Comment Graph Neural Network (TPC-GCN) and Decomposed TPC-GCN (DTPC-GCN) have also been further applied in this field. Among them, TPC-GCN integrates the graph structure and content of topics, posts and comments for dispute discrimination at the post level. DTPC-GCN splits the features related to topics and those not related to topics and dynamically merges them. At the same time, the combined modeling of CNN, RNN and GNN has also developed rapidly. For example, Li et al. innovatively proposed a novel multi-view structure-time learning framework MiSTR for false public opinion information discrimination, which jointly learns the temporal features of forwarding dynamics, the structural features of the propagation graph and the text features of the source tweets, and comprehensively represents the propagation structure between users from the micro and meso perspectives. Wei et al. designed a GNN with RNN components to handle the dialogue structure in social networks and learn content embedding through multi-task.

[0257] Most of the existing GNN models for false public opinion information identification require the input of the entire propagation tree graph for classification. However, it is crucial to effectively identify false public opinion information as early as possible before it causes serious social chaos. At present, research on early identification of false public opinion information has just started. For example, Song et al. regard all replies to false public opinion candidates as a sequence and propose a credible early public opinion identification model (CED, Credible Early Detection), thereby seeking an early time point for credible prediction; in addition, the introduction of models such as dynamic GCN with attention mechanism, GCN adjusted along the time dimension, and content graph joint recognition based on user preference perception have further enriched the methods of capturing the propagation structure and time information of false public opinion information, and further developed the research on early identification of information.

[0258] It should be noted that this field currently lacks early and accurate methods for analyzing false public opinion information in online social networks.

[0259] Current research focuses on using deep learning methods such as CNN, RNN, and GNN to study and judge each piece of public opinion information in online social networks. The research perspective is more focused on the microscopic perspective of the characteristics of a single piece of information itself. At the same time, due to the continuous development of online social networks, new forms of communication such as comments and forwarding have gradually emerged. Some scholars have begun to introduce the public opinion information dissemination structure diagram and apply it to judge the information source itself. However, most of the current research requires a complete dissemination structure diagram for judgment, which means that each judgment must wait until the end of the dissemination before being identified, which makes the public opinion strategy have a certain time lag. Therefore, how to identify non-public opinion information as early as possible and accurately during the dissemination process has become an urgent problem to be solved.

[0260] In this embodiment, in order to explore in detail the mechanism for analyzing and judging false public opinion information on online social networks, an early analysis method for false public opinion information on online social networks is proposed with the help of a reinforcement learning framework.

[0261] In this embodiment, in order to shorten the response time to false public opinion events on online social networks, this embodiment proposes an early judgment method for false public opinion information on online social networks (based on reinforcement learning), and learns a stable and efficient public opinion information discrimination strategy through uninterrupted interaction with the false public opinion information propagation simulation environment on online social networks. This embodiment realizes early false public opinion information judgment by exploring the structured information of the information source text and its forwarding and comments during the public opinion propagation process. Through the actor-evaluator reinforcement learning framework, the false public opinion information propagation process is represented as a dynamic graph structure and GNN is used to extract spatiotemporal features, and then the PPO algorithm is used to enable the information source to be accurately discriminated during its propagation process. Combined with the evaluation, the model proposed in this embodiment is superior to the neural network learned by the supervised learning method in terms of accuracy and discrimination speed. Through specific case analysis, the application of this model in rapid response to public opinion is specifically analyzed, which is conducive to shortening the response time to public opinion events.

[0262] It should be noted that false public opinion information on online social networks is mostly transmitted in text, and its characteristics conform to the specific mode of "feature extraction + classification model" in the learning-based text classification method. Therefore, this embodiment uses this method (i.e., the learning-based text classification method) to extract features from the online social network public opinion information database using the GCN network, and uses the PPO algorithm to accurately identify the source of false public opinion information.

[0263] In addition, in one embodiment, a specific example is provided for experiment and analysis:

[0264] The training results of the early identification method of false public opinion information on online social networks on a real data set will be presented to explore the effectiveness of the model. In order to prove that the model has excellent early identification function, the model is compared with the existing false public opinion information identification method, which fully verifies the identification effect of the proposed model.

[0265] Datasets and comparison methods:

[0266] Dataset:

[0267] Based on the fact that Sina Weibo is the most popular online social network platform in China in recent years, this paper uses a real dataset from Sina Weibo to explore the effectiveness of the early identification method of false public opinion information in online social networks based on reinforcement learning. This Weibo dataset uses binary labels to mark public opinion information, namely "real public opinion information" and "false public opinion information". All false public opinion information in the dataset comes from the Sina Community Management Center, which publishes various types of erroneous information in real time. On the basis of determining the false public opinion information, Sina Weibo's API can capture all forwarding or comments on the false public opinion information, thereby forming a complete dataset of the dissemination process of false public opinion information. In addition, the real public opinion information is collected by capturing the published content that has not been reported as false public opinion information. The details of the Sina Weibo dataset are shown in Table 1:

[0268] Table 1 Dataset statistics

[0269] statistics Sina Weibo Number of users 2746818 Total Weibo 3805656 Number of source microblogs 4664 Number of real public opinion information sources 2313 Number of sources of false public opinion information 2351 Average time span of source microblogs 2460.7 hours Average number of forwardings of source microblogs 816 Maximum number of forwardings of source microblogs 59318 Minimum number of forwardings of source microblogs 10

[0270] Comparison method:

[0271] We selected the early identification method of false public opinion information on online social networks, and the false public opinion information identification model based on RNN, GCN, Graph Attention Network (GAT) and Graph SAGE (Graph Sample AndAGGregate) as the experimental comparison method. In addition to the proposed (reinforcement learning-based) early identification method of false public opinion information on online social networks, the detailed descriptions of other comparison methods are as follows:

[0272] (1) RNN: The public opinion information is discriminated by combining RNN with CNN. That is, the information is vectorized through CNN and input into the RNN model for discrimination.

[0273] (2) GCN: Considering the correlation between the information released by online social network accounts in public opinion events, the hidden layer representation of the nodes is obtained through GCN model learning, and then the nodes are classified to finally complete the judgment of false public opinion information.

[0274] (3) GraphSAGE: Based on the idea of ​​GCN neighbor aggregation, it does not aggregate all neighbors, but only aggregates some neighbors. Its essence is to sample and generate small subgraphs for training and local update.

[0275] (4) GAT: Based on the GNN model to generate and identify the public opinion information propagation topology map, it learns the different levels of attention to information text.

[0276] In addition, to simplify the description, the early analysis methods for false public opinion information on online social networks proposed below are all simplified using RL.

[0277] Evaluation indicators:

[0278] The model's discrimination of true and false public opinion information will be evaluated separately, and the specific evaluation indicators are accuracy, precision, recall and F1-Score, and the above evaluation indicators are calculated by the confusion matrix elements. The proposed RL model has three types of decisions: judged as a true public opinion information source, judged as a false public opinion information source, and unverified public opinion information source. Since there are only two possibilities for public opinion information sources in reality, namely true or false, all information sources in the model that cannot be judged at present and need to be further verified are predicted errors, that is, the confusion matrix size of the model is 2×2. On this basis, the confusion matrix elements of the RL model are TP, TN, FP, and FN. Specifically, TP is True Positive, which means that the prediction is a positive example (P) and the prediction is correct (T); TN is True Negative, which means that the prediction is a negative example (N) and the prediction is correct (T); FP is False Positive, which means that the prediction is a positive example (P) but the prediction is wrong (F); FN is False Negative, which means that the prediction is a negative example (N) but the prediction is wrong (F). The above positive examples and negative examples are relative, that is, if the model's discriminant effect on real public opinion information is evaluated, the positive example is the real public opinion information, and the negative example is the false public opinion information. The evaluation of the discriminant effect on false public opinion information is just the opposite. The calculation methods of the five evaluation indicators are as follows:

[0279] (1) Accuracy

[0280] Accuracy refers to the proportion of the model's correct predictions to the total number of classified samples, and is a statistic for all samples. This evaluation index satisfies the following formula:

[0281]

[0282] Accuracy can determine the overall correctness of the model, and the higher the accuracy, the better the model performance. However, although accuracy is the simplest and most intuitive evaluation indicator in classification problems, it still has obvious defects. If there is a sample imbalance problem in the data set, such as 90% positive samples and 10% negative samples, the serious imbalance of sample data means that the model only needs to predict all samples as positive samples to get a high accuracy of 90%, which makes the high accuracy result of the model actually inaccurate, and the category with a large proportion of data samples often becomes the most important factor affecting the accuracy. Therefore, other evaluation indicators are still needed to comprehensively evaluate the performance of the model.

[0283] (2) Precision

[0284] Precision is defined as the ratio of samples that are actually positive to all samples that are predicted to be positive. This evaluation index is a statistic for some samples, focusing on the statistics of data predicted by the model as positive. The formula for calculating this evaluation index is as follows:

[0285]

[0286] The calculation methods of precision and accuracy are similar, but the concepts of the two are completely different. Precision represents the accuracy of the model's prediction of the positive sample prediction results, while accuracy describes the accuracy of the model's prediction of the entire data sample, including both positive samples and negative samples.

[0287] (3) Recall

[0288] The recall rate indicates the ratio of the number of correctly predicted positive samples to the number of true positive samples. This evaluation index is also a statistic for some samples, focusing on the statistics of true positive samples. The calculation formula of the recall rate can be expressed as:

[0289]

[0290] The higher the recall rate of the model, the higher the probability that the sample that is actually positive is predicted. As can be seen from the above, precision and recall are a pair of contradictory and unified indicators. Generally speaking, in order to improve the precision rate, the model needs to predict the sample as a positive sample when it has "higher confidence" as much as possible, but it often misses many "uncertain" positive samples because the model is too conservative, which leads to a decrease in recall rate. Therefore, it can be inferred that when the precision rate of the model is high, its recall rate will be relatively low; conversely, when the recall rate of the model is high, its precision rate will be low. In addition, when the model implements some simple tasks, it may also have high precision and high recall rates at the same time.

[0291] (4)F1-Score

[0292] F-Score is a combination of precision and recall, and is a suitable metric for measuring the quality of a binary classification model. Its expression is as follows:

[0293]

[0294] Where β is the adjustment parameter of the importance of precision and recall.

[0295] It can be seen that F-Score is the harmonic mean of precision and recall, and the importance of precision and recall is adjusted by parameter β. Let β = 1, that is, precision and recall are equally important in the RL model. The F-Score indicator with a β value of 1 is called F1-Score. When the model precision or recall rate decreases, the value of F1-Score will decrease; conversely, when any value of the model precision and recall rate increases, the value of F1-Score will also increase.

[0296] Parameter settings:

[0297] The experiment used Nvidia RTX3070 graphics card and Pytorch deep learning framework for training. The dataset was divided into training dataset and test dataset in a 1:1 ratio, and Spacy was used to preprocess the dataset. During the preprocessing process, the word embeddings of all public opinion information texts in the dataset were matched one by one with the word vectors in the corresponding domain corpus. The RL model was mainly implemented using PyTorch. The training parameters of the model are shown in Table 2:

[0298] Table 2 Model parameters

[0299]

[0300]

[0301] The main structures of the comparison methods RNN, GCN, GAT, and GraphSAGE are all composed of a graph neural network layer, a maximum pooling layer, and a fully connected layer. By changing the internal structures of different graph neural network layers, PyTorch is used to implement the above comparison methods.

[0302] Experimental results:

[0303] Based on the above parameter settings, the proposed RL model is trained, and the training effect is as follows Fig.10 As shown in the figure, the initial reward value of the model is -0.5. This is because the initial strategy in the model is randomly generated. As the reinforcement learning training process continues, its cumulative reward value will gradually increase with the ongoing strategy learning and converge in about 2000 seconds. The cumulative reward value of the model always remains around 0.95 and is relatively stable until the end of learning.

[0304] The proposed RL model and its comparative method will be evaluated from two aspects: macro detection performance evaluation and early detection performance evaluation. The macro detection performance evaluation indicators are mainly the accuracy, precision, recall and F1-Score, while the early detection performance evaluation is reflected by the impact of changes in the number of public opinion information judgment observations and the number of public opinion information source comments on the performance of the model. Among them, the number of public opinion information judgment observations indicates the number of times the model needs to correctly judge the authenticity of the information continuously. Only when the number of consecutive correct judgments reaches the number of judgment observations, the judgment result will be truly output. The number of public opinion information source comments refers to the number of comments that have been generated during the dissemination of the information source. The two performance evaluations are specifically shown as follows:

[0305] Macroscopic detection performance evaluation:

[0306] Fig.11 The accuracy of each model in identifying public opinion information is demonstrated. It can be seen that the proposed RL model has the highest recognition accuracy. Among the comparison methods, GCN's accuracy performance is relatively better than the other three methods, GAT and GraphSAGE have the same accuracy, and RNN's accuracy is the lowest among the five methods. Based on the accuracy index, although the proposed RL model is better than various comparison methods, its advantage is not very obvious. Since the accuracy index is a relatively broad standard, it can only simply reflect the detection level of the model and has obvious defects, so other evaluation indicators are still needed to assist in further evaluating the performance of the model.

[0307] On this basis, the specific values ​​of precision, recall, F1-Score and other indicators of the RL model and its comparison method on the Sina Weibo dataset were calculated, and the performance of all models in discriminating public opinion information was divided into two categories: true and false public opinion information for evaluation. The results are shown in Tables 3 and 4 respectively.

[0308] Table 3 Experimental results of real public opinion information

[0309] Comparison Method Accuracy Recall F1-Score RL 1.0000 0.8729 0.9322 GCN 1.0000 0.8560 0.9224 GAT 1.0000 0.8505 0.9192 GraphSAGE 1.0000 0.8516 0.9199 RNN 1.0000 0.8090 0.8944

[0310] The experimental results show that the proposed RL model has the best performance in the process of discriminating real public opinion information in terms of precision, recall and F1-Score, which are 1, 0.8729 and 0.9322 respectively. It can be found that the model has relatively high discrimination accuracy and is not too conservative to miss too much real public opinion information. This is because the model has added a reinforcement learning architecture, which not only learns the text's own features but also further simulates the propagation process of public opinion information in the reinforcement learning interaction mechanism, which enables the model to better collect all features in the information propagation process and then accurately discriminate public opinion information. It is worth noting that the precision of the proposed RL model and its comparison method is 1, which means that all models have no false public opinion information being misjudged as real public opinion information. In contrast, GCN has achieved relatively excellent index values ​​in its recall and F1-Score among many comparison models, while the RNN method has the worst recognition performance. This is because the RNN method does not consider the information propagation topological graph structure, but focuses more on extracting the text features of the information itself, which limits its accuracy in information discrimination.

[0311] Table 4 Experimental results of false public opinion information

[0312] Comparison Method Accuracy Recall F1-Score RL 0.8600 0.9060 0.8824 GCN 0.8476 0.8941 0.8702 GAT 0.8446 0.9020 0.8723 GraphSAGE 0.8464 0.9007 0.8727 RNN 0.8060 0.9369 0.8665

[0313] As can be seen from Table 4, in terms of discriminating false public opinion information, the RL model proposed in this chapter is still ahead of all the comparison methods in terms of precision and F1-Score indicators, but the recall rate indicator is slightly lower than the RNN method. This phenomenon shows that the proposed model still has better discrimination performance than all the comparison methods, but it is slightly more conservative than the RNN method in the process of discriminating false public opinion information. Furthermore, unlike the discrimination of real public opinion information, the performance of the comparison methods does not show a one-sided trend, that is, GCN still maintains an advantage in precision, while GraphSAGE's F1-Score indicator is even better. Although RNN performs best in terms of recall rate, its precision and F1-Score are still far behind other comparison methods.

[0314] Comparing the experimental results in Table 3 and Table 4, it can be seen that when different types of public opinion information are used as the discrimination target, the performance of each model is different. Among them, the precision and F1-Score of the discrimination of true public opinion information are higher than those of false public opinion information, but the recall rate is relatively low, which shows that the RL model and its comparison method are relatively aggressive in the discrimination of false public opinion information, that is, if the model believes that the information has a tendency to become false public opinion information, it will be directly judged as false public opinion information, which will cause some real public opinion information to be judged as false. This result verifies the central idea of ​​the mentioned early judgment model, that is, the public opinion information judgment model should strive to shorten the judgment time while trying to identify all false public opinion information as much as possible, so as to achieve "all judgments should be made". In summary, the proposed RL model performs better than all the comparison methods, which fully verifies the effectiveness of its recognition performance.

[0315] Early Detection Performance Evaluation:

[0316] The number of discriminant observations and the number of comments on public opinion information sources are selected as the influencing factors of early research and judgment performance, and sensitivity analysis is used to explore their respective impacts on model performance. The performance of the model can be expressed using the above-mentioned macro-detection performance evaluation indicators. The details are as follows:

[0317] (1) Sensitivity analysis of the number of comments on public opinion information sources

[0318] The impact of the number of comments on public opinion information sources on model performance can be further understood as the proportion of correct judgments that the model can make before the propagation reaches the specified number of comments on public opinion information sources. Under the specified number of comments on public opinion information sources, the higher the accuracy of the model's judgment; or under the same accuracy requirement, the fewer comments on information sources the model needs, both indicate that the model's early judgment performance is better. The precision rate in the macro-detection performance evaluation index is selected to measure the impact of the number of comments on public opinion information sources on the discrimination of true public opinion information and false public opinion information. The specific performance is as follows: Fig.12 shown.

[0319] Depend on Fig.12 (a) It can be seen that the proposed RL model is lower than all the comparison methods in the initial stage of public opinion propagation. It does not perform better than GCN until the propagation scale reaches 175 public opinion information source comments. As the number of public opinion information source comments increases, it approaches GAT and GraphSAGE, but is still slightly lower than RNN. The possible reasons for this phenomenon are as follows:

[0320] (1) The proposed model contains three actions, namely, true public opinion information sources, false public opinion information sources, and unverified public opinion information sources. In the process of identification, some false public opinion information sources will be identified as unverified. The model will wait for the advancement of the propagation process to collect more features and re-evaluate and judge, and this mechanism pays more attention to the accuracy of identification.

[0321] (2) In fact, whether in the dataset or in real scenarios, the number of comments on online social network public opinion information sources is mostly higher than 200. Fig.12 (a) It can be observed that when the number of comments on the public opinion information source is greater than 200, the accuracy of the proposed model is not much different from that of the comparison method.

[0322] The above two aspects show that although the proposed RL model will be partially affected in the initial stage of public opinion information dissemination, this impact is not significant in its actual application.

[0323] In contrast, Fig.12 (b) The trend shown is consistent with Fig.12 (a) Completely different. The proposed RL model performs much better than all the comparison methods in discriminating real public opinion information from the beginning of the spread, and its accuracy will increase slightly with the increase of the number of comments on the public opinion information source. As the spread of public opinion expands, when the number of comments on the public opinion information source reaches about 100, the accuracy of the model begins to remain stable. In addition, the growth trends of GCN, GAT, and GraphSAGE in the comparison methods are similar. They all begin to maintain stability after the number of comments on the public opinion information source reaches about 150, and their accuracy rates are not much different; while the performance of RNN is relatively poor, and its accuracy rate is still far behind the above four methods.

[0324] From the above results, it can be seen that as the number of comments on the public opinion information source increases, the discrimination of false public opinion information is generally better than that of true public opinion information. This is because the comments of false public opinion information sources have relatively obvious characteristics. With the same number of source comments, the basis for discrimination is clearer, making the source discrimination easier and more accurate. However, the sensitivity of the detection of true public opinion information sources to the number of comments is relatively low, and it will not increase significantly with the increase of the number of comments. Therefore, the time for discriminating public opinion information can be further shortened.

[0325] (2) Sensitivity analysis of discriminant observation number

[0326] The number of discriminant observations can be further interpreted as the strictness of the model's judgment performance. A high number of discriminant observations means that the model needs to correctly judge the information source more times in succession before it can output the judgment result. This also means that the model's judgment time is longer and the recognition output is more difficult. Accuracy, precision, recall and F1-Score are used to describe the impact of the number of public opinion information discrimination observations on the discrimination of true public opinion information and false public opinion information.

[0327] Fig.13The accuracy of the proposed RL model and other comparison methods under different discriminant observation numbers is shown. The RL model outperforms all comparison methods in terms of accuracy. Even though its accuracy steadily decreases with the increase in the number of discriminant observations, it is still much higher than all other comparison methods. Fig.14 The paper shows the effect of the number of observations on the precision, recall and F1-Score of the RL model and other comparison methods, with the goal of judging true and false public opinion information. Fig.14 (a) and Fig.14 (b) It can be seen that as the number of judgment observations τ increases, the accuracy of the false public opinion information of the proposed RL model gradually decreases from about 0.98 to about 0.8; the accuracy of the true public opinion information remains unchanged at 1. The recall rate of the RL model is exactly the opposite of the accuracy rate. The recall rate of the false public opinion information only decreases slightly with the increase of the number of judgment observations, but the recall rate of the true public opinion information decreases by about 8% with the increase of the number of judgment observations. Fig.14 (e) and Fig.14 (f) It can be seen that the F1-Score index of the proposed RL model for both true and false public opinion information decreases slowly with the increase of the number of discriminant observations, among which the judgment performance of false public opinion information is more sensitive to the number of discriminant observations. In contrast, the precision, recall and F1-Score index of the GCN, GAT, GraphSAGE and RNN methods first increase and then decrease with the increase of the number of discriminant observations, showing a hillock-shaped trend and the floating range does not exceed 1%. It can be seen that all the comparison methods are insensitive to the number of discriminant observations. This is because compared with the proposed RL model, the training process of these methods discretizes the public opinion information propagation process of online social networks and inputs all the propagation topology graphs of all time steps into the neural network. There is a lack of reinforcement learning process learning, so it is impossible to learn the dynamic characteristics of the information propagation process, resulting in its relatively weak early judgment ability.

[0328] In this implementation, the proposed identification method is compared with the existing public opinion identification algorithm to verify that the proposed judgment model is superior to the graph neural network and recurrent neural network learned by supervised learning methods in terms of accuracy and judgment speed. On this basis, the real-time and accuracy of false public opinion information judgment is further optimized.

[0329] In this embodiment, it can be seen from the above experiments that the proposed early analysis model for false public opinion information in online social networks is able to accurately identify public opinion information in the early stages of dissemination, and is simultaneously compared with GCN, GAT, Graph-SAGE and RNN methods, which further verifies the excellent performance and stability of the model, and has strong versatility, confirming the possibility of further application and promotion in reality.

[0330] In addition, in one implementation, a specific example is provided, taking a public opinion event (an explosion at a city port) as a case, analyzing the impact of the analysis and judgment mechanism on the dissemination of public opinion, and pointing out how the proposed analysis and judgment model can better help reduce the response time of public emergencies.

[0331] Case selection and description:

[0332] Reasonableness of case selection:

[0333] The case of a particularly serious fire and explosion accident at a dangerous goods warehouse of a company in a certain city port was selected as a representative case. In this explosion incident at a certain city port, social network media played a role while official authoritative information was missing, leading to the spread of false public opinion information and the spread of panic.

[0334] In response to the explosion, the technical defects of the information collection and discrimination methods in the social network environment are consistent with the problem of information source research. Therefore, for the core research problem, this case is typical and effective, which can well prove the significance of the proposed model.

[0335] Case public opinion evolution trend:

[0336] A dangerous goods explosion occurred in a company in a development zone in a new district of a city, resulting in 165 deaths and 798 injuries. Due to the special nature of the incident and its severe consequences, it attracted widespread attention from users of social platforms. A large amount of false public opinion information on online social networks spread rapidly on social media. By querying and sorting out relevant reports from official authoritative media, the evolution of public opinion on the incident was restored.

[0339] Incubation period: The first piece of news about the "port explosion in a certain city" was first released by a social platform and immediately attracted intensive attention from users of the social platform. Subsequently, mainstream media took the lead in reporting it in full. The online public opinion began to show a crisis.

[0340] Outbreak period: Generally, crisis public relations is most effective within 4 hours of an emergency. Otherwise, the negative impact of the incident will be amplified, and related opinions will erupt in a short period of time. At this time, the dissatisfaction of social platform users reaches a breaking point, and they express an urgent need to clarify the cause and result of the incident, and ask questions and doubts.

[0341] Diffusion period: After the accident, various false public opinion information emerged, seriously affecting the credibility of the party and causing it to fall into the "Tacitus Trap".

[0342] Decline period: One week after the accident, timely disclosure of information dispelled some false public opinion information. The intervention of mainstream media and clarification of related doubts made the emotions and attitudes of social platform users flat and decline, and negative comments continued to decline. The online public opinion crisis was thus alleviated and gradually dissipated.

[0343] Questions and reflections:

[0344] From the practice of false public opinion on online social networks during the "port explosion in a certain city", we can see that the main reasons for the widespread spread of false public opinion information are:

[0345] (1) The hollowing out of official accounts is quite serious

[0346] In terms of value, although the number of official accounts has increased significantly with the increasing popularity of the Internet, daily maintenance and operation are still not taken seriously, and utilitarianism and hollowing out are serious. This phenomenon has resulted in the official account being unable to serve as a bridge for daily communication and interaction with social platform users when sudden public opinion events such as the "port explosion in a certain city" occur, and lacking information interaction with social platform users. While being unable to follow up on the progress of the event in a timely manner, there is also the risk of intensifying conflicts and increasing the spread of false public opinion.

[0347] (2) The technology for automatically analyzing hot public opinion issues is not yet mature

[0348] On the technical level, as a newly developed media technology, official accounts provide a new way of communication, but there are still technologies such as automated identification, judgment and analysis of public opinion responses that need to be developed and improved. Only official accounts with corresponding technical guarantees can play their due functionality and value. In the "port explosion in a certain city", there was a lack of corresponding technical means to conduct early judgment on the public opinion information released by users of various social platforms in online social networks and the comment information under official accounts, resulting in the local media taking action nearly 4 hours after the public emergency occurred. The failure to discover the focus of social platform users within the critical time allowed the widespread spread of false public opinion information, which in turn caused the panic of social platform users to spread.

[0349] (3) The organic connection between hot public opinion analysis and rapid guidance mechanism needs to be improved

[0350] As the "information gathering place" of the new era, online social networks have the function of timely collecting a large number of opinions from social platform users and establishing effective dialogues. Due to the lack of means to collect information on emergencies, the speed of related investigations is far behind, and the postponement of press conferences has reduced the effectiveness of clarifying and responding to public opinion issues, and the effect of public opinion guidance is not good.

[0351] Management and control strategies:

[0352] Based on the above reflections on the case, combined with the proposed model (method) for early identification of false public opinion information on online social networks, this paper specifically analyzes how the model should be organically connected with the rapid response mechanism for sudden public opinion events and applied to shorten the response time of public opinion events. Fig.15 shown.

[0353] The rapid response mechanism for sudden public opinion incidents that integrates the analysis and judgment method mainly consists of four parts, namely information collection, public opinion analysis and judgment, analysis and presentation, and crisis management. The details are as follows:

[0354] (1) Information Collection

[0355] Since false public opinion information has the characteristics of time, space, diversity, and mobility, it is necessary to follow the principles of effectiveness, continuity, and exhaustiveness, and focus limited energy on collecting continuous network information under hot public opinions on social platforms. In recent years, with the continuous change of online social network information processing methods, data mining, artificial intelligence, and big data analysis can be used to obtain public opinion information from social network platforms such as search engines, online forum communities, social platform public accounts, and news media, providing a huge data foundation for public opinion judgment.

[0356] (2) Public Opinion Analysis

[0357] An effective public opinion analysis mechanism can provide efficient support for decision-making and assist in timely response to emergencies.

[0358] The process of the public opinion analysis mechanism mainly includes inputting the collected public opinion information, judging the authenticity of the public opinion information through the proposed online social network false public opinion information early analysis model, evaluating the accuracy of the judgment, and updating the result to the early analysis model data set. The implementation of the public opinion analysis mechanism needs to be embedded in major forums by building false public opinion information early analysis terminals. At present, many companies and platforms have taken the lead in realizing the deployment and application of public opinion analysis. However, most public opinion information monitoring is based on information collection and text mining technology to realize online social network public opinion status monitoring and news hot spot tracking, news supervision and other functional requirements, and the analysis of false public opinion information is relatively lacking. This is also the purpose of studying the false public opinion information analysis mechanism, that is, to provide reliable technical support for realizing a clear online social network environment.

[0359] (3) Analysis and presentation

[0360] When obtaining the results of the false public opinion information analysis on online social networks, the subject content and spatiotemporal characteristics of the false public opinion information will also be obtained. Personnel with network expertise should be arranged to closely monitor the "false public opinion information data" sent back by the false public opinion monitoring terminal, analyze the theme, core issues, development timeline, positions and views of relevant parties, and the credibility and authority of the analysis results. Finally, a corresponding analysis report is formed and submitted to the establishment of goals for crisis management.

[0361] (4) Crisis Management

[0362] Establishment of disposal goals. In the traditional model, the establishment of disposal goals is mostly based on personal experience or subjective judgment, lacking real-time goal setting. Under this mechanism, the online social network false public opinion information early analysis model proposed in this paper can grasp the status of public opinion in real time, making the task establishment and disposal direction both proactive.

[0363] Formulate emergency response plans. According to the established public opinion handling goals, it is necessary to clarify the specific handling methods, the strategic tools used, resource allocation and scheduling plans, etc. In addition, a collaborative management and control mechanism can be established, which can provide relevant knowledge, cases and local experience in real time through the system platform, thereby further supplementing and improving the existing emergency response plans.

[0364] Plan execution and feedback. According to the emergency plan for public opinion control of emergencies, online social networks can be used to carry out online and offline joint emergency response. Online response includes using official accounts to declare problems, educate social platform users, and dialogue with social platform users. In contrast, offline response includes post-event aftermath of emergencies, holding press conferences, and community publicity and guidance. Through the false public opinion information dissemination model, the development trend of public opinion is simulated, the plan is adjusted in a timely manner and feedback is given to each action subject to ensure that the public opinion can be handled quickly and accurately.

[0365] In this implementation, the application of the model in rapid response to public opinion is specifically analyzed through the case analysis of "an explosion incident at a port in a certain city", which is conducive to shortening the response time to public opinion incidents.

[0366] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation modes and equivalent substitutions based on the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An early analysis method for false public opinion information on online social networks, characterized in that: The method comprises the following steps: Steps for building a detection framework for early detection models of false public opinion information on online social networks; The detection framework includes an agent module and a system environment module, wherein: The system environment module is used to generate a new source text or comment of public opinion information in each interaction process; it is also used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback; The intelligent agent module is used to compile the input public opinion information source text and comments in the online social network into a propagation topology graph; The intelligent agent module is used to transfer the compiled propagation topology map to the strategy network, and classify the data of the propagation topology map until a final judgment is made; Steps for constructing the state space to support the reinforcement learning process; Steps for configuring the action space of the model for early identification of false public opinion information on online social networks; Steps for configuring system environment modules; Steps for configuring a policy network for an early identification model of false public opinion information on online social networks; A step for obtaining a configured online social network false public opinion information early analysis model according to the above steps, and obtaining a final analysis result.

2. The method for early identification of false public opinion information in online social networks according to claim 1 is characterized in that: The steps for constructing a state space to support a reinforcement learning process include the following steps: It is used to build a data preprocessing and graph construction pipeline, and compile a propagation topology graph from an online social network public opinion information text dataset; The data preprocessing and graph construction pipeline includes an input module, a preprocessing module, a topology graph construction module, and an output module; The input module is used to input public opinion information source text and comments; The preprocessing module is used to preprocess the source text and comments of public opinion information to obtain the node features of the propagation topology graph; The topology map construction module is used to construct a propagation topology map according to the node characteristics of the propagation topology map; The output module is used to output the propagation topology map as a state input to the policy network.

3. The method for early identification of false public opinion information in online social networks according to claim 2 is characterized in that: The preprocessing module includes the following submodules: Feature extraction submodule: used to clean the text and comments of public opinion information sources and delete mixed special symbols, expressions and punctuation marks; Feature encoding submodule: used to separate the meaningful words in the source text and comments of the cleaned public opinion information. Each separated source text or comment text will be transferred to an ordered word list; Feature embedding submodule: used to embed all separated words into a vector space according to the pre-trained dictionary and generate a feature matrix, in which each column is the feature vector of each word; Feature averaging submodule: used to reduce data dimension, use feature averaging processing to compress the feature matrix into a feature vector with an average function, and obtain the node features of the propagation topology diagram of public opinion information.

4. The method for early identification of false public opinion information in online social networks according to claim 1 is characterized in that: The steps for configuring the action space of the online social network false public opinion information early analysis model include the following steps: Set up three discrete actions: true public opinion information sources, false public opinion information sources, and unverified public opinion information sources; Get the propagation topology of the public opinion information currently obtained from the state space as the state input, and determine its action type: If it is identified as an unverified public opinion information source, the propagation process of the public opinion event will not be terminated, but will be rolled to the propagation topology state of the next timestamp and continue to be iterated and identified by the reinforcement learning algorithm; As the propagation time progresses, when it is judged to be a true public opinion information source or a false public opinion information source, the input iteration of the propagation topology map state will stop; the learning process of the early analysis model will be reset and restarted by another public opinion event to be judged.

5. The method for early identification of false public opinion information in online social networks according to claim 1 is characterized in that: The steps for configuring the system environment module include the following steps: Steps for configuring run rules: The reinforcement learning model randomly selects real public opinion events or false public opinion events from the training data set; When the selected event propagation ends or the agent makes a final judgment, the analysis model will terminate all judgment processes of the current event and start the next public opinion event to be judged; The starting time of online social network public opinion events is determined randomly, and a threshold is set to limit the maximum size of the propagation topology graph; Steps for configuring the reward equation: The reward R is composed of two parts: decision reward and decision time delay reward ; The reward R is expressed as: In the formula, α is the unity and Parameters set by the unit; The maximum number of nodes in the current public opinion event propagation topology graph; use To punish each additional decision step in the agent's judgment process, this value is relatively small in the early judgment model. = 0.

01.

6. The method for early identification of false public opinion information in online social networks according to claim 1 is characterized in that: The step of configuring a policy network for an early analysis model of false public opinion information on online social networks; The policy network consists of two modules: a feature extractor and an actor-evaluator architecture; The feature extractor consists of a graph neural network module and a multi-layer fully connected neural network module; The graph neural network module is used to extract the main features of the public opinion event propagation topology graph, which is implemented by two graph convolutional neural network layers; The multi-layer fully connected neural network module contains two fully connected layers, which are used to further process the features extracted by the graph neural network module; The actor-evaluator framework consists of two parts: the Actor part and the Critic part, where: The Actor part is used to predict the action distribution in the early analysis model of false public opinion information on online social networks. The agent uses the policy gradient method to learn the strategy and then output the corresponding action; The Critic part is used to estimate the value function and is responsible for evaluating the action strategy.

7. An early analysis device for false public opinion information on online social networks, characterized in that: The device comprises the following modules: A module for building a detection framework for early analysis of false public opinion information on online social networks; The detection framework includes an agent module and a system environment module, wherein: The system environment module is used to generate a new source text or comment of public opinion information in each interaction process; it is also used to evaluate the actions of the intelligent agent in the process of public opinion information dissemination and give the intelligent agent a certain reward as feedback; The intelligent agent module is used to compile the input public opinion information source text and comments in the online social network into a propagation topology graph; The intelligent agent module is used to transfer the compiled propagation topology map to the strategy network, and classify the data of the propagation topology map until a final judgment is made; Modules for constructing state spaces to support the reinforcement learning process; A module for configuring the action space of the early detection model of false public opinion information on online social networks; Modules for configuring system environment modules; A module for configuring the policy network of the early identification model of false public opinion information on online social networks; A module used to obtain the configured online social network false public opinion information early analysis model according to the above steps and obtain the final analysis result.

8. A computer device comprising: A processor and a memory, characterized in that the memory is used to store executable instructions of the processor, and the processor is configured to execute the method for early analysis of false public opinion information in an online social network as described in any one of claims 1 to 6 by executing the executable instructions.

9. A computer storage medium, characterized in that The storage medium stores a computer program, and when the computer program is run, the method for early analysis of false public opinion information in an online social network as described in any one of claims 1 to 6 is executed.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method for early analysis of false public opinion information in online social networks as described in any one of claims 1 to 6 are implemented.

Images