Network rumor automatic intervention and propagation control method based on evolutionary game model

Abstract

The invention discloses a network rumor automatic intervention and propagation control method based on an evolutionary game model. The method comprises the following steps: firstly, tagging behavior characteristics and psychological power of network public opinion subjects, and constructing a subject-object-power ternary factor interaction atlas through emotion and activeness analysis of user historical speaking data; on this basis, in combination with an evolutionary multi-agent game, a rumor propagation process is dynamically simulated, and key propagation nodes with strong influence and high mentality plasticity are positioned in real time through a sharp change node detection algorithm. And for the identified key nodes, a scene self-adaptive intervention content generation algorithm is adopted to automatically generate personalized clarification, guidance or pacification type speech, and intervention is implemented through multi-channel accurate pushing. By means of the method, efficient, dynamic and intelligent intervention on network rumor spreading can be achieved, rumor spreading is effectively restrained, and the network space public opinion governance capacity is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of sociology, and more particularly relates to a network rumor automatic intervention and propagation control method based on an evolutionary game model. BACKGROUND

[0002] In recent years, with the rapid development of the Internet and social media, the speed and breadth of information dissemination in the network space have greatly improved. However, the openness and anonymity of the network space also lead to the frequent emergence and rapid spread of rumor information, which seriously interferes with public order and affects social stability. Traditional network rumor governance methods mainly rely on manual monitoring, keyword filtering, public opinion analysis and other means, and have problems such as response lag, single intervention means, and lack of individualization, which makes it difficult to meet the governance needs of the increasingly complex network public opinion environment. Some existing technologies attempt to combine machine learning or deep learning methods for public opinion recognition and rumor detection, but still have deficiencies in rumor propagation path prediction, key node identification, and individualized intervention content generation, and cannot achieve dynamic and accurate control of the whole process of rumor propagation. In recent years, evolutionary game theory and multi-agent interaction model have been widely used in complex social network analysis, providing a new technical path for network public opinion guidance and rumor governance. However, how to autonomously identify the propagation critical node, dynamically adjust the strategy and generate targeted intervention content to achieve active interception and diffusion inhibition of network rumors is still a difficult point and challenge that current technology needs to break through. Therefore, it is urgent to propose a new network rumor automatic intervention and propagation control method that integrates evolutionary game model, multi-dimensional behavior feature analysis and intelligent intervention pushing, to achieve efficient and accurate governance of rumor information in the network space. SUMMARY

[0003] The present application aims to solve the problems of response lag, single intervention means, lack of accurate identification of key propagation nodes and individualized intervention in existing network rumor governance technology, and proposes a comprehensive control method that can integrate network public opinion subject behavior characteristics and psychological motivation recognition, dynamically construct subject-object interaction atlas, adaptively adjust intervention strategies based on evolutionary game model, and achieve automatic identification of rumors, accurate positioning of critical nodes, and intelligent generation and automatic pushing of individualized intervention content, so as to realize active, dynamic and accurate intervention of the network rumor propagation process, and effectively improve the intelligent and automated level of network rumor governance.

[0004] In order to achieve the above-mentioned purpose, the present application is realized by adopting the following technical scheme: the method comprises: Network public opinion subject behavior characteristics and psychological motivation labeling, collecting historical speech data of public opinion subjects on a specific event, generating individual psychological motivation and behavior migration labels according to text content and sentiment analysis; Construct a dynamic public opinion object-subject interaction map, map all posts and comments as public opinion object nodes, and users as subject nodes to establish a subject-object-dynamic three-element factor graph; Construct an evolutionary multi-agent game intervention model. According to the evolution of psychological dynamic tags, surrounding group attitudes, and external events, game strategies will dynamically adapt and adjust in the game. Automatic rumor recognition and precise positioning of intervention nodes, using evolutionary game critical node detection algorithm: in the game evolution sequence, real-time detection of information transmission power mutation critical point, sorting nodes according to influence, mentality plasticity, and community structure importance, automatically determining the best intervention object; Personalized intervention measures generation and automatic push, for the selected intervention nodes, according to their dynamic, transmission tendency and historical response, using scenario adaptive intervention content generation algorithm, automatically generating speech templates for targeted clarification, guidance, counter-question or emotional pacification, integrating current affairs popularity and audience psychology, implementing hierarchical content push.

[0005] In one scheme, the network public opinion subject behavior characteristics and psychological dynamic tagging specifically includes: assuming that the user The set of speeches in T time is denoted as ; for each speech , use the sentiment analysis model to obtain its emotional score and theme relevance vector ; On this basis, each user will also calculate its activity , influence , and speech timing sensitivity ; At the same time, for the user's position in the network social structure, construct the adjacency matrix , and use the PageRank method to determine the structural influence score in the social network; Combine these basic features to define the multi-element psychological dynamic vector of each public opinion subject u , where each component has the following meaning: : crowd tendency coefficient, indicating its tendency to follow the mainstream and be easily influenced by group opinions; : suspicion coefficient, indicating its degree of suspicion of information authenticity; : incitement coefficient, reflecting its tendency to aggressively speak out or promote emotional diffusion; : order maintenance tendency, measuring its preference for rational speech, pursuit of clarification, and social order motivation.

[0006] In one approach, the construction of a dynamic public opinion object-subject interaction graph specifically includes: First, all posts and comments related to a specific event on the internet platform must be discretized into public opinion "object" nodes, with each specific posting behavior being recorded as an object. All network users act as primary nodes. Based on this structure, we introduce the multiple psychodynamic vectors of each subject. And abstract it as a "dynamic" node. This allows for the construction of a "subject-motivation-object" tripartite factor graph, used to express complex interactive relationships; The graph model is as follows ,in For the main node set, For the set of object nodes, For the set of psychodynamic nodes, It is a set of edges of various types; For each user comment, establish a system starting from the main body. Pointing to the object Behavioral side Then, trace the behavior back to the corresponding psychodynamic dimension. ; In this way, each propagation action is mapped to a path. ,in Based on The maximum value of the central psychodynamic component is determined; Edge weights in dynamic graphs Depending on the user's current psychological motivation, the intensity of their past statements, and the content characteristics of the subject of the opinion, the edge weight formula is set as follows: in It is currently the dominant component of psychological dynamics. It's user activity. For the object's heat, diffusion rate, or emotional polarity, This refers to the coupling characteristics between interaction moments and the event evolution sequence.

[0007] In one approach, the construction of an evolutionary multi-agent game intervention model specifically includes: each agent... Holding a psychodynamic vector that changes over time As an intrinsic driving force for behavioral decision-making; In each round of the evolutionary game, all agents, as participants, determine their own dynamic state and the aggregation of strategies among other agents in the surrounding network. ) and external factors of event development are denoted as , jointly game, select the next step behavior; The strategy space of the subject is The subject At time t, the probability of selecting strategy Is jointly modulated by the psychological motivation weight, the surrounding influence and the external input, and the calculation formula is as follows: Among them, the utility function The psychological motivation weight vector for strategy s is made an inner product with To capture the power bias; Measure the weighted sum of the strategy distribution in the neighbor group; Reflect the systematic thrust of external public opinion and event development on the strategy, Optimize the hyperparameters; The immediate income generated by the game Not only by the spread of information, but also by the complex combination of rumor disposal positive reward, punishment for questioning and psychological fit degree.

[0008] In one scheme, the rumor automatic identification and intervention node accurate positioning includes: along the time axis t, the psychological motivation mean And the propagation key indicators of all subjects at each time are segmented by sliding window, and the mutation point detection mechanism is called; let the whole rumor power change sequence be , wherein The node influence weight is The incitement component is The propagation strategy indication function is An adaptive entropy increase rate factor is introduced to automatically judge the critical point of power sharp rise; Through the change of the power and strategy of the subject node in the mutation window, the quantitative indicators of the rapid transition from observation or clarification to propagation are sorted, and the suspected key nodes with the fastest power response and the most drastic strategy transformation are selected; Then, all candidate nodes are further assigned fine weight scores, considering their network influence, mental plasticity and community structure importance, and the top K are taken as the best intervention objects and pushed to the system.

[0009] In one scheme, the personalized intervention measure generation and automatic pushing include: firstly, collecting the latest psychological motivation vector, current strategy state, propagation tendency score and historical intervention response record of each node for the screened key intervention node set; according to the information, calling a scene adaptive intervention content generation algorithm to dynamically determine the optimal intervention content type according to the node individual motivation portrait, community structure characteristics, current external public opinion hot spot; According to the information fusion result, each type of intervention template is selected and customized. Once the intervention content is generated, the personalized statement is automatically delivered to the node through the platform port according to the priority distribution pushing scheme, or the precise intervention is implemented by means of customized private message, comment insertion, community top-pinning multi-level aggregation.

[0010] The present application has the following advantages: The present application introduces the evolutionary game theory, combines the multi-dimensional behavior characteristics of network public opinion subjects and psychological motivation analysis, realizes the automatic identification and accurate positioning of the core propagation dynamics and key nodes of network rumors, and overcomes the disadvantages of response lag and intervention generalization of traditional methods. By constructing a dynamic public opinion subject-object-dynamic three-factor interaction map, and collecting user historical speech data for emotion and activity modeling, the description of rumor propagation mechanism and the accuracy of key node intervention decision are effectively improved. In addition, the present application uses a critical node detection algorithm, which can identify high-influence, high-malleability community key users at the first time when the information propagation dynamics changes, and realize real-time monitoring of potential propagation paths. Based on the scene adaptive intervention content generation algorithm, personalized, hierarchical clarification, guidance or emotional pacification information is customized for different types of intervention objects, and is automatically pushed by the platform, which greatly enhances the pertinence and effectiveness of the intervention measures. Overall, the present application can realize active, dynamic and intelligent intervention of network rumor propagation, thereby effectively curbing the spread of rumors, maintaining the good order of network space, and improving the efficiency and automation level of social public opinion governance. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 The network public opinion system block diagram is shown in Figure 1. Figure 2 The method flowchart of the present application is shown in Figure 2. DETAILED DESCRIPTION

[0012] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The drawings show typical embodiments of the present application. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0013] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0014] like Figure 1 As shown, the online public opinion system refers to the sum of the beliefs, attitudes, opinions, and emotions expressed by public opinion subjects on the Internet within a certain time and space, under specific psychological conditions, regarding a certain social event. It can influence the development of the event itself.

[0015] Composition of the public opinion system The online public opinion system consists of public opinion subjects, public opinion objects, and public opinion driving forces, such as... Figure 1 Public opinion systems exist within specific system environments, separated from their environment by system boundaries. There are certain inputs and outputs between the public opinion system and its environment. The subjects of public opinion in an online public opinion system include those who express opinions on events and the parties involved in those events, such as netizens. The objects of public opinion are the sum of the beliefs, attitudes, opinions, and emotions expressed by the subjects, such as the posts published by netizens. The driving force of public opinion is the psychology of the subjects, the internal driving force behind their expression of different objects; for example, resentment towards the wealthy is the underlying psychological reason for some netizens to post strongly worded posts. The environment of the public opinion system is the continuous development of events; it is the input to the system and also the external driving force propelling its development. The boundaries of the public opinion system are a certain space and time, and specific events. The impact of the public opinion system on events is the system's output, which has two essentially different types: one is the direct result caused by the system, and the other is a conclusion with more universal applicability.

[0016] The connections between the various elements of a public opinion system constitute the system's structure. First, the subjects of public opinion influence each other. The views of some subjects influence those of others, acting as opinion leaders. Second, the objects of public opinion are interdependent in time and space. Some posts exert a preconceived influence on others in terms of time; while information asymmetry between different forums creates spatial isolation between objects. Third, there is an action-reaction relationship between the subjects and objects of public opinion. Subjects post different things under different public opinion dynamics, and these different posts influence the subjects' psychology, thus affecting the subjects themselves. Fourth, system dynamics act as the connector, transmitter, and promoter of the relationships between subjects and objects within the system. It is the internal driving force behind the system's development and change. The interactions between subjects, between objects, and between subjects and objects are all completed under the impetus of this system dynamic of subject psychology. A qualitative description of the composition, structure, and function of the online public opinion system provides a theoretical basis for the quantitative calculation of online public opinion intervention mechanisms. Intervention in online public opinion can be based on the characteristics of the online public opinion system, adjusting its composition and structure, and altering its functions to achieve the intervention goal of promoting positive interaction between public opinion and events and guiding them towards a direction conducive to social harmony.

[0017] like Figure 2 As shown, an automatic intervention and propagation control method for online rumors based on an evolutionary game model is described, with the following specific implementation steps: Step 1: Labeling the behavioral characteristics and psychological dynamics of online public opinion subjects Collect historical data on online users' (public opinion subjects) statements on specific events, and automatically generate individual psychological dynamics and behavioral migration tags based on text content and sentiment analysis, combined with users' historical activity, influence, social structure, and timing of statements.

[0018] Introducing a multidimensional psychodynamic evolution labeling algorithm: For each subject, an evolutionary multidimensional psychological vector (including herd mentality, skepticism, inflammatory behavior, order maintenance tendency, etc.) is established and coupled with conventional group propagation parameters for real-time dynamic adjustment.

[0019] In the detailed implementation of step 1, "Labeling of Behavioral Characteristics and Psychological Dynamics of Online Public Opinion Subjects," the first step is to collect historical speech data from public opinion subjects for specific online events. This process typically includes text crawling, user feature extraction, and multi-dimensional behavioral statistics. For the text content, Natural Language Processing (NLP) techniques are used to extract topics and perform sentiment analysis on users' historical speeches. Assuming that for users... The set of speeches within time period T is denoted as For each comment Sentiment scores were obtained using a sentiment analysis model. (e.g., positive / negative probabilities) and topic relevance vectors .

[0020] In addition, each user's activity level will be calculated. Influence (such as the number of followers and the number of reposts / comments weighted), and the sensitivity of the timing of their statements. (e.g., whether they were among the first to speak out, or their activity level during periods of heightened public opinion). Simultaneously, an adjacency matrix is ​​constructed based on the user's position within the online social structure. The PageRank method was used to determine the structural influence score in social networks. .

[0021] Based on these fundamental characteristics, we define the multi-dimensional psychological dynamic vector of each opinion subject u. The meanings of each component are as follows: Conformity coefficient: This indicates the tendency to conform to the mainstream and be easily influenced by group opinions. The skepticism coefficient represents the degree of doubt about the authenticity of information. Agitativity coefficient reflects the tendency to make radical statements or promote the spread of emotions; Orderliness measures a person's motivation to prefer rational speech, seek clarification, and maintain social order.

[0022] Each component of the multivariate psychodynamic vector is dynamically assigned a value according to the following function, for example: in, The sigmoid normalization function. The weight coefficients for training or setting each component. For the user's neighbor set, To account for similarity with neighboring nodes or propagation effects, The average sentiment of users' historical posts. As a characteristic of orderliness, The speaker's emotional fluctuations.

[0023] Ultimately, the multidimensional psychodynamic vector of each user It not only reflects users' behavioral tendencies in the public opinion field, but also provides dynamic label inputs for strategy selection and propagation probability adjustment in the evolutionary game process. Each component is updated in real time with the progress of time and events (or feedback on external intervention / propagation results). This parameterized psychodynamic evolution labeling algorithm breaks through the modeling of single-group characteristics and lays the foundation for refined evolutionary game and automatic intervention of heterogeneous subjects.

[0024] Step 2: Constructing a dynamic interaction graph of public opinion object and subject. Map all posts and comments as object nodes of public opinion, and users as subject nodes, to establish a subject-object-dynamic ternary factor graph.

[0025] The adaptive causal inference method is used to infer the feedback effect of each object on the subject's psychology, and the stimulating or inhibiting effect of the object on the subject's motivation is modeled hierarchically.

[0026] In the detailed implementation of step 2, "Constructing a Dynamic Interaction Graph of Public Opinion Objects and Subjects," the first step is to discretize all posts and comments related to a specific event on the internet platform into public opinion "object" nodes, with each specific posting behavior being recorded as an object. All network users act as primary nodes. Based on this structure, the multivariate psychodynamic vectors of each subject obtained in step 1 are introduced. And abstract it as a "dynamic" node. This creatively constructs a "subject-motivation-object" tripartite factor graph to express complex interactive relationships.

[0027] In terms of form, the map can be modeled as ,in For the main node set, For the set of object nodes, For the set of psychodynamic nodes, It is a set of edges of various types. For each user's message, a framework is established from the main body... Pointing to the object Behavioral side Then, trace the behavior back to the corresponding psychodynamic dimension. (For example, if a post is inflammatory, then it is associated with an inflammatory node.) In this way, each propagation action can be mapped to a path. ,in Based on step 1 The maximum value of a certain psychodynamic component is determined.

[0028] Edge weights in dynamic graphs Depending on the user's current psychological dynamics (such as inflammatory nature, herd mentality, etc.), the intensity of their past statements, and the content characteristics of the subject matter, the edge weight formula can be set as follows: in It is currently the dominant component of psychological dynamics. It's user activity. For the object's heat, diffusion rate, or emotional polarity, This describes the coupling characteristics between interaction moments and the event evolution sequence. (Function) Linear or nonlinear normalization can be combined with multimodal fusion (such as attention mechanisms) to distinguish different interaction intensities.

[0029] To delve deeper into the reverse psychological feedback of objects (posts / comments) to the subject, an adaptive causal inference method (such as Granger causality, DoWhy, or extensions of structural equation modeling) is employed to analyze the object's psychological response at multiple time steps. Did the change in attributes "lead to" the subject? A certain psychological dynamic component Significant changes. Formally, in It is the magnitude of the change of the dynamic component over time (t+1). Characteristic object The external effects of the dynamics within this time window, It is the causal weight (which can be obtained through training in causal regression or counterfactual reasoning). This represents the system residual.

[0030] Furthermore, the ternary factor graph can be stratified by time to form a dynamic graph. Through recursive updates, each user's operation on the object and input from the external system is fed back to the subject's psychological vector in real time via the linkage of network structure, node states, and edge weights. In this way, the innovatively introduced dynamic layer nodes not only characterize deep psychological motivations but also realize the inverse modeling of the stimuli from the object on the subject's psychological evolution, thus laying the theoretical and structural foundation for the accurate deduction and node selection of subsequent evolutionary game theory and intelligent intervention mechanisms.

[0031] Step 3: Construct an evolutionary multi-agent game intervention model Based on the dynamic graph in step 2, a self-evolving multi-agent heterogeneous game model is designed: the agents evolve according to their psychological dynamic labels, the attitudes of the surrounding groups, and external events, and their game strategies (spreading / questioning / clarifying / observing, etc.) will be dynamically and adaptively adjusted in the game. The payoff vector is not only affected by the group attitude, but also modulated by the dynamic factors.

[0032] In the detailed implementation of step 3, "Evolutionary Multi-Agent Game Intervention Simulation," based on the agent-dynamic-object ternary dynamic graph generated in step 2, each agent... Holding a psychodynamic vector that changes over time This serves as the intrinsic driving force for behavioral decisions. In each round of the evolutionary game in the system, all agents, as participants in the game, base their decisions on their current dynamic state and the aggregation of strategies of other agents in the surrounding network (such as the average strategy of neighbors). ) and external factors in the development of the event (such as clarification of facts, public opinion outburst, etc., denoted as ( ), engaging in a collaborative game, choosing their next action. The strategy space of the agents is . .main body Choose a strategy at time t The probability is modulated by psychodynamic weights, surrounding influences, and external inputs, and is calculated as follows: Among them, utility function It is the psychodynamic weight vector for strategy s, and To capture dynamic bias by performing inner product calculations; Measure the weighted sum of the distribution of this strategy in the neighborhood group (e.g., the amplification of the propagation decision by herd mentality). Reflecting the systemic driving force of external public opinion and event developments on strategy. To optimize hyperparameters.

[0033] Immediate gains from the game The dissemination effect is not only influenced by the spread of information, but also by a complex combination of positive rewards from handling rumors, punishment for questioning, and the degree of psychological alignment. The innovation lies in the fact that after each round of the game, the psychological dynamics of the participants are dynamically adjusted in reverse based on their current choices and gains, achieving proactive adaptive evolution. in Evolution rate, It is an external disturbance term. This reflects the sensitivity of behavioral rewards to the feedback of motivational components. Thus, each agent in the model, based on policy responses and environmental feedback, will self-reinforce or weaken its psychological motivations and future behavioral tendencies, thereby enabling the overall evolutionary process to reflect the diversity of agents and the self-organizing evolutionary mechanism of group behavior under real-world conditions of online rumor dissemination and multi-round intervention. This innovative coupling of multi-motivational empowerment, multi-round benefit feedback, and agent psychological self-evolution provides accurate, high-dimensional, and highly dynamic theoretical and algorithmic support for intelligent intervention and automatic evolutionary control.

[0034] Step 4: Precise Location of Automatic Rumor Identification and Intervention Nodes An evolutionary game-theoretic change node detection algorithm is adopted: in the game evolution sequence, the critical point of information propagation dynamic change (i.e. the fault point where rumors are about to break out) is detected in real time, and the nodes are sorted according to their influence, psychological plasticity, and importance of the community structure, and the best intervention target is automatically determined.

[0035] Specifically, along the time axis t, the mean psychodynamics of all subjects at each moment. We can segment the data using a sliding window method based on key dissemination indicators (such as total dissemination probability and the spread of rumor content) and invoke a mutation point detection mechanism. We can define the overall dynamic change sequence of the rumor as follows: ,in Weighting the influence of nodes As an inflammatory element, This serves as the propagation strategy indicator function. By utilizing mutation point algorithms such as CUSUM (cumulative sum detection) or innovatively introducing an adaptive entropy growth factor, the system can automatically determine the critical point of a sharp increase in momentum. in This is an empirical or dynamic learning threshold. In parallel, we adjust the threshold for the main node within the mutation window. Changes in intrinsic motivation and strategy (such as the absolute rate of change of psychodynamic components) The strategy was ranked by quantitative indicators (the rate at which the strategy quickly shifted from observation or clarification to dissemination), and suspected key nodes with the fastest dynamic response and the most dramatic strategy transformation were selected.

[0036] Subsequently, all candidate nodes were further assigned fine-grained weight scores, taking into account their network influence. (e.g., PageRank / degree centrality), mental malleability (e.g., psychodynamic variance, or historical sensitivity to "post-intervention dynamic regression") and the importance of community structure. (Such as the modularity of the community, the number of cross-community edges, etc.): in The system automatically adjusts the weights for multiple objectives. Mental malleability can be defined as follows: This reflects the sensitivity of the node's psychological state to external intervention or the evolution of events. Ultimately, all key suspected nodes... according to Sort the data and select the top K as the best intervention targets to push to the system.

[0037] This algorithm can accurately identify the nodes where sudden changes in group psychology and communication behavior first take effect, seizing the intervention window that is most likely to influence the diffusion chain and the current state of the group, thus achieving efficient "source braking" in the early stages of public opinion outbreaks. By innovatively combining evolutionary dynamic mutations, node heterogeneity, and community structure coupling, it significantly improves the scientific rigor, timeliness, and efficiency of intervention resource utilization in automatic rumor identification and intervention positioning.

[0038] Step 5: Generation and Automatic Delivery of Personalized Intervention Measures For the selected intervention nodes, based on their motivation, dissemination tendency and historical response, a scenario-adaptive intervention content generation algorithm is used to automatically generate targeted clarification, guidance, questioning or emotional reassurance speech templates, integrating current events and audience psychology to implement tiered content delivery.

[0039] The intervention effect is fed back to the game model in real time, updating the overall structure and enabling the method to self-optimize and evolve.

[0040] First, regarding the set of key intervention nodes selected in step 4 of the previous step... Collect the latest psychodynamic vector of each node. Current strategy state The system uses dissemination tendency scores and historical intervention response records. Based on this information, the system invokes a scenario-adaptive intervention content generation algorithm. This algorithm uses individual node dynamic profiles, community structural characteristics, and current external public opinion hotspots (denoted as...). The optimal intervention content type is dynamically determined by taking the input as an example.

[0041] The decision-making model for selecting intervention document types can be formalized as follows: For each node, define the intervention type. Then the node At any moment Preferred Its decision score is in, To adapt weight vectors to the dynamics of different intervention types, The sensitivity of the current policy of the expression node to this type of intervention. As a factor that integrates public opinion heat and the timeliness of intervention content, These are the parameter tuning coefficients. Final .

[0042] For each type of intervention template, the system selects and customizes it based on the information fusion results. For example, for highly inflammatory ( For nodes with high (high) and tendency to spread information, the system will prioritize customizing templates that focus on fact-clarification and rational guidance, while minimizing emotional conflict; for nodes with high skepticism ( High) or high order maintenance tendency ( For nodes with high (high) levels, it is more suitable to push logically challenging or authoritative content to stimulate rational judgment and self-correction. This generates the final intervention content. Taking into account the language style of the community where the node is located and current trending language (daily memes, authoritative announcements, mainstream online narratives, etc.), adaptive templates and real-time natural language generation are dynamically invoked under a layered architecture (community, individual, and mainstream across the entire network). Specific content generation can be represented as follows: Describing node profiles and content fit can be based on deep semantic embedding (BERT similarity), the acceptance of content by historically similar nodes, or the evaluation score of the content generation model.

[0043] Once intervention content is generated, a push plan is assigned according to priority (impact, psychological activity level, criticality of the dissemination path, etc.). Personalized statements are automatically delivered to nodes through the platform port, or precise intervention is implemented through multi-level aggregation methods such as customized private messages, comment insertion, and community pinning. The system simultaneously records node feedback (such as changes in speech, emotional fluctuations, and strategy switching), forming a feedback loop. Feedback indicators representing changes in psychological dynamics and behavior after intervention at representative nodes are dynamically labeled and fed back to the evolutionary game model in real time: in The feedback reinforcement coefficients are used to automatically update the individual nodes and the overall structure of the model, driving psychodynamic evolution and adaptive optimization of the propagation path. Ultimately, this method achieves a fully closed-loop intelligent evolution system from key node identification, content customization, and hierarchical push to model self-learning iteration, effectively improving the accuracy, efficiency, and sustainability of early intervention, propagation interception, and social mental health guidance for online rumors.

[0044] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0045] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic intervention and propagation control of online rumors based on an evolutionary game model, characterized in that: The method includes: The behavioral characteristics and psychological dynamics of online public opinion subjects are labeled by collecting historical speech data of public opinion subjects on specific events and generating individual psychological dynamics and behavioral transfer labels based on text content and sentiment analysis. Construct a dynamic interaction graph of public opinion object and subject, mapping all posts and comments as public opinion object nodes and users as subject nodes, and establish a subject-object-dynamic ternary factor graph; An evolutionary multi-agent game intervention model is constructed, in which the game strategy will dynamically and adaptively adjust during the game based on psychodynamic labels, attitudes of surrounding groups and external events. The system automatically identifies and precisely locates nodes for intervention in rumor detection, employing an evolutionary game theory-based algorithm for detecting sudden changes in nodes. In the game evolution sequence, it detects the critical point of sudden changes in the dynamics of information dissemination in real time, sorts nodes by influence, psychological malleability, and importance of the community structure, and automatically determines the best intervention target. Personalized intervention measures are generated and automatically pushed. For selected intervention nodes, based on their motivation, dissemination tendency and historical response, a scenario-adaptive intervention content generation algorithm is used to automatically generate targeted clarification, guidance, questioning or emotional reassurance speech templates. The algorithm integrates current events and audience psychology to implement tiered content push.

2. The method for automatic intervention and propagation control of online rumors based on an evolutionary game model as described in claim 1, characterized in that: The aforementioned labeling of the behavioral characteristics and psychological dynamics of online public opinion subjects specifically includes: assuming that users... The set of speeches within time period T is denoted as For each comment Sentiment scores were obtained using a sentiment analysis model. and topic relevance vector ; In addition, each user's activity level will be calculated. Influence and sensitivity to the timing of vocalization ; Simultaneously, an adjacency matrix is ​​constructed based on the user's position within the network social structure. The PageRank method was used to determine the structural influence score in social networks. ; Based on these fundamental characteristics, we define the multi-dimensional psychological dynamic vector of each opinion subject u. The meanings of each component are as follows: Herd mentality coefficient: This indicates the tendency to conform to the mainstream and be easily influenced by the group's views. The skepticism coefficient indicates the degree of doubt about the authenticity of the information. : Inflammatory coefficient, reflecting the tendency to voice radical opinions or promote the spread of emotions; The tendency to maintain order measures the motivation of those who prefer rational speech, seek clarification, and maintain social order.

3. The method for automatic intervention and propagation control of online rumors based on an evolutionary game model according to claim 1, characterized in that: The construction of a dynamic public opinion object-subject interaction graph specifically includes: First, all posts and comments related to a specific event on the internet platform must be discretized into public opinion "object" nodes, with each specific posting behavior being recorded as an object. All network users act as primary nodes. Based on this structure, we introduce the multiple psychodynamic vectors of each subject. And abstract it as a "power" node. This allows for the construction of a "subject-motivation-object" tripartite factor graph, used to express complex interactive relationships; The graph model is as follows ,in For the main node set, For the set of object nodes, For the set of psychodynamic nodes, It is a set of edges of various types; For each user comment, establish a system starting from the main body. Pointing to the object Behavioral side Then, trace the behavior back to the corresponding psychodynamic dimension. ; In this way, each propagation action is mapped to a path. ,in Based on The maximum value of the central psychodynamic component is determined; Edge weights in dynamic graphs Depending on the user's current psychological motivation, the intensity of their past statements, and the content characteristics of the subject of the opinion, the edge weight formula is set as follows: ; in It is currently the dominant component of psychological dynamics. It's user activity. For the object's heat, diffusion rate, or emotional polarity, This refers to the coupling characteristics between interaction moments and the event evolution sequence.

4. The method for automatic intervention and propagation control of online rumors based on an evolutionary game model according to claim 1, characterized in that: The construction of the evolutionary multi-agent game intervention model specifically includes: each agent Holding a psychodynamic vector that changes over time As an intrinsic driving force for behavioral decision-making; In each round of the evolutionary game, all agents, as participants, determine their own dynamic state and the aggregation of strategies among other agents in the surrounding network. ) and external factors in the development of the event are denoted as Through joint competition, they choose their next course of action; The subject's strategy space is ,main body Choose a strategy at time t The probability is modulated by psychodynamic weights, surrounding influences, and external inputs, and is calculated as follows: ; Among them, utility function ; It is the psychodynamic weight vector for strategy s, and To capture dynamic bias by performing inner product calculations; Measure the weighted sum of the distributions of this strategy within the neighborhood group; Reflecting the systemic driving force of external public opinion and event developments on strategy. To optimize hyperparameters; Immediate gains from the game It is influenced not only by the dissemination effect of information spread, but also by a complex combination of positive rewards from handling rumors, punishment for questioning, and psychological alignment.

5. The method for automatic intervention and propagation control of online rumors based on an evolutionary game model according to claim 1, characterized in that: The aforementioned automatic rumor identification and intervention node precise positioning includes: along the time axis t, the mean psychodynamic values ​​of all subjects at each moment. The system segments key indicators of dissemination using a sliding window approach and invokes a mutation point detection mechanism to make the overall dynamic change sequence of rumors... ,in Weighting the influence of nodes As an inflammatory element, The function is the propagation strategy indicator. An adaptive entropy growth factor is introduced to automatically determine the critical point of a sharp increase in momentum. By ranking the quantitative indicators of the rapid shift from observation or clarification to propagation of the dynamics and strategies of the main nodes within the mutation window, the suspected key nodes with the fastest dynamic response and the most dramatic strategy transformation are screened out. Subsequently, all candidate nodes were further assigned fine-grained weight scores, taking into account their network influence, psychological malleability, and importance of community structure, and the top K nodes were selected as the best intervention targets and pushed to the system.

6. The method for automatic intervention and propagation control of network rumors based on an evolutionary game model according to claim 1, characterized in that: The personalized intervention measures generation and automatic push include: first, for the selected set of key intervention nodes, collecting the latest psychodynamic vector, current strategy status, dissemination tendency score and historical intervention response records for each node; based on this information, calling the scenario-adaptive intervention content generation algorithm, using the node individual dynamic profile, community structure characteristics and current external public opinion hotspots, to dynamically determine the optimal intervention content type; For each type of intervention template, selection and customization are based on the information fusion results; Once the intervention content is generated, a push plan is assigned according to priority, and personalized statements are automatically delivered to nodes through the platform port, or precise intervention is implemented through multi-level aggregation methods such as customized private messages, comment insertion, and community pinning.

Images