Social simulator cognitive intervention method and system

Abstract

The application relates to the technical field of network security, and discloses a social simulator cognitive intervention method and system, which comprises the following steps: generating a group topology based on user interaction data of a real social network; dynamically adjusting the configuration proportion of an intelligent agent node based on an initial evaluation signal and a Bayesian optimization method; inferring the edge probability of each intelligent agent node in the group topology based on a probabilistic graph model, screening candidate intervention nodes according to the edge probability, and deploying intervention-type intelligent agent nodes based on the adjusted configuration proportion to implement intervention on the candidate intervention nodes; evaluating the intervention result to generate an evaluation signal, and feeding back the evaluation signal to the dynamic adjustment step to replace the initial evaluation signal. The application can effectively organize intervention subjects with different roles based on large-scale data of a complex network environment, dynamically optimize the delivery proportion, and simulate and predict the nodes to be intervened based on probabilistic graph reasoning.

Description

Technical Field

[0001] This invention relates to the field of cybersecurity technology, specifically to a social simulator cognitive intervention method and system. Background Technology

[0002] Traditional cognitive intervention research has significant limitations in practical applications. Traditional intervention modeling methods mainly include four types of numerical modeling frameworks: Independent Cascade Model, Linear Threshold Model, Game-Theoretic Model, and Epidemics Model. While these models can theoretically describe information dissemination and intervention mechanisms, they generally employ overly simplified assumptions about the dissemination environment and participant behavior, abstracting the intervention process into numerical calculations, thus failing to fully reflect the complex interactions in real social networks. Furthermore, these models have a large number of hyperparameters, making them prone to overfitting and reducing their generalization ability and practicality.

[0003] In recent years, social simulation systems based on Large Language Models (LLMs) have begun to be applied to research on network event propagation and information intervention. These systems generate text content for information delivery to simulate different intervention scenarios. However, the intervention methods of this type are often limited to simple information injection, lacking specificity and strategy, and are difficult to adapt to the intervention needs of multi-role, multi-level, and multi-strategy collaboration in social networks. Furthermore, these simulation systems generally neglect the appropriate intervention ratio during the intervention process, and the predictability of unspoken nodes believing false information in the propagation path (usually, intervention content is delivered after each agent has finished speaking), failing to achieve dynamic optimization and accurate prediction of intervention effects.

[0004] In complex social network environments, the effectiveness of intervention strategies is influenced by a variety of factors, including group structural characteristics, individual cognitive states, and information propagation dynamics. Existing research lacks automated and scientific support for strategy evaluation, leading to intervention decisions often relying on experience or manual settings, resulting in a lack of repeatability and quantifiability. This situation urgently requires a comprehensive cognitive intervention method and system that is controllable, dynamically optimized, and accurately predictive, providing theoretical basis and data support for strategy design and implementation. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a social simulator cognitive intervention method and system. This method can effectively organize intervention subjects with different roles based on large-scale data in complex network environments, dynamically optimize the deployment ratio, and perform simulation prediction of the intervention nodes based on probabilistic graphical reasoning, providing real-time feedback of diverse intervention indicators, thereby achieving efficient, controllable, and strategic cognitive intervention simulation.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a cognitive intervention method using a social simulator, comprising:

[0008] A group topology is generated based on user interaction data from real social networks. The group topology contains multiple intelligent agent nodes playing different roles, including at least one interventionist intelligent agent node that can have a negative effect on the spread of misinformation.

[0009] Dynamic adjustment step: Based on the initial evaluation signal and Bayesian optimization method, the configuration ratio of at least one intervention agent node in the population topology is dynamically adjusted;

[0010] Based on the probabilistic graphical model, the marginal probability of each agent node in the group topology accepting false information is inferred. Candidate intervention nodes are selected according to the marginal probability, and intervention agent nodes are deployed according to the adjusted configuration ratio to intervene in the candidate intervention nodes. The intervention agent nodes are configured to generate intervention text content.

[0011] The intervention results are evaluated to generate an evaluation signal, which is then fed back into the dynamic adjustment step to replace the initial evaluation signal.

[0012] In one embodiment, generating the group topology based on user interaction data from real social networks specifically includes:

[0013] Acquire user interaction data, which includes the connection relationships and status information of user nodes in real social networks; the status information includes the structural features, behavioral features, content features, and credibility features of user nodes;

[0014] The user interaction data is clustered using a Gaussian mixture model to obtain the feature distribution of multiple role types and the initial role proportion vector;

[0015] The number of agent nodes playing each role is determined based on the total number of agent nodes and the initial role proportion vector, and the node features of the corresponding agent nodes are generated by sampling according to the feature distribution of different roles.

[0016] The group connection preference mechanism is used to construct the edge relationships between the agent nodes to form the group topology, and the group connection preference mechanism is used to make the topological features of agent nodes of each role type conform to the corresponding feature distribution.

[0017] In one embodiment, the step of constructing the edge relationships between the agent nodes according to a group connection preference mechanism to form the group topology, wherein the group connection preference mechanism is used to ensure that the topological features of agent nodes of each role type conform to the corresponding feature distribution, specifically including:

[0018] For any intelligent agent node The role to be assigned is The group connection preference mechanism is defined as the following scoring function. Optimization issues:

[0019] ;

[0020] in, and For intelligent agent nodes out-degree and in-degree, and For the role In structural features The mean of the probability distribution of out-degree and the mean of the probability distribution of in-degree in the given data. and For the role In structural features The standard deviations of the out-degree and in-degree are in the range. This is a set constant.

[0021] In one embodiment, the dynamic adjustment of the configuration ratio of at least one interventionist agent node in the population topology based on the initial evaluation signal and Bayesian optimization method specifically includes:

[0022] Define system feedback function ,in, A role proportion vector that includes the configuration ratio of interventionist agent nodes. To evaluate the signal, Let λ be the intervention cost function, and λ be the trade-off parameter.

[0023] The system feedback function is optimized using the Bayesian optimization method. The prior probability distribution is modeled as a Gaussian process, and calculations are performed based on historical observation datasets. The posterior probability distribution;

[0024] The expected gain function is constructed based on the system feedback function and the corresponding posterior probability distribution, by maximizing the expected gain function. Let be the expected gain function of the variable, for Update.

[0025] In one embodiment, the system feedback function is optimized using a Bayesian method. The prior probability distribution is modeled as a Gaussian process, and calculations are performed based on historical observation datasets. The posterior probability distribution specifically includes:

[0026] Based on Bayesian optimization theory, the system feedback function is first... The prior probability distribution is approximately a Gaussian process:

[0027] ;

[0028] in, Represents a Gaussian process. and These represent two different character proportion vectors. It is a mean function. It is the covariance function;

[0029] Under the assumptions of the Gaussian process mentioned above, based on Bayesian theory, Estimate the posterior probability in Observational data were obtained in the intervention simulation of the wheel. ,in, This represents the role proportion vector used in the t-th round of intervention simulation of the system. This indicates that the role proportion vector is used in the t-th round of intervention simulation. The feedback value afterward; The posterior probability distribution is estimated as follows Assuming Mean function in prior probability If the mean is zero, then the mean of the posterior probability distribution is... , Expressing expectations, Indicates transpose. This represents the custom observation noise variance. for The covariance vector of the vector representing the proportion of historical roles, ,in For kernel function, For the kernel matrix, elements in , It is the identity matrix. For the observed vector, ; Variance of the posterior probability distribution , Represents the variance of observation noise. for The autocovariance.

[0030] In one embodiment, the construction of the expected gain function based on the system feedback function and the corresponding posterior probability distribution maximizes the expected gain function. Let be the expected gain function of the variable, for The update includes:

[0031] Based on system feedback function Define the expected gain function to be optimized. ,in, Expressing expectations, , Let represent the i-th feedback value in the observed data; based on Bayesian optimization theory, the analytical expression of the expected gain function is obtained:

[0032] ;

[0033] intermediate variables , and These are the probability distribution function and probability density function of the standard normal distribution, respectively. for The mean and standard deviation of the posterior probability distribution; the expected gain function is optimized to obtain... Role proportion vector of agent nodes at each time step :

[0034] , ;

[0035] in, show One-dimensional simplex space with unit probability. This represents the total number of roles of the agent nodes, i.e. , ; The given first number is restricted by experts The upper bound of the maximum deployment ratio of role-based intelligent agent nodes. This represents the index of the r-th type of role.

[0036] In one embodiment, the step of inferring the marginal probability of each agent node in the group topology accepting false information based on a probabilistic graphical model specifically includes:

[0037] A deep neural network model is trained based on a real event dataset to calculate the conditional propagation probability between two connected agent nodes based on their state information, which is then used as the weight of the corresponding directed edge in the group topology.

[0038] Based on the belief propagation algorithm and the conditional propagation probability, the iterative information transmitted by the neighboring nodes of each agent node is iteratively calculated to calculate the marginal probability of each agent node accepting false information.

[0039] The iterative information is calculated based on the agent node's prior probability of believing false information, the edge potential function, and the iterative messages passed by other neighbor nodes. The edge potential function is defined by the conditional propagation probability.

[0040] In one embodiment, the step of filtering candidate intervention nodes based on the marginal probability and deploying intervention-type intelligent agent nodes based on the adjusted configuration ratio to intervene in the candidate intervention nodes, wherein the intervention-type intelligent agent nodes are configured to generate intervention text content, specifically includes:

[0041] From the population topology, agent nodes with edge probabilities higher than a set probability threshold are selected to form a high-risk node set.

[0042] Based on the adjusted configuration ratio of intelligent agent nodes, determine the deployment quantity of at least one type of intervention-type intelligent agent node, which includes misinformation skeptical intelligent agent nodes, anti-misinformation dissemination intelligent agent nodes, and information blocker intelligent agent nodes.

[0043] A hierarchical ranking method is used to sort candidate intervention nodes in the high-risk node set based on their marginal probabilities and structural characteristics. These structural characteristics include the out-degree, in-degree, PageRank value of the agent nodes, and the network distance between the candidate intervention nodes and the source of misinformation. A predetermined number of candidate intervention nodes at the top of the ranking are identified as high-risk or misinformation source types; a predetermined number of candidate intervention nodes at the bottom of the ranking are identified as highly susceptible or potential propagation types; and other candidate intervention nodes are identified as high-influence or core propagation types. Based on the categories of the candidate intervention nodes, corresponding intervention agent nodes are matched and deployed, wherein:

[0044] For candidate intervention nodes classified as highly susceptible or potentially propagating, the misinformation skeptic agent node is deployed. The misinformation skeptic agent node is configured to generate intervention text content containing questioning or reminder content based on the first prompt word template.

[0045] For candidate intervention nodes classified as high-influence or core-propagation type, the anti-misinformation propagator intelligent agent node is deployed. The anti-misinformation propagator intelligent agent node is configured to generate intervention text content that includes refutation by citing authoritative sources based on the second prompt word template.

[0046] For candidate intervention nodes classified as high-risk or sources of misinformation, the information blocker intelligent agent node is deployed. The information blocker intelligent agent node is configured to generate intervention text content containing account or content restriction prompts based on a third prompt word template.

[0047] In one embodiment, the step of evaluating the intervention result to generate an evaluation signal and feeding the evaluation signal back to the dynamic adjustment step to replace the initial evaluation signal specifically includes:

[0048] Set of indexes of successfully intervened agent nodes ; Indicates the first Each intelligent agent node The stance value on misinformation at the time step. This indicates a strong disbelief in the false information. This indicates an extreme belief in false information. To simulate the start time of the system, This is a simulated time interval; This represents the index of the i-th agent;

[0049] Evaluation indicators include:

[0050] Intervention transmission rate : ;

[0051] Intervention in the depth of transmission : ;in This represents the propagation path length between the successfully intervened agent node and the intervention source agent node. These are the indices of the successfully intervened agent node and the intervention source agent node, respectively;

[0052] Intervention spread : ; The total number of all agent nodes;

[0053] Opinion conversion rate : ;

[0054] Sentiment Score Tendency : ; It is a set of indices for all agent nodes;

[0055] Final evaluation signal for: ; An index for evaluation metrics;

[0056] in, Representative evaluation indicators The normalized value, for The corresponding weights.

[0057] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.

[0058] Compared with the prior art, the beneficial technical effects of the present invention are:

[0059] This invention addresses the reliance on numerical modeling frameworks in traditional methods by designing multi-agent roles and generating group topologies. It also combines the semantic generation and theoretical reasoning capabilities of large language models to construct customized intervention agents, thereby significantly improving the simulation realism and reasoning rigor of cognitive intervention research.

[0060] This invention addresses the limitations of existing intervention studies based on large language models, which often rely on singular intervention methods and lack automated optimization. It proposes a dynamic optimization scheme for intervention ratios. This scheme utilizes controllable system feedback signals to adaptively adjust intervention ratios, effectively enhancing the flexibility and adaptability of the intervention process.

[0061] This invention addresses the shortcomings of existing methods in predicting states where false information is believed, and proposes a probabilistic graphical reasoning framework for intervention. This method characterizes the complete information flow propagation process in an analytical, non-black-box manner, predicts the marginal probability of nodes believing false information, accurately identifies high-risk nodes, and implements targeted interventions, thereby significantly improving the accuracy and interpretability of interventions. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0063] Figure 2 This is a schematic diagram of the algorithm for adjusting the intervention ratio in this invention.

[0064] Figure 3 This is a schematic diagram of the multi-agent cognitive intervention algorithm of the present invention. Detailed Implementation

[0065] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0066] like Figure 1 As shown, a social simulator cognitive intervention method of the present invention includes the following steps:

[0067] S1. Generate a group topology based on user interaction data from real social networks. The group topology contains multiple intelligent agent nodes playing different roles, including at least one interventionist intelligent agent node that can have a negative effect on the spread of misinformation.

[0068] S2, Dynamic adjustment step: Based on the initial evaluation signal and Bayesian optimization method, dynamically adjust the configuration ratio of at least one interventionist agent node in the population topology;

[0069] S3, based on the probabilistic graphical model, infer the marginal probability of each intelligent agent node in the group topology accepting false information, select candidate intervention nodes according to the marginal probability, and deploy intervention intelligent agent nodes according to the adjusted configuration ratio, and implement intervention on the candidate intervention nodes. The intervention intelligent agent nodes are configured to generate intervention text content.

[0070] S4, Automatedly evaluate the intervention results to generate an evaluation signal, and feed the evaluation signal back to the dynamic adjustment step to replace the initial evaluation signal.

[0071] The main parts of the present invention will be described in detail below.

[0072] I. Group topology generation.

[0073] This section generates an agent swarm structure driven by a large language model, featuring multiple simulated roles, to replicate the heterogeneous user distribution in a social network. It serves as the initialization module for subsequent intervention ratio adjustment. By modeling social network nodes as agent roles, they are categorized into k types. The configuration ratios of agent nodes with different roles form a role proportion vector. , This represents the proportion of agent nodes of type r among all agent nodes. T is the transpose symbol. ,and , Let be the empirical upper bound for the proportion of agent nodes in role r. Roles can be categorized as follows: disinformation spreader, radical ignorant person, conservative ignorant person, disinformation skeptic, anti-disinformation spreader, and information blocker, with the latter three types being interventionist roles. The specific process is as follows.

[0074] First, we need to obtain a dataset of real user interactions during a specific event to automatically generate a group topology reflecting user heterogeneity based on the existing distribution. Assume the number of real users in the available data is... The input data includes The edges and state information of each user node are described. Regarding edge relationships, each edge represents the historical behavioral interaction between nodes (e.g., if user A forwards, replies to, or references user B, there is a directed edge from B to A to indicate the information flow). The state information of each user node includes structural features S and behavioral features. Content characteristics Credibility characteristics Four main categories of characteristics. Users The status information includes:

[0075] Structural features Such as out-degree, in-degree, and PageRank value;

[0076] Behavioral characteristics Such as posting frequency, forwarding frequency, reply frequency, and citation frequency;

[0077] Content Features Such as discussion topics, stances, emotional polarity, and discourse style based on the LDA model;

[0078] Credibility features Such as suspiciousness (e.g., whether the published content is frequently forwarded or has duplicate text, whether the username in the account information is random and irregular, whether the avatar or introduction is missing, etc.) and source reliability (the proportion of commonly used external link domains in the trusted list).

[0079] Based on the above real user interaction data, a Gaussian mixture model is used for clustering to fit the statistical distribution parameters of various roles across different features, approximating a Gaussian distribution. For the first... For intelligent agent nodes of the same type, each feature will approximately satisfy a probability distribution. , , , , Indicates the first The structural characteristics, behavioral characteristics, content characteristics, and credibility characteristics of intelligent agent nodes of the role type. Represents a normal distribution; Representing structural features Behavioral characteristics Content characteristics With credibility features The mean, Representing structural features Behavioral characteristics Content characteristics With credibility features The variance of these parameters can be statistically calculated based on massive amounts of real user data. Furthermore, clustering operations can yield the first-order variance of real events. Role proportion vector of agent nodes of the same type .

[0080] Obtain the character percentage vector Then, the characteristics of agent nodes and the group topology in the social network simulation can be generated. Assume the number of agent nodes is... , The initial number of agent nodes for each role is: , Features of each agent node (structural features) Behavioral characteristics Content characteristics With credibility features That is, sampling is performed based on the distribution of role characteristics in real data.

[0081] A group topology is generated based on a group connectivity preference mechanism to fit the real social structure. This mechanism aims to ensure that the topological characteristics (in-degree and out-degree) of nodes of each role type in the network conform as closely as possible to the empirical distribution fitted after clustering. For any agent node... The role type to be assigned is , The group connection preference mechanism is defined by the following scoring function. Optimization issues:

[0082] ;

[0083] in, and For intelligent agent nodes out-degree and in-degree, and For the role The mean of the probability distribution of out-degree and the mean of the probability distribution of in-degree in structural features. and For the role The standard deviations of out-degree and in-degree in structural features A relatively small constant is set to avoid division by zero errors in the calculation. This scoring function measures the deviation between the structural characteristics of simulated nodes and the distribution of real roles, thereby enabling population topology generation based on real-world data.

[0084] In-degree represents the number of edges pointing to the current agent node, that is, how many other agent nodes have interacted with the agent node through historical behavior by forwarding, replying or referencing. A high in-degree means that the user's content is widely disseminated or referenced and has a high content influence in the current social network.

[0085] Out-degree represents the number of edges pointing from the current agent node to other agent nodes, that is, how many other agent nodes the agent node actively forwards, replies to, or references. A high out-degree means that the user is active, frequently participates in the process of information re-propagation, and plays the role of an information disseminator.

[0086] II. Adjustment of intervention ratio.

[0087] like Figure 2 As shown, this part, based on Bayesian optimization, dynamically adjusts the deployment ratio of different intervention roles after receiving evaluation signals from the automated evaluation part. This gradually adjusts the role distribution during the intervention process to an optimal or sub-optimal state, improving intervention efficiency and strategy flexibility. The optimization of the intervention ratio is based on the following system feedback function. :

[0088] ;

[0089] in, The evaluation signals fed back by the automated evaluation component can be viewed as a role proportion vector. The black-box function uses an initialization evaluation signal upon first use. An intervention cost function defined by domain experts based on experience. To balance the parameters, based on Bayesian optimization theory, we first... The prior probability distribution is approximately a Gaussian process:

[0090] .

[0091] in, and These represent the role percentage vectors of agent nodes with two different roles; This represents a Gaussian process, i.e., for input... For any set, the corresponding output follows a multivariate normal distribution, which is uniquely determined by the mean function and the covariance function. It is the mean function (assumed to be 0); The covariance function is usually . A kernel function, including Automatic Correlation Detection (ARD), automatically identifies the importance of input dimensions:

[0092] ;

[0093] ;

[0094] in, The signal variance that determines the overall amplitude for and The Euclidean distance between them A scaling factor to control similarity decay. and This represents the proportion of the two different role proportion vectors in the d-th role category. Indicates the total number of character categories. This represents the scaling factor that controls similarity decay based on the proportion of characters in class d. If the proportion of a certain type of character... A large value indicates that small changes in this dimension have little impact on the prediction results; conversely, a small value indicates that this dimension is important and requires attention. It is a hyperparameter in the Gaussian process, which is automatically optimized based on historical data using the Automatic Correlation Discrimination (ARD) mechanism, and is used to quantify the contribution of each role. This represents the index of the d-th character in the character proportion vector.

[0095] Under the assumptions of the Gaussian process mentioned above, based on Bayesian theory, Estimate the posterior probability in Observational data were obtained in the intervention simulation of the wheel. ,in, This represents the role proportion vector used in the t-th round of intervention simulation of the system. This indicates that the role proportion vector is used in the t-th round of intervention simulation. The feedback value afterward; The posterior probability distribution is estimated as follows Assuming Mean function in prior probability If the mean is zero, then the mean of the posterior probability distribution is... , Expressing expectations, Indicates transpose. This represents the custom observation noise variance. for The covariance vector of the vector representing the proportion of historical roles, ,in For kernel function, For the kernel matrix, elements in , It is the identity matrix. For the observed vector, ; Variance of the posterior probability distribution , Represents the variance of observation noise. for The autocovariance.

[0096] To balance the excessive variance or mean during the optimization process, based on the system feedback function... Define the expected gain function to be optimized. ,in This indicates the optimal value of the currently known observations.

[0097] Bayesian optimization theory can prove that the expected gain function has an analytical expression:

[0098] ;

[0099] in, , and These are the probability distribution function and probability density function of the standard normal distribution, respectively. To iterate the role proportion vector for the next time step... The expected gain function is optimized and solved.

[0100] , ;

[0101] in, show The one-dimensional simplex space of probability, i.e. , ; The given first number is restricted by experts The maximum upper bound for the deployment ratio of role-based intelligent agent nodes.

[0102] Thus, the updated character proportion vector value is obtained. The data is then input into the multi-agent cognitive intervention component for intervention probability inference, and new observations are obtained through the automated evaluation component. Update observation data The system can continuously optimize and iterate the intervention ratio strategy based on this module until convergence, thereby achieving dynamic optimization of the intervention effect.

[0103] III. Multi-agent cognitive intervention.

[0104] like Figure 3 As shown, this part combines probabilistic graphical models and large language models to model the mechanism by which intelligent agent nodes believe in the spread of false information, effectively simulating and predicting the marginal probability of users believing in false information in social networks; at the same time, it can configure intervention-type intelligent agents to automatically generate intervention texts, achieving efficient intervention and guidance.

[0105] The simulated social network has been constructed as a directed graph-based population topology. In this topology, each agent node represents a user, and its binary variables... Indicates whether or not false information is accepted. (Agent node) Pointing to the agent node edge Represents intelligent agent nodes For agent nodes The graph uses the forwarding, replying, or citation relationships of posts to represent the information flow that might influence the stance of agent nodes towards false information. Each edge in the directed graph must have a weight. , , That is, intelligent agent nodes The neighboring nodes are weighted to represent the conditional propagation probability of each edge, thereby constructing the joint distribution of the entire directed graph and deriving the marginal probability of a node believing false information.

[0106] Specifically, to align the conditional propagation probability settings with real-world data and generalize to simulated social networks, a deep neural network is used to train the conditional propagation probability, making the estimates more consistent with real-world patterns. The training process for the conditional propagation probability can be represented as follows: Among them, multidimensional feature vectors For network input, i.e., agent nodes A concatenated vector comprising four dimensions: structural features, behavioral features, content features, and credibility features. This represents the constructed deep neural network model. The parameters are optimized. This can accurately fit the information propagation probability in the real world. In label construction, it is based on whether the label is generated by an intelligent agent node in the real event dataset. Triggered the agent node The phrase "believe misinformation" is used to construct a binary classification label (1 for positive examples and 0 for negative examples). Since causal relationships are difficult to observe in real social network datasets, the labeling of such propagation events often employs approximate inference, that is, judging the agent nodes based on the chronological order of believing misinformation. Does the state of believing false information affect the intelligent agent node? Impact. Based on the actual occurrence of transmission. For example, the specific rules for tag construction are as follows:

[0107] Positive sample label: if and only if agent node Becoming a node that believes false information, and an intelligent agent node. At the node Later, when nodes believe false information, there will be directed edges. The spread of the label is recorded as a positive example.

[0108] Negative sample label: A directed edge will be added when any of the following conditions are met. The propagation label is recorded as a negative example. 1) Propagation failure scenario: agent node For nodes that believe false information, but within the observation period, agent nodes 1) Not becoming a node that believes false information; 2) No causal situation: agent node A node that believes false information, but its neighboring intelligent agent nodes Not becoming a node that believes false information, i.e., a node The state change is not caused by the node trigger.

[0109] After training on a real-world event dataset, each agent node in the simulated social network can input its feature vectors into the deep neural network model. To obtain the conditional propagation probability That is, the edge weights between nodes in the social topology. Unlike traditional graph neural networks that rely solely on neighbor feature aggregation, this module, after outputting the conditional propagation probability, can further infer the marginal probability of each agent node believing false information, thus characterizing the information flow propagation process in a more analytical and non-black-box manner. The specific design is as follows.

[0110] In the simulated social topology, a probabilistic graphical inference method based on belief propagation can be used to calculate the marginal probability of believing false information by iterating through messages. Specifically, for neighboring agent nodes... To the agent node to be estimated Iteration information It can be represented as:

[0111] .

[0112] in, For intelligent agent nodes The prior probability (the probability that an agent node's own feature "is prone to believing false information" without considering neighbor propagation can be used to train an additional prior classifier based on real-world data). For Bernoulli probabilistic boundary potential functions:

[0113] .

[0114] This function only applies to two connected agent nodes. Its value equals the effective conditional propagation probability only when both parties are in a state of believing false information; otherwise, it is a neutral value. ; Represents intelligent agent nodes In addition to intelligent agent nodes By simulating multi-hop propagation of information through multiple iterations, neighboring intelligent agent nodes can cover the influence of indirect neighbors and reflect a more complete information flow.

[0115] Based on the above iterative information Agent nodes to be estimated Believing in false information has a marginal probability This can be expressed as,

[0116] ;

[0117] in, To be classified into one item, For intelligent agent nodes Believing in the prior probability of false information The multiplication symbol represents the aggregation of all agent nodes. Iterative information about neighbors. For large-scale node graphs in simulated social networks, the above message passing algorithm can be executed in parallel on multiple GPUs using a professional graph computing framework (such as PyTorch Geometric) to parsely derive the marginal probability of believing false information for each agent node to be estimated.

[0118] After performing a marginal probability analysis on each agent node regarding the belief in false information, a set of high-risk nodes is selected as candidate intervention nodes based on whether the probability exceeds a set threshold. After screening, candidate intervention nodes in the high-risk node set are ranked based on two factors: first, the probability of their current belief in misinformation; and second, structural characteristics (such as in-degree / out-degree, PageRank value, and distance from the source of misinformation). A hierarchical ranking method is used here, first ranking by marginal probability, and then by structural characteristics if the probabilities are the same. A predetermined number of candidate intervention nodes at the top of the ranking are identified as high-risk or misinformation source types, a predetermined number of candidate intervention nodes at the bottom of the ranking are identified as highly susceptible or potential spread types, and other candidate intervention nodes are identified as high-influence or core spread types. "Top" means that the node is the most urgent intervention target at the current time, while "bottom" means that the node has lower risk or is in a passive state. Nodes at the top have a higher marginal probability; when marginal probabilities are equally high, nodes ranked higher have stronger structural characteristics for spread (higher in-degree / out-degree, higher PageRank, or closer to the source of misinformation).

[0119] After ranking the candidate intervention nodes, the optimized vector of agent node role proportions for different roles is used. Calculate the number of deployed intelligent agent nodes for each role. , The system then strategically deploys intervention-oriented intelligent agent nodes within the existing topology. For these agents, a pre-defined prompt template is configured so that corresponding intervention text is generated when the agent is invoked. The specific deployment method is as follows: Based on the marginal probability of believing misinformation and topological characteristics, candidate intervention nodes can be categorized into three types: high-harm or misinformation source type, high-influence or core propagation type, and highly susceptible or potential propagation type. Correspondingly, intervention can be conducted using three types of intervention-oriented intelligent agents: information blockers, anti-misinformation propagators, and misinformation skeptics, designed as follows.

[0120] Skeptics of misinformation: Implicit behavioral guidance is provided to highly susceptible or potentially disseminating candidate intervention points, such as by questioning or reminding them to weaken their willingness to believe misinformation;

[0121] The prompt template is as follows:

[0122] You are a regular user who finds the content of this message suspicious after seeing it.

[0123] - Tone: Non-confrontational, curiosity-driven, asking questions rather than directly refuting;

[0124] -Task: Generate one or more sentences of slightly questioning or reminder information to make the target user hesitate;

[0125] - Style: Use interrogative / hinting sentences to avoid excessive criticism;

[0126] Example output:

[0127] Is the source of this information reliable?

[0128] "It seems like the data doesn't have a source; should we double-check it?"

[0129] "I haven't seen similar information elsewhere, so I may need to verify it again."

[0130] Counter-spreaders of misinformation: Personalized persuasion of high-influence or core dissemination candidate intervention nodes, and release credible rebuttal information to offset their diffusion effect;

[0131] The prompt template is as follows:

[0132] You are a rational and influential social media user who has seen someone spreading misinformation.

[0133] - Tone: calm and rational, citing authoritative information;

[0134] -Task: Generate a concise and persuasive response, providing credible sources or facts to refute the misinformation;

[0135] - Style: It conveys a certain degree of authority while avoiding emotional attacks;

[0136] Example output:

[0137] "According to a certain media report, this statement is not true and has now been clarified...";

[0138] "Statistical data shows that this phenomenon has not occurred, and the screenshots in the rumors have been spliced ​​together."

[0139] "If you look at the announcements from authoritative institutions, you will find that such claims are misleading."

[0140] Information blockers: Block or downgrade candidate intervention nodes that pose a high risk or are the source of false information, thereby reducing the spread of false information at its source;

[0141] The prompt template is as follows:

[0142] You play the role of a platform administrator or an automated detection system;

[0143] - Tone: authoritative, clear, and non-negotiable;

[0144] -Task: Generate blocking or demotion prompts for high-risk nodes;

[0145] - Style: Simple, formal, and permissive.

[0146] Example output:

[0147] "This account has had its forwarding function restricted due to repeatedly posting false information."

[0148] "This content has been marked as false information by the platform, and its dissemination is restricted."

[0149] "Illegal dissemination behavior has been detected. The account has been placed on the watch list and its influence has been reduced."

[0150] After the intervention agent is deployed and measures are implemented, the remaining agents in the network generate content based on the event, simulating the evolution of the event discussion. After all agents have finished speaking, a scorer based on a large language model assigns a continuous stance value to the stance of the statements. The value is +1, which indicates that the information is extremely false and -1 indicates that the information is extremely false.

[0151] IV. Automated evaluation.

[0152] This section is used for multi-dimensional, end-to-end automated evaluation of intervention results to quantify the effectiveness of intervention strategies and provide data feedback for strategy optimization. (Regarding the system simulation start time...) Simulation time interval Set of nodes successfully intervened It can be represented as:

[0153] ;

[0154] in Indicates the first Each intelligent agent node The attitude value towards false information at each time step: -1 indicates extreme disbelief in false information, and +1 indicates extreme belief in false information.

[0155] The evaluation of intervention outcomes includes the following five indicators:

[0156] Intervention transmission rate This indicates the speed at which effective intervention information spreads within a social network per unit of time.

[0157] .

[0158] Intervention in the depth of transmission This represents the average propagation path depth of effective intervention information in social networks, where... This represents the propagation path length between the successfully intervened agent node and the agent node that initiated the intervention:

[0159] .

[0160] Intervention spread This represents the proportion of nodes in the social network covered by intelligent agent nodes that have effectively intervened. Total number of agent nodes:

[0161] .

[0162] Opinion conversion rate This indicates the proportion of target agent nodes whose cognitive stance changes after intervention:

[0163] .

[0164] Sentiment Score Tendency This represents the overall emotional polarity of all intelligent agent nodes in information interaction, where... It is the set of all agent nodes:

[0165] .

[0166] The final intervention effect is a linearly weighted average of the normalized indicators mentioned above, with the weights of each indicator flexibly configured by expert users according to the key intervention objectives.

[0167] .

[0168] in, Representing the Evaluation indicators The normalized value, for Corresponding weights. Evaluation signal. Feedback is fed back to the second part of the intervention ratio adjustment, and the intervention effect is gradually iterated to reach the optimal or suboptimal state.

[0169] Through comprehensive analysis and signal feedback in this section, the system can achieve the formulation of optimal intervention strategies through human-machine collaboration. That is, experts formulate the key points of the strategy, and the system automatically completes data evaluation, feedback, and optimization, thereby improving the pertinence, timeliness, and sustainability of the intervention.

[0170] This invention proposes a controllable, dynamically optimized, and accurately predictive cognitive intervention framework, which covers four parts: group topology generation, intervention ratio adjustment, multi-agent cognitive intervention, and automated evaluation, providing systematic support for existing cognitive intervention applications.

[0171] In the group topology generation part, this invention automatically generates diverse role configurations based on the distribution of real social network data in scenarios of false information dissemination, thereby more closely resembling the complex propagation environment of actual social networks.

[0172] In the intervention ratio optimization section, this invention utilizes feedback signals from an automated evaluation module, combined with a Bayesian optimization method, to continuously update the ratio configuration of different intervention roles, enabling the role distribution to gradually converge to an optimal or suboptimal state, thereby improving the flexibility and efficiency of intervention.

[0173] In the multi-agent cognitive intervention part of this invention, the invention predicts the marginal probability of each agent node believing false information through probabilistic graphical reasoning, and then deploys intervention agents in a targeted manner to achieve differentiated and targeted intervention measures.

[0174] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0175] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0176] In one embodiment, the present invention provides a computer device, which may be a server. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.

[0177] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0179] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A cognitive intervention method using a social simulator, characterized in that, include: A group topology is generated based on user interaction data from real social networks. The group topology contains multiple intelligent agent nodes playing different roles, including at least one interventionist intelligent agent node that can have a negative effect on the spread of misinformation. Dynamic adjustment step: Based on the initial evaluation signal and Bayesian optimization method, the configuration ratio of at least one intervention agent node in the population topology is dynamically adjusted; Based on the probabilistic graphical model, the marginal probability of each agent node in the group topology accepting false information is inferred. Candidate intervention nodes are selected according to the marginal probability, and intervention agent nodes are deployed according to the adjusted configuration ratio to intervene in the candidate intervention nodes. The intervention agent nodes are configured to generate intervention text content. The intervention results are evaluated to generate an evaluation signal, which is then fed back into the dynamic adjustment step to replace the initial evaluation signal.

2. The social simulator cognitive intervention method according to claim 1, characterized in that, The generation of community topology based on user interaction data from real social networks specifically includes: Acquire user interaction data, which includes the connection relationships and status information of user nodes in real social networks; the status information includes the structural features, behavioral features, content features, and credibility features of user nodes; The user interaction data is clustered using a Gaussian mixture model to obtain the feature distribution of multiple role types and the initial role proportion vector; The number of agent nodes playing each role is determined based on the total number of agent nodes and the initial role proportion vector, and the node features of the corresponding agent nodes are generated by sampling according to the feature distribution of different roles. The group connection preference mechanism is used to construct the edge relationships between the agent nodes to form the group topology, and the group connection preference mechanism is used to make the topological features of agent nodes of each role type conform to the corresponding feature distribution.

3. The social simulator cognitive intervention method according to claim 2, characterized in that, The process of constructing edge relationships between agent nodes based on a group connection preference mechanism to form the group topology, wherein the group connection preference mechanism is used to ensure that the topological features of agent nodes of each role type conform to the corresponding feature distribution, specifically including: For any intelligent agent node The role to be assigned is The group connection preference mechanism is defined as the following scoring function. Optimization issues: ； in, and For intelligent agent nodes out-degree and in-degree, and For the role In structural features The mean of the probability distribution of out-degree and the mean of the probability distribution of in-degree in the given data. and For the role In structural features The standard deviations of the out-degree and in-degree are in the range. This is a set constant.

4. The social simulator cognitive intervention method according to claim 1, characterized in that, The method of dynamically adjusting the configuration ratio of at least one interventionist agent node in the population topology based on the initial evaluation signal and Bayesian optimization includes: Define system feedback function ,in, A role proportion vector that includes the configuration ratio of interventionist agent nodes. To evaluate the signal, Let λ be the intervention cost function, and λ be the trade-off parameter. The system feedback function is optimized using the Bayesian optimization method. The prior probability distribution is modeled as a Gaussian process, and calculations are performed based on historical observation datasets. The posterior probability distribution; The expected gain function is constructed based on the system feedback function and the corresponding posterior probability distribution, by maximizing the expected gain function. Let be the expected gain function of the variable, for Update.

5. The social simulator cognitive intervention method according to claim 4, characterized in that, The system feedback function is optimized using the Bayesian optimization method. The prior probability distribution is modeled as a Gaussian process, and calculations are performed based on historical observation datasets. The posterior probability distribution specifically includes: Based on Bayesian optimization theory, the system feedback function is first... The prior probability distribution is approximately a Gaussian process: ； in, Represents a Gaussian process. and These represent two different character proportion vectors. It is a mean function. It is the covariance function; Under the assumptions of the Gaussian process mentioned above, based on Bayesian theory, Estimate the posterior probability in Observational data were obtained in the intervention simulation of the wheel. ,in, This represents the role proportion vector used in the t-th round of intervention simulation of the system. This indicates that the role proportion vector is used in the t-th round of intervention simulation. The feedback value afterward; The posterior probability distribution is estimated as follows Assuming Mean function in prior probability If the mean is zero, then the mean of the posterior probability distribution is... , Expressing expectations, Indicates transpose. This represents the custom observation noise variance. for The covariance vector of the vector representing the proportion of historical roles, ,in For kernel function, For the kernel matrix, elements in , It is the identity matrix. For the observed vector, ; Variance of the posterior probability distribution , Represents the variance of observation noise. for The autocovariance.

6. The social simulator cognitive intervention method according to claim 4, characterized in that, The expected gain function is constructed based on the system feedback function and the corresponding posterior probability distribution, by maximizing the... Let be the expected gain function of the variable, for The update includes: Based on system feedback function Define the expected gain function to be optimized. ,in, Expressing expectations, , Let represent the i-th feedback value in the observed data; based on Bayesian optimization theory, the analytical expression of the expected gain function is obtained: ； intermediate variables , and These are the probability distribution function and probability density function of the standard normal distribution, respectively. for The mean and standard deviation of the posterior probability distribution; the expected gain function is optimized to obtain... Role proportion vector of agent nodes at each time step : ， ； in, show One-dimensional simplex space with unit probability. This represents the total number of roles of the agent nodes, i.e. , ; The given first number is restricted by experts The upper bound of the maximum deployment ratio of role-based intelligent agent nodes. This represents the index of the r-th type of role.

7. The social simulator cognitive intervention method according to claim 1, characterized in that, The inference of the marginal probability of each agent node in the population topology accepting false information based on the probabilistic graphical model specifically includes: A deep neural network model is trained based on a real event dataset to calculate the conditional propagation probability between two connected agent nodes based on their state information, which is then used as the weight of the corresponding directed edge in the group topology. Based on the belief propagation algorithm and the conditional propagation probability, the iterative information transmitted by the neighboring nodes of each agent node is iteratively calculated to calculate the marginal probability of each agent node accepting false information. The iterative information is calculated based on the agent node's prior probability of believing false information, the edge potential function, and the iterative messages passed by other neighbor nodes. The edge potential function is defined by the conditional propagation probability.

8. The social simulator cognitive intervention method according to claim 1, characterized in that, The process involves selecting candidate intervention nodes based on the marginal probabilities and deploying intervention-type intelligent agent nodes according to the adjusted configuration ratio to intervene in the candidate intervention nodes. The intervention-type intelligent agent nodes are configured to generate intervention text content, specifically including: From the population topology, agent nodes with edge probabilities higher than a set probability threshold are selected to form a high-risk node set. Based on the adjusted configuration ratio of intelligent agent nodes, determine the deployment quantity of at least one type of intervention-type intelligent agent node, which includes misinformation skeptical intelligent agent nodes, anti-misinformation dissemination intelligent agent nodes, and information blocker intelligent agent nodes. A hierarchical ranking method is used to sort candidate intervention nodes in the high-risk node set based on their marginal probabilities and structural characteristics. These structural characteristics include the out-degree, in-degree, PageRank value of the agent nodes, and the network distance between the candidate intervention nodes and the source of misinformation. A predetermined number of candidate intervention nodes at the top of the ranking are identified as high-risk or misinformation source types; a predetermined number of candidate intervention nodes at the bottom of the ranking are identified as highly susceptible or potential propagation types; and other candidate intervention nodes are identified as high-influence or core propagation types. Based on the categories of the candidate intervention nodes, corresponding intervention agent nodes are matched and deployed, wherein: For candidate intervention nodes classified as highly susceptible or potentially propagating, the misinformation skeptic agent node is deployed. The misinformation skeptic agent node is configured to generate intervention text content containing questioning or reminder content based on the first prompt word template. For candidate intervention nodes classified as high-influence or core-propagation type, the anti-misinformation propagator intelligent agent node is deployed. The anti-misinformation propagator intelligent agent node is configured to generate intervention text content that includes refutation by citing authoritative sources based on the second prompt word template. For candidate intervention nodes classified as high-risk or sources of misinformation, the information blocker intelligent agent node is deployed. The information blocker intelligent agent node is configured to generate intervention text content containing account or content restriction prompts based on a third prompt word template.

9. A social simulator cognitive intervention method according to claim 1, characterized in that, The step of evaluating the intervention results to generate an evaluation signal and feeding the evaluation signal back to the dynamic adjustment step to replace the initial evaluation signal specifically includes: Set of indexes of successfully intervened agent nodes ; Indicates the first Each intelligent agent node The stance value on misinformation at the time step. This indicates a strong disbelief in the false information. This indicates an extreme belief in false information. To simulate the start time of the system, This is a simulated time interval; This represents the index of the i-th agent; Evaluation indicators include: Intervention transmission rate : ; Intervention in the depth of transmission : ;in This represents the propagation path length between the successfully intervened agent node and the intervention source agent node. These are the indices of the successfully intervened agent node and the intervention source agent node, respectively; Intervention spread : ; The total number of all agent nodes; Opinion conversion rate : ; Sentiment Score Tendency : ; It is a set of indices for all agent nodes; Final evaluation signal for: ; An index for evaluation metrics; in, Representative evaluation indicators The normalized value, for The corresponding weights.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images