Complex social behavior simulation and public opinion deduction method based on multi-agent

Abstract

The present invention relates to the technical field of social network analysis and generative artificial intelligence, and discloses a method for simulating complex social behavior and deducing public public opinion based on multiple agents. Social groups are divided, and users in a social network are divided into key opinion leader communities and ordinary user communities; key opinion leader agents and ordinary user agents driven by a large language model are respectively constructed; a dynamic weighted directed network is constructed, and each key opinion leader agent is modeled as a node. The neighborhood of the network node is weightedly calculated by the opinion index and the node influence, and the sentiment score is calculated based on the text sentiment index. The neighborhood of the network node is dynamically adjusted based on the sentiment score, and the method is iterated several times to complete the deduction and simulation of network public opinion. The method solves the problem of difficulty in modeling the dissemination of public opinion in a cross-domain dynamic context, and can more accurately simulate the differences among individual public opinion users and complex social interaction relationships. The method is practical and feasible.

Description

Technical Field

[0001] The present invention relates to the field of social network analysis and generative artificial intelligence technology, and specifically to a method for simulating complex social behaviors and deducing public opinion based on multiple intelligent agents. Background Art

[0002] Online public opinion management primarily analyzes historical public opinion, sentiment, and trends expressed on various social media platforms to predict future trends and potential opinions. Its goal is to effectively provide businesses and social organizations with a window into public opinion and assist in policy formulation. Public opinion prediction is a key technical approach to online public opinion analysis.

[0003] Currently, public opinion deduction tasks are mainly faced with problems such as low effective data density, diverse expression methods, and strong real-time requirements. Existing public opinion deduction methods can be roughly divided into two categories: statistical learning-based methods and simulation-based methods.

[0004] Statistical learning-based methods primarily analyze historical data and construct feature engineering to make probabilistic predictions about factors such as speech text, social behavior, and user influence. However, these methods are less effective when dealing with complex social networks and dynamic scenarios, and their models lack sufficient interpretability, making them difficult to meet the deeper needs of practical applications.

[0005] Simulation-based approaches primarily include multi-agent-based approaches, social network analysis, and a combination of the two. These approaches typically construct rule-based agents, formulate rules to simulate their interactions on complex social networks, and thus simulate the dynamic development of public opinion. However, the most significant challenges with these approaches are rule sensitivity and node staticity. For example, in models of opinion diffusion dynamics, individual behavior is often assumed to follow standardized patterns. While this simplification reduces the computational complexity of the model, it also ignores the impact of individual differences in reality, resulting in limited ability to model complex social interactions. Furthermore, rule-based agents are sensitive to initial parameter settings, making it difficult to accurately simulate the diverse interactions found in real social networks.

[0006] However, public opinion deduction faces the following difficulties: (1) how to accurately quantify and simulate the process of individual cognitive changes; (2) how to build a framework that can accurately simulate real social interactions; and (3) how to achieve effective prediction in a complex social network environment with high-frequency dynamic changes. Existing methods have difficulty in simulating the differences between individual public opinion contributors and accurately modeling complex social interaction relationships.

[0007] During the dissemination of opinions, individual cognitive differences among public opinion holders lead to dynamic and nonlinear evolution of their opinions, influenced by a variety of factors. By deeply studying the evolution of public cognition from the perspective of individual differences and constructing effective models, and mitigating the impact of overly homogenized treatments in traditional methods, the overall performance of complex social behavior simulation and public opinion deduction can be more effectively improved. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for simulating complex social behaviors and deducing public opinion based on multiple intelligent agents.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] The method for simulating complex social behaviors and inferring public opinion based on multi-agents specifically includes the following steps:

[0011] S1. Divide social groups: Divide users in social networks into key opinion leader communities and ordinary user communities;

[0012] S2. Building Agents: By designing different prompt words for the KOL community and the general user community, we construct KOL agents and general user agents based on different base models and driven by large language models.

[0013] S3. Build a social network. Construct a dynamic weighted directed network. Model each key opinion leader agent and ordinary user agent as a node, model social behavior as a directed edge, and model the influence of behavior as edge weight. Simulate the local information propagation of social media with high information cohesion and simulate long-distance connections in social networks.

[0014] S4, weighted calculation of the network node neighborhood by using the opinion index and the node weight corresponding to the ordinary user agent;

[0015] S5. Evaluate the text sentiment index based on SnowNLP to calculate the sentiment score;

[0016] S6. Dynamically adjust the weight of the network node opinion index and the threshold of neighborhood judgment, i.e., the size of influence, based on the level of sentiment score;

[0017] S7. Iterate the steps S4-S6 several times to complete the deduction and simulation of online public opinion.

[0018] In the present invention, a multi-agent system (Multi Agent System) driven by LLM is constructed to realize multi-agent simulation, combine the dynamics of opinion propagation and information processing theory to study the cognitive changes of the target population, and provide guidance strategies for specific events. First, based on the classic BDI architecture, a large language model-driven agent (LLM-Agent) with long and short-term memory capabilities is constructed to create a Muilt-Agent pool. Subsequently, since the user participation in social networks is roughly power-law distributed, that is, the main body of public opinion is usually a minority of key opinion leaders (KOLs), the Muilt-Agent pool is divided into key opinion leader communities and ordinary user communities, constructing a multi-layer network with the key opinion leader community as the core layer and the ordinary user community as the access layer.

[0019] In the present invention, the core of public opinion deduction is to construct an intelligent agent to simulate the social behavior of the user, the subject of public opinion, and then build a specific indicator system to evaluate the deduction effect. In combination with traditional ABM, a social user intelligent agent is constructed. Each intelligent agent includes a sensor, a controller, and an effector. The input of the sensor is structured information extracted from knowledge. The controller controls the intelligent agent to perform corresponding social behaviors based on the information provided by the sensor. The effector controls the action trend of the intelligent agent, that is, provides the intention of cognitive evolution.

[0020] In the present invention, preferably, after constructing the intelligent agent, the opinion pool of the social group is modeled as a neighborhood:

[0021] ,

[0022] in x i and x j Represents the roles played by the intelligent agents i and j The opinion index of the agent is x When the difference is less than the threshold ε, the corresponding two agents are neighbors of each other, and all shared neighboring agents are collectively called neighborhoods, which are expressed as , which is the neighbor set of role i.

[0023] In the present invention, preferably, a dynamic weighted directed network is constructed, with each key opinion leader agent modeled as a node, social behavior modeled as a directed edge, and behavioral influence modeled as edge weight, to simulate the local information propagation of social media with high information cohesion. Social media includes Weibo super topics, WeChat group chats, and Telegram groups, etc. Ordinary user agents are modeled as single access nodes to simulate long-distance connections in social networks. The constructed dynamic weighted directed network is iterated. In each iteration process, the information of the access node is the text and corresponding parameters generated by each round of iteration of the core layer. Each ordinary user agent is then combined with open source intelligence related to the deduced target topic to freely express opinions in the simulated Weibo space. Subsequently, public opinion is gradually formed in multiple rounds of iterations with reference to the ideas of social selection theory. The policy publisher can formulate corresponding communication guidance suggestions based on the formed opinions.

[0024] In the present invention, preferably, the weighted calculation of the network node neighborhood by the opinion index and the node influence specifically includes:

[0025] Sentiment analysis is performed on text generated by the large language model to obtain a sentiment index, which is then mapped to the dynamic weight of the intelligent agent within a dynamic weighted directed network neighborhood. The user influence is quantified using the emotional contagion theory to efficiently achieve authenticity and effectiveness between the large language model and humans.

[0026] Based on the dynamics of opinion propagation, a dynamic attenuation mechanism is added to the neighborhood judgment to simulate the self-service mechanism;

[0027] Based on the classic Hegselmann-Krause algorithm, the opinion convergence mechanism is simulated, that is, the weighted average of opinions within the neighborhood is calculated to simulate the transmission of information and the formation of cognitive opinions;

[0028] Network differences are used to simulate high locality and long-distance connectivity to divide key opinion leaders (KOLs) and ordinary users into two major groups. Specifically, directed relationships within the network are used to simulate different groups. For the highly localized KOL group, because information dissemination within this group is highly clustered and typically spreads within the community, relationships within this group are bidirectional, with no outgoing directed relationships with the outside world. For the ordinary user group with long-distance connectivity, because the connection between this group and the KOL group is less stable, each individual has a unidirectional directed relationship with the KOL, but no relationships exist within the group.

[0029]

[0030] In the formula w j Neighborhood role j The weight of x i yest The opinion index within a time period, the initial value can be randomly selected or preset based on a specific event or target strategy.

[0031] In the present invention, preferably, the influence of high-weighted agents in the neighborhood is increased based on the weighted average formula.

[0032] ,

[0033] In the formula Neighborhood role i exist t -1 moment weight, ε is a hyperparameter, ε (t) for t Threshold for time neighborhood determination, hyperparameters α Controlling the influence of weights on the trust range, hyperparameters β Controlling the decay rate. At time t in each iteration, a check is performed on the neighborhood Ni of character i. Let T be the preset total number of iterations (a preset hyperparameter, i.e., the number of iterations). t / T is the "decay rate," which increases with time t. By "subtracting 1," the overall threshold gradually decays. The value of the beta parameter controls the decay rate, as can be seen from the formula. We model an individual's trust in others using a "neighborhood" (as shown in the previous formula). All individuals trusted by an individual are considered "neighbors" (i.e., those whose opinion index x is less than the threshold). The set of all "neighbors" is the "neighborhood." However, in reality, individual trust in others is unequal, and many factors are involved. We model this by assuming that people with "intense emotions" are more likely to "infect others" (i.e., in reality, our opinions may differ significantly, but if your emotions are more high and exciting, I am more likely to accept your opinion, such as a politician's passionate speech). The weight w is calculated in the next formula. This formula indicates that people with high emotion scores are more likely to infect others, and their "weight" in the neighborhood is higher (higher emotion scores correspond to higher weight w). The overall model is built by "easier to enter the neighborhood", that is, people with high emotions will have higher weights, which will lead to a higher threshold for "neighborhood" judgment. The hyperparameter alpha is used to control the impact of the w weight on the "trust range" or "threshold".

[0034] In the present invention, weights are preferably introduced into the calculation of the trust range for the agents in the constructed neighborhood, with key opinion leader agents receiving higher weights than ordinary user agents. This ensures that highly weighted agents are more trusted by neighborhood users, raising the threshold for neighborhood judgment, i.e., influence. First, exponential decay is used to ensure that the trust range decreases with increasing iteration rounds, simulating the tendency of users to maintain their fixed views (i.e., a self-serving mechanism) in long discussions in real social networks. This inclusion of weights ensures that highly weighted agents can, to a certain extent, resist the effects of round decay.

[0035] In the present invention, preferably, for weight w The value selection and dynamic adjustment scheme is to align the intelligent agent driven by the large language model with humans, using natural language as the medium and calculating the sentiment score based on the text sentiment index. Specifically, the value of the intelligent agent's weight is dynamically adjusted:

[0036] ,

[0037] In the formula s Represents the sentiment score of the current agent, which is calculated using the SnowNLP tool to generate the text. The sentiment score s ranges from 0 to 1, with values ​​closer to 1 indicating more positive sentiment and values ​​closer to 0 indicating more negative sentiment. Based on other research experience, if s>0.6, it is classified as positive; if s<0.4, it is classified as negative; and if 0.4≤s≤0.6, it is classified as neutral. The manipulation weight is dynamically increased or decreased in the direction of emotion. δ To prevent excessive pessimism from causing weight collapse, the lower limit of the weight is set to 0.5.

[0038] In the present invention, since the discussion between agents is driven by emotions, in order to prevent the generated natural language from falling into excessive emotions and causing the weight to grow rapidly, a nonlinear saturation function is used to smooth the growth of the weight. Dynamically adjusting the neighborhood of network nodes based on the emotion score includes using a sigmoid function to adjust the weight of the agent. w limited to between 0.5 and 2.0,

[0039] ,

[0040] The parameters k Controls the steepness of the adjustment, c is the center point of sigmoid.

[0041] In the present invention, preferably, after using the sigmoid function to smooth the change of weights, an excitement suppression mechanism is added after each round of iteration to prevent excessive excitement from causing uncontrolled growth and causing the generated natural language to lose statistical significance.

[0042] ,

[0043] in γ is the cooling coefficient, which is used to control the recovery speed of each round of weight. w 0 is the target steady-state value of the weight, which can be selected based on the sensitivity of the specific event or target strategy.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This paper utilizes multi-agent systems, large language models, opinion diffusion dynamics, and information processing theory to address the difficulty of modeling public opinion dissemination in a cross-domain dynamic context. It can also more accurately simulate the differences between individual opinion influencers and complex social interactions. The method is practical and feasible. By establishing a complex social network to divide opinion leaders and ordinary user communities, and simulating the behavioral characteristics of different groups, it achieves relatively good public opinion simulation analysis results. The method requires relatively few computing resources and data, eliminating the need for complex upfront resource preparation, and lays a good foundation for subsequent strategic impact assessment research. The method can be flexibly applied to online public opinion analysis research scenarios in a variety of fields, and has reference significance for the design of solutions to related social computing and communication problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the method for simulating complex social behaviors and inferring public opinion based on multiple agents described in the present invention.

[0047] Figure 2 This is a schematic diagram of the structure of the intelligent agents in the multi-agent based complex social behavior simulation and public opinion deduction method described in the present invention.

[0048] Figure 3 This is a simulation result diagram for empirical analysis in the cultural field.

[0049] Figure 4 Simulation results for empirical analysis in the fields of science and art.

[0050] in Figure 3 (a) and (d) are the change graphs of the opinion index of key opinion leaders. Figure 3 (b) and (e) are the change graphs of the opinion index of ordinary users. Figure 3 (c) and (f) are the group sentiment variance change diagrams of ordinary users.

[0051] Figure 4 (a), (b), and (c) are respectively the opinion index change graphs of key opinion leaders in the technology field, the opinion index change graphs of ordinary users, and the group sentiment variance change graphs of ordinary users. Figure 4(d)(e)(f) are the opinion index change graphs of key opinion leaders in the art field, the opinion index change graphs of ordinary users, and the group sentiment variance change graphs of ordinary users, respectively. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0054] See Figure 1 A preferred embodiment of the present invention provides a method for simulating complex social behaviors and inferring public opinion based on multiple agents. By integrating traditional opinion propagation dynamics and generative artificial intelligence technology and utilizing the emergent capability of the large language model (LLM), it is possible to simulate and analyze the social behaviors of users in the process of public opinion propagation, thus solving the problem of difficulty in quantifying and simulating the process of individual cognitive changes. In response to the difficulties in public opinion governance under complex dynamic backgrounds, a multi-level dynamic agent network architecture is constructed, and a dynamic weight evolution strategy and an emotional calming mechanism driven by text emotions are introduced. Based on the improved weighted Hegselmann-Krause algorithm, it is possible to simulate the behavioral characteristics of different groups in the process of public opinion propagation by accurately dividing opinion leaders and ordinary user communities. This method has high computational efficiency and strong adaptability, does not require a large amount of pre-processed data, and can lay a good foundation for subsequent strategic impact evaluation research. At the same time, it provides new theoretical and technical support for the study of complex social systems.

[0055] First, open-source intelligence (OSS) is obtained through media, social platforms, and online channels, such as news websites and social media, generating large amounts of text data. This data is then cleaned and structured into a readable, structured format. A specific dissemination strategy is specified through manual summarization and input into the constructed LLM-based agent in the form of structured text data. The number of LLM-based agents is then interactively determined based on factors such as the scale of the event and cost control. Neighborhoods are created to define the domain, and the agents are further divided into regular user agents and key opinion leader agents according to a power-law distribution. Subsequently, a dynamic neighborhood dissemination simulation is performed using a modified weighted Hegselmann-Krause algorithm based on the theory of opinion dissemination dynamics, based on the input dissemination strategy (structured text data) and randomly generated opinion indices. The LLM agent in the key opinion leader community then outputs text data. Its opinion index is updated to a new value through adjustments to the neighborhood judgment threshold and opinion weight, serving as the opinion index input for the LLM agent in the next iteration. This value is a string of numbers representing the LLM agent's current opinion orientation. The input to the ordinary user agent is the text data output by the key opinion leader agent, and the output is also text data, with the opinion index value selected in the same way. Finally, after the iteration is complete, the user analyzes the evolution of public opinion by reading the output text data of the ordinary user community and observing the changing trends of the opinion index. This last step is designed based on social selection theory, which states that the opinions of the group are "selected and filtered" by ordinary users. This is modeled as users observing the changes in the opinion index and text data of the large language model of the ordinary user community to summarize the evolution of public opinion.

[0056] In this embodiment, since the key opinion leader agent is mainly responsible for the analysis of opinion propagation dynamics in this article, the ordinary user agent is mainly used for information processing and opinion formation, and the number of key opinion leaders is far greater than that of ordinary users, and considering the cost and evolution efficiency, a large language model with better performance and stronger contextual ability is selected as the base model of the key opinion leader agent, and a large language model with higher reasoning efficiency is selected as the base model of the ordinary user agent. The module design of the agent is mainly completed by utilizing the role-playing ability of the large language model, and the preset position is introduced by constructing role information. In terms of system construction, the multi-agent tool LangChain is used to complete the management of long-term memory, the context function of the large language model is used to realize the short-term memory of the agent, and the external parameter assembly of the agent is completed by constructing and assembling the prompt project. The detailed agent structure is shown in the figure. Figure 2 shown.

[0057] In this implementation, we selected the events of "Nezha 2 topped the box office", "Black Myth: Wukong" was released", and "deepseek released the inference model R1" from the three fields of art, culture, and technology. We extracted keywords and summarized the existing public opinion data on social media, sorted out the event outlines and the basic views of netizens, and determined the inference strategy (see Table 1).

[0058] Table 1

[0059]

[0060] In terms of deduction parameters, take the initial trust range ε 0 is 0.2, with a total of 10 iterations. Starting from the third round, the social behavior "debate" is unlocked, with an initial weight of 0.5, and the default starting state is rational speech. For the large language model, the temperature parameter, which controls the randomness and diversity of the text generated by the large language model, is set to 0.7, improving the random generalization ability of each agent. For community segmentation, a total of 21 agents were created, divided into key opinion leader agents and ordinary user agents according to a power-law distribution ratio of 3:7, namely 6 key opinion leaders and 15 ordinary users. The key opinion leaders' profiles were equally divided into two opposing positions. Taking the "Black Myth: Wukong" release event in the cultural field as an example, the communication strategies were first determined as "cultural output" and "cultural concept." Then, the opinion index of each large language model agent was randomly initialized by calling the random library in Python. Subsequently, a prompting process was established, and each large language model agent, based on the input event summary, communication strategy, and opinion index, began a communication simulation based on the different communities. In each round of communication simulation, the key opinion leaders can see the speeches of all other key opinion leaders within the group, unlock the social behavior "debate" and randomly select one person to make a rebuttal speech; the general user group can see the speeches of all key opinion leaders but do not have the permission to view each other. A posting simulation is conducted every two rounds. In addition to the event summary, communication strategy and opinion index, the input also includes the historical speeches of all key opinion leaders. After each round of iteration, the neighborhood threshold, opinion weight and opinion index of all large language model agents are adjusted, and then the updated parameters are used as input to enter the next round of iteration until the number of iterations is reached. The results of public opinion deduction are shown in Figure 3 It can be found that the emotional variance of the ordinary user group strictly conforms to the periodic characteristics of public opinion evolution, and the evolution of the opinions of the key opinion leaders group shows basically positive consistency. The reason may be that excellent culture, due to its own diversity, can gain strong emotional resonance from the public in the online public opinion space.

[0061] Simulation results in science, technology and art Figure 4 . Figure 4(a)(b)(c) are the public opinion deductions for the event "deepseek releases inference model R1". It can be found that the evolution of opinions of key opinion leaders and ordinary users shows a complex and fragmented tendency. The reason is that the technology field has a high degree of heterogeneity due to its high complexity and threshold, which poses a high barrier to the majority of netizens with low cognitive levels. Figure 4 (d)(e)(f) are the public opinion deductions for the event of "Nezha 2 topping the box office". It can be found that the evolution trend of the opinions of the general user group is highly correlated with that of the key opinion leaders. It can be considered that when dealing with similar public opinion events, it is necessary to strengthen the content guidance of the key opinion leaders.

[0062] In some other preferred embodiments of the present invention, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method described in the above embodiment.

[0063] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0064] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for simulating complex social behaviors and inferring public opinion based on multi-agents, characterized by: Divide social groups: divide users in social networks into key opinion leader communities and ordinary user communities; Building intelligent agents: By designing different prompt words for key opinion leader communities and general user communities, we construct key opinion leader intelligent agents and general user intelligent agents based on different base models and driven by large language models. Building a social network: Build a dynamic weighted directed network, modeling each key opinion leader agent and ordinary user agent as a node, social behavior as a directed edge, and behavioral influence as edge weight. Simulate the local information propagation of social media with high information cohesion and simulate long-distance connections in social networks. The network node neighborhood is calculated by weighting the opinion index and the node weights corresponding to ordinary user agents. The sentiment index is obtained by sentiment analysis of the text generated by the large language model, and is mapped to the dynamic weight of the agent individual in the dynamic weighted directed network neighborhood. The user influence is quantified by combining the emotional contagion theory. Based on the dynamics of opinion propagation, a dynamic attenuation mechanism is added to the threshold of neighborhood judgment to simulate the self-service mechanism. The opinion convergence mechanism is simulated based on the classic Hegselmann-Krause algorithm. The network differences are used to simulate high locality and long-distance connectivity to divide the two major groups into key opinion leaders and ordinary users. After constructing the agent, the opinion index pool of the social group is constructed as a neighborhood, which is expressed as: ， in x i and x j Represents the roles played by the intelligent agents i and j The opinion index of the agent is x When the difference is less than the threshold ε, the corresponding two agents are neighbors, and all shared neighboring agents are collectively called neighborhoods, which are expressed as , The opinion index represents the opinion tendency of the agent, which is expressed as: ， In the formula w j Neighborhood role j The weight of x j （t） It's a role j exist t Sentiment index for the time period; Evaluate text sentiment index based on SnowNLP to calculate sentiment score; The weight of the opinion index of the network node and the threshold of the neighborhood judgment are dynamically adjusted according to the level of the sentiment score. The dynamic adjustment of the weight of the opinion index is expressed as: ， In the formula s Represents the current agent's sentiment score, δ is the step value of increase or decrease; The sigmoid function is used to weight the agent's opinion index w It is limited to between 0.5 and 2.0 to prevent the generated natural language text from falling into excessive emotion and causing the weight to grow rapidly. ， The parameters k Controls the steepness of the adjustment, c is the center point of sigmoid, After using the sigmoid function to smooth the changes in weights, an excitement suppression mechanism is added after each round of iteration. ， in γ is the cooling coefficient, w 0 is the target stable value of the weight; Dynamically adjusting the threshold of domain judgment is expressed as increasing the influence of high-weighted agents in the neighborhood based on the weighted average formula, that is: ， In the formula Neighborhood role i exist t -1 moment weight, ε is a hyperparameter, ε (t) for t Threshold for time neighborhood determination, hyperparameters α Controlling the influence of weights on the trust range, hyperparameters β Control the decay rate, T is the total number of iterations, t / T is the decay rate, The cycle is iterated several times to complete the deduction and simulation of online public opinion.

2. The method for simulating complex social behaviors and inferring public opinion based on multiple agents according to claim 1 is characterized in that: The intelligent agents all include sensors, controllers and effectors. The input of the sensors is structured information after knowledge extraction. The controller controls the intelligent agents to perform corresponding social behaviors based on the information provided by the sensors, and the effectors control the action trends of the intelligent agents.

3. The method for simulating complex social behaviors and inferring public opinion based on multiple agents according to claim 1 is characterized in that: The constructed neighborhood is iterated. During the iteration process, the information of the access node is the text and corresponding parameters generated by each round of iteration of the core layer. Each ordinary user intelligent agent then combines the open source intelligence related to the deduced target topic and freely expresses opinions in the simulated Weibo space. Then, referring to the ideas of social choice theory, public opinion is gradually formed in multiple rounds of iteration.

4. The method for simulating complex social behaviors and inferring public opinion based on multiple agents according to claim 3 is characterized in that: Weights are introduced in the calculation of the threshold for neighborhood judgment, where the weight for key opinion leader agents is higher than that for ordinary user agents.

Images