A method and system for large-scale dynamic social simulation

Abstract

This invention relates to the field of computer software technology and discloses a method and system for large-scale dynamic social simulation. The method includes: constructing intelligent agents; randomly assigning basic demographic attributes to the agents and initializing opinion preference values; generating interest and Big Five personality traits based on the basic demographic attributes and a large language model; forming opinion leader nodes through a priority connection mechanism; strengthening internal community connections through a triangular connection mechanism; generating a power-law distribution network with free scale; calculating an information confusion index based on a hierarchical collaborative network; and distinguishing core agents from regular agents based on the information confusion index. Core agents interact using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values, thereby realizing social simulation. This invention significantly improves the scale and efficiency of simulation.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and more specifically to a method and system for large-scale dynamic social simulation. Background Technology

[0002] To gain a deeper understanding of the behavior and evolution of social systems, large-scale dynamic social simulation methods have emerged. Social simulation is a research method based on computer models and algorithms that predicts and analyzes social phenomena and their changes by simulating the behavioral interactions of individuals and groups in society. It utilizes mathematical models, computer simulations, and data analysis to transform individuals, groups, and their behavioral rules within a social system into models for experimentation and observation in a virtual environment. Large-scale dynamic social simulation is of great significance, enabling researchers to conduct systematic analyses of complex social problems; revealing underlying patterns and mechanisms by simulating market behavior, consumer decisions, and price fluctuations; and in sociology, revealing the evolution of social structures and the formation of group behavior. With the continuous development of big data technology, artificial intelligence, and computing power, the scale and accuracy of social simulations are gradually improving, allowing for the simulation of increasingly diverse social factors and dynamic changes.

[0003] Currently, significant progress has been made in multi-user simulation technologies, which can be mainly divided into the following categories: agent-based simulation, data-driven simulation, and hybrid model-based simulation.

[0004] Agent-based simulation techniques are among the most common types of social simulations. This technique designs individualized behavioral rules for each simulated object (or "agent"), enabling it to make decisions within the simulated environment based on the environment and the behavior of other agents. Agent models can simulate the interactive behavior of individuals in complex social systems and are widely used in fields such as economic markets and social networks. The behavior of each agent is typically based on a specific strategy or algorithm, such as game theory, decision trees, or genetic algorithms.

[0005] Data-driven simulation techniques rely on extracting patterns from large amounts of historical data and using machine learning or deep learning algorithms to predict and simulate the behavior of multiple users. These techniques typically employ big data analytics, combining historical user behavior data, social interaction data, and more to build dynamic user behavior models. By training the model, the system can adjust its predictions in real time during the simulation process, adapting to complex and ever-changing user behaviors. This approach is particularly suitable for applications such as social media analytics and consumer behavior prediction.

[0006] Hybrid model-based simulation techniques combine the advantages of agent-based and data-driven models, and are typically applied in complex multi-user systems. By combining traditional rule-driven and data-driven models, hybrid models can more flexibly handle various types of user behavior and changes in the system environment. This approach has demonstrated powerful simulation capabilities in fields such as smart cities, public safety, and financial markets, capable of simultaneously considering the impact of multiple factors on user behavior and making dynamic adjustments. These existing technologies provide strong support for multi-user simulation, enabling the simulation of diverse user behaviors and complex social interactions in different application scenarios.

[0007] While existing multi-user simulation technologies have achieved some success in simulating social dynamics, several shortcomings remain. First, agent-based methods are generally simplistic, typically employing pre-defined behavioral rules that fail to realistically reflect the complex and dynamically changing individual behaviors and interactions within society. This static and limited model design restricts the capture and simulation of real-world social dynamics. Second, recent attempts to drive agent behavior using large-scale pre-trained language models have improved the intelligence of individual decision-making, but these methods largely lack the ability to dynamically adjust the simulation framework. This makes it difficult for the simulation system to adapt flexibly to environmental changes and emerging social phenomena, thus affecting the scalability and effectiveness of the simulation. Furthermore, existing simulation methods generally neglect the close connections within local communities and their crucial role in information dissemination. Particularly in event propagation simulations, the structure and interaction frequency of local communities play a decisive role in the rapid spread and amplification of events. However, many simulation models fail to construct credible and detailed network topologies, resulting in an inability to accurately reflect the real paths and speeds of large-scale event propagation. This not only weakens the credibility of the simulation results but also limits the guiding value of the models in practical applications such as social governance and event prevention. Therefore, improving the dynamic adaptability of simulation models and their realistic portrayal of community structures are important challenges that current large-scale dynamic social simulation methods urgently need to address. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a large-scale dynamic social simulation method and system that integrates a large language model and a traditional agent model. It employs a dynamic interaction strategy and a hierarchical collaborative network to dynamically allocate agents and determine their interaction patterns. Agents are divided into two categories: core agents and regular agents. Core agents, driven by a large language model, possess complex reasoning and decision-making capabilities and are used to simulate the behavior of key users in positions of information conflict. Regular agents, based on mathematical models, are used to simulate ordinary users in a stable environment, significantly affected by the information cocoon effect. The dynamic interaction strategy dynamically identifies core agents based on information confusion and diversity and determines the interaction patterns between agents, thereby effectively simulating large-scale agent behavior. The hierarchical collaborative network is constructed based on priority connection and triangular connection mechanisms, generating network topologies with community clustering and short average node distances, more realistically simulating the propagation path and speed of information in social networks.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for large-scale dynamic social simulation, comprising:

[0011] The system constructs intelligent agents, including a core intelligent agent for simulating key users and a regular intelligent agent for simulating ordinary users; the core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses continuously changing viewpoints based on a mathematical model; the core intelligent agent has user profiles and a memory bank.

[0012] Randomly assign basic demographic attributes to all agents and initialize opinion preference values. Generate interest and Big Five personality traits based on basic demographic attributes and a large language model. The basic demographic attributes, interests, and Big Five personality traits constitute the user profile.

[0013] The following process is iterated to conduct social simulation: opinion leader nodes are formed through a priority connection mechanism, and internal community connections are strengthened through a triangular connection mechanism to generate a power-law distribution network with free scale as a hierarchical collaborative network; the information confusion index is calculated based on the hierarchical collaborative network, and core agents and regular agents are distinguished according to the information confusion index; core agents interact with each other using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values.

[0014] In one embodiment, the user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, wherein name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, wherein personal experience memory is used to record one's own past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

[0015] In one embodiment, the conventional agent employs the Deffuant model, where opinion bias is represented by continuous values ​​[-1, 1].

[0016] In one embodiment, the process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes:

[0017] A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

[0018] In one embodiment, the initialization of the fully connected seed graph involves gradually adding agents as new nodes to the seed graph, using existing nodes in the seed graph as target nodes, and new nodes performing a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. Specifically, this includes:

[0019] Select n agents to initialize a fully connected seed graph ,in, These are the vertex set and edge set of the initial seed graph, respectively. , ,satisfy:

[0020] ;

[0021] Among them, the number of nodes in the initial seed graph Less than the number of agents , This represents the node corresponding to the i-th agent. This represents the edge between the nodes corresponding to the i-th agent and the j-th agent;

[0022] In the a-th iteration, a new node Through the following connection mechanism Added to the seed map output during the (a-1)th iteration middle:

[0023] ;

[0024] New node Using a priority connection mechanism with probability p Establish connections with existing nodes in the seed graph, using a priority connection mechanism. middle, With existing nodes in the seed graph Connection probability and nodes degree Proportional;

[0025] New node Using a triangular connection mechanism with probability 1-p Connecting nodes to existing nodes in the seed graph, in the triangular connection mechanism, Random and an existing node in the seed graph Build a connection, and then with A connection is established with a random neighbor node, thus forming a triangular structure.

[0026] In one embodiment, the step of calculating the information confusion index based on the hierarchical cooperative network and distinguishing between core agents and regular agents according to the information confusion index specifically includes:

[0027] The information confusion index is composed of the agent's information similarity score and information diversity score; the information similarity score is:

[0028] ;

[0029] in, This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the number of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent in the process;

[0030] The information diversity score is:

[0031] ;

[0032] in, This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: ;

[0033] Information Confusion Index for:

[0034] ;

[0035] in, Indicates to Taken from the natural constant Exponential operations with base 0. This is a hyperparameter; the information confusion index is taken in each iteration. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

[0036] Secondly, the present invention provides a large-scale dynamic social simulation system, comprising:

[0037] The intelligent agent construction module is used to construct intelligent agents, including a core intelligent agent for simulating key users and a regular intelligent agent for simulating ordinary users; the core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses continuously changing viewpoints based on a mathematical model; the core intelligent agent has user profiles and a memory bank;

[0038] The agent initialization module randomly assigns basic demographic attributes to all agents and initializes opinion preference values. Based on the basic demographic attributes and the big language model, it generates interest and Big Five personality traits. The basic demographic attributes, interests, and Big Five personality traits constitute the user profile.

[0039] The social simulation module forms opinion leader nodes through a priority connection mechanism and strengthens internal community connections through a triangular connection mechanism, generating a power-law distribution network with free scale as a hierarchical collaborative network. Based on the hierarchical collaborative network, it calculates the information confusion index and distinguishes core agents from regular agents according to the information confusion index. Core agents interact with each other using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values. The above process is iteratively run to achieve social simulation.

[0040] In one embodiment, the user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, wherein name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, wherein personal experience memory is used to record one's own past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

[0041] In one embodiment, the process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes:

[0042] A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

[0043] In one embodiment, the step of calculating the information confusion index based on the hierarchical cooperative network and distinguishing between core agents and regular agents according to the information confusion index specifically includes:

[0044] The information confusion index is composed of the agent's information similarity score and information diversity score; the information similarity score is:

[0045] ;

[0046] in, This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the number of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent in the process;

[0047] The information diversity score is:

[0048] ;

[0049] in, This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: ;

[0050] Information Confusion Index for:

[0051] ;

[0052] in, Indicates to Taken from the natural constant Exponential operations with base 0. This is a hyperparameter; the information confusion index is taken in each iteration. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

[0053] Compared with the prior art, the beneficial technical effects of the present invention are:

[0054] 1. Significantly Improved Efficiency and Realism in Large-Scale Social Network Simulation: By dynamically distinguishing between core agents with complex reasoning capabilities and rule-based ordinary agents, this invention achieves a rational allocation of computing resources. This allows key users to play a central role in the evolution of public opinion, while a large number of ordinary users operate efficiently with a low-cost model, significantly improving the scale and efficiency of the simulation. This strategy supports dynamic interactions among millions of agents, ensuring both detailed characterization of individual behaviors and the scalability of the overall system, realistically reflecting the complex and ever-changing event propagation process in real society.

[0055] 2. Enhanced Realism and Efficiency of Social Network Information Dissemination: The hierarchical collaborative topology network designed in this invention, combining priority connection and triangular connection mechanisms, effectively shapes the network structure of opinion leaders and close-knit communities. This structure not only shortens the information dissemination path and improves network connectivity but also strengthens close ties within communities, promoting the rapid spread of events within local groups. This network topology more closely resembles the characteristics of real-world social networks, enhancing the credibility and timeliness of information dissemination simulation and providing a solid foundation for social network public opinion analysis and intervention. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram of a large-scale dynamic social simulation framework in an embodiment of the present invention. Detailed Implementation

[0058] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0059] like Figure 1 As shown, this invention provides a large-scale dynamic social simulation method, comprising the following steps:

[0060] S1, Construct intelligent agents, including a core intelligent agent for simulating key users and a regular intelligent agent for simulating ordinary users; the core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses continuously changing viewpoints based on a mathematical model; the core intelligent agent has user profiles and a memory bank;

[0061] S2, randomly assign basic demographic attributes to all agents and initialize opinion preference values, generate interest and Big Five personality traits based on basic demographic attributes and big language model; basic demographic attributes, interest and Big Five personality traits constitute the user profile;

[0062] S3, iterate through the following process to conduct social simulation: form opinion leader nodes through a priority connection mechanism, strengthen internal community connections through a triangular connection mechanism, and generate a power-law distribution network with free scale as a hierarchical collaborative network; calculate the information confusion index based on the hierarchical collaborative network, and distinguish core agents and regular agents according to the information confusion index; core agents interact with each other using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values.

[0063] The iteration stopping condition can be: reaching the set number of iterations or achieving the expected social simulation task.

[0064] In one embodiment, the user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, wherein name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, wherein personal experience memory is used to record one's own past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

[0065] In one embodiment, the conventional agent employs the Deffuant model, where opinion bias is represented by continuous values ​​[-1, 1].

[0066] This invention aims to construct a hierarchical intelligent agent architecture to achieve efficient and realistic simulation of the spread of events (such as rumors) in large-scale social networks. This architecture distinguishes between two types of intelligent agents: core agents and regular agents. The core agent is driven by a large language model, possessing powerful natural language understanding and generation capabilities, and can simulate complex human social behaviors and decision-making processes. The core agent is equipped with detailed user profiles, including name, gender, age, occupation, interests, and personality traits (based on the Big Five personality model), making it highly personalized and behaviorally diverse. The core agent also includes a memory bank, divided into personal experience memory and environmental memory, recording its past behaviors and observations of its environment, supporting memory retrieval and reflection, and helping to make appropriate action decisions in different situations. The core agent can perform various social media actions, such as original posting, forwarding, replying, and liking, generating natural language content to enrich the realism of the simulation. In contrast, the regular agent is based on a traditional rule-based agent model, using mathematical methods (such as the Deffuant model) to express the continuous changes in its opinions, using simplified numerical forms to represent opinions, and simulating the interaction behavior between ordinary users. Conventional intelligent agents have low computational costs, making them suitable for simulating large-scale user groups and ensuring the scalability of the simulation system. The two types of agents work collaboratively through a hierarchical architecture: the core agent focuses on key opinion formation and dissemination nodes, while the conventional agents simulate a large amount of general user behavior. This design effectively balances computational efficiency and simulation detail, making it particularly suitable for simulating social networks on a scale of millions, breaking through the limitations of traditional models in terms of scale and performance.

[0067] Agent information initialization. The initialization process for each agent is meticulously designed to ensure diversity and realism in their attributes. First, basic demographic attributes are randomly assigned to each core agent, including name, gender, age, and occupation. Age follows a truncated normal distribution to better reflect real-world population structure; occupation and gender are randomly determined according to preset proportions. Based on these basic attributes, a large language model is used to infer and assign 3 to 5 potential interests to the agent, enhancing the realism of behavioral diversity and social attributes. The agent's personality traits are encoded according to the Big Five personality model, including five dimensions: openness, conscientiousness, extraversion, agreeableness, and emotional stability. These personality traits directly influence the agent's viewpoint updates and behavioral tendencies during event dissemination. Furthermore, an initial opinion bias value is assigned to the agent. This opinion bias value is a continuous numerical value representing the agent's inclination towards a particular event, typically ranging from -1 to 1. Negative values ​​represent doubt or opposition, positive values ​​represent belief, and zero values ​​represent neutrality or uncertainty. The initial opinion bias distribution reflects the diverse reactions of people to information in reality. The core agent is also equipped with a memory bank, including personal experience memory, recording its past behaviors and interactions, and environmental memory, containing observed neighbor behaviors and environmental information, serving as the basis for dynamic behavior. This initial multi-dimensional information provides the foundation for subsequent complex interactions and event propagation dynamics between agents, ensuring that the simulation process realistically reflects the cognitive heterogeneity and behavioral diversity of social reality.

[0068] In one embodiment, the process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes:

[0069] A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

[0070] In one embodiment, the initialization of the fully connected seed graph involves gradually adding agents as new nodes to the seed graph, using existing nodes in the seed graph as target nodes, and new nodes performing a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. Specifically, this includes:

[0071] Select n agents to initialize a fully connected seed graph ,in, These are the vertex set and edge set of the initial seed graph, respectively. , ,satisfy:

[0072] ;

[0073] Among them, the number of nodes in the initial seed graph Less than the number of agents , This represents the node corresponding to the i-th agent. This represents the edge between the nodes corresponding to the i-th agent and the j-th agent;

[0074] In the a-th iteration, a new node Through the following connection mechanism Added to the seed map output during the (a-1)th iteration middle:

[0075] ;

[0076] New node Using a priority connection mechanism with probability p Establish connections with existing nodes in the seed graph, using a priority connection mechanism. middle, With existing nodes in the seed graph Connection probability and nodes degree It is directly proportional; the degree of the target node refers to the number of neighboring nodes that the target node currently has.

[0077] New node Using a triangular connection mechanism with probability 1-p Establish connections with existing nodes in the seed graph. In a triangular connection, Randomly with an existing node Build a connection, and then with A connection is formed with a random neighbor, thus creating a triangular structure.

[0078] A hierarchical collaborative network is constructed. As a topological structure for interaction among intelligent agents, the hierarchical collaborative network can simulate the complex connection characteristics of real-world social networks. It employs two key mechanisms: a priority connection mechanism and a triangular connection mechanism, simulating opinion leaders and close-knit communities in real-world social networks. The priority connection mechanism encourages newly added nodes to connect to already highly connected nodes, generating a Matthew effect and forming a small number of highly connected opinion leaders, effectively shortening the average path length and increasing the speed of information flow. The triangular closure mechanism prioritizes establishing new connections between existing connected neighbors, strengthening the density and clustering characteristics within communities, reflecting strong connections and local groups in social circles. This mechanism increases the network clustering coefficient, enhances the coherence of the community structure, and promotes the rapid spread of information within local communities. Specifically, the network starts with an initially fully connected seed graph and gradually adds new nodes. New nodes execute the priority connection mechanism with probability p and the triangular closure mechanism with probability 1-p. The probability of node connection depends on the degree of the target node or random selection of neighbors, thus achieving a scale-free power-law distribution network. This structure possesses small-world characteristics and high clustering, enabling it to efficiently and realistically reflect the propagation paths and speeds of events in real social networks, providing a solid network foundation for information dissemination among intelligent agents.

[0079] In one embodiment, the step of calculating the information confusion index based on the hierarchical cooperative network and distinguishing between core agents and regular agents according to the information confusion index specifically includes:

[0080] The information confusion index is composed of the agent's information similarity score and information diversity score; the information similarity score is:

[0081] ;

[0082] in, This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the number of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent in the process;

[0083] The information diversity score is:

[0084] ;

[0085] in, This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: .

[0086] Information Confusion Index for:

[0087] ;

[0088] in, Indicates to Taken from the natural constant Exponential operations with base 0. Indicates to Taken from the natural constant Exponential operations with base 0. This is a hyperparameter; the information confusion index is taken in each iteration. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

[0089] like Figure 2 As shown, this invention models agent behavior through a dynamic interaction strategy to dynamically adapt to the complex environment of event propagation. Based on the information cocoon theory, the dynamic interaction strategy divides agents into two states: information conflict and information cocoon. Agents in the information conflict state have diverse information sources and multiple viewpoints. These agents, as key nodes in information dissemination and opinion formation, employ core agents driven by a large language model, supporting complex natural language communication and reasoning capabilities. Agents in the information cocoon state mainly encounter homogeneous information, and their viewpoints tend to assimilate. They use low-cost rule-based conventional agents to quickly update numerical viewpoints. The dynamic interaction strategy includes two key modules: an adaptive grouping module and a dynamic communication module. The adaptive grouping module calculates the information confusion index by calculating the similarity and diversity indicators between agents and their neighbors, dynamically identifying key core agents and ensuring that system resources are focused on the most influential nodes. The dynamic communication module flexibly adjusts the interaction method according to the agent category: core agents use natural language dialogue to simulate complex interactions; core agents and conventional agents achieve influence transmission by mapping natural language content to continuous viewpoint tendency values; and conventional agents directly exchange viewpoint tendency values ​​based on mathematical models. This strategy balances computational efficiency and simulation detail, enabling the system to achieve dynamic and hierarchical information dissemination simulations on a scale of millions, accurately capturing the evolution of individual behavior and public opinion dynamics.

[0090] The invention will be further illustrated by a case study involving a large-scale social simulation of an event propagation scenario using a large language model to drive 10,000 agents. The large language model adopted is the Llama-3.3-70B-Instruct model.

[0091] In this embodiment, a hierarchical agent architecture is first constructed to support large-scale simulation of event propagation for ten thousand agents. The core agents are driven by the Llama-3.3-70B-Instruct large language model, possessing powerful natural language understanding and generation capabilities, and able to simulate complex human social behaviors and diverse decision-making processes. Each core agent is equipped with a detailed user profile to ensure personalized and realistic individual behavior. The core agents have built-in personal experience and environmental memories, enabling them to retrieve and reflect on past interaction information, thereby making reasonable behavioral decisions in different scenarios. The core agents support various social actions such as original posting, forwarding, replying, and liking, generating natural language content to enrich the simulation effect. The regular agents adopt a rule-based Deffuant model, using continuous numerical values ​​to express changes in opinion tendencies, simulating the interaction behavior of a large number of ordinary users, and possessing low computational resource consumption.

[0092] The system randomly assigns basic demographic attributes to each agent, including name, gender, age, and occupation. Age follows a truncated normal distribution, more closely resembling real-world population structure. Occupation and gender are randomly assigned according to a preset ratio. The Llama-3.3-70B-Instruct model is used to infer and assign agents 3 to 5 potential interest domains, enhancing behavioral diversity. Agent personality traits are encoded based on the Big Five personality model, including five dimensions: openness, conscientiousness, extraversion, agreeableness, and emotional stability. These traits directly influence the agent's opinion updates and behavioral performance. Each agent is assigned an initial opinion score, continuously ranging from -1 to 1, representing skepticism, support, and neutrality, respectively. The initial opinion distribution reflects the diversity of real-world populations. Agents are equipped with a personal experience memory bank and an environmental memory bank. The personal experience memory bank stores past behaviors and interactions; the environmental memory bank contains neighbor behaviors and external information, serving as a basis for dynamic decision-making. This multi-dimensional initialization provides a solid foundation for subsequent complex interactions and dynamic event propagation among agents, ensuring cognitive heterogeneity and behavioral diversity during the simulation process.

[0093] This embodiment employs a hierarchical collaborative network to construct an agent relationship network, combining a priority connection mechanism and a triangular connection mechanism. The initial network is a small, fully connected seed graph containing 5 nodes, with new nodes added progressively. During each node addition, priority connections are performed with probability p, making it more likely that the new node will connect to nodes with higher degrees, forming a minority of opinion leaders and shortening the average network path. Triangular connections are performed with probability 1-p, prioritizing connections between neighbors of neighbors, enhancing the clustering coefficient and connectivity within communities, and forming a tight community structure. This structure exhibits a power-law distribution, possessing small-world characteristics and high clustering. The constructed hierarchical collaborative network effectively simulates the rapid propagation path of information among opinion leaders and local communities, closely approximating the propagation patterns of real social networks, laying a solid foundation for information exchange and event propagation between agents, and supporting efficient interaction among millions of agents.

[0094] This embodiment simulates agent behavior through a dynamic interaction strategy, dynamically classifying agents into information conflict states or information cocoon states. Core agents represent key users in information conflict zones, driven by the Llama-3.3-70B-Instruct model, supporting complex natural language communication and multi-turn reasoning. Regular agents represent users in information cocoons, using a rule-based Deffuant model to quickly update their viewpoints. An adaptive grouping module calculates the similarity and diversity between agents and their neighbors, forming an information confusion index to dynamically identify core agents and ensure reasonable resource allocation. A dynamic communication module adjusts the interaction mode according to agent category. Core agents simulate complex interactions through natural language dialogue, while influence is transmitted between core and regular agents through a mapping of natural language to opinion bias values. Regular agents directly exchange opinion bias values. This strategy balances computational efficiency and behavioral complexity, enabling dynamic, hierarchical information dissemination simulation at a scale of millions, accurately capturing individual behavior and the evolution of group public opinion in event propagation. Through continuous execution of agent interactions, a large-scale social simulation is ultimately completed.

[0095] This invention designs a multi-agent dynamic interaction strategy, which dynamically distinguishes between core agents with complex reasoning capabilities and rule-based ordinary agents, thereby rationally allocating computing resources. Core agents represent key users in positions of information conflict, capable of influencing public opinion; ordinary agents represent users within information cocoons. This strategy can effectively support event propagation simulations on a scale of millions.

[0096] This invention designs a hierarchical collaborative topology network to more realistically simulate the spread of information on large-scale social networks. This network employs a priority connection mechanism to construct opinion leaders, thereby effectively shortening the information dissemination path; simultaneously, it relies on a triangular connection method to further enhance the cohesion of local communities. This network structure can accelerate the spread of information within communities and better reflect the characteristics of information dissemination in reality.

[0097] 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.

[0098] Based on the description of the above method embodiments, the present invention also provides a system. The system may be a system that uses software (applications), modules, components, servers, clients, etc., using the methods described in the embodiments of this specification, combined with necessary implementation hardware. Based on the same innovative concept, the systems in one or more embodiments provided in this disclosure are as described in the following embodiments. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can refer to the implementations of the foregoing methods, and repeated details will not be repeated. As used below, the terms "module" or "module group" refer to a combination of software and / or hardware capable of implementing a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, hardware implementations, or a combination of software and hardware, are also possible and contemplated.

[0099] A large-scale dynamic social simulation system, comprising:

[0100] The intelligent agent construction module is used to construct intelligent agents, including a core intelligent agent for simulating key users and a regular intelligent agent for simulating ordinary users; the core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses continuously changing viewpoints based on a mathematical model; the core intelligent agent has user profiles and a memory bank;

[0101] The agent initialization module randomly assigns basic demographic attributes to all agents and initializes opinion preference values. Based on the basic demographic attributes and the big language model, it generates interest and Big Five personality traits. The basic demographic attributes, interests, and Big Five personality traits constitute the user profile.

[0102] The social simulation module forms opinion leader nodes through a priority connection mechanism and strengthens internal community connections through a triangular connection mechanism, generating a power-law distribution network with free scale as a hierarchical collaborative network. Based on the hierarchical collaborative network, it calculates the information confusion index and distinguishes core agents from regular agents according to the information confusion index. Core agents interact with each other using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values. The above process is iteratively run to achieve social simulation.

[0103] In one embodiment, the user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, wherein name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, wherein personal experience memory is used to record one's own past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

[0104] In one embodiment, the process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes:

[0105] A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

[0106] In one embodiment, the step of calculating the information confusion index based on the hierarchical cooperative network and distinguishing between core agents and regular agents according to the information confusion index specifically includes:

[0107] The information confusion index is composed of the agent's information similarity score and information diversity score; the information similarity score is:

[0108] ;

[0109] in, This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the number of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent in the process;

[0110] The information diversity score is:

[0111] ;

[0112] in, This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: ;

[0113] Information Confusion Index for:

[0114] ;

[0115] in, Indicates to Taken from the natural constant Exponential operations with base 0. This is a hyperparameter; the information confusion index is taken in each iteration. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

[0116] 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.

[0117] 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.

[0118] 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 large-scale dynamic social simulation method, characterized in that, include: Construct intelligent agents, including core intelligent agents for simulating key users and regular intelligent agents for simulating ordinary users; The core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses opinions based on mathematical models that tend to change continuously; the core intelligent agent has user profiles and a memory bank; Randomly assign basic demographic attributes to all agents and initialize opinion preference values. Generate interest and Big Five personality traits based on basic demographic attributes and a large language model. The basic demographic attributes, interests, and Big Five personality traits constitute the user profile. The following process is iterated to conduct social simulation: opinion leader nodes are formed through a priority connection mechanism, and internal community connections are strengthened through a triangular connection mechanism to generate a power-law distribution network with free scale as a hierarchical collaborative network; The information confusion index is calculated based on the hierarchical collaborative network, and the core agents and regular agents are distinguished according to the information confusion index. The core agents interact with each other using natural language, the core agents and regular agents interact through opinion bias value mapping, and the regular agents directly exchange opinion bias values. The information confusion index is composed of the agent's information similarity score and information diversity score; ； This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent; ； This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: ; Information Confusion Index ; This is a hyperparameter.

2. The large-scale dynamic social simulation method according to claim 1, characterized in that, The user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, where name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, where personal experience memory records one's past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

3. The large-scale dynamic social simulation method according to claim 1, characterized in that, The conventional agent adopts the Deffuant model, with continuous values ​​[-1, 1] representing opinion bias.

4. The large-scale dynamic social simulation method according to claim 1, characterized in that, The process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes: A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

5. The large-scale dynamic social simulation method according to claim 4, characterized in that, The initialization of the fully connected seed graph involves gradually adding agents as new nodes to the seed graph, using existing nodes in the seed graph as target nodes. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. Specifically, this includes: Select n agents to initialize a fully connected seed graph ,in, These are the vertex set and edge set of the initial seed graph, respectively. , ,satisfy: ； Among them, the number of nodes in the initial seed graph Less than the number of agents , This represents the node corresponding to the i-th agent. This represents the edge between the nodes corresponding to the i-th agent and the j-th agent; In the a-th iteration, a new node Through the following connection mechanism Added to the seed map output during the (a-1)th iteration middle: ； New node Using a priority connection mechanism with probability p Establish connections with existing nodes in the seed graph, prioritizing connections. middle, With existing nodes in the seed graph Connection probability and nodes degree Proportional; New node Using a triangular connection mechanism with probability 1-p Connecting nodes to existing nodes in the seed graph, in the triangular connection mechanism, Random and an existing node in the seed graph Build a connection, and then with A connection is established with a random neighbor node, thus forming a triangular structure.

6. The large-scale dynamic social simulation method according to claim 1, characterized in that, The calculation of the information confusion index based on the hierarchical cooperative network, and the differentiation between core agents and regular agents based on the information confusion index, specifically includes: In each iteration, the information confusion index is taken. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

7. A large-scale dynamic social simulation system, characterized in that, include: The agent building module is used to build agents, including core agents for simulating key users and regular agents for simulating ordinary users; The core intelligent agent is driven by a large language model and has the ability to understand and generate natural language; the regular intelligent agent expresses opinions based on mathematical models that tend to change continuously; the core intelligent agent has user profiles and a memory bank; The agent initialization module randomly assigns basic demographic attributes to all agents and initializes opinion preference values. Based on the basic demographic attributes and the big language model, it generates interest and Big Five personality traits. The basic demographic attributes, interests, and Big Five personality traits constitute the user profile. The social simulation module forms opinion leader nodes through a priority connection mechanism and strengthens internal community connections through a triangular connection mechanism, generating a power-law distribution network with free scale as a hierarchical collaborative network. The information confusion index is calculated based on a hierarchical collaborative network, and core agents and regular agents are distinguished according to the information confusion index. Core agents interact with each other using natural language, core agents and regular agents interact through opinion preference value mapping, and regular agents directly exchange opinion preference values. The above process is iteratively run to realize social simulation. The information confusion index is composed of the agent's information similarity score and information diversity score; ； This represents the information similarity score of the i-th agent at time t. This represents the set of neighboring agents of the i-th agent. This represents the attitude score of the i-th agent at time t; express The j-th intelligent agent; ； This represents the information diversity score of the i-th agent at time t. This represents the average attitude score of all of agent i's neighbors at time t: ; Information Confusion Index ; This is a hyperparameter.

8. A large-scale dynamic social simulation system according to claim 7, characterized in that, The user profile includes name, gender, age, occupation, interests, and personality traits based on the Big Five personality model, where name, gender, age, and occupation constitute the basic demographic attributes; the memory bank includes personal experience memory and environmental memory, where personal experience memory records one's past behaviors and interactions, and environmental memory contains observed behaviors of other agents and environmental information.

9. A large-scale dynamic social simulation system according to claim 7, characterized in that, The process of forming opinion leader nodes through a priority connection mechanism and strengthening internal community connections through a triangular connection mechanism to generate a power-law distribution network with free scale, serving as a hierarchical collaborative network, specifically includes: A fully connected seed graph is initialized. Agents are gradually added to the seed graph as new nodes. Each existing node in the seed graph is used as a target node. New nodes execute a priority connection mechanism with probability p and a triangular connection mechanism with probability 1-p. The priority connection mechanism means that the connection probability between a new node and a target node is proportional to the degree of the target node. The triangular connection mechanism means that a new node establishes a connection with a random target node in the seed graph and constructs a connection with a random neighbor node of the target node, thereby forming a triangular structure.

10. A large-scale dynamic social simulation system according to claim 7, characterized in that, The calculation of the information confusion index based on the hierarchical cooperative network, and the differentiation between core agents and regular agents based on the information confusion index, specifically includes: In each iteration, the information confusion index is taken. The top 1% of agents are designated as core agents, while the remaining agents are designated as regular agents.

Images