Large-scale social network simulation method and system based on dynamic topological structure

Abstract

The invention relates to the technical field of computer software, and discloses a large-scale social network simulation method and system based on a dynamic topological structure. The method comprises the following steps: distributing a user portrait and a memory to a corresponding agent, and constructing a social network for the agent; distinguishing a core agent from a conventional agent according to the influence measurement; the core agent is driven by a large language model and interacts with other core agents through a natural language, so that the attitude of a conventional agent to an event is influenced; the conventional agent updates the attitude through the trust weight, and adjusts the social network by using a dynamic link prediction engine; updating the memory of the intelligent agent, pushing a library and a social network; and continuously performing a dynamic evolution process at each time step until the deduction of the event is completed. According to the method, detailed description of individual behaviors is ensured, the expandability of the whole system is also considered, and the decision-making efficiency and the simulation precision of the super-large-scale social network simulator are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer software technology, and in particular to a large-scale social network simulation method and system based on a dynamic topology structure. Background Art

[0002] Social network simulations explore social phenomena such as information dissemination mechanisms, the evolution of group attitudes, and information cocoons by simulating the interactions of individuals and groups. Leveraging mathematical models, computer simulation, and data analysis techniques, social network simulations transform the complex individual interactions and group coordination rules found in real-world social platforms into virtual simulation models. These models then recreate and analyze the dynamic evolution of social networks in a controlled virtual environment, enabling predictions of social dynamics and the discovery of patterns in social behavior.

[0003] Social network simulators have important application value in the following aspects: 1) Sociological research and the revelation of the mechanism of public opinion formation: Social network simulators can reveal the formation mechanism of public opinion and the laws of event propagation by simulating the formation and evolution of various group emotions; 2) Event prediction and early warning: By simulating the evolution of social network events, simulators can predict the direction of public opinion or events, helping relevant personnel to provide early warnings in events and formulate response strategies; 3) Conducting cognitive intervention experiments: Different public opinion and cognitive intervention guidance strategies can be simulated and verified in a virtual environment, and their effectiveness can be evaluated, thereby providing a reference for public opinion guidance in reality.

[0004] Previous social network simulations based on agent-based models (ABMs) designed predefined behavioral rules for simulated entities, enabling them to make decisions based on the environment and the behavior of other agents within the simulated environment. This approach focused on macro-level modeling and mechanistic analysis of group interaction patterns, ignoring the micro-level driving effects of heterogeneous individual behaviors on the development of events. This approach, however, exhibited limited adaptability in complex social environments. In recent years, large language model (LLM)-driven social network simulators have emerged, through fine-grained modeling of user behavior and decision-making on real social platforms, accurately reproducing event evolution patterns and group behavior phenomena, opening up new avenues for studying the operational mechanisms of social platforms. The scale of agents has become a core issue in the research of large language model-driven social network simulators, driving exponential growth in simulation scale. For example, existing technologies have successfully captured news dissemination patterns by leveraging the human-like capabilities of large language models in perception, reasoning, and behavior, enabling initial exploration of small-scale social network simulations. Existing technologies have also constructed environments containing thousands of agents, simulating their interactions such as posting, forwarding, and commenting to replicate individual behavioral decisions and the evolution of group attitudes. Faced with the inefficiency and high cost of large-scale user simulation, many studies have focused on improving simulation efficiency. These studies typically employ a hierarchical design involving core agents and regular agents, selectively activating a small number of core agents at each time step to engage in human-like interactions, thereby successfully reproducing larger-scale group behavior.

[0005] While the aforementioned existing technologies have made significant progress in social network simulation, blindly expanding the number of agents is unwise. Existing research largely relies on static social networks, ignoring the dynamic evolution of social relationships. Modeling dynamic social networks presents significant challenges. First, during event propagation, user behavioral decisions involve more than just interactions; they also involve adaptive adjustments to their social relationships. The factors influencing the evolution of social relationships are multidimensional and complex. For example, users tend to connect with those who share similar views, reduce interaction with those who oppose them, and even disconnect. Furthermore, user influence, one-way attention, and unequal interactions lead to asymmetric social relationships. Content quality and timeliness determine the appeal and reach of information, thus influencing the formation and evolution of social relationships. The interplay of these factors significantly increases the complexity of modeling dynamic social networks. Second, to balance simulation efficiency and scale, existing large-scale simulators typically select a fixed set of core agents with human-like interaction capabilities. Core agents serve as key nodes in event propagation, driving the evolution of events. However, as events unfold, user roles change. Core agents inherently possess propagation characteristics such as significant influence, making random or static selection strategies unsuitable. Therefore, effectively quantifying these key features and dynamically identifying core agents at different time steps becomes the key path to aligning with the real event propagation situation. In summary, modeling the dynamics of social networks is an important challenge that needs to be addressed in current large-scale social network simulations. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a large-scale social network simulation method and system based on a dynamic topology structure. The present invention specifically adopts the following technical solutions: In a first aspect, the present invention provides a large-scale social network simulation method based on a dynamic topology structure, which is used to deduce attitudes about events occurring in a target user group, comprising: Assign user profiles and memories to corresponding agents, and build a social network for the agent based on the user's outgoing and incoming followers on the social platform; Dynamic evolution process: The influence measure of each agent is calculated based on the number of second-order incoming followers corresponding to each agent and the variance of the incoming followers' attitudes towards the event. The core agents are distinguished from conventional agents based on the influence measure. The core agents are driven by a large language model and interact with other core agents through natural language, influencing the attitudes of conventional agents towards the event. The core agents use user portraits, memory mechanisms, and information flow-based link prediction methods to decide on interactive behaviors with other tweets or agents and adjust the social network. Conventional agents are driven by a subject model and interact with other conventional agents. Conventional agents update their attitudes through trust weights and adjust the social network using a dynamic link prediction engine. The agent's memory, tweet library, and social network are updated. The dynamic evolution process is continued at each time step until the deduction of the event is completed and the user's attitude towards the event is obtained.

[0007] In one embodiment, the influence measure of each agent is calculated by the number of second-order inbound followers corresponding to each agent and the variance of the inbound followers' attitudes toward the event, specifically including: The influence metric quantifies the agent's communication ability by the number of second-order inbound followers, and measures the content diversity of the agent's tweets by the variance of the inbound followers' attitudes towards the event: ; in, represents the influence measure of agent i at time t, represents the attitude of agent i towards the event at time t, Indicates whether agent i pays attention to agent j, Indicates that agent i does not pay attention to agent j, Indicates that agent i pays attention to agent j; represents the incoming follower list of agent i; Indicates whether agent j pays attention to agent k, Indicates that agent j does not pay attention to agent k, Indicates that agent j pays attention to agent k.

[0008] In one embodiment, the distinguishing between core agents and regular agents based on influence metrics specifically includes: At each time step, the top K agents with the highest influence metrics are selected as core agents, and the other agents are used as regular agents.

[0009] In one embodiment, the core agent uses user profiling, memory mechanisms, and information flow-based link prediction methods to decide on interactive behaviors with other tweets or agents and adjust social networks, specifically including: Calculate the recommendation score of core agent i and candidate tweet j , Integrates the content representation of the latest tweets of core agent i and content representation of candidate tweets The content matching degree between them, the life cycle factor of the candidate tweet and the content influence of the candidate tweet j : ; in, represents the cosine similarity, is a natural constant, represents the release time of candidate tweet j, represents the index at time t, Indicates the decay rate of the life cycle factor; Form a recommendation list for core agent i based on the recommendation score: select agents that meet the following conditions among the agents that are not followed by core agent i: the number of inbound followers exceeds the set threshold or the outbound followers of the outbound followers of core agent i; use the tweets posted by the selected agents as candidate tweets, and select the N tweets with the highest recommendation scores from the candidate tweets to form a recommendation list; The core agent i makes decisions based on user portraits, memories, messages received from neighboring agents, and personalized information flows based on recommendation lists, obtains interactive behaviors with other tweets or agents, and adjusts its own attention relationships.

[0010] In one embodiment, the core agent i makes decisions based on user profile, memory, messages received from neighboring agents, and personalized information flow based on recommendation lists to obtain interactive behaviors with tweets or other agents, specifically including:

[0011] For the decision-making process, User portrait for the core agent i, represents the memory of core agent i at time t-1, is the personalized information flow of core agent i at time t, is the message received by the core agent i from the neighboring agent at time t-1; represents the set of actions performed at time t, which includes uploading tweets, forwarding tweets, commenting on tweets, following other agents, unfollowing other agents, liking tweets uploaded by others, and doing nothing.

[0012] In one embodiment, the conventional agent updates its attitude by trust weight, specifically including: Based on trust, the attitude interaction between regular agents and other agents is quantified. Agents with higher trust or more inbound followers are given higher weights: ; ; in, Represents the attitude of conventional agent i towards the event at time t Neighboring Agents The weight of influence, Indicates trust in weight The proportion of represents the degree of trust that agent i has in agent j, is the set of neighboring agents of regular agent i at time t, is the incoming follower list of agent j, Update the weights for the attitudes.

[0013] In one embodiment, adjusting the social network using a dynamic link prediction engine specifically includes: Missing link prediction scores by regular agent i and agent j and false link identification score Modeling the creation and disconnection of social relationships, The highest agent j is the newly followed outgoing follower of regular agent i, The highest agent j is the most recently unfollowed outgoing follower of regular agent i.

[0014] In one embodiment, the missing link prediction score and false link identification score The calculation method is: ; ; in, represents the attitude of conventional agent i towards the event at time t, represents the attitude of agent j towards the event at time t, is the incoming follower list of agent j, represents the set of candidate follower agents of regular agent i in the social network. The elements of the candidate follower agent set include the outgoing followers of regular agent i and agents whose number of incoming followers exceeds the set value. Indicates whether agent j pays attention to regular agent i, is the outgoing follower list of regular agent i, Indicates the degree of trust that agent i has in agent j.

[0015] In one embodiment, the trust level of agent i in agent j According to the agent k acting as the middleman in all trust propagation paths: ; ; in, Represents the trust relationship between agents i and j established based on agent k in the trust propagation path.

[0016] In a second aspect, the present invention provides a computer system comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of any one embodiment of the first aspect when executing the computer program.

[0017] Compared with the prior art, the beneficial technical effects of the present invention are: 1. Significantly Improved Efficiency of Large-Scale Social Network Simulations: Through a hierarchical design, this invention dynamically distinguishes between core agents with human-like decision-making capabilities and regular agents operating based on predetermined rules, achieving optimal allocation of computing resources. This allows core users with key characteristics to play a central role in the evolution of public opinion, while a large proportion of regular users can operate efficiently using a low-cost model. This strategy supports the dynamic interaction of very large-scale agents, ensuring detailed characterization of individual behavior while balancing the scalability of the overall system. This significantly improves the decision-making efficiency and simulation accuracy of very large-scale social network simulators, enabling them to realistically reflect the complex and ever-changing public opinion dissemination processes in real society.

[0018] 2. Superior dynamic evolution and simulation capabilities compared to static networks: Compared to static networks, dynamic social topologies effectively capture the dynamics of public attitudes, demonstrating superior accuracy, stability, and adaptability across different events. Dynamic social networks better reflect real-world communication, accelerate the polarization of attitudes, and exhibit distinct clustering patterns in newly added attention relationships. Furthermore, dynamic social networks can predict the growth of inbound followers during event propagation, providing empirical support for the cultivation of opinion leaders.

[0019] In summary, the present invention provides a high-fidelity experimental platform for simulating event propagation dynamics and reproducing group behavior phenomena. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 1 is a flow chart of a method in an embodiment of the present invention.

[0021] Figure 2 It is a schematic diagram of the framework structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, a large-scale social network simulation method based on a dynamic topology structure in the present invention is used to deduce events occurring in a target user group, comprising the following steps: S1, assigns user portraits and memories to corresponding agents, and builds a social network for the agent based on the user's outgoing followers and incoming followers in the social platform.

[0024] S2, dynamic evolution process: the influence measure of each agent is calculated by the number of second-order incoming followers corresponding to each agent and the variance of the incoming followers' attitudes towards the event, and the core agents and conventional agents are distinguished according to the influence measure; the core agent is driven by a large language model and interacts with other core agents through natural language to influence the attitudes of conventional agents towards the event. The core agent decides on the interactive behavior with tweets or other agents and adjusts the social network through user portraits, memory and information flow-based link prediction methods; conventional agents are driven by subject models and interact with other conventional agents; conventional agents update their attitudes through trust weights and use dynamic link prediction engines to adjust the social network; the agent's memory, tweet library and social network are updated.

[0025] S3, continuing the dynamic evolution process at each time step until the iterative simulation time step reaches the set number of times, completing the deduction of the event, and finally obtaining the user's attitude towards the event.

[0026] Outbound followers are other users followed by the current user; inbound followers are other users who follow the current user, which are commonly referred to as "fans" on social platforms.

[0027] To overcome the limitations of existing social simulation methods, this paper divides users into core agents based on large language models (LLMs) and conventional agents driven by agent-based models (ABMs), which correspond to opinion leaders and ordinary users in the real world, respectively.

[0028] This paper innovatively introduces a dynamic hierarchy to select core agents by quantifying the dissemination potential and content diversity at each time step. This enables core agents to be adaptively switched based on key features, thereby balancing efficiency and accuracy in large-scale social simulations.

[0029] In addition, the present invention designs different dynamic social relationship modeling strategies for different types of agents. The core agent adopts an information flow-based link prediction method that evaluates attitude similarity, content timeliness, and tweet influence to recommend potential like-minded non-neighboring agents. This enables the core agent to independently decide whether to follow or unfollow other agents, thereby simulating real homogeneous connection behavior and dynamic relationship evolution. Conventional agents construct a behavioral decision-making process for unequal interactions. By introducing the concept of trust to quantify unequal interactions between agents and using a dynamic link prediction engine to simulate relationship evolution driven by multiple factors, it reflects how core agents and local neighbors influence the passive behavior of most agents in the real world.

[0030] See also Figure 2 , the technical solution of the present invention is described in detail in several parts below.

[0031] 1. Task request and environment server initialization.

[0032] Upon receiving a request for an event situation deduction task, the environment server first initializes the user profile and social network topology of the target user group. The environment server's primary function is to maintain the state and data of the social media platform, such as user information, historical tweets, and user relationships. The environment server consists of four main modules: the user module, the tweet module, the relationship module, and the recommendation list. The user module is responsible for storing each user's user profile and memory; the tweet module contains all user-posted tweets and records detailed information such as comments, number of likes, and creation time for each tweet; the relationship module is used to store the structure of the social network, including each user's outbound and inbound follower lists; and the recommendation list module uses an information flow-based link prediction method to generate personalized recommendation lists for each user based on the user's memory and tweet information.

[0033] 2. Dynamic hierarchical modules identify core agents.

[0034] In order to ensure high accuracy and high efficiency in large-scale user simulation, the dynamic stratification module constructs a communication potential and content richness evaluation system, identifies the core agents that play a key role in event propagation, and adjusts the information interaction mode between agents according to the agent type. The dynamic stratification module includes two parts: core agent selection and dynamic interaction. In terms of core agent selection, core agents usually have higher communication potential and content diversity. The former represents the breadth and depth of information dissemination, and the latter reflects the agent's knowledge level, thinking depth, and willingness to express opinions. In order to effectively identify core agents at different stages of dissemination, the present invention designs an influence metric. , and at each time step select the The top K agents in the network serve as core agents. Dynamically and adaptively switching core agents with key characteristics at different time steps allows computing resources to focus on the core users with the greatest communication influence, thus achieving a balance between efficiency and accuracy in large-scale social simulations. Regarding dynamic interactions between agents, the interaction method depends on the agent type. Core agents interact through natural language dialogue, while conventional agents use agent-based models (ABMs) to transfer information. Furthermore, the content generated by core agents is converted into attitude scores using a large language model (LLM), and influences the attitude updates of conventional agents through agent-based models (ABMs), while the attitudes of core agents are unaffected by conventional agents. The dynamic layering module balances computational efficiency and simulation details, accurately reproducing the communication dynamics in reality while achieving dynamic and hierarchical simulations of large-scale social networks.

[0035] 3. Core intelligent agent behavior decision-making.

[0036] The core agent driven by the large language model (LLM) corresponds to the opinion leaders in real-world social networks. The core agent is equipped with a user portrait and memory mechanism. The user portrait includes name, gender, age, occupation, interest set, and personality traits, which makes it highly personalized and behaviorally diverse, and ensures the rationality of the user portrait. The memory mechanism takes into account the personal experience and environmental interaction memory of the agent. Personal experience memory represents the user's historical behavior record, and environmental interaction memory captures the attitudes and behavioral responses of neighboring agents to specific events. The memory mechanism includes retrieval, update, and reflection operations, and the core agent can make appropriate behavioral decisions based on the most relevant and urgent memories. In addition, in order to promote the core agent to think about the evolution of social relationships, the present invention designs a link prediction method based on information flow, which calculates the recommendation score between the core agent i and the candidate tweet j by quantifying attitude similarity, content timeliness, and tweet influence. , used to recommend potential like-minded non-neighboring agents, autonomously decide on interactive behaviors and adjust social relationships, realistically simulating homogeneous connection behavior and dynamic relationship evolution. The core agent's behavioral decisions reflect its attitude towards specific events. Specific actions include uploading tweets, retweeting tweets, commenting on tweets, following other agents, unfollowing other agents, liking tweets uploaded by others, and taking no action. Each action is closely related to the user's social interaction and information dissemination.

[0037] 4. Conventional agent behavior decision-making.

[0038] To balance computational efficiency and scale, conventional intelligent agents use agent-based models (ABMs) to describe the interaction process between users. These can be uniformly expressed using selection functions, attitude update functions, and information transfer functions. However, these traditional agent-based models (ABMs) are not adaptable to the unequal interactions and dynamic social networks found in real social networks. Therefore, this paper constructs a behavioral decision-making model for conventional intelligent agents that targets unequal social relationships. This model introduces the concept of trust to quantify unequal attitudinal interactions between users and utilizes a dynamic link prediction engine to capture social relationship adjustment patterns driven by multiple factors, such as attitude similarity and unidirectional attention. This model better adapts to the unequal interactions and social relationship adjustments found in real social networks, simulating the behavior of most nodes in real social networks, where they are passively influenced by core nodes and local neighbors.

[0039] 5. Environmental server updates and cyclic iterations.

[0040] After each simulation, the environment server updates the agent's memory, tweet repository, and social network. This server provides a dynamic, scalable infrastructure for the entire simulation framework, ensuring data flow and state updates throughout the simulation, which influence the agent's decisions in the next round. By continuously iterating at a preset time step, the social network simulator is capable of performing large-scale event simulations and simulating group behavior.

[0041] In one embodiment, the present invention designs an influence metric for the dynamic tiering module. to distinguish core agents from regular agents and select the agent with the highest The first K agents are used as core agents. Influence measurement The communication ability of the agent is quantified by the number of second-order inbound followers, and the content diversity is measured by the variance of the inbound followers' attitudes. The calculation method can be expressed as: ; in, represents the attitude score of agent i at time t, Represents the incoming follower list of agent i. Indicates whether agent i pays attention to agent j, Indicates that agent i does not pay attention to agent j, Indicates that agent i pays attention to agent j.

[0042] In one embodiment, in the core agent behavior decision, the link prediction method based on information flow calculates the recommendation score between agent i and candidate tweet j This score integrates the content representation of the agent's latest tweets and content representation of candidate tweets Content matching between , the life cycle factor of candidate tweets and their content influence The recommendation score is given by the following formula: ; in, represents the release time of candidate tweet j, The influence of content is quantified by weighting the number of likes, reposts, comments and followers. The life cycle factor follows The exponential decay law of the rate. To ensure Not zero, influence of original content Add 1 to the basis.

[0043] Based on the recommendation score, a recommendation list of core agent i is formed. The candidate tweets of the recommendation list are selected from the tweets posted by agents that are not followed by core agent i. The publishers of the candidate tweets are all agents whose number of inbound followers exceeds the set threshold (the threshold for the number of inbound followers in a 100,000-scale agent can be set to 1,000) or the outbound followers of the outbound followers of core agent i (i.e., friends of friends). The recommendation list is composed of N tweets with the highest recommendation scores from the candidate tweets.

[0044] Based on user portrait , historical memory , messages received from neighboring agents and personalized information flow , the decision-making process of agent i It can be modeled as: ; in, represents the set of actions executed at time t. The core agent integrates new information and personalized information flows from neighbors through behavioral interactions, thereby promoting the dynamic update of attitudes and the evolution of social relationships within the social network.

[0045] In one embodiment, the present invention introduces trust when conventional intelligent agents make behavioral decisions. The concept of quantifies the unequal attitude interactions among users, assigning higher weights to agents with high trust or a large number of inbound followers, and thus defining agent attitudes Neighboring Agent Set Impact: ; .

[0046] in, Represents the attitude of conventional agent i towards the event at time t Neighboring Agents The weight of influence, represents the proportion of trust in the weight, It is the weight of historical attitude in the attitude updating process.

[0047] In one embodiment, the present invention uses a dynamic link prediction engine to model the way a regular agent adjusts social relationships. The adjustment of social relationships is influenced by multiple factors such as user attitude similarity, one-way attention, user influence, and trust. Based on the above considerations, the present invention defines the following missing link prediction score: and false link identification score , used to model the addition and disconnection of social relationships, The highest agent j is the new follower of agent i, Similar to this.

[0048] ; .

[0049] in, represents the outgoing follower list of regular agent i in the social network, represents the set of candidate follower agents of a regular agent i in a social network, which only includes friends of friends and users with a large number of incoming followers. In addition, when a user adds a new follower, a trust relationship needs to be established between agents i and j, and the trust relationship can be established through the middleman k. Combining all potential trust propagation paths, the trust relationship between agents It can be expressed as, ; .

[0050] The dynamic layering module of this invention designs an influence metric to quantify the communication potential and content diversity of agents. It identifies core agents with key characteristics at each time step and constructs a dynamic layered design of core agents and regular agents. The dynamic layering module balances computational efficiency with simulation detail, enabling dynamic, hierarchical simulation of large-scale social networks while accurately reproducing real-world communication dynamics.

[0051] This paper designs an information flow-based link prediction method for core intelligent agents. By quantifying attitude similarity, content timeliness and tweet influence, it recommends potential like-minded non-neighbor intelligent agents, enabling core intelligent agents to independently decide on interactive behaviors and adjust social relationships, truly simulating homogeneous connection behaviors and dynamic relationship evolution, thereby accurately aligning with the attitude dynamics and relationship evolution in the real world.

[0052] This paper constructs a behavioral decision-making module for unequal interactions for conventional intelligent agents, introduces the concept of trust to quantify the unequal attitude interaction relationship between users, and uses a dynamic link prediction engine to capture the social relationship adjustment pattern driven by multidimensional factors such as attitude similarity and one-way attention, so as to better adapt to unequal interactions and social relationship adjustments in real social networks.

[0053] The technical solution of the present invention is described in more detail below with reference to an embodiment.

[0054] This embodiment simulates the attitude evolution and behavioral decision-making of 100,000 agents in a social network towards a specific event over 12 days, of which 2,000 are core agents and the rest are regular agents.

[0055] Upon receiving an event scenario simulation task request, the environment server first initializes the user profile and social network of the target user group. Each agent is assigned basic user profile attributes, including name, age, gender, and occupation. Age uses a truncated normal distribution that accurately reflects the demographic structure, while occupation and gender are assigned to each core agent according to a preset ratio. Combining these basic user profile attributes, the large-scale model infers and assigns three to five potential areas of interest to each agent, enhancing their heterogeneity and behavioral diversity. Personality traits are encoded based on the BigFive personality model, encompassing five personality descriptors: openness, conscientiousness, extraversion, agreeableness, and emotional stability. These traits serve as internal drivers of attitude updates and behavioral performance. Each agent receives an initial attitude score for each event request to be evolved, ranging from -1 to 1, where -1 indicates skepticism and +1 indicates support. Values ​​closer to 0 indicate a more neutral attitude. This initial attitude distribution reflects the diversity of the real population. The agent is equipped with a personal experience memory bank to store past behavior records and interaction content. At the same time, the environment interaction memory records neighbor behavior and external recommendation information, allowing the agent to make appropriate behavioral decisions based on the most relevant and urgent memory. This multi-level initialization can provide a foundation for subsequent complex interactions and rumor propagation between agents, while ensuring that cognitive differences and behavioral diversity are reflected in the simulation. In addition, the environment server initializes the social network topology, using a directed graph Represents the attention relationship between agents. In different time steps, agent i follows the social network Receives incoming follower lists and outgoing followers list messages, allowing social networks (for brevity, The representation of time t is omitted) changes adaptively over time.

[0056] To ensure high accuracy and efficiency in large-scale user simulations, the dynamic stratification module constructs a system for evaluating communication potential and content richness, identifying core agents critical to event propagation and adjusting inter-agent information interaction based on agent type. The dynamic stratification module selects core agents with higher communication potential and content diversity from regular agents. Core agents are driven by a large language model (LLM), supporting complex natural language interaction and reasoning, endowed with human-like behavioral decision-making capabilities. Regular agents are driven by agent-based models (ABMs), enabling rapid attitude updates. Dynamically and adaptively switching core agents with key event propagation characteristics at different time steps allows computing resources to be focused on the most influential core users, thus balancing efficiency and accuracy in large-scale social simulations. Core agents interact through natural language dialogue, while regular agents utilize agent-based models (ABMs) to transmit information. Furthermore, content generated by core agents is converted into attitude scores using the large language model (LLM). These influence the attitude updates of regular agents through agent-based models (ABMs), while the attitudes of core agents remain unaffected by regular agents. The dynamic layering module takes into account both computational efficiency and simulation details, achieving dynamic and hierarchical simulation of large-scale social networks while accurately restoring the real-world communication dynamics.

[0057] The core agent makes behavioral decisions based on user profiles, memory mechanisms, and information flow link prediction. First, user profiles provide the core agent with personalized context, helping it understand its own personality and factor in individual needs and preferences when making decisions, resulting in more diverse and rational behavior. Second, the memory mechanism enables the agent to make decisions based on past behavior records and interactions with neighbors. The agent possesses a personal experience memory, allowing it to review its own historical behavior. It can also capture the attitudes and behavioral responses of neighboring agents from environmental interaction memories. These memories can be retrieved, updated, and reflected upon, helping the agent make more targeted and timely decisions. Finally, information flow link prediction methods help the agent analyze the evolution of social relationships. By quantifying attitude similarity, content timeliness, and tweet influence, this process helps the agent identify potential social connections and interaction opportunities, and recommend appropriate tweets or interaction partners. By integrating the latest information flow, personalized needs and historical memory, the core agent can flexibly make seven types of behavioral decisions based on the current situation, such as uploading tweets, forwarding tweets, commenting on tweets, following other agents, unfollowing other agents, liking tweets uploaded by others, and doing nothing. This ultimately promotes interaction and information dissemination in social networks, while prompting dynamic changes in social relationships and attitudes.

[0058] The behavioral decision-making module of conventional agents for non-equal social relationships mainly solves the problems of unequal interactions and dynamic social relationship adjustment in real social networks by introducing the concept of trust and a dynamic link prediction engine. First, in order to quantify the unequal attitude interaction between users, the present invention introduces the concept of trust. Trust is used to measure the degree of trust between conventional agent i and its neighboring agent j. Agents with higher trust or more inbound followers will have a greater impact on the decision-making process. In the attitude update process, the attitude of conventional agent i is Based on the neighboring agent set The influence of each neighbor is adjusted by a weight This weight takes into account trust and the number of inbound followers in the social network. Secondly, the dynamic link prediction engine simulates the dynamic adjustment of social relationships through multi-dimensional factors such as attitude similarity, one-way attention, user influence, and trust level. Based on these factors, the dynamic link prediction engine predicts the possible addition or disconnection of attention relationships between agents. By calculating the missing link prediction score and false link identification score , the dynamic link prediction engine can evaluate the interaction possibility between agents and adjust the links in the social network accordingly. In particular, when a new social relationship is established, agents need to establish a new trust relationship through the trust propagation path. The time wheel starts, press Link prediction is performed periodically, and the probability of adding a new attention relationship is , the probability of unfollowing is In summary, the behavioral decision-making module for unequal social relationships enables conventional agents to better adapt to unequal interactions in the real world, while accurately capturing the evolution of social relationships and precisely aligning with the real-world communication dynamics.

[0059] After completing a simulation, the environment server updates the agent's memory, tweet library, and social network. Based on the core agent's interactive behavior, it updates detailed information such as comments, number of likes, and creation time for each tweet. The tweet content is then sent to incoming followers to update the agent's memory. Based on the agent's follow and unfollow behavior, the social network topology is updated. The environment server provides a dynamic and scalable infrastructure for the entire simulation framework, ensuring data flow and state updates during the simulation, which influence the agent's decision-making in the next round. By continuously iterating at a preset time step, the social network simulator is capable of performing large-scale event situation simulations and group behavior simulations.

[0060] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps 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 multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times. The order of execution of these steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0061] In one embodiment, the present invention provides a computer system, which may be a server. The computer system includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer system is used to provide computing and control capabilities. The memory of the computer system includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer system is used to store data used in the above method. The network interface of the computer system is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0062] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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.

[0063] 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 embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0064] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A large-scale social network simulation method based on dynamic topology structure, characterized by: Used to deduce attitudes towards events occurring in the target user group, including: Assign user profiles and memories to corresponding agents, and build a social network for the agent based on the user's outgoing and incoming followers on the social platform; Dynamic evolution process: The influence measure of each agent is calculated based on the number of second-order incoming followers corresponding to each agent and the variance of the incoming followers' attitudes towards the event. The core agents are distinguished from conventional agents based on the influence measure. The core agents are driven by a large language model and interact with other core agents through natural language, influencing the attitudes of conventional agents towards the event. The core agents use user portraits, memory mechanisms, and information flow-based link prediction methods to decide on interactive behaviors with other tweets or agents and adjust the social network. Conventional agents are driven by a subject model and interact with other conventional agents. Conventional agents update their attitudes through trust weights and adjust the social network using a dynamic link prediction engine. The agent's memory, tweet library, and social network are updated. The dynamic evolution process is continued at each time step until the deduction of the event is completed and the user's attitude towards the event is obtained.

2. The large-scale social network simulation method based on dynamic topology structure according to claim 1, characterized in that: The influence measure of each agent is calculated by the number of second-order incoming followers corresponding to each agent and the variance of the incoming followers' attitudes towards the event, specifically including: The influence metric quantifies the agent's communication ability by the number of second-order inbound followers, and measures the content diversity of the agent's tweets by the variance of the inbound followers' attitudes towards the event: ； in, represents the influence measure of agent i at time t, represents the attitude of agent i towards the event at time t, Indicates whether agent i pays attention to agent j, Indicates that agent i does not pay attention to agent j, Indicates that agent i pays attention to agent j; represents the incoming follower list of agent i; Indicates whether agent j pays attention to agent k, Indicates that agent j does not pay attention to agent k, Indicates that agent j pays attention to agent k.

3. The large-scale social network simulation method based on dynamic topology structure according to claim 2, characterized in that: The distinction between core agents and conventional agents based on influence metrics specifically includes: At each time step, the top K agents with the highest influence metrics are selected as core agents, and the other agents are used as regular agents.

4. The large-scale social network simulation method based on dynamic topology structure according to claim 1, characterized in that: The core agent uses user profiling, memory mechanisms, and information flow-based link prediction methods to decide on interactions with other tweets or agents and adjust social networks. Specifically, Calculate the recommendation score of core agent i and candidate tweet j , Integrates the content representation of the latest tweets of core agent i and content representation of candidate tweets The content matching degree between them, the life cycle factor of the candidate tweet and the content influence of the candidate tweet j : ； in, represents the cosine similarity, is a natural constant, represents the release time of candidate tweet j, represents the index at time t, Indicates the decay rate of the life cycle factor; Form a recommendation list for core agent i based on the recommendation score: select agents that meet the following conditions among the agents that are not followed by core agent i: the number of inbound followers exceeds the set threshold or the outbound followers of the outbound followers of core agent i; use the tweets posted by the selected agents as candidate tweets, and select the N tweets with the highest recommendation scores from the candidate tweets to form a recommendation list; The core agent i makes decisions based on user portraits, memories, messages received from neighboring agents, and personalized information flows based on recommendation lists, obtains interactive behaviors with other tweets or agents, and adjusts its own attention relationships.

5. The large-scale social network simulation method based on dynamic topology structure according to claim 4, characterized in that: The core agent i makes decisions based on user profile, memory, messages received from neighboring agents, and personalized information flow based on recommendation lists, and obtains interactive behaviors with tweets or other agents, specifically including: For the decision-making process, User portrait for the core agent i, represents the memory of core agent i at time t-1, is the personalized information flow of core agent i at time t, is the message received by the core agent i from the neighboring agent at time t-1; represents the set of actions performed at time t, which includes uploading tweets, forwarding tweets, commenting on tweets, following other agents, unfollowing other agents, liking tweets uploaded by others, and doing nothing.

6. The large-scale social network simulation method based on dynamic topology structure according to claim 1, characterized in that: The conventional agent updates its attitude through trust weights, specifically including: Based on trust, the attitude interaction between regular agents and other agents is quantified. Agents with higher trust or more inbound followers are given higher weights: ； ； in, Represents the attitude of conventional agent i towards the event at time t Neighboring Agents The weight of influence, Indicates trust in weight The proportion of represents the degree of trust that agent i has in agent j, is the set of neighboring agents of regular agent i at time t, is the incoming follower list of agent j, Update the weights for the attitudes.

7. The large-scale social network simulation method based on dynamic topology structure according to claim 1, characterized in that: The use of a dynamic link prediction engine to adjust a social network specifically includes: Missing link prediction scores by regular agent i and agent j and false link identification score Modeling the creation and disconnection of social relationships, The highest agent j is the newly followed outgoing follower of regular agent i, The highest agent j is the most recently unfollowed outgoing follower of regular agent i.

8. The large-scale social network simulation method based on dynamic topology structure according to claim 7, characterized in that: The missing link prediction score and false link identification score The calculation method is: ； ； in, represents the attitude of conventional agent i towards the event at time t, represents the attitude of agent j towards the event at time t, is the incoming follower list of agent j, represents the set of candidate follower agents of regular agent i in the social network. The elements of the candidate follower agent set include the outgoing followers of regular agent i and agents whose number of incoming followers exceeds the set value. Indicates whether agent j pays attention to regular agent i, is the outgoing follower list of regular agent i, Indicates the degree of trust that agent i has in agent j.

9. The large-scale social network simulation method based on dynamic topology structure according to claim 8, characterized in that: The degree of trust that agent i has in agent j According to the agent k acting as the middleman in all trust propagation paths: ； ； in, Represents the trust relationship between agents i and j established based on agent k in the trust propagation path.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images