A method and system for predicting propagation trends based on social network simulation

Abstract

The application relates to the technical field of computer software, and discloses a propagation trend prediction method and system based on social network simulation. The method comprises the following steps: constructing a social network simulation sandbox, the social network simulation sandbox comprising a plurality of intelligent agents based on a large language model, which are used to simulate the interaction of social media users in a social network topology; performing intelligent agent interaction based on a social mean field mechanism, representing the social network state through a social mean field state comprising a text mean field and a numerical mean field; performing multi-source information aggregation, integrating heterogeneous social media metadata and propagation characteristics into a text representation; and based on a multi-modal large model, inputting the image of user-generated content and the text representation to predict the propagation trend of the user-generated content. The application overcomes the limitations of traditional historical data induction methods and can simulate the propagation process of user-generated content in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and more specifically to a method and system for predicting propagation trends based on social network simulation. Background Technology

[0002] With the rapid development of social networks, social media has become an important platform for people to obtain information and interact. Social media trend prediction, a technique that predicts the popularity of information or content by analyzing user behavior and content dissemination patterns, has become an important tool for studying the dynamics of social networks. Trend prediction not only helps businesses, media outlets, and other organizations understand changes in public interest but also predicts the trends in information dissemination. The core of social media trend prediction lies in understanding and simulating the path and speed of information dissemination within social networks. By analyzing user behavior, social relationships, and interaction patterns, it predicts which content will receive widespread attention and spread rapidly, and which factors play a key role in the popularity of content. For example, user behaviors such as forwarding, commenting, and liking reflect the popularity of content, while factors such as the community structure of social networks and the influence of information sources directly determine the scope and speed of content dissemination. Social media trend prediction has significant practical implications. In the media industry, predicting popular content helps improve the accuracy and influence of news dissemination, attracting more readers or viewers. Therefore, social media trend prediction is not only an in-depth exploration of the patterns of content dissemination within social networks but also an important component of modern social information management and decision support systems. With the continuous development of technologies such as big data, artificial intelligence, and deep learning, the methods and systems for predicting social media trends will become increasingly accurate, providing more scientific data support for decision-makers in various industries.

[0003] Existing social media dissemination trend prediction technologies can be broadly categorized into three types: feature engineering-based methods, graph neural network-based methods, and multimodal generative model-based methods. First, feature engineering-based techniques extract a series of features (such as text, image features, and user behavior) from social media content and then utilize traditional machine learning algorithms (such as support vector regression and decision trees) to predict dissemination trends. These methods typically rely on manually selected features, building a model to capture the relationship between different features and content dissemination trends. For example, the HyFea method designs and selects multiple features and uses the CatBoost algorithm to optimize model performance. Second, graph neural network-based methods focus on simulating the complex relationships between users and content in social networks. By constructing graph structures and encoding information such as user interactions and content dissemination paths into the graph, graph neural networks can capture high-order dependencies in social networks. For example, the NIPA method uses graph convolutional neural networks to learn structural features in the social graph, thereby more accurately predicting content dissemination trends. Furthermore, in recent years, dissemination trend prediction methods based on multimodal generative models have gradually emerged. These methods typically combine text and image data, using multimodal generative models for joint modeling to capture the complex semantic information of multimodal content. The BLIP model learns features of social content from both visual and textual perspectives through language-visual pre-training techniques, while the RAG-Trans model enhances the expressive power of user-generated content (UGC) features through augmented hypergraph representations. These methods generally employ end-to-end training, enabling them to jointly optimize propagation trend prediction results across multiple modalities. These existing techniques have made significant progress in simulating social network dynamics and capturing content propagation patterns, providing a wealth of models and methods for propagation trend prediction.

[0004] Existing methods for predicting social media trends generally rely on inductive paradigms, that is, reasoning and prediction based on statistical patterns in historical data. These methods attempt to predict future content dissemination trends through inductive analysis of historical data. However, the core problem with this approach is that it can only infer from past events and user behavior, making it ineffective in dealing with the dynamically changing social network environment. In actual social media, information dissemination is not only influenced by historical behavior but also closely related to current hot topics, individual interactions, and real-time changes in the social network. Therefore, relying solely on historical data for inductive prediction ignores the timeliness and dynamism of information dissemination in social networks and fails to capture changes in events or user behavior in a timely manner. Furthermore, historical data itself may gradually become outdated over time, especially on rapidly changing social platforms, where past data cannot fully reflect current dissemination trends and the evolution of user interests. Due to these limitations, inductive methods based on historical data often exhibit insufficient adaptability when facing new social media trends and cannot provide accurate dissemination trend predictions. Therefore, although these methods are effective under certain conditions, their limitation lies in their inability to capture the dynamic characteristics of information dissemination, resulting in low accuracy and real-time performance in predicting social media content dissemination trends. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for predicting propagation trends based on social network simulation.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting propagation trends based on social network simulation, comprising: A social network simulation sandbox is constructed, which contains multiple agents based on a large language model. Each agent has an independent user profile, memory module, and behavioral pattern to simulate the interaction of social media users in the social network topology. The interaction of intelligent agents is carried out based on the social mean field mechanism. The state of the social network is represented by the social mean field state, which includes text mean field and numerical mean field. The text mean field summarizes the generation of all intelligent agent behaviors through a large language model. The numerical mean field simulates the dynamic opinion of intelligent agents on user-generated content based on the Deffuant model. The intelligent agent makes behavioral decisions based on the social mean field state and updates the social mean field state in reverse. The social mean field state is used as a propagation feature. Multi-source information aggregation is performed to integrate heterogeneous social media metadata and dissemination characteristics into a text representation; Based on a multimodal large model, the propagation trend of user-generated content is predicted by taking the image and text representation of user-generated content as input.

[0007] In one embodiment, the user profile of the agent includes name, gender, age, occupation, and interests; the agent's memory module is used to record and review the agent's interaction history with the environment and other agents; the agent's behavior patterns include posting content, forwarding, liking, and commenting; each agent dynamically adjusts its own behavior patterns based on its interactions with neighboring agents and changes in the social average field.

[0008] In one embodiment, the text mean field is generated and updated in the following manner: ; in, for The mean field of the text at time t, It is a function that uses the behavior of an agent to update the average field of the text; For intelligent agents A textual description of the behavior at time t is transformed using a fixed template; To guide the large language model in summarizing the cue words of the entire social network state.

[0009] In one embodiment, the numerical mean field simulates the dynamics of an agent's opinions on user-generated content based on the Deffuant model, specifically including: In each round of interaction, the agent's attitude score towards user-generated content is updated according to the following method: ; ; in, For updating the agent's attitude score, For intelligent agents The attitude score towards user-generated content at time t. It is an intelligent agent The degree of change in attitude scores from time t to time t+1 For intelligent agents Influence weight, For time t, the agent can be... The set of agent indices that exert influence: ; in, It is an intelligent agent The set of indices of neighboring agents; It is a threshold; and They are intelligent agents and intelligent agents Index; Numerical mean field at time t It is obtained by averaging the attitude scores of all agents towards user-generated content: ; This represents the total number of intelligent agents.

[0010] In one embodiment, the agent makes behavioral decisions based on the social average field state, specifically including: ; in, For the prompt function, use a large language model for reasoning agent. Behavior at time t+1 And estimate the attitude score at time t+1. , As a prompt word, Mean field of text Sum of numerical mean fields The social average field state that constitutes it. For the memory module of the intelligent agent, This represents the text representation after integrating heterogeneous social media metadata and dissemination characteristics.

[0011] Secondly, the present invention provides a propagation trend prediction system based on social network simulation, comprising: Sandbox building module: Constructs a social network simulation sandbox. The social network simulation sandbox contains multiple agents based on a large language model. Each agent has an independent user profile, memory module, and behavior pattern, which are used to simulate the interaction of social media users in the social network topology. Interaction Module: Based on the social mean field mechanism, the module executes agent interaction. The social network state is represented by the social mean field state, which includes text mean field and numerical mean field. The text mean field summarizes the generation of all agent behaviors through a large language model. The numerical mean field simulates the dynamic opinion of agents on user-generated content based on the Deffuant model. Agents make behavioral decisions based on the social mean field state and update the social mean field state in reverse. The social mean field state is used as a propagation feature. Integration module: Aggregates multi-source information, integrating heterogeneous social media metadata and dissemination characteristics into a text representation; Prediction module: Based on a multimodal large model, it takes the image and text representation of user-generated content as input to predict the spread trend of user-generated content.

[0012] In one embodiment, the user profile of the agent includes name, gender, age, occupation, and interests; the agent's memory module is used to record and review the agent's interaction history with the environment and other agents; the agent's behavior patterns include posting content, forwarding, liking, and commenting; each agent dynamically adjusts its own behavior patterns based on its interactions with neighboring agents and changes in the social average field.

[0013] In one embodiment, the text mean field is generated and updated in the following manner: ; in, for The mean field of the text at time t, It is a function that uses the behavior of an agent to update the average field of the text; For intelligent agents A textual description of the behavior at time t is transformed using a fixed template; To guide the large language model in summarizing the cue words of the entire social network state.

[0014] In one embodiment, the numerical mean field simulates the dynamics of an agent's opinions on user-generated content based on the Deffuant model, specifically including: In each round of interaction, the agent's attitude score towards user-generated content is updated according to the following method: ; ; in, For updating the agent's attitude score, For intelligent agents The attitude score towards user-generated content at time t. It is an intelligent agent The degree of change in attitude scores from time t to time t+1 For intelligent agents Influence weight, For time t, the agent can be... The set of agent indices that exert influence: ; in, It is an intelligent agent The set of indices of neighboring agents; It is a threshold; and They are intelligent agents and intelligent agents Index; Numerical mean field at time t It is obtained by averaging the attitude scores of all agents towards user-generated content: ; This represents the total number of intelligent agents.

[0015] In one embodiment, the agent makes behavioral decisions based on the social average field state, specifically including: ; in, For the prompt function, use a large language model for reasoning agent. Behavior at time t+1 And estimate the attitude score at time t+1. , As a prompt word, Mean field of text Sum of numerical mean fields The social average field state that constitutes it. For the memory module of the intelligent agent, This represents the text representation after integrating heterogeneous social media metadata and dissemination characteristics.

[0016] The system and method in this invention correspond to each other; the specific technical solutions applicable to the method are also applicable to the system.

[0017] Compared with the prior art, the beneficial technical effects of the present invention are: The accuracy of propagation trend prediction is significantly improved based on the simulation paradigm: By introducing a social network simulation paradigm, this invention effectively overcomes the limitations of traditional historical data inductive methods and can simulate the propagation process of user-generated content (UGC) in a dynamic environment. This method, by constructing an intelligent agent system based on a Large Language Model (LLM), captures the nonlinear dynamics and complex interactions of information propagation, providing more accurate propagation trend predictions. Through comprehensive simulation of agent behavior, social network interactions, and propagation paths, the model can reflect real-time changes in hot topics on social media, greatly improving the accuracy and adaptability of propagation trend prediction.

[0018] Multi-source information aggregation enhances the multi-dimensional understanding of dissemination trend prediction: This invention integrates heterogeneous social media data (such as user information, user-generated content, and dissemination characteristics) into a semantically rich unified text input through a multi-source information aggregation method, thereby enhancing the model's ability to understand the dynamics of user-generated content dissemination. This method effectively solves the problems of information silos and semantic sparsity in traditional dissemination trend prediction, enabling the prediction model to more comprehensively consider the complex relationships between user behavior, social context, and content features, thus improving the robustness and accuracy of dissemination trend prediction. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the large-scale dynamic social simulation framework proposed in this invention. Detailed Implementation

[0021] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, a propagation trend prediction method based on social network simulation in this invention includes the following steps: S1, Construct a social network simulation sandbox. The social network simulation sandbox contains multiple intelligent agents based on a large language model. Each intelligent agent has an independent user profile, memory module, and behavior pattern, which is used to simulate the interaction of social media users in the social network topology. S2, intelligent agent interaction is performed based on the social average field mechanism. The social network state is represented by the social average field state, which includes text average field and numerical average field. The text average field summarizes the generation of all intelligent agent behaviors through a large language model. The numerical average field simulates the dynamic opinion of intelligent agents on user-generated content based on the Deffuant model. The intelligent agent makes behavioral decisions based on the social average field state and updates the social average field state in reverse. The social average field state is used as a propagation feature. S3 aggregates multi-source information, integrating heterogeneous social media metadata and dissemination characteristics into a text representation; S4, based on a multimodal large model, uses the image and text representation of user-generated content as input to predict the propagation trend of user-generated content.

[0023] This invention divides the prediction of dissemination trends into two stages: simulation and prediction. In the simulation stage, this invention constructs a social network sandbox simulation to model the dissemination process of user-generated content (UGC). The social network sandbox can include 1000 agents based on a large language model (LLM), each with an independent user profile, memory module, and action module, adaptively activated according to social participation inequality rules to simulate social media user behavior. To effectively simulate the large-scale, rapid, dynamic dissemination process of UGC, this invention proposes an agent interaction mechanism based on a social mean field, such as... Figure 2As shown, this mechanism, based on mean-field theory, maintains a social mean field in both textual and numerical forms to encode a group-centric, continuously evolving social network state representation. The textual mean field, taking the behavior of all agents as input, uses a large language model to generate a textual representation of the propagation state of user-generated content (UGC), capturing emerging social behaviors and providing strong flexibility and adaptability. The numerical mean field based on the Deffuant model simulates the dynamics of agents' opinions on UGC, providing a more accurate description of agent interactions. Agents make decisions and interact based on the social mean field state, and their behavior, in turn, affects the updating of the social mean field. This invention uses the social mean field state as a propagation feature for subsequent propagation trend prediction. In the prediction phase, this invention trains a prediction model to comprehensively analyze multimodal user-generated content and propagation features for propagation trend prediction. Due to the heterogeneity and discreteness of user-generated content text and propagation features, this invention employs a multi-source aggregation method to convert social metadata and user-generated content propagation features into semantically rich text for unified feature processing.

[0024] The technical solution of the present invention will be described in detail below in several parts.

[0025] 1. Construct a social network simulation sandbox.

[0026] This section aims to build a social network simulation sandbox based on a large language model agent to achieve efficient and realistic simulation of the spread of user-generated content (UGC) and thus assist in the prediction of subsequent spread trends.

[0027] A social network simulation sandbox consists of agents and their environment. Its construction mainly includes the design of the agent architecture and the simulation environment.

[0028] First, the agent architecture is designed to simulate user behavior in social networks. Each agent represents a social media user and has an independent user profile, memory module, and behavioral patterns. The agent's user profile includes social attributes such as name, gender, age, occupation, and interests, which determine the agent's behavioral tendencies and participation levels in social interactions. Furthermore, the agent is equipped with a memory module to record and review its interaction history with the environment and other agents, thereby guiding its future behavioral decisions. Each agent can perform a series of social media actions, such as posting content, forwarding, liking, and commenting, simulating the user's interaction process on the social platform.

[0029] Secondly, the sandbox simulates user relationships and information dissemination paths in real social media platforms by designing a social network topology. A small-world model is used to initialize the social network topology, ensuring the complexity and diversity of interactions between agents. Agents dynamically interact within the sandbox based on their user profiles and social backgrounds, simulating the spread of user-generated content within the social network. The sandbox simulates how information flows, diffuses, and evolves through users in the social network, providing rich dissemination characteristics and behavioral data for subsequent dissemination trend prediction. The core objective of this step is to provide a reliable infrastructure for subsequent social network simulation and dissemination trend prediction by constructing a realistic and efficient simulation sandbox environment.

[0030] 2. Perform agent interaction based on the social average field mechanism.

[0031] This section aims to model the interaction process between agents using the Social Mean Field (SMF) mechanism to efficiently capture multiple dynamic features in the dissemination of user-generated content and provide accurate input data for predicting dissemination trends.

[0032] Specifically, this invention uses textual mean field and numerical mean field to characterize the state of a social network. The textual mean field is a textual description summarizing the behaviors and states of all agents, capable of capturing emerging group behaviors and reflecting the dynamic changes in the dissemination of user-generated content within the social network. The numerical mean field, on the other hand, uses the Deffuant model to simulate the dynamic changes in agents' opinions on user-generated content. This is a mathematical model for simulating information dissemination; by simulating the mutual influence between agents, it updates each agent's interest value or attitude towards specific user-generated content, reflecting the mutual influence of opinions and the evolution of viewpoints in social networks.

[0033] Each agent makes behavioral decisions using a large language model based on the current state of the social mean field. An agent's actions (such as posting content, forwarding, and commenting) are influenced not only by its own memory and current viewpoint but also by the state of the social mean field. Furthermore, each agent's actions and updated viewpoints affect the state of the social mean field, thus driving the evolution of user-generated content dissemination throughout the social network. Through this two-way interactive mechanism, the social mean field can accurately simulate the complex relationship between individual and group behaviors during information dissemination.

[0034] The agent's behavior is adaptive, dynamically adjusting its actions based on interactions with neighbors and changes in the social mean field. This mechanism allows agents to reflect behavioral differences across various contexts, thus more realistically reproducing the dynamics of user-generated content (UGC) propagation in social networks. Within the framework of the social mean field mechanism, agent behaviors are not isolated but intertwined and mutually influential. By dynamically modeling the behavior of each agent in the entire network, we can obtain a global view of UGC propagation, capturing the propagation effect from individual users to the entire group. Through this stage of interactive simulation, all agents, driven by the social mean field mechanism, can complete a comprehensive simulation of UGC propagation, using the social mean field state as a feature of UGC propagation for subsequent propagation trend prediction. This method effectively enhances the understanding of the dynamics of UGC propagation in social networks, thereby providing strong data support for accurately predicting content propagation trends.

[0035] In agent-to-agent interactions based on the social mean field mechanism, the process of updating the text mean field using agent behavior information can be represented as: ; in, for The mean field of the text at time t, It is a function that uses the behavior of an agent to update the average field of the text. For intelligent agents The textual description of the behavior at time t is transformed using a fixed template. To guide the large language model in summarizing the state of the entire social network using cue words.

[0036] In agent interaction based on the social mean field mechanism, this invention employs the Deffuant model to simulate the dynamics of agents' opinions on user-generated content in a social network. This model captures the process of mutual influence between neighboring agents. In each round of interaction, agents update their attitude scores towards user-generated content (range 0-10, where 10 indicates strong interest and 0 indicates disinterest) according to the following equation: ; ; in, A function to update the agent's viewpoint. For intelligent agents The attitude score towards user-generated content at time t. It is an intelligent agent The degree of change in attitude scores from time t to time t+1 For intelligent agents Influence weight, For time t, the agent can be... The set of agent indices that exert influence: ; in, It is an intelligent agent The set of indices of neighboring agents, It is a threshold used to limit and The acceptable difference in influence between them. This means that only when agents... and When attitudes toward user-generated content are sufficiently similar, Only then will it be right It produces an assimilation effect.

[0037] The numerical mean field at time t is obtained by averaging the attitude scores of all agents towards the user-generated content: .

[0038] The process described above, in which the agent makes decisions and generates behavior and attitude scores for the next moment, can be represented as follows: ; in, As a prompt function, a large language model is used to infer the agent's behavior in the next time step based on existing information. And estimate attitude scores , As a prompt word, Mean field of text Sum of numerical mean fields The social average field state formed by this. It serves as the agent's memory, recording the agent's historical behavior.

[0039] 3. Multi-source information aggregation.

[0040] Metadata generated on social media and user-generated content propagation features generated during the simulation phase are typically heterogeneous and stored in a structured format, making them unsuitable for direct use in propagation trend prediction. Therefore, in the prediction phase, to unify feature processing and enhance its representation, this invention integrates multi-source heterogeneous social media metadata, providing semantically richer social user-generated content features for propagation trend prediction. Specifically, this invention uses a specific template to expand field names in social media metadata and merge them with their corresponding field values, thereby transforming all information into coherent, semantically rich text. This process addresses the semantic sparsity problem in the original social media data, providing more contextual information for subsequent propagation trend prediction based on a large language model, enabling a better understanding of the meaning of each feature.

[0041] 4. Forecasting of transmission trends.

[0042] A multimodal large-scale prediction model is trained to predict the spread trend of user-generated content by integrating text representations obtained from multi-source information aggregation and propagation features. Input samples composed of prompt words and actual propagation trends are used as training data to fine-tune the large language model, thereby improving the agent's predictive performance. The training objective is to minimize the discrepancy between the model's predicted popularity score and the actual popularity metric associated with each social media post, achieved through the cross-entropy function.

[0043] This invention proposes a method for predicting dissemination trends based on social network simulation, replacing traditional inductive prediction methods based on historical data with a simulation paradigm. In the simulation phase, a social network sandbox is constructed, utilizing an intelligent agent system based on a large language model to simulate the dissemination process of user-generated content on social media. By simulating agent behavior and the dynamics of information dissemination within social networks, this method can capture the complexity and nonlinear characteristics of user-generated content dissemination, overcoming the limitations of traditional methods that can only predict from historical data, thus providing more accurate dynamic data support for dissemination trend prediction.

[0044] This invention designs an innovative social mean field mechanism to simulate the dynamic characteristics of user-generated content (UGC) propagation in large-scale social networks using both textual and numerical forms of social mean fields. The textual mean field takes the behavior of all agents as input and uses a large language model to generate a textual representation of the UGC propagation state, capturing social dynamics from group behavior. The numerical mean field, on the other hand, simulates the dynamics of agent opinions using a Deffuant model, describing the evolution of individuals' attitudes towards UGC in the social network. Agents make behavioral decisions based on the social mean field state and update the social mean field state in reverse through a two-way interaction mechanism, thereby realizing the mutual influence between individual and group behaviors. This mechanism provides an efficient computational framework for simulating large-scale UGC propagation and offers accurate propagation characteristics for subsequent propagation trend prediction.

[0045] This invention proposes a multi-source information aggregation method that integrates heterogeneous social media metadata with user-generated content (UGC) propagation features to generate semantically richer UGC features. Specifically, a specific template is used to merge metadata fields with their corresponding values, transforming them into a coherent and semantically rich text representation. This process addresses the semantic sparsity problem in the original data, enhancing the model's ability to understand the propagation features of UGC in social networks. Through this aggregation method, the model can more accurately grasp the meaning of each feature, providing strong support for propagation trend prediction based on large language models, ultimately improving the accuracy and robustness of propagation trend prediction.

[0046] Example: In this embodiment, a large language model is used to drive one thousand agents and construct a social network simulation sandbox to simulate the propagation of user-generated content, obtain propagation characteristics, and predict propagation trends. The large language model adopts the Llama3.1-8B-Instruct model.

[0047] In this embodiment, a social network simulation sandbox based on a large language model agent is first constructed. The agents in this sandbox are built using the Llama3.1-8B-Instruct large language model, possessing powerful language understanding and generation capabilities, and can realistically simulate complex human social behaviors and diverse decision-making patterns. Each agent contains complete user profile information, such as name, gender, age, occupation, interests, and personality traits, making its behavior more personalized and realistic. The agents integrate personal experience memory and environmental memory systems, enabling them to retrieve and reflect on past interactions to make reasonable judgments in different situations. The agents support various social operations such as original posting, content forwarding, interactive replies, and liking, further enhancing the realism of the simulation through natural language generation. After constructing 1000 agents, a social network simulation can be performed. In each round of interaction, the agents take turns acting, calculating the activation probability based on the similarity between their own profile information and the input user-generated content information (post content), and determining whether to be activated. If activated, they execute subsequent actions.

[0048] Agent interaction is executed based on a social average field mechanism. First, the text average field is initialized to an empty string, and the attitude scores of all agents are set to 0. Agents take turns acting; the activated agents generate attitude scores for user-generated content and their subsequent actions based on a large language model. After all agents have acted, the text average field is updated using the large language model. All agent actions are input into the large language model, which is then prompted to summarize the current attitudes of agents in the social network towards user-generated content. The content generated by the large language model is used as the text average field. The numerical average field uses a definitive propagation model to spread the activated agent's opinion to all other users, updating the attitude scores of all other agents. The average of all agent attitude scores is used as the numerical average field state. In subsequent interactions, agents use both the text and numerical social average field content as input information before making decisions. After 10 rounds of interaction, the social average field state will be used as a feature of user-generated content propagation to assist in predicting subsequent propagation trends.

[0049] This step transforms user-generated content, other social information, and dissemination characteristics into semantically rich text. A specific transformation template is used to convert this information, expanding field names and merging them with their corresponding values ​​to convert all information into coherent, semantically rich text. The transformation template is shown in Table 1.

[0050] Table 1, Conversion Template:

[0051] The fields in Table 1, representing social media metadata, are explained as follows: Uid represents the user's unique identifier, Photo_count represents the total number of photos uploaded by the user, Photo_first_datetaken represents the time when the user's first photo was taken, Timezone_id represents the user's time zone, Ispro represents whether the user is a professional, Pid represents the unique identifier for the photo / post, Mediatype represents the media type, Title represents the post title, Postdate represents the posting time, Category represents the main category, Concept represents the core theme concept, Subcategory represents the subcategory, Pathalias represents the user's personalized domain alias, Alltags represents all relevant tags, Longitude represents longitude, Latitude represents latitude, Geoaccuracy represents geolocation accuracy, and Avg_score represents the average interest score of all simulated users.

[0052] It should be noted that all data involved in this application are fully authorized by all parties, and the collection, use and processing of the relevant data comply with relevant laws and regulations.

[0053] This embodiment uses Llama-3.2-11B-Vision-Instruct as the multimodal prediction model. This model takes the images containing social media metadata and propagation features, as well as user-generated content, obtained earlier, as input, and directly predicts the propagation trend score of this content. A higher propagation trend score indicates that the user-generated content is likely to have high popularity in real social networks, meaning it can spread rapidly.

[0054] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0055] 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.

[0056] 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. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can be found in the implementations of the foregoing methods, and repeated details will not be elaborated upon.

[0057] 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.

[0058] 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.

[0059] 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 method for predicting propagation trends based on social network simulation, characterized in that, include: A social network simulation sandbox is constructed, which contains multiple agents based on a large language model. Each agent has an independent user profile, memory module, and behavioral pattern to simulate the interaction of social media users in the social network topology. The interaction of intelligent agents is carried out based on the social mean field mechanism. The state of the social network is represented by the social mean field state, which includes text mean field and numerical mean field. The text mean field summarizes the generation of all intelligent agent behaviors through a large language model. The numerical mean field simulates the dynamic opinion of intelligent agents on user-generated content based on the Deffuant model. The intelligent agent makes behavioral decisions based on the social mean field state and updates the social mean field state in reverse. The social mean field state is used as a propagation feature. Multi-source information aggregation is performed to integrate heterogeneous social media metadata and dissemination characteristics into a text representation; Based on a multimodal large model, the propagation trend of user-generated content is predicted by taking the image and text representation of user-generated content as input.

2. The method for predicting propagation trends based on social network simulation according to claim 1, characterized in that, The user profile of an intelligent agent includes name, gender, age, occupation, and interests; the agent's memory module is used to record and review the agent's interaction history with the environment and other intelligent agents; the agent's behavior patterns include posting content, forwarding, liking, and commenting; each intelligent agent dynamically adjusts its own behavior patterns based on its interactions with neighboring intelligent agents and changes in the social average field.

3. The method for predicting propagation trends based on social network simulation according to claim 1, characterized in that, The text mean field is generated and updated in the following manner: ； in, for The mean field of the text at time t, It is a function that uses the behavior of an agent to update the average field of the text; For intelligent agents A textual description of the behavior at time t is transformed using a fixed template; To guide the large language model in summarizing the cue words of the entire social network state.

4. The method for predicting propagation trends based on social network simulation according to claim 1, characterized in that, The numerical mean field is based on the Deffuant model to simulate the dynamics of an agent's opinions on user-generated content, specifically including: In each round of interaction, the agent's attitude score towards user-generated content is updated according to the following method: ； ； in, For updating the agent's attitude score, For intelligent agents The attitude score towards user-generated content at time t. It is an intelligent agent The degree of change in attitude scores from time t to time t+1 For intelligent agents Influence weight, For time t, the agent can be... The set of agent indices that exert influence: ； in, It is an intelligent agent The set of indices of neighboring agents; It is a threshold; and They are intelligent agents and intelligent agents Index; Numerical mean field at time t It is obtained by averaging the attitude scores of all agents towards user-generated content: ； This represents the total number of intelligent agents.

5. The method for predicting propagation trends based on social network simulation according to claim 1, characterized in that, The agent makes behavioral decisions based on the social average field state, specifically including: ； in, For the prompt function, use a large language model for reasoning agent. Behavior at time t+1 And estimate the attitude score at time t+1. , As a prompt word, Mean field of text Sum of numerical mean fields The social average field state that constitutes it. For the memory module of the intelligent agent, This represents the text representation after integrating heterogeneous social media metadata and dissemination characteristics.

6. A propagation trend prediction system based on social network simulation, characterized in that, include: Sandbox building module: Constructs a social network simulation sandbox. The social network simulation sandbox contains multiple agents based on a large language model. Each agent has an independent user profile, memory module, and behavior pattern, which are used to simulate the interaction of social media users in the social network topology. Interaction Module: Based on the social mean field mechanism, the module executes agent interaction. The social network state is represented by the social mean field state, which includes text mean field and numerical mean field. The text mean field summarizes the generation of all agent behaviors through a large language model. The numerical mean field simulates the dynamic opinion of agents on user-generated content based on the Deffuant model. Agents make behavioral decisions based on the social mean field state and update the social mean field state in reverse. The social mean field state is used as a propagation feature. Integration module: Aggregates multi-source information, integrating heterogeneous social media metadata and dissemination characteristics into a text representation; Prediction module: Based on a multimodal large model, it takes the image and text representation of user-generated content as input to predict the spread trend of user-generated content.

7. The propagation trend prediction system based on social network simulation according to claim 6, characterized in that, The user profile of an intelligent agent includes name, gender, age, occupation, and interests; the agent's memory module is used to record and review the agent's interaction history with the environment and other intelligent agents; the agent's behavior patterns include posting content, forwarding, liking, and commenting; each intelligent agent dynamically adjusts its own behavior patterns based on its interactions with neighboring intelligent agents and changes in the social average field.

8. The propagation trend prediction system based on social network simulation according to claim 6, characterized in that, The text mean field is generated and updated in the following manner: ； in, for The mean field of the text at time t, It is a function that uses the behavior of an agent to update the average field of the text; For intelligent agents A textual description of the behavior at time t is transformed using a fixed template; To guide the large language model in summarizing the cue words of the entire social network state.

9. A propagation trend prediction system based on social network simulation according to claim 6, characterized in that, The numerical mean field is based on the Deffuant model to simulate the dynamics of an agent's opinions on user-generated content, specifically including: In each round of interaction, the agent's attitude score towards user-generated content is updated according to the following method: ； ； in, For updating the agent's attitude score, For intelligent agents The attitude score towards user-generated content at time t. It is an intelligent agent The degree of change in attitude scores from time t to time t+1 For intelligent agents Influence weight, For time t, the agent can be... The set of agent indices that exert influence: ； in, It is an intelligent agent The set of indices of neighboring agents; It is a threshold; and They are intelligent agents and intelligent agents Index; Numerical mean field at time t It is obtained by averaging the attitude scores of all agents towards user-generated content: ； This represents the total number of intelligent agents.

10. A propagation trend prediction system based on social network simulation according to claim 6, characterized in that, The agent makes behavioral decisions based on the social average field state, specifically including: ； in, For the prompt function, use a large language model for reasoning agent. Behavior at time t+1 And estimate the attitude score at time t+1. , As a prompt word, Mean field of text Sum of numerical mean fields The social average field state that constitutes it. For the memory module of the intelligent agent, This represents the text representation after integrating heterogeneous social media metadata and dissemination characteristics.

Images