A method and system for simulating social media user interaction behavior based on real-world behavior alignment and dynamic environmental constraints.
By constructing locally visible contexts and real-time updated simulation environments in social media, the problem that existing technologies struggle to reflect differences in information environments and form closed loops in user behavior simulation is solved, achieving accurate simulation of event propagation structures and temporal evolution trajectories.
Patent Information
- Application Number
- CN202611457360.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-17
AI Technical Summary
Existing social media simulation technologies struggle to form accurate event propagation structures and time evolution trajectories through continuous individual interactions. They are limited by platform constraints, making it difficult to collect user behavior data, reflect differences in users under different information environments, and form a continuous feedback loop between user operations and the information environment.
We adopt a social media user interaction behavior simulation method based on real behavior alignment and dynamic environment constraints. By constructing a locally visible context during the user operation iteration process, we first determine the behavior type, interaction goal and emotional attributes, and then perform operations in the current simulation environment and update them in real time to form an event propagation structure and time evolution trajectory.
It enables the reflection of differences in user information environment in social media context, reduces the interference of text generation on operational decisions, forms a continuous feedback loop between user behavior and information environment, and improves the ability to reconstruct the dynamic dissemination process of social media events.
Smart Images

Figure CN123117350A_ABST
Abstract
Description
Technical Field
[0001] This application pertains to a behavior simulation method, specifically relating to a social media user interaction behavior simulation method and system based on real behavior alignment and dynamic environmental constraints. Background Technology
[0002] Public events on social media typically spread continuously through interactions such as original posts, reposts, comments, and multiple layers of replies. The information users encounter during an event is influenced by factors such as follower relationships, platform recommendations, content ordering, and page actions. Furthermore, previous users' comments and reposts alter the information other users subsequently access. Therefore, the event propagation process is a dynamic process formed by the continuous interaction between user behavior and the information environment. Simulating social media events requires not only simulating user-generated text content but also reflecting the actions users take within a specific information environment, the interaction objects they interact with, and the emotional attributes they exhibit. It also necessitates demonstrating the impact of continuous user interactions on the event propagation process.
[0003] Existing social media simulation technologies mainly include rule-driven, cue-driven, and real-data-supervised approaches. Rule-driven approaches describe user activities through predefined rules, decision trees, or dynamic processes; cue-driven approaches utilize large language models for role setting and natural language inference to generate user responses; and real-data-supervised approaches use user information, behavior records, or social relationships to learn user responses, interaction goals, or group activities to simulate user behavior and responses during event propagation. However, most existing methods are limited by platform constraints, making it difficult to collect user behavior data and thus hindering end-to-end behavior supervision. Some solutions mix behavior selection with text generation, making it difficult to separately constrain user behavior types, interaction goals, and emotional attributes. Furthermore, using fixed content pools or providing globally accessible information fails to reflect the local information differences formed by different users on the current page, content order, followed content, and message notifications. Moreover, when new user interactions are not immediately reflected and the simulation environment is not updated, subsequent users still respond based on an environment that does not reflect the results of previous operations, making it difficult to form a closed loop of continuous feedback between user actions and the information environment. Consequently, it is difficult to reconstruct the event propagation structure and its temporal evolution from continuous individual interactions. Summary of the Invention
[0004] This application addresses the technical problem that existing social media simulation methods struggle to form accurate event propagation structures and temporal evolution trajectories through continuous individual interactions. It proposes a social media user interaction behavior simulation method and system based on real-world behavior alignment and dynamic environmental constraints.
[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a method for simulating social media user interaction behavior based on alignment with real-world behavior and dynamic environmental constraints, including: Obtain the event background and initial simulation environment of the event to be simulated; Using the initial simulation environment as the current simulation environment, the currently executing user is determined during the user operation iteration process, and a local visible context for the currently executing user is constructed based on the current time, the event background, the user information of the currently executing user, and the current simulation environment. The user information of the currently executing user, the event background, and the locally visible context are input into the behavior decision module to obtain the behavior type, interaction target, and emotional attribute of the current operation; wherein, if the behavior type involves text expression, the content generation module is called to generate corresponding text content based on the user information of the currently executing user, the event background, the locally visible context, the behavior type, the interaction target, and the emotional attribute. Based on the behavior type, the interaction target, and the corresponding text content when the behavior type involves text expression, the current operation is executed in the current simulation environment, and the current simulation environment is updated in real time based on the result of the current operation; The next user operation iteration is performed based on the updated current simulation environment until the simulation task is completed. Based on the interaction records generated from each user operation iteration, the event propagation structure and time evolution trajectory of the event to be simulated are output.
[0006] Furthermore, the event background and initial simulation environment of the event to be simulated are obtained, including: Determine the initial participating users, initial execution order, and initial time of the event to be simulated; load the user information, follow relationships, and historical posts of the initial participating users; and obtain the initial event content of the event to be simulated. The event background is generated based on the initial event content available at the initial moment, without introducing the subsequent development content of the event to be simulated; The initial simulation environment is constructed based on the initial event content.
[0007] Furthermore, the types of behavior include browsing, reading, posting, forwarding, commenting, replying, viewing messages, and exiting; The emotional attributes include negative, neutral, and positive. The interaction target is represented by the post index or user index of the current page in the locally visible context.
[0008] Furthermore, the behavior decision module includes a decision encoder, a decision model, and a decision decoder; The decision encoder is used to convert the user information of the currently executing user, the event background, and the locally visible context into input lexical representations; The decision model is used to perform decision reasoning based on the input lexical representation and generate the current user's decision lexical sequence. The decision decoder is used to parse the decision lexical sequence into behavior type, interaction target, and emotional attribute.
[0009] Furthermore, the content generation module includes a language encoder, a language model, and a language decoder; The language encoder is used to convert the user information, locally visible context, behavior type, interaction goal and emotion attribute of the current executing user into an input lexical representation; The language model is used to perform language content reasoning based on the input lexical representation and generate the current user's content lexical sequence; The language decoder is used to convert the content word sequence into the text content.
[0010] Furthermore, the method for obtaining the event propagation structure includes: The event propagation structure is constructed based on the interaction records generated by each user operation iteration and the parent-child relationships between the interaction records; the event propagation structure retains the executing user, interaction target, text content and occurrence time corresponding to each behavior.
[0011] Furthermore, the time evolution trajectory includes the trajectory of group cognitive emotions and the trajectory of public attention; The method for obtaining the group cognitive emotion trajectory includes: dividing the text content generated by each user operation iteration into negative, neutral and positive emotions, and summarizing them according to a unified time interval to form the group cognitive emotion trajectory. The method for obtaining the public attention trajectory includes: determining the public attention trajectory based on the number of posts, comments, and reposts related to the event to be simulated within each time interval.
[0012] Secondly, this application proposes a social media user interaction behavior simulation system based on real-world behavior alignment and dynamic environmental constraints, comprising: The basic module is used to obtain the event background and initial simulation environment of the event to be simulated; The context module is used to take the initial simulation environment as the current simulation environment, determine the current executing user during the user operation iteration process, and construct a local visible context for the current executing user based on the current time, the event background, the user information of the current executing user, and the current simulation environment. The calling module is used to input the user information of the currently executing user, the event background, and the locally visible context into the behavior decision module to obtain the behavior type, interaction target, and emotional attribute of the current operation; wherein, if the behavior type involves text expression, the content generation module is called to generate corresponding text content based on the user information of the currently executing user, the locally visible context, the behavior type, the interaction target, and the emotional attribute. The execution module is used to perform the current operation in the current simulation environment according to the behavior type, the interaction target, and the corresponding text content when the behavior type involves text expression, and to update the current simulation environment in real time according to the result of the current operation; to perform the next user operation iteration based on the updated current simulation environment until the simulation task is completed; and to output the event propagation structure and time evolution trajectory of the event to be simulated according to the interaction records formed by each user operation iteration.
[0013] Furthermore, the basic module includes: Determine the initial participating users, initial execution order, and initial time of the event to be simulated; load the user information, follow relationships, and historical posts of the initial participating users; and obtain the initial event content of the event to be simulated. The event background is generated based on the initial event content available at the initial moment, without introducing the subsequent development content of the event to be simulated; The initial simulation environment is constructed based on the initial event content.
[0014] Furthermore, the behavior decision module includes a decision encoder, a decision model, and a decision decoder; The decision encoder is used to convert the user information of the currently executing user, the event background, and the locally visible context into input lexical representations; The decision model is used to perform decision reasoning based on the input lexical representation and generate the current user's decision lexical sequence. The decision decoder is used to parse the decision lexical sequence into behavior type, interaction target, and emotional attribute.
[0015] Compared with the prior art, this application has the following beneficial effects: This application proposes a social media user interaction behavior simulation method based on real-world behavior alignment and dynamic environmental constraints. By constructing a locally visible context based on the current time, event background, user information of the current user, and the current simulation environment during user operation iterations, behavioral decisions are grounded in the locally accessible information available to the user. This reflects the differences in information environments for different users and avoids insufficient local information constraints caused by using fixed or globally accessible content. The behavior decision module first determines the behavior type, interaction goal, and emotional attribute of the current operation. Only when the behavior type involves text expression is the content generation module called to generate corresponding text content, decoupling behavior selection from text generation. This separately constrains what operation the user takes, what object it acts upon, and what emotion it expresses, reducing the interference of text generation results on operation decisions. By executing user operations in the current simulation environment and updating the simulation environment in real time based on the operation results, the content, interaction relationships, and environmental state changes generated by previous user operations can enter the locally visible context upon which subsequent user operations are based. This forms a closed loop of continuous feedback between user behavior and the information environment, and allows continuous individual interactions to gradually accumulate to form the propagation structure and temporal evolution trajectory of the simulated event, improving the ability to recreate the dynamic propagation process of social media events.
[0016] This application also proposes a social media user interaction behavior simulation system based on real behavior alignment and dynamic environmental constraints, which has the advantages of social media user interaction behavior simulation methods based on real behavior alignment and dynamic environmental constraints. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart illustrating the overall process of simulating social media user interaction behavior based on real-world behavior alignment and dynamic environmental constraints in the embodiments of this application. Figure 2 This is a schematic diagram illustrating the potential behavior path inference and label construction in the embodiments of this application; Figure 3 This is a structural diagram of behavioral decision-making and on-demand content generation in the embodiments of this application; Figure 4 This is a closed-loop diagram illustrating the construction of locally visible contexts and the real-time updating of the simulated environment in an embodiment of this application. Figure 5This is a simulation result diagram of the public attention trajectory in the embodiments of this application; Figure 6 This is a simulation result diagram of the group cognitive emotion trajectory in the embodiments of this application; Figure 7 This is a schematic diagram of the social media behavior simulation system in the embodiments of this application. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0021] This application proposes a social media user interaction behavior simulation method based on real behavior alignment and dynamic environment constraints. It is used to continuously simulate the behavior of different users in the social media environment for the event to be simulated, with each user operation as the iterative unit. The changes caused by each operation to the simulation environment can be incorporated into the information environment on which subsequent user operations are based. The event propagation structure and time evolution trajectory of the event to be simulated are gradually formed by continuous individual interactions.
[0022] As an embodiment of the social media user interaction behavior simulation method based on real behavior alignment and dynamic environment constraints in this application, the following method can be specifically adopted: (1) Obtain the event background and initial simulation environment of the event to be simulated.
[0023] The event context provides relevant information about the event to be simulated, informing subsequent behavioral decisions. The initial simulation environment represents the state of the social media environment at the start of the simulation task. Together, the event context and the initial simulation environment form the basis for subsequent user action iterations, with the initial simulation environment continuously changing as user actions are performed.
[0024] (2) Use the initial simulation environment as the current simulation environment and enter the user operation iteration process.
[0025] In each user operation iteration, the currently executing user is determined, and the current time and corresponding user information are obtained. Based on the current time, event context, user information of the currently executing user, and the current simulation environment, a locally visible context corresponding to the currently executing user is constructed. This locally visible context reflects the range of information that the currently executing user can use to make operational decisions at the current moment and in the current simulation environment state. Therefore, the locally visible contexts corresponding to different currently executing users or the same user in different environment states may be different.
[0026] (3) Input the user information, event background and local visible context of the current user into the behavior decision module, and the behavior decision module obtains the behavior type, interaction goal and emotional attribute of the current operation.
[0027] Behavior type determines the type of action performed by the user, interaction goal determines the object of the action, and emotional attribute characterizes the emotional tendency corresponding to the action. By first forming a structured decision result that includes behavior type, interaction goal, and emotional attribute, it is possible to determine whether a user performs an action, where the action is performed, and the corresponding emotional constraints of the action before the specific textual expression.
[0028] When the behavior decision module determines that the behavior type involves text expression, it then calls the content generation module. The content generation module generates text content corresponding to the current operation based on the current user's information, event background, local visible context, behavior type, interaction goal, and emotional attributes. Therefore, in operations requiring text expression, the text content is simultaneously constrained by the current user's own information, the current local information environment, and the already determined structured behavior decision. For operations not involving text expression, there is no need to generate corresponding text.
[0029] (4) Based on the behavior type and interaction goal, and when the behavior type involves text expression, combine the corresponding text content to perform the current operation in the current simulation environment.
[0030] After the current operation is completed, the current simulation environment is updated immediately based on the result of the current operation, so that the environmental changes caused by the current operation are written into the current simulation environment before the next user operation iteration begins. Thus, the content changes, interaction changes, or other environmental state changes caused by previous user operations can become the basis for constructing locally visible contexts for subsequent operations.
[0031] (5) After a current simulation environment update is completed, the next user operation iteration is entered based on the updated current simulation environment.
[0032] In the next user operation iteration, the current user is identified again, and the locally visible context corresponding to the current user in the new environment is reconstructed. Behavioral decisions, on-demand text generation, operation execution, and real-time environment updates continue. This process continues until the predetermined simulation task is completed.
[0033] (6) Form corresponding interaction records during each user operation iteration.
[0034] After the simulation task is completed, the event propagation structure and temporal evolution trajectory of the simulated event are output based on the interaction records formed by each user operation iteration. Since each user operation acts on the current simulation environment, and the result of the operation enters the environment state immediately before the next iteration, the output event propagation structure and temporal evolution trajectory are gradually accumulated by continuous individual operations and their environmental feedback. This addresses the problem that existing social media simulation methods struggle to form event propagation structures and temporal evolution trajectories through continuous individual interactions, thus establishing a continuous state correlation between individual behavior and group propagation results.
[0035] The present application will be further described in detail below through some more detailed embodiments.
[0036] In one embodiment of this application, social media behavior simulation continuously unfolds around the information accessible to the currently executing user, the actions performed by the currently executing user, and the environmental changes after the actions are completed. The information accessible to users in real social media is influenced by factors such as following relationships, platform recommendations, content ranking, the current page, and message notifications. Furthermore, a single comment, reply, repost, or other interaction may alter the information subsequently accessible to other users. Therefore, this embodiment maintains a simulated environment that changes instantly with each user action and reconstructs the locally visible context before each action based on the currently executing user and the current environmental state, creating a closed loop between individual behavior and subsequent information exposure.
[0037] like Figure 1 As shown, in a preferred embodiment of this application, the overall processing procedure may include: S1, Construction of a traceable corpus dataset.
[0038] First, public events, user data, and behavioral records are acquired and correlated to form traceable behavioral samples. Optionally, a traceable corpus dataset is constructed before training the behavior decision-making module and the content generation module. Specifically, initial event content, along with related posts, reposts, comments, citations, and multi-level replies, are collected on a per-public-event basis. The executing user, behavior type, interaction target, text content, timestamp, and position in the propagation structure are recorded for each observable behavior. The propagation relationships between content and the behavioral relationships between users are uniformly organized at the event level, ensuring that each behavioral record can be traced back to a specific user and corresponding propagation branch.
[0039] Event trajectories are divided into behavior instances in chronological order, and each behavior instance retains an event identifier, upstream content, and interaction target to maintain contextual relationships between behavior instances. Therefore, the observable final results such as posting, commenting, forwarding, quoting, and replying can not only be used to generate monitoring tags, but also be associated with user information, dissemination location, and the information environment in which the behavior occurred, providing constraints for subsequently recovering intermediate behaviors such as browsing, reading, and page switching that are often missing from public records.
[0040] User information can consist of public attributes, follower relationships, and historical posts prior to the event. Optionally, public attributes include information such as profile, gender, authentication status, and birthday; follower relationships are collected within the publicly accessible scope, and recent historical posts are retrieved. Historical posts used to construct user information are all earlier than the time the user first participated in the corresponding event, ensuring that the user information used for behavioral simulation remains prior and avoiding the reverse introduction of post-event expressions into prior behavior modeling.
[0041] S2, inference of potential behavioral paths.
[0042] Based on the sample, potential behavioral paths are constructed and screened, and the behavioral records that originally only reflected the final result are expanded into multi-step behavioral supervision labels arranged in chronological order.
[0043] Optionally, user information, initial event background, and corresponding environmental information at the time of the behavior are extracted for each traceable behavior sample. The event background is summarized from the content of the initial event, and candidate posts in the environment are organized into consecutive pages according to the recommendation and ranking rules of the corresponding platform to reconstruct the local information that the user may have encountered before the behavior occurred. At the same time, the interaction goals and behavior results in the actual records are used as result constraints for the candidate path, and a multi-step user operation path that may lead to the actual result is inferred from the initial page.
[0044] When the current page does not yet contain the target post or related content corresponding to the actual record, the candidate path leads to the next page through a browsing operation; after reaching the relevant content, it then connects to reading and commenting, forwarding, or other interactive operations consistent with the actual record. This inference process allows for the formation of multiple candidate paths and does not require the intermediate operation to be uniquely determined solely by the final interaction result.
[0045] When jointly screening candidate paths, two constraint levels are retained simultaneously: environmental accessibility and result consistency. Environmental accessibility requires that page browsing, content access, and interactive operations within the path conform to the platform's permitted operation methods and enable reaching the final interactive goal from the initial page according to the corresponding operational relationships. Result consistency requires that the final interactive objects and behavioral results corresponding to the candidate path are consistent with the actual records. Paths that cannot reach the goal or whose final behavioral results cannot be reproduced are eliminated. When multiple feasible paths meet the above conditions, user interest matching can be further used as a path selection criterion to choose the path that best matches the user's historical interests from among the feasible paths.
[0046] The preserved paths are converted into behavior, goal, content, and emotion tags according to the actual sequence of operations, thus restoring a single final behavior record to a continuous decision sequence including intermediate browsing and reading processes. This processing ensures that the model training samples not only contain the observable actions ultimately performed by the user, but also the intermediate operation chain from local information exposure to the final interaction.
[0047] Figure 3 The diagram illustrates the process of inferring potential behavioral paths and constructing labels. First, user information, event background, and observed interaction goals and behaviors are input. Then, candidate local information pages are reconstructed according to recommendations and content sorting. Based on the reconstructed pages, candidate paths encompassing browsing, reading, and the final interaction are formed from the initial pages. Next, joint filtering is performed using environmental accessibility and result consistency, discarding paths that cannot reach the goal according to platform operation relationships or cannot form observation results. Finally, the retained paths are converted into a multi-step label sequence of behavior, goal, content, and emotion according to the order of operations. Figure 2 The top-to-bottom approach reflects the processing relationship from "the observed final result" to "a continuous chain of behaviors that can be used for supervised learning," rather than directly using the final text as the sole training target.
[0048] S3 decouples behavioral decisions from content generation.
[0049] The behavior decision-making module and the content generation module are trained separately to decouple the determination of behavior type, interaction goal and emotional attribute from the text content generation.
[0050] In a preferred embodiment of this application, the behavior alignment agent includes a behavior decision module and a content generation module. The behavior decision module first determines the behavior type, interaction target, and emotional attribute of the current operation; the content generation module is invoked only when the current behavior type involves text expression, and uses the already formed behavior decision result as the text generation condition. Through the above-mentioned chain relationship, operations such as browsing, reading, and viewing messages that do not require text expression do not need to go through the content generation process, while operations that require text expression generate content only after the structured behavior decision is determined.
[0051] For users During operation The input to the behavior decision-making module consists of user information, event background, and the local visible context of the current user. The input relationship is expressed as follows:
[0052] In equation (1), Input to the model, For users, For the operation time, For user information, As the initial event background, This is the local visible context for the current user.
[0053] The behavior decision-making module includes a decision encoder, a decision model, and a decision decoder. The decision encoder converts user information, event context, and locally visible context into input lexical representations. The decision model then infers and generates decision lexical units based on these representations. The decision decoder parses the complete decision lexical sequence to obtain the behavior type, interaction goal, and emotional attribute. The given decision output relationship is represented as follows:
[0054] in, Represents a behavioral decision-making model. Represents the trainable parameters of the model. and These represent the decision encoder and decision decoder in the behavioral decision model, respectively. Indicates input The input lexical representation obtained after processing by the decision encoder, The model represents the first time. Decision morphemes generated at each position, to Indicates generation The previously obtained decision word sequence, This indicates the number of valid decision tokens generated before the ending token. The decision decoder will be... to The complete decision lexical sequence is parsed into behavior type. Interaction goals and emotional attributes .
[0055] Behavior types include browsing, reading, posting, forwarding, commenting, replying, viewing messages, and exiting; emotional attributes include negative, neutral, and positive; interaction targets are represented by the post index or user index of the current page within the locally visible context, limiting the operation object selected by the decision module to the range of objects that can be located on the current page. When performing a posting operation, optional online content can be retrieved based on specified keywords to generate supplementary information related to the user's interests for the text. This retrieval is a possible enhancement method and is not required for every post.
[0056] Optionally, the behavior decision module uses structured labels consisting of behavior type, interaction target, and emotional attributes as supervision signals to fine-tune the self-trained MediaPilot-mini model, resulting in the behavior decision model DecideGen-Decide. MediaPilot-mini and DecideGen-Decide are specific model implementations, and the behavior decision module is not limited to this specific model name.
[0057] When the behavior type involves textual expression, the content generation module is invoked. The content generation module includes a language encoder, a language model, and a language decoder. The language encoder converts the current user's information, locally visible context, behavior type, interaction goal, and emotional attributes into input lexical representations. The language model infers and generates a content lexical sequence based on the input lexical representations. The language decoder converts the complete content lexical sequence into text content. In a specific expression, X in equation (1)... (u,t) As part of the content generation conditions, its conditional dependency is represented as follows:
[0058] in, Represents a content generation language model. Represents the trainable parameters of the model. and These represent the language encoder and language decoder in the content generation model, respectively. This represents the input word representation obtained after the input information has been processed by the language encoder. The model represents the first time. Content tokens generated at each location, to Indicates generation The previously obtained content word sequence, This indicates the number of valid content tokens generated before the ending token. The language decoder will be... to Convert the complete content word sequence into text content The content generation model only operates when the behavioral decision outcome requires textual representation.
[0059] Optionally, the self-trained MediaPilot model can be fine-tuned using reference text to obtain the content generation model DecideGen-Generator. Similar to the behavior decision module, this specific model name represents a preferred implementation.
[0060] During the training phase, the behavior decision module can use the cross-entropy loss between predicted decision terms and reference decision terms, and the content generation module can use the cross-entropy loss between predicted content terms and reference content terms. For a single training sequence:
[0061] in, This represents the word-level cross-entropy loss of the behavior decision module. Indicates the length of the reference decision lexical sequence. Indicates the position of a word in the sequence. Indicates the first One reference decision term, to Indicates that it is located at Previous reference decision lexical sequence, This represents the input word representation obtained after the corresponding sample input has been processed by the decision encoder. The parameter is The behavior decision model outputs conditional probabilities. During training, the model calculates the negative log-likelihood of the reference decision word at each position, conditioned on the input word representation and the current true word sequence, and performs supervised fine-tuning of the behavior decision model by minimizing this loss.
[0062]
[0063] in, This represents the word-level cross-entropy loss of the content generation module. Indicates the length of the reference content word sequence. Indicates the position of a word in the sequence. Indicates the first One reference content terminology, to Indicates that it is located at Previous reference content lexical sequence, The parameter is The conditional probability output by the content generation model.
[0064] Figure 3 This diagram illustrates the connection between behavioral decisions and on-demand content generation. In the diagram, user information, event background, and locally visible context are first input into the behavioral decision module, which outputs the behavior type, interaction goal, and emotional attribute. When the decision result corresponds to an operation that does not require text, the process directly proceeds to browsing, viewing, or similar operations without calling the content generation module. When the decision result corresponds to an operation that requires textual expression, the context information, along with the determined behavior type, interaction goal, and emotional attribute, are input into the content generation module, which outputs the text content. The two branches ultimately converge into executable structured operations and text content generated when needed. Figure 3 This reflects the conditional dependency that behavioral selection precedes text generation, and text generation is constrained by structured decision-making. S4, Locally Visible Environment Initialization and Construction.
[0065] For the event to be simulated, the participating users, execution order, and initial time are determined. User information, follow relationships, and historical posts of the participating users are loaded, and the initial event content of the event to be simulated is obtained. An initial simulation environment is constructed based on the initial event content available at the initial time. The event background is generated based on the initial event content available at the initial time, without introducing subsequent developments of the event that actually occur after the initial time, to avoid influencing the current user's behavioral decisions with real information prematurely. In other embodiments of this application, a network search can also be performed based on target keywords to supplement the event background.
[0066] The initial simulation environment is used as the current simulation environment. In each round of user operation iteration, the current user is selected according to a determined execution order, and a locally visible context for that user is constructed based on the current time, event background, user information, and the current simulation environment. The current simulation environment can store states such as content, relationships, sorting, pages, and messages. The results of previous operations, after being written to the current simulation environment, can change the locally visible context of subsequent users.
[0067] Optionally, notifications generated by actions such as commenting, replying, and forwarding are written to the user's message box to reflect the impact of platform notifications on subsequent actions. When a user who received the notification becomes the current user in a subsequent iteration, they can choose to enter the message box and view the corresponding message as a locally visible context for the user.
[0068] In a preferred embodiment of this application, the simulated environment organizes posts according to the content presentation mechanisms of different platforms. In a microblogging-style environment, the content ranking is determined by combining the initial post, popular posts, content from followed users, and reverse-time sorting results; in a Twitter-style environment, the content ranking is determined by combining recommendation sorting and reverse-time sorting. The sorted content is divided into a fixed number of pages, and each post is assigned a corresponding post index, content, and interaction information for the behavior decision module to specify interaction goals. Users can only access the current page and content accessible through platform operations such as pagination, viewing messages, expanding full text, accessing the original post, and viewing repost relationships, thus constraining the locally visible context by page position and platform operation relationships.
[0069] S5, operation-by-operation closed-loop simulation and real-time updates.
[0070] The system executes behavioral decisions in the current environment, generates text on demand when textual expression is required, updates the environment immediately after each operation, and then proceeds to the next operation based on the updated environment.
[0071] At the start of a user action, the user information of the current user, the event background, and the current locally visible context are input into the behavior decision module to obtain the behavior type, interaction goal, and emotional attribute of the next action. If the behavior type involves text expression, the content generation module is further invoked to generate corresponding text content based on the user information of the current user, the locally visible context, and the behavior type, interaction goal, and emotional attribute; if the behavior type does not involve text expression, the action is executed directly based on the behavior type and interaction goal.
[0072] Browsing, reading, and viewing messages are used to change the state of the page or partial information that the currently executing user can subsequently access; posting, commenting, replying, and forwarding can add content or establish new dissemination connections. The operation target is located through the post index or user index of the current page, so that the behavioral decision result is actually executed within the currently visible environment.
[0073] After each operation, the simulation environment immediately writes the new content generated by the current operation and synchronously updates user messages, propagation relationships, and post interaction states, so that the next user operation can reconstruct the locally visible context based on the updated environment. This state transition relationship is represented as follows:
[0074] In equation (12), Indicates the first The state of the simulation environment before the next operation is executed. This represents the state transition function that executes user actions and updates the simulation environment. Indicates the user who performed the operation. Indicates the operation type. Indicate the interaction target, This indicates the text content generated when the operation requires textual expression; for operations that do not involve textual expression... Empty.
[0075] Depend on Towards The state transition occurs immediately after the current operation is completed. The updated state is used when constructing the locally visible context in the next iteration. Therefore, the content, messages, post status, or propagation relationships created by the current user can alter the information subsequently received by other users. The process iteratively executes steps including determining the current user, constructing locally visible contexts, making behavioral decisions, generating content on demand, executing operations, and updating the environment in real-time, until the predetermined simulation task is completed.
[0076] Figure 4 This diagram illustrates a closed loop formed by locally visible context and real-time environment updates. The initial event content and user execution order serve as simulation inputs. The current simulation environment maintains content, relationships, ordering, and message states. In each round of operations, the currently executing user is selected, and a locally visible page for that user is constructed based on the current environment. Subsequently, behavioral decisions are made sequentially, text is generated when necessary, and user actions are executed. After the operation is completed, newly added content, messages, post states, and propagation relationships are immediately written back to the current simulation environment. The looping arrows in the diagram indicate that the updated environment re-enters the context construction process for the next operation. As the loop continues, interaction records are gradually accumulated and used to output the propagation structure, group cognitive emotion trajectory, and public attention trajectory. Figure 4 The loop shown indicates that environmental feedback occurs immediately after a single user action, rather than being updated uniformly after all user actions are completed.
[0077] S6 outputs the event propagation structure and time evolution trajectory.
[0078] The event propagation structure, collective cognitive emotion trajectory, and public attention trajectory are output based on continuous interaction records. After the simulation task is completed, the event propagation structure is constructed based on the interaction records formed by each user operation iteration and the parent-child relationships between these records. The event propagation structure retains the executing user, interaction target, text content, and occurrence time corresponding to each behavior, enabling the event propagation branches to be traced along the interaction relationships. For derivative records formed by operations such as comments, replies, and reposts, their parent-child relationships are used to represent the propagation connection between the corresponding behavior and upstream content.
[0079] The time evolution trajectory includes the trajectory of collective cognitive emotion and the trajectory of public attention. For the trajectory of collective cognitive emotion, the text content generated by each user operation iteration is divided into negative, neutral and positive emotions, and the texts of each type are summarized according to a unified time interval. The summary results of each time interval are arranged in chronological order to form the trajectory of collective cognitive emotion.
[0080] For the public attention trajectory, under the same time division caliber, the number of posts, comments, and reposts related to the event to be simulated within each time interval is counted, and the public attention trajectory is formed based on the corresponding numbers for each time interval. Optionally, when evaluating the simulation results and real event records, the real public attention sequence and the simulated public attention sequence are compared after applying the same logarithmic transformation to maintain consistency in the data processing caliber of the two types of sequences.
[0081] In a preferred implementation of this application, the event propagation structure, the trajectory of group cognition and emotion, and the trajectory of public attention are all gradually accumulated through continuous user operations, without pre-setting a macro-level target curve that individual behaviors need to fit. Thus, the continuous actions of individual users within a locally visible context gradually form a propagation result at the group level through real-time environmental feedback.
[0082] As an application and experimental example, public event data from Weibo and Twitter were used to form a corpus for simulation verification. Weibo events were selected from the trending topics list in May 2025, and Twitter events were selected from global trending topics from September to November 2025. Search keywords were determined based on the original topic tags and their temporal changes, and GPT was used to supplement and filter the keywords. The data retrieval range was set to 7 days before and 30 days after the event first entered the trending list.
[0083] Twitter data is collected recursively along reply and citation relationships to preserve multi-level nested structures. For large-scale microblogging events, relevant content is sampled, retaining comments, retweets, and multi-level derivative relationships within the sample range, and supplementing with popular posts to reconstruct the main dissemination structure. User information's public attributes include 22 items, which may include bio, gender, verification status, and birthday. User follow relationships are collected within the publicly accessible scope, and up to 50 recent historical posts for each user are obtained. The historical posts used to construct user information are earlier than the time the user first participated in the corresponding event.
[0084] The size of the corpus dataset instances is shown in Table 1.
[0085] Table 1. Corpus Dataset Implementation Scale
[0086] The data scale in Table 1 is used to illustrate a specific corpus construction example and does not limit the simulation method to using the same number of events, users, follow relationships, or historical posts.
[0087] In this experimental example, the model was trained using eight NVIDIA A100 GPUs with a learning rate set to 5×10⁻⁶. -6 The training process lasts for 2 epochs; a cosine learning rate is used for scheduling, with a warm-up ratio of 0.1; the batch size per device is 8, the gradient accumulation step is 2, the global effective batch size is 128, and a fixed random seed is used.
[0088] When evaluating the simulation results, the accuracy and Weighted-F1 scores for behavior type, interaction goal, and emotion attribute were compared, and the generated content was evaluated using BERT-Sim and emotion consistency. Weighted-F1 was used to sum the weights of each category equally; BERT-Sim was used to calculate the cosine similarity between the generated text and the reference text representation; and emotion consistency was used to indicate whether the emotion classification labels of the two texts were the same. The experimental statistical results are shown in Table 2.
[0089] Table 2 Statistical Results of Simulation Experiment
[0090] In this experiment, DecideGen's weighted-F1 score for behavior type was 0.756, for interaction target it was 0.776, for emotion it was 0.661, for semantic similarity to BERT-Sim it was 0.814, and for emotion consistency it was 0.818.
[0091] like Figure 5 As shown, the simulation results of public attention trajectories for multiple public events are presented. Each subplot corresponds to a public event, and the actual and simulated sequences are compared along the same time axis. Horizontal variations reflect the event's propagation over time, while the fluctuations and peaks of the curves reflect changes in the amount of event-related activity within the corresponding time intervals. The actual and simulated sequences are processed using the same data processing methods, recording the results after the same logarithmic transformation for comparison. As can be seen from the subplots, the simulated and actual propagation sequences exhibit high consistency in overall trends, stage fluctuations, and major peak positions. Figure 5 This is used to display the time-varying results of public attention trajectories obtained from continuous user interaction records over multiple events.
[0092] like Figure 6As shown, the simulation results of the collective cognitive emotion trajectory for multiple public events are presented, corresponding to the collective cognitive emotion trajectory described in this application. Each subgraph in the figure compares the real emotion sequence with the simulated emotion sequence along the time axis, and the changes in the curves over time reflect the phased evolution of collective emotion during the event's propagation. The simulated trajectory in each subgraph can follow the real sequence, exhibiting corresponding overall changes and local fluctuations, thus demonstrating that the simulation method can reflect the dynamic changes of collective emotion during the propagation of different public events. Figure 6 Primarily used to demonstrate the temporal evolution of group emotional output, and Figure 5 The changes in public attention shown together constitute a case study verifying the process of group communication.
[0093] like Figure 7 The diagram shown is a schematic of a social media behavior simulation system, which may include: The basic module is used to obtain the event background and initial simulation environment of the event to be simulated; The context module is used to take the initial simulation environment as the current simulation environment, determine the current executing user during the user operation iteration process, and construct a local visible context for the current executing user based on the current time, the event background, the user information of the current executing user, and the current simulation environment. The calling module is used to input the user information of the currently executing user, the event background, and the locally visible context into the behavior decision module to obtain the behavior type, interaction target, and emotional attribute of the current operation; wherein, if the behavior type involves text expression, the content generation module is called to generate corresponding text content based on the user information of the currently executing user, the locally visible context, the behavior type, the interaction target, and the emotional attribute. The execution module is used to perform the current operation in the current simulation environment according to the behavior type, the interaction target, and the corresponding text content when the behavior type involves text expression, and to update the current simulation environment in real time according to the result of the current operation; to perform the next user operation iteration based on the updated current simulation environment until the simulation task is completed; and to output the event propagation structure and time evolution trajectory of the event to be simulated according to the interaction records formed by each user operation iteration.
[0094] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.
[0095] Furthermore, the modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0096] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for simulating social media user interaction behavior based on real behavior alignment and dynamic environment constraints, characterized in that, include: Obtain the event background and initial simulation environment of the event to be simulated; Using the initial simulation environment as the current simulation environment, the currently executing user is determined during the user operation iteration process, and a local visible context for the currently executing user is constructed based on the current time, the event background, the user information of the currently executing user, and the current simulation environment. The user information of the currently executing user, the event background, and the locally visible context are input into the behavior decision module to obtain the behavior type, interaction target, and emotional attribute of the current operation; wherein, if the behavior type involves text expression, the content generation module is called to generate corresponding text content based on the user information of the currently executing user, the event background, the locally visible context, the behavior type, the interaction target, and the emotional attribute. Based on the behavior type, the interaction target, and the corresponding text content when the behavior type involves text expression, the current operation is executed in the current simulation environment, and the current simulation environment is updated in real time based on the result of the current operation; The next user operation iteration is performed based on the updated current simulation environment until the simulation task is completed. Based on the interaction records generated from each user operation iteration, the event propagation structure and time evolution trajectory of the event to be simulated are output.
2. The method of claim 1, wherein, Obtain the event background and initial simulation environment for the event to be simulated, including: Determine the initial participating users, initial execution order, and initial time of the event to be simulated; load the user information, follow relationships, and historical posts of the initial participating users; and obtain the initial event content of the event to be simulated. The event background is generated based on the initial event content available at the initial moment, without introducing the subsequent development content of the event to be simulated; The initial simulation environment is constructed based on the initial event content.
3. The method of claim 1, wherein, The behavior types include browsing, reading, posting, forwarding, commenting, replying, viewing messages, and exiting; The emotional attributes include negative, neutral, and positive. The interaction target is represented by the post index or user index of the current page in the locally visible context.
4. The method for simulating social media user interaction behavior based on real behavior alignment and dynamic environment constraints according to claim 1, characterized in that, The behavioral decision-making module includes a decision encoder, a decision model, and a decision decoder; The decision encoder is used to convert the user information of the currently executing user, the event background, and the locally visible context into input lexical representations; The decision model is used to perform decision reasoning based on the input lexical representation and generate the current user's decision lexical sequence. The decision decoder is used to parse the decision lexical sequence into behavior type, interaction target, and emotional attribute.
5. The method for simulating social media user interaction behavior based on real behavior alignment and dynamic environment constraints according to claim 1, wherein, The content generation module includes a language encoder, a language model, and a language decoder; The language encoder is used to convert the user information, locally visible context, behavior type, interaction goal and emotion attribute of the current executing user into an input lexical representation; The language model is used to perform language content reasoning based on the input lexical representation and generate the current user's content lexical sequence; The language decoder is used to convert the content word sequence into the text content.
6. The method for simulating social media user interaction behavior based on real behavior alignment and dynamic environment constraints according to claim 1, characterized in that, The method for obtaining the event propagation structure includes: The event propagation structure is constructed based on the interaction records generated by each user operation iteration and the parent-child relationships between the interaction records; the event propagation structure retains the executing user, interaction target, text content and occurrence time corresponding to each behavior.
7. The method for simulating social media user interaction behavior based on real behavior alignment and dynamic environment constraints according to claim 1, characterized in that, The time evolution trajectory includes the trajectory of group cognition and emotion and the trajectory of public attention; The method for obtaining the group cognitive emotion trajectory includes: dividing the text content generated by each user operation iteration into negative, neutral and positive emotions, and summarizing them according to a unified time interval to form the group cognitive emotion trajectory. The method for obtaining the public attention trajectory includes: determining the public attention trajectory based on the number of posts, comments, and reposts related to the event to be simulated within each time interval.
8. A social media user interaction behavior simulation system based on real behavior alignment and dynamic environment constraints, characterized in that, include: The basic module is used to obtain the event background and initial simulation environment of the event to be simulated; The context module is used to take the initial simulation environment as the current simulation environment, determine the current executing user during the user operation iteration process, and construct a local visible context for the current executing user based on the current time, the event background, the user information of the current executing user, and the current simulation environment. The calling module is used to input the user information of the currently executing user, the event background, and the locally visible context into the behavior decision module to obtain the behavior type, interaction target, and emotional attribute of the current operation; wherein, if the behavior type involves text expression, the content generation module is called to generate corresponding text content based on the user information of the currently executing user, the locally visible context, the behavior type, the interaction target, and the emotional attribute. The execution module is used to perform the current operation in the current simulation environment according to the behavior type, the interaction target, and the corresponding text content when the behavior type involves text expression, and to update the current simulation environment in real time according to the result of the current operation; to perform the next user operation iteration based on the updated current simulation environment until the simulation task is completed; and to output the event propagation structure and time evolution trajectory of the event to be simulated according to the interaction records formed by each user operation iteration.
9. A social media user interaction behavior simulation system based on real-world behavior alignment and dynamic environmental constraints according to claim 8, characterized in that, The basic module includes: Determine the initial participating users, initial execution order, and initial time of the event to be simulated; load the user information, follow relationships, and historical posts of the initial participating users; and obtain the initial event content of the event to be simulated. The event background is generated based on the initial event content available at the initial moment, without introducing the subsequent development content of the event to be simulated; The initial simulation environment is constructed based on the initial event content.
10. The social media user interaction behavior simulation system based on real behavior alignment and dynamic environment constraints according to claim 8, characterized in that, The behavioral decision-making module includes a decision encoder, a decision model, and a decision decoder; The decision encoder is used to convert the user information of the currently executing user, the event background, and the locally visible context into input lexical representations; The decision model is used to perform decision reasoning based on the input lexical representation and generate the current user's decision lexical sequence. The decision decoder is used to parse the decision lexical sequence into behavior type, interaction target, and emotional attribute.
