Public opinion multi-scenario situation deduction method and system based on SDE and DBGM

Abstract

The application discloses a public opinion multi-scene situation deduction method and system based on SDE and DBGM. The method performs time window alignment on multi-source time sequence data of a target public opinion event, and constructs a public opinion state sequence containing emotion, propagation, subject, topic, clarification, intervention, time lag, reflux, rekindling and action cooling. According to a disposal target and a constraint condition, a counter-factual end state is generated, a scenario condition is coded, and a candidate situation trajectory is generated by DBGM according to SDE containing a terminal bridging traction item. Then, the unreachable trajectory is removed, the risk cost and the trajectory fingerprint are screened, a multi-scene situation trajectory set is obtained, the intervention response quantity is inversely solved, and the execution time, the action object and the action strength of the disposal action are output. The method improves the target constraint, the propagation rationality and the disposal executability of the public opinion deduction.

Description

Technical Field

[0001] This invention relates to the fields of public opinion situation inference, stochastic dynamics modeling, diffusion generation model and artificial intelligence-assisted decision-making technology, and in particular to a public opinion multi-scenario situation inference method and system based on SDE and DBGM.

[0002] In this invention, SDE stands for Stochastic Differential Equation and DBGM stands for Diffusion Bridge Generation Model. This invention achieves continuous simulation of public opinion trends across various scenarios, including natural evolution, enhanced propagation, topic reversal, and intervention suppression, by constructing states, generating counterfactual endpoints, encoding scenario conditions, generating diffusion bridge trajectories, eliminating unreachable trajectories, and resolving intervention response quantities from multi-source time-series data of target public opinion events. Background Technology

[0003] With the rapid development of social media, news platforms, short video platforms, forum platforms, and instant messaging platforms, online public opinion regarding public emergencies, corporate brand incidents, product quality incidents, and social hot topics exhibits characteristics such as multi-source heterogeneity, rapid dissemination, complex participating entities, intense emotional fluctuations, frequent topic shifts, and significant cross-platform backflow. Public opinion trend prediction technology is typically used to predict the intensity of public opinion, changes in sentiment, the scope of dissemination, and subsequent development trends, providing technical support for public opinion early warning, risk assessment, and decision-making.

[0004] Existing methods for predicting public opinion mainly include trend prediction methods based on time series, spread range prediction methods based on propagation dynamics, follow-up event inference methods based on event graphs, state transition prediction methods based on Markov chains, and public opinion evolution prediction methods based on deep learning models. These methods can reflect the intensity, sentiment, or thematic changes in public opinion to some extent, but their prediction results typically focus on single-path predictions and are difficult to simultaneously generate continuous trajectory data under different scenarios such as natural evolution, intensified propagation, topic reversal, and intervention / suppression, all from the same starting point and under the same public opinion state.

[0005] Meanwhile, most existing methods for predicting public opinion trends extrapolate future trends based on historical data. They lack a counterfactual endpoint state jointly defined by the handling objectives, dissemination constraints, emotional constraints, credible dissemination lag, cross-platform backflow intensity, and resurgence trigger rate. This makes it difficult to answer the question of "how public opinion might evolve under the premise of meeting a certain handling objective and constraint boundary." For actual public opinion handling, simply predicting future heat or emotional changes is insufficient to support decision-making. It is also necessary to determine how the public opinion state reaches the target endpoint under different handling objectives, as well as the potential risks of dissemination rebound, topic shift, cross-platform backflow, and resurgence during the process of reaching the target endpoint.

[0006] Furthermore, while ordinary generative models or general diffusion time-series generative models can generate time-series samples, they typically lack bridging constraints between the current starting state and the target ending state, making it easy to generate situational trajectories that do not meet the disposal objectives or conform to the causal logic of propagation. For example, unreasonable trajectories may occur, such as a premature decline in negative sentiment before the facts have been clarified, a sudden shift in the topic when there is no reachable propagation path in the propagation subject structure, or an abnormal decrease in the propagation heat when the platform diffusion relationship remains unchanged.

[0007] Furthermore, existing public opinion projection methods have a weak ability to express response actions. They can usually only output trend judgments, risk levels, or sentiment changes, but they are unable to extract the required intervention response at each moment from the projection trajectory, nor can they further map them into specific response actions such as clarification releases, authoritative reposts, de-escalation guidance, interception of false information, topic diversion, or platform traffic restriction. Therefore, existing technologies suffer from insufficient coverage of multiple scenarios, insufficient target constraints, insufficient causal rationality of generated trajectories, insufficient executable response actions, and insufficient adaptability to cross-platform backflow, resurgence risks, and action cooling phenomena.

[0008] Therefore, it is necessary to provide a method and system for multi-scenario public opinion situation inference based on SDE and DBGM. By constructing a continuous data processing chain that includes public opinion state sequence, counterfactual endpoint state, scenario conditions, diffusion bridge candidate situation trajectory, causal reachability screening, risk evolution cost screening, and intervention response quantity inverse solution, the goal constraint, dissemination rationality, trajectory diversity, and handling feasibility of multi-scenario public opinion inference can be improved. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for multi-scenario public opinion situation inference based on SDE and DBGM, in order to solve the following problems existing in the current public opinion inference technology: it is difficult to generate continuous situation trajectories of multiple scenarios under the same starting state; it is difficult to transform the handling target, propagation constraints, emotion constraints, credible propagation time delay, cross-platform backflow intensity and resurgence trigger degree into constrained counterfactual endpoint states; it is difficult to ensure that the generated trajectory simultaneously satisfies the target endpoint constraints and the causal logic of public opinion propagation; it is difficult to eliminate abnormal trajectories that are numerically feasible but whose propagation paths are unreachable; and it is difficult to reverse-engineer specific executable intervention actions from the inferred trajectory.

[0010] This invention represents a public opinion event as a sequence of public opinion states, including emotional intensity, dissemination heat, subject influence, topic shift, factual clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factor. It then generates scenario conditions by combining the dissemination subject structure, platform diffusion relationship, topic migration direction, cross-platform backflow intensity, and intervention action intensity. Next, it uses SDE with endpoint bridging traction term to drive DBGM to generate candidate situational trajectories connecting the current starting state and the counterfactual endpoint state. Finally, it filters by causal reachability, endpoint constraint satisfaction, risk evolution cost, and trajectory fingerprint similarity, and reverse solves the intervention response volume, thereby realizing situational deduction for multi-scenario public opinion evolution and executable disposal actions.

[0011] To achieve the above objectives, this invention provides a method for multi-scenario public opinion situation inference based on SDE and DBGM, comprising: S1. Align the multi-source time-series data of the target public opinion event with a unified time window, extract the intensity of emotion, the heat of dissemination, the degree of influence of the subject, the degree of topic shift, the degree of clarification of facts, the amount of intervention response, the credible dissemination time lag, the intensity of cross-platform backflow, the degree of re-ignition triggering and the action cooling factor, form a public opinion state sequence, and generate the dissemination subject structure and platform diffusion relationship; S2. Determine the current starting state from the public opinion state sequence, and generate multiple counterfactual endpoint states based on the handling target, propagation constraints, emotional constraints, credible propagation time delay, cross-platform backflow intensity, and resurgence trigger degree; S3. Encode the scenario conditions by the aforementioned dissemination subject structure, platform diffusion relationship, topic migration direction, cross-platform backflow intensity, and intervention action intensity. S4. Couple the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM, and the DBGM will generate candidate situation trajectories according to the SDE containing the ending bridging traction term. S5. Based on the directed dependency relationship between the dissemination subject structure, topic migration direction, fact clarification degree and intervention response quantity, unreachable trajectories are eliminated, and a multi-scenario situation trajectory set is obtained by filtering according to the endpoint constraint satisfaction degree, risk evolution cost and trajectory fingerprint similarity. S6. Extract the intervention response quantities at each moment from the set of multi-scenario situation trajectories, map them to the execution time, target and intensity of the action, and output the multi-scenario situation deduction results.

[0012] This invention also provides a public opinion multi-scenario situational inference system based on SDE and DBGM, including: The state construction module is used to align multi-source time-series data of target public opinion events according to a unified time window, extract emotion intensity, dissemination heat, subject influence, topic shift, fact clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, re-ignition trigger degree, and action cooling factor, form a public opinion state sequence, and generate the dissemination subject structure and platform diffusion relationship; The endpoint generation module is used to determine the current starting state from the public opinion state sequence, and generate multiple counterfactual endpoint states based on the handling objectives, dissemination constraints, emotional constraints, credible dissemination time delay, cross-platform backflow intensity, and resurgence trigger degree. The scenario coding module is used to encode scenario conditions based on the structure of the dissemination subject, the platform diffusion relationship, the direction of topic migration, the intensity of cross-platform backflow, and the intensity of intervention actions. The diffusion bridge simulation module is used to couple the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM, and the DBGM generates candidate situation trajectories by SDE containing the ending bridging traction term. The trajectory filtering module is used to eliminate unreachable trajectories based on the directed dependency relationship between the structure of the dissemination subject, the direction of topic migration, the degree of factual clarification, and the amount of intervention response, and to filter and obtain a set of multi-scenario situational trajectories based on the degree of end-point constraint satisfaction, the cost of risk evolution, and the similarity of trajectory fingerprints. The results output module is used to inversely solve the intervention response quantities at each moment from the multi-scenario situation trajectory set, map them into the execution time, target and intensity of the action, and output the multi-scenario situation inference results.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects:

[0014] 1. Improve the completeness of public opinion status expression. This invention represents target public opinion events as a sequence of public opinion statuses that includes emotional intensity, dissemination heat, subject influence, topic shift, factual clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factors. This makes public opinion statuses no longer limited to a single heat or a single emotional indicator, but can more completely express the dissemination, subject, topic, clarification, intervention, backflow, resurgence, and action cooling factors in the evolution of public opinion.

[0015] 2. Enhance the target constraint of the extrapolation results. This invention generates multiple counterfactual endpoint states based on the handling target, propagation constraints, emotional constraints, credible propagation time delay, cross-platform backflow intensity, and resurgence trigger degree. This ensures that the candidate situation trajectory is guided by the target endpoint and constrained by the boundary during the generation stage, avoiding the problem that existing trend prediction methods cannot meet the public opinion handling target by extrapolating only historical data.

[0016] 3. Enhanced ability to continuously extrapolate across multiple scenarios. This invention couples the current starting state, the counterfactual ending state, and scenario conditions into the DBGM, and generates candidate situational trajectories through SDE containing the ending bridging term. It can generate continuous evolution paths under different scenarios such as natural evolution, enhanced propagation, topic reversal, and intervention suppression under the same current public opinion state, thereby improving the coverage of multi-branch public opinion evolution.

[0017] 4. Improve the rationality of the generated trajectory dissemination. This invention eliminates unreachable trajectories based on the directed dependency relationship between the dissemination subject structure, the direction of topic migration, the degree of factual clarification, and the amount of intervention response. It can identify and exclude unreasonable trajectories such as premature decline in sentiment before factual clarification takes effect, sudden topic migration due to unreachable subject dissemination path, and abnormal decline in dissemination heat due to mismatch of platform diffusion relationship, thereby improving the causal rationality of the inference results.

[0018] 5. Improve the accuracy and diversity of trajectory selection. This invention uses endpoint constraint satisfaction, risk evolution cost, and trajectory fingerprint similarity to jointly select candidate situation trajectories. This not only retains trajectories that meet the disposal objectives and have low risk costs, but also avoids outputting a large number of similar evolution paths, making the multi-scenario situation trajectory set more comparable and valuable for decision-making.

[0019] 6. Enhance adaptability to cross-platform backflow and resurgence risks. This invention introduces credible propagation delay, cross-platform backflow intensity, and resurgence triggering degree into the processes of counterfactual endpoint generation, scenario condition encoding, and risk evolution cost calculation. It can identify the resurgence risk caused by re-propagation on other platforms, re-participation by high-influence entities, or re-shifting of topics after public opinion has cooled down, thereby improving the adaptability of the simulation results to the real online public opinion propagation environment.

[0020] 7. Improve the ability to handle action cooling-off phenomena. This invention introduces an action cooling-off factor into the public opinion state sequence and scenario conditions, and determines whether continuous repetitive actions need to be replaced with alternative action combinations based on the action cooling-off factor during the intervention response mapping process. This can avoid the marginal effect of similar intervention actions decreasing or triggering negative feedback after continuous execution, and improve the continuous effectiveness of the intervention path.

[0021] 8. Improve the feasibility of public opinion management actions. This invention decomposes the intervention response quantities at each moment from a set of multi-scenario situational trajectories and maps them to the execution time, target, and intensity of the management actions. This allows the simulation results to not only be expressed as trend curves or risk levels, but also to further form executable management solutions such as clarification releases, authoritative forwarding, de-heating guidance, interception of false information, topic diversion, or platform traffic restriction.

[0022] 9. Enhance closed-loop decision-making capabilities for public opinion situation analysis. This invention sequentially couples state construction, endpoint generation, scenario coding, diffusion bridge analysis, trajectory screening, and result output, making the output of the previous step the data basis for the next step, forming a closed-loop analysis process centered on the goal of public opinion management, thereby improving the technical support capabilities for public opinion risk assessment, scenario comparison, and management decision-making. Attached Figure Description

[0023] Figure 1 This is a flowchart of the public opinion multi-scenario situation inference method based on SDE and DBGM provided in the first embodiment of the present invention.

[0024] Figure 2 This is a block diagram of the public opinion multi-scenario situation inference system based on SDE and DBGM provided in the second embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to specific embodiments. This embodiment provides a method and system for multi-scenario public opinion situation inference based on SDE and DBGM, wherein SDE stands for Stochastic Differential Equation and DBGM stands for Diffusion Bridge Generative Model. This embodiment takes multi-source time-series data of the target public opinion event as input, and forms a continuous processing chain from public opinion data input to multi-scenario situation inference result output through unified time window alignment, public opinion state sequence construction, counterfactual endpoint state generation, scenario condition encoding, diffusion bridge candidate situation trajectory generation, unreachable trajectory elimination, trajectory screening, and intervention response quantity inverse solution.

[0026] In this implementation, the target public opinion event can be a public emergency, a corporate brand event, a product quality event, a public safety event, a social hotspot event, or other events with characteristics of online dissemination. Multi-source time-series data can originate from news platforms, social media platforms, short video platforms, forum platforms, Q&A platforms, instant messaging platforms, fact-checking platforms, and action recording systems. By uniformly processing the above data, this implementation can generate continuous situational trajectories for various scenarios, including natural evolution, intensified dissemination, topic reversal, and intervention suppression, all from the same starting point. Furthermore, it can output the execution time, target, and intensity of actions taken in each scenario.

[0027] In some implementations, to enhance subsequent trajectory extrapolation and result verification capabilities, the system can also synchronously generate a state evidence package for the public opinion state vector within each time window. The state evidence package records the original data source, platform identifier, subject identifier, text fragment fingerprint, propagation link fragment, factual verification basis, and handling action record that formed the public opinion state vector. The state evidence package does not change the dimension of the public opinion state vector, but establishes a one-to-one correspondence with it, used for subsequent unreachable trajectory elimination, trajectory interpretation, handling action tracing, and manual verification. Through the state evidence package, the problem of the model only outputting numerical trajectories without data sources can be avoided, making the situation extrapolation results traceable.

[0028] First Embodiment

[0029] like Figure 1 As shown, the first embodiment provides a method for multi-scenario situational inference of public opinion based on SDE and DBGM, including the following steps.

[0030] S1. Align multi-source time-series data of the target public opinion event according to a unified time window, and extract sentiment intensity, dissemination heat, subject influence, topic shift, factual clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factor to form a public opinion state sequence, and generate the dissemination subject structure and platform diffusion relationship.

[0031] In this step, the first step is to acquire multi-source time-series data of the target public opinion event. This multi-source time-series data includes text content data, comment and repost data, interactive dissemination data, subject relationship data, platform diffusion data, fact-checking data, and action data.

[0032] The data includes the following: text content data: news headlines, news articles, user comments, forwarded text, short video titles, short video subtitles, forum posts, and Q&A content. Comment and forwarding data: number of comments, forwards, likes, favorites, citations, and comment growth rate. Interactive dissemination data: comment relationships, forwarding relationships, citation relationships, and interaction frequency among users. Subject relationship data: relationships between publishing entities, forwarding entities, commenting entities, authoritative entities, media entities, and key opinion entities. Platform diffusion data: the time of appearance, dissemination order, reposting links, and changes in popularity of the same public opinion event across different platforms. Fact verification data: debunking information, authoritative statements, fact-clarification content, verification conclusions, verification release time, and dissemination reach of fact-clarification content. Action data: execution time, target, and intensity of actions such as clarification releases, authoritative forwarding, de-escalation guidance, interception of false information, topic diversion, or platform traffic restriction.

[0033] After acquiring multi-source time-series data, the data is aligned according to a unified time window. The unified time window can be five minutes, ten minutes, thirty minutes, one hour, or other time granularities set according to the spread speed of the target public opinion event. The system maps data generated from different sources, platforms, and entities to the corresponding time windows, enabling the joint calculation of text content, dissemination behavior, entity relationships, fact verification, and handling actions at the same moment.

[0034] After time window alignment, the text content data undergoes deduplication, word segmentation, entity recognition, event element extraction, sentiment recognition, and topic clustering. Comment forwarding and interactive dissemination data are filtered for anomalies, removing bot-generated views, repeated forwarding within a short period, obviously irrelevant comments, and invalid interaction records. Subject relationship data is merged, mapping accounts of the same subject on different platforms to the same subject node or related subject nodes. Platform diffusion data undergoes propagation link identification to determine the order of diffusion between different platforms. Fact verification data and action data are normalized according to publication time, dissemination arrival time, target audience, and feedback effect.

[0035] In some implementations, the system can also generate content fingerprints for similar texts, video titles, comment templates, and forwarding phrases within the same target public opinion event, and determine whether a common-source diffusion relationship exists between different platforms based on these fingerprints. For common-source diffusion relationships, the system not only records the forward diffusion path but also the backflow path from a later-stage platform influencing an earlier-stage platform, marking this backflow path as a backflow foldback edge. The backflow foldback edge is used to express the "spillover-re-propagation-backflow" propagation relationship of the same issue across different platforms. Through the backflow foldback edge, the platform diffusion relationship can not only express unidirectional diffusion but also the risk of cross-platform resurgence.

[0036] After completing the above preprocessing, the following state features are extracted respectively.

[0037] Emotional intensity is determined by the proportion of negative expressions, the degree of emotional polarization, the intensity of emotional words, the frequency of aggressive expressions, and the degree of group emotional fluctuation in the text content data. Dissemination popularity is determined by the number of comments, reposts, exposure, interaction frequency, dissemination speed, and changes in platform popularity. Subject influence is determined by the subject node's fan coverage, reposting ability, centrality index, cross-platform diffusion ability, and historical influence level. Topic deviation is determined by the semantic deviation of the current topic cluster relative to the initial or benchmark topic cluster. Fact clarification is determined by the credibility of the fact-verified content, the level of authority of the source, dissemination reach, citation frequency, and coverage of the main topic cluster. Intervention response volume is determined by the type of action taken, execution time, target audience, action intensity, and dissemination feedback effect of the actions already implemented. Credible dissemination lag is determined by the fact-verified content publication time, the time of first citation by a high-influence subject, the time of first entry into the main topic cluster, and the emotional change response time. Cross-platform return intensity is determined by the frequency of recurrence of the same topic across different platforms, reposting links, inter-platform dissemination direction, and changes in popularity after returning to the original platform. The resurgence trigger factor is determined by the rebound in dissemination intensity from a low point, a renewed increase in topic shift, renewed participation of high-influence entities, and a downward trend in fact-clarification. The action cooling factor is determined by the frequency of action execution, the number of consecutive repetitions, user feedback text, and subsequent changes in dissemination intensity.

[0038] In some implementations, the system can also construct a clarification absorption gating. This gating is used to determine whether factual clarification content has been absorbed by major dissemination entities, major platforms, and major topic clusters. When factual clarification content has been published but has not yet been cited by high-influence entities, or has not yet entered the major topic cluster, or is only spreading on low-profile platforms, the system reduces the immediate impact of factual clarification on the decrease in emotional intensity. When factual clarification content has been cited by major dissemination entities, covers the major topic cluster, and there is an effective dissemination path in the corresponding platform dissemination relationship, the system increases the weight of factual clarification on subsequent changes in emotional intensity and dissemination popularity. Through the clarification absorption gating, the model can avoid directly equating "the clarification content has been published" with "the clarification has taken effect."

[0039] In this step, the public opinion state vector in the public opinion state sequence is represented as: , in, For a moment The public opinion state vector, For a moment The intensity of emotions, For a moment The popularity of the spread For a moment The degree of influence of the main body For a moment The degree of topic shift, For a moment Clarity of facts For a moment Intervention response volume For a moment The reliable propagation time delay, For a moment Cross-platform backflow strength, For a moment The degree of re-ignition triggering, For a moment Action cooling factor, This is a time for observing or predicting public opinion.

[0040] A public opinion state sequence is formed by arranging public opinion state vectors from multiple consecutive moments in chronological order. This public opinion state sequence is used to subsequently determine the current starting state, generate counterfactual ending states, encode scenario conditions, generate candidate situation trajectories, and calculate the risk evolution cost.

[0041] Simultaneously, a communication subject structure is generated based on subject relationship data and interactive communication data. This structure includes subject nodes, communication direction between subjects, communication intensity, and reachable communication paths. Subject nodes can be ordinary users, key opinion subjects, media accounts, authoritative subjects, or organized communication subjects. Communication direction indicates the direction in which public opinion content spreads from one subject to another; communication intensity indicates communication frequency, forwarding weight, or impact strength; and reachable communication paths determine whether changes in a particular issue or sentiment are supported by pathways between communication subjects.

[0042] Platform diffusion relationships are generated based on platform diffusion data. These relationships represent the order of diffusion, reposting relationships, backflow direction, and changes in platform popularity of a target public opinion event across news platforms, social media platforms, short video platforms, forum platforms, and Q&A platforms. In implementations that include backflow loops, platform diffusion relationships also represent the paths through which a cooled-down platform is reactivated by related topics on other platforms.

[0043] When it comes to model training, this step can utilize an emotion recognition model, an issue clustering model, a subject influence assessment model, a fact clarification assessment model, and an action cooling-off assessment model. The emotion recognition model is trained using historical public opinion texts and their emotion tags. Inputs include text content, emotion words, syntactic features, and contextual semantics; outputs are emotion category and emotion intensity. The issue clustering model is trained using text semantic vectors from historical events and manually labeled issue clusters; outputs the main issue clusters and issue offset directions within the current time window. The subject influence assessment model is trained using historical subject dissemination records, subject interaction relationships, and dissemination results; outputs the subject influence. The fact clarification assessment model is trained using historical fact verification texts, authority source levels, and clarification dissemination effects; outputs the fact clarification degree. The action cooling-off assessment model is trained using historical action records, action repetition frequency, user feedback text, and subsequent dissemination heat changes; outputs an action cooling-off factor.

[0044] In some implementations, the clarification absorption gating can be trained using historical fact-clarification events. Training samples include the release time of the fact-clarification, the citation time by major subjects, the coverage time of major issue clusters, the dissemination time across different platforms, the time of decrease in emotional intensity, and the time of decrease in dissemination popularity. During training, the system learns the absorption conditions between the release of fact-clarification content and its actual impact, enabling the model to distinguish between the states of "clarification has been released but has not yet taken effect" and "clarification has been absorbed by the dissemination chain and has begun to take effect."

[0045] In this embodiment, text content data simultaneously participates in the generation of sentiment intensity, topic shift, content fingerprint, and factual clarification; comment and forwarding data, interactive dissemination data, and platform diffusion data jointly participate in the generation of dissemination heat; subject relationship data and platform diffusion data jointly participate in the generation of subject influence and dissemination subject structure; fact verification data, subject relationship data, platform diffusion relationship, and text content data jointly participate in the generation of credible dissemination lag, factual clarification, and clarification absorption gating; handling action data, fact verification data, and dissemination feedback jointly participate in the generation of intervention response volume and action cooling factor; platform diffusion data, content fingerprint, and subject relationship data jointly participate in the generation of cross-platform backflow intensity, backflow foldback edge, and platform diffusion relationship; dissemination heat, topic shift, subject influence, and factual clarification jointly participate in the generation of resurgence trigger degree; state evidence package and corresponding public opinion state vector are generated synchronously and used for subsequent trajectory interpretation and result verification.

[0046] This step transforms scattered, heterogeneous, and asynchronous public opinion data into a computable, continuous sequence of public opinion states, forming a dissemination subject structure and platform diffusion relationships. This allows subsequent projections to move beyond relying solely on a single trending or sentiment indicator, simultaneously considering the coupled effects between sentiment, dissemination, subjects, issues, fact-clarification, intervention response, credible dissemination lag, cross-platform backflow, resurgence triggering, action cooling, clarification absorption, and backflow reversal. This provides a unified data foundation for the subsequent generation of counterfactual endpoint states and candidate situational trajectories.

[0047] S2. Determine the current starting state from the aforementioned public opinion state sequence, and generate multiple counterfactual endpoint states based on the handling objective, propagation constraints, emotional constraints, credible propagation delay, cross-platform backflow intensity, and resurgence trigger degree.

[0048] In this step, the current starting state is determined based on the public opinion state sequence formed in step S1. The current starting state can be selected from the public opinion state vector at the current moment, or it can be obtained by smoothing the public opinion state vectors from multiple recent time windows. The current starting state is used to represent the overall situation of the target public opinion event at the start of the simulation.

[0049] For example, when the target public opinion event is still in the rising stage of dissemination, the current starting state may include high dissemination heat, high emotional intensity, large subject influence, and low factual clarification; when the target public opinion event is in the stage of decline after clarification, the current starting state may include declining dissemination heat, gradually decreasing emotional intensity, rising factual clarification, and a certain degree of resurgence trigger.

[0050] After determining the current starting state, the system generates multiple counterfactual endpoint states based on the handling objective, dissemination constraints, emotional constraints, credible dissemination lag, cross-platform backflow intensity, and resurgence trigger rate. The handling objective represents the desired outcome of public opinion management at the endpoint, which may include reducing dissemination intensity, reducing the intensity of negative emotions, controlling topic deviation, improving factual clarification, reducing the spread impact of high-influence entities, or reducing the risk of resurgence. Dissemination constraints limit the upper bounds of dissemination intensity, dissemination scope, cross-platform spread, or high-influence entity dissemination at the endpoint. Emotional constraints limit the upper bounds of emotional intensity, emotional polarization, or the proportion of negative expression at the endpoint. Credible dissemination lag limits the reasonable delay between the release of factual clarification information and its absorption by the main dissemination entities. Cross-platform backflow intensity corrects for the impact of backflow dissemination across different platforms on the endpoint dissemination intensity. The resurgence trigger rate retains a corresponding risk level for the endpoint state when public opinion appears to have cooled down but there is a risk of resurgence.

[0051] In some implementations, a layered approach to endpoint credibility can be introduced when generating counterfactual endpoint states. This layered approach categorizes candidate endpoint states into immediately reachable endpoints, delayed reachable endpoints, and risk-reserved endpoints. Immediately reachable endpoints represent states that can be directly reached at the endpoint given the current propagation structure, platform diffusion relationships, and the intensity of the action. Delayed reachable endpoints represent states that require fact-clarification propagation absorption or cross-platform diffusion attenuation to reach. Risk-reserved endpoints represent states where, although the propagation heat and emotional intensity have decreased, the resurgence trigger level or cross-platform backflow intensity still needs to be retained. This layered approach avoids directly setting all action targets to the same type of endpoint state, improving the authenticity and executability of counterfactual endpoint states.

[0052] In this step, the counterfactual endpoint state is generated as follows: , in, For the first End point under similar scenario conditions The counterfactual endpoint state, This is the candidate endpoint state. For the first Under similar scenario conditions, the current starting state and disposal targets The target state obtained by mapping, This is the current starting state. For the first Targets for handling similar scenarios Candidate endpoint status Penalties for violating communication and emotional constraints. Candidate endpoint status Penalty for violating the trusted propagation time delay boundary. Candidate endpoint status Penalties for violating cross-platform backflow boundaries Candidate endpoint status The penalty for underestimating the amount of reignition retention. To determine the weights of the propagation constraints and emotional constraints as punishment items, The weight of the reliable propagation time delay boundary penalty term, The weight of the cross-platform backflow boundary penalty term, The weight of the penalty term for the amount of re-ignition retention. Assign scenario numbers, The endpoint.

[0053] The aforementioned counterfactual endpoint states are not simple extrapolations of historical data, but rather constrained endpoint states generated based on the current starting state and according to different scenario conditions and action objectives. For the same current starting state, counterfactual endpoint states can be generated under natural evolution scenario, reinforced propagation scenario, issue reversal scenario, and intervention and suppression scenario, respectively.

[0054] In a natural evolutionary scenario, the counterfactual endpoint is primarily generated based on current propagation trends and the natural decay patterns of similar historical events, with intervention response levels remaining low. This endpoint is used to extrapolate the state of public opinion that might be reached without significantly increasing intervention measures.

[0055] In a heightened dissemination scenario, the counterfactual endpoint reflects the risk boundary of further escalation in dissemination intensity, subject influence, or emotional intensity. This endpoint is used to predict high-risk evolutionary directions when key subjects continue to voice their opinions, platform recommendations are enhanced, or cross-platform backflow is strengthened.

[0056] In a topic reversal scenario, the counterfactual endpoint reflects a significant increase in topic deviation, a marked change in the direction of topic migration, and a failure to promptly cover new topics with factual clarification. This endpoint is used to predict the evolution of public opinion from discussions of the original event to disputes over responsibility, group conflicts, regulatory questions, or other derivative topics.

[0057] In intervention and suppression scenarios, the counterfactual endpoint is characterized by a decrease in dissemination intensity, a decrease in emotional intensity, an increase in factual clarification, a decrease in the resurgence trigger, and an action cooling factor within an acceptable range. This endpoint is used to deduce feasible intervention paths that can meet the intervention objectives.

[0058] In some implementations, the system can also perform evidence consistency verification on the counterfactual endpoint state based on the state evidence package. When the candidate endpoint state requires a significantly increased degree of factual clarification, but the state evidence package lacks authoritative sources of clarification, records cited by major entities, or records covered by major issue clusters, the system increases the penalty for credible propagation time lag boundaries. When the candidate endpoint state requires a significantly decreased degree of propagation intensity, but the state evidence package still contains records of continued propagation by high-influence entities or a continuously strengthening backflow foldback edge, the system increases the penalty for cross-platform backflow boundaries. When the candidate endpoint state requires a low resurgence trigger value, but the state evidence package contains records of re-propagation of old content, disclosure of new evidence, or re-participation by high-influence entities, the system increases the penalty for resurgence retention. Through evidence consistency verification, the counterfactual endpoint state can not only meet numerical constraints but also be supported by data sources.

[0059] When training the model, this step can train the endpoint state mapping model. The endpoint state mapping model uses historical starting states, historical handling objectives, historical propagation constraints, historical sentiment constraints, historical credible propagation lags, historical cross-platform backflow intensity, historical resurgence triggering degree, and historical endpoint states from historical public opinion events as training samples. During training, the historical starting states and historical handling objectives are input into the model, and the predicted endpoint states are output. The predicted endpoint states are compared with the historical true endpoint states or artificially constructed target endpoint states to calculate the endpoint state error. Simultaneously, constraint penalty errors are calculated based on propagation constraints, sentiment constraints, credible propagation lag boundaries, cross-platform backflow boundaries, and resurgence retention. By jointly reducing the endpoint state error and constraint penalty error, model parameters capable of generating counterfactual endpoint states are trained.

[0060] In some implementations, the endpoint state mapping model can also incorporate hierarchical training based on endpoint credibility. Training samples are categorized into immediately accessible samples, delayed accessible samples, and risk-reserved samples according to how endpoint states are formed in historical events, enabling the model to learn the formation conditions for different endpoint states. Through this training, the model can automatically determine whether an immediately accessible endpoint, delayed accessible endpoint, or risk-reserved endpoint should be generated when faced with a target, thereby reducing unrealistic endpoint settings.

[0061] In this embodiment, the current starting state originates from the public opinion state sequence formed in step S1; the target value of the handling target, propagation constraint, and emotion constraint is applied to the candidate endpoint state; the credible propagation delay is used to limit the reasonable increase in fact clarification at the endpoint; the cross-platform backflow intensity and backflow foldback edge are used to limit the global decrease in propagation heat at the endpoint; the re-ignition trigger degree is used to retain the re-ignition risk amount for the propagation heat and topic offset in the endpoint state; the state evidence package is used to verify whether the candidate endpoint state has data support; the generated counterfactual endpoint state serves as the endpoint traction target for the candidate situation trajectory generated by DBGM in step S4.

[0062] This step transforms the goals of public opinion management and dissemination constraints into calculable counterfactual endpoint states. This allows subsequent simulations to not only predict how public opinion might develop, but also how it might develop when the management goals and constraints are met. Furthermore, it reduces the likelihood of overly optimistic endpoint states due to factors such as ignoring factual clarification delays, cross-platform backflow, the risk of resurgence, and insufficient evidence.

[0063] S3. Encode the scenario conditions based on the aforementioned dissemination subject structure, platform diffusion relationships, topic migration direction, cross-platform backflow intensity, and intervention action intensity.

[0064] In this step, based on the propagation subject structure, platform diffusion relationships, cross-platform backflow intensity generated in step S1, and the counterfactual endpoint state determined in step S2, scenario conditions are encoded in conjunction with the topic migration direction and intervention action intensity. Scenario conditions are used to control DBGM to generate different candidate trend trajectories under different public opinion evolution assumptions.

[0065] The scenario conditions include subject amplification factor, platform diffusion factor, emotional disturbance factor, issue migration factor, intervention action factor, backflow trigger factor, clarification delay factor, and action cooling factor.

[0066] The subject amplification factor, derived from the subject's influence and the structure of the dissemination subject, represents the amplifying effect of high-influence subjects, key opinion subjects, media subjects, or organized dissemination subjects on public opinion dissemination. The platform diffusion factor, derived from dissemination heat and platform diffusion relationships, represents the speed, scope, and return direction of the target public opinion event's spread across different platforms. The emotional disturbance factor, derived from emotional intensity, represents the disruptive effect of negative emotions, polarized emotions, or group emotional fluctuations on the public opinion situation. The topic migration factor, derived from topic deviation and topic migration direction, represents the trend of public opinion topics shifting from the original event towards accountability, group conflict, product safety, regulatory questioning, or other directions. The intervention action factor, derived from the intervention response volume and factual clarification degree, represents the impact of actions such as clarification releases, authoritative reposts, cooling-down guidance, interception of false information, topic diversion, or platform traffic restriction on future trends. The return trigger factor, derived from cross-platform return intensity, represents the possibility of a topic that has cooled down on one platform returning after spreading again on another platform. The clarification lag factor, derived from the credible propagation lag, indicates the degree to which factual clarification has been published but has not yet been absorbed by the main propagation path. The action cooling factor, derived from the frequency of action execution and propagation feedback, indicates the degree to which the marginal effect decreases or negative feedback is triggered after consecutive execution of similar intervention actions.

[0067] When generating scenario conditions, the system first identifies high-influence subjects, bridging subjects, and marginal diffusion subjects based on the structure of the dissemination subjects, and determines the achievable dissemination paths between these subjects. Then, based on platform diffusion relationships, it determines the diffusion order, dissemination intensity, backflow direction, and backflow foldback edges of public opinion content across different platforms. Next, it determines the topic migration direction based on changes in topic clusters within a continuous time window, and determines the intensity of intervention actions based on action data and changes in fact-clarification levels. Finally, it combines subject amplification factors, platform diffusion factors, emotional disturbance factors, topic migration factors, intervention action factors, backflow triggering factors, clarification delay factors, and action cooling-off factors to form the scenario conditions.

[0068] In some implementations, the system can also construct a scenario conflict matrix. This matrix records the exclusion and coexistence relationships between different scenario condition factors. For example, there is an exclusion relationship between strong intervention action factors and completely natural evolution factors; high backflow triggering factors can coexist with intervention suppression scenarios to simulate cross-platform resurgence after intervention; and there is an exclusion relationship between high clarification lag factors and rapid fact clarification. When generating scenario conditions, the system corrects contradictory scenario factors based on the scenario conflict matrix to avoid contradictory combinations such as setting both strong intervention suppression and no intervention action in the same scenario condition, or setting both clarification not yet spread and immediate emotional decline.

[0069] In the natural evolution scenario, the subject amplification factor, platform diffusion factor, emotional disturbance factor, and issue migration factor in the scenario conditions extend according to the current observed trends, the intervention action factor remains at a low level, and the backflow trigger factor and action cooling factor are retained according to the current state. This scenario condition is used to extrapolate the natural change path of public opinion without significant intervention.

[0070] In the enhanced dissemination scenario, the subject amplification factor, platform diffusion factor, emotional disturbance factor, and backflow trigger factor are set to relatively high levels to simulate situations where key subjects speak out, platform recommendations are enhanced, cross-platform reposting increases, or the spread of negative emotions accelerates. This scenario is used to extrapolate high-risk dissemination paths.

[0071] In the topic reversal scenario, the topic migration factor and sentiment disturbance factor are set to relatively high levels, while the factual clarification level covers a lower degree of the new topic. This scenario is used to simulate the shift of public opinion from the original event to a derivative controversy. This scenario condition is used to extrapolate the impact of topic shift on the popularity and intensity of sentiment.

[0072] In the intervention suppression scenario, the intervention action factor and factual clarification level are set to relatively high levels, while the action cooling factor and clarification lag factor are also considered to simulate a situation where the intervention action has a suppressive effect on the spread of information and emotional intensity, but there is a spread delay and action cooling. This scenario condition is used to deduce executable intervention paths.

[0073] In some implementations, the system can also generate a scenario evidence chain for each scenario condition. The scenario evidence chain includes key nodes supporting the scenario condition, platform diffusion paths, issue migration segments, action records, and fact-checking records. The scenario evidence chain is associated with the state evidence package and used to explain why a candidate situation trajectory belongs to a natural evolution, reinforced propagation, issue reversal, or intervention suppression scenario when the results are output. The scenario evidence chain improves the interpretability of the scenario condition coding results.

[0074] When training the model, this step can train a scenario coding model. The scenario coding model uses the dissemination subject structure, platform diffusion relationships, topic migration direction, cross-platform backflow intensity, historical response intensity, and historical evolution category in historical public opinion events as training samples. During training, the historical dissemination subject structure, historical platform diffusion relationships, and historical topic migration direction are input into the model, outputting a scenario condition vector. The classification error is calculated based on the natural evolution, intensified dissemination, topic reversal, or intervention suppression category ultimately exhibited by the historical event. Simultaneously, the scenario factor error is calculated based on the actual changes in subject dissemination, platform diffusion, and intervention response in the historical trajectory. By reducing classification error and scenario factor error, the scenario coding model can generate scenario conditions that can distinguish different public opinion evolution hypotheses.

[0075] In some implementations, the scenario coding model can also be trained using a scenario conflict matrix. During training, historical real-world scenario conditions and artificially constructed contradictory scenario conditions are distinguished, enabling the model to learn to exclude scenario factor combinations that do not conform to propagation logic. In this way, the scenario coding model can automatically reduce the weight of conflict factors when generating scenario conditions, improving the stability of subsequent trajectory generation.

[0076] In this embodiment, the subject amplification factor originates from the subject's influence and the structure of the propagation subject; the platform diffusion factor originates from the propagation heat, platform diffusion relationship, and backflow foldback edge; the emotional disturbance factor originates from the emotional intensity; the issue migration factor originates from the issue offset and issue migration direction; the intervention action factor originates from the intervention response volume and fact clarification degree; the backflow trigger factor originates from the cross-platform backflow intensity; the clarification lag factor originates from the credible propagation lag and clarification absorption gating; the action cooling factor originates from the frequency of action execution and propagation feedback; the scenario conflict matrix is ​​used to correct contradictory scenario factors; and the scenario evidence chain is used to record the basis for the formation of scenario conditions. The above scenario conditions correspond one-to-one with the counterfactual endpoint state generated in step S2 and are jointly input into the DBGM in step S4.

[0077] This step transforms the dissemination structure, platform diffusion relationships, topic migration direction, cross-platform backflow intensity, and intervention action intensity into scenario conditions usable by DBGM. This ensures that candidate situational trajectories are not generated randomly and unconditionally, but rather reflect the combined influence of main dissemination, platform diffusion, emotional disturbance, topic migration, intervention actions, cross-platform backflow, delayed fact clarification, action cooling, scenario conflict constraints, and scenario evidence chains on the direction of public opinion evolution.

[0078] S4. Couple the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM. The DBGM then generates candidate situation trajectories using the SDE (Situational Dependency Analysis) which includes the ending bridging term.

[0079] In this step, the current starting state and the corresponding counterfactual ending state obtained in step S2, along with the scenario conditions obtained in step S3, are coupled and input into DBGM. DBGM describes the continuous random evolution of public opinion state over time through SDE, and uses the endpoint bridging term to make the generated candidate situation trajectory start from the current starting state and gradually approach the counterfactual ending state under the corresponding scenario.

[0080] The DBGM generates candidate situational trajectories according to the following SDE: , in, For the first Time under similar scenario conditions The increase in public opinion status For the first Time under similar scenario conditions The public opinion state vector, For model parameters The following corresponding scenario conditions The public opinion drift function, These are the model parameters for DBGM. For the first Scenario conditions, For the first Time under similar scenario conditions The bridging traction coefficient, For the first End point under similar scenario conditions The counterfactual endpoint state, For time infinitesimal elements, For model parameters The following corresponding scenario conditions The diffusion intensity function, For a moment Brownian motion increment, Assign scenario numbers, The endpoint.

[0081] In the aforementioned SDE, the sentiment drift function Used to express the main trends in the changing state of public opinion under the current circumstances and conditions; bridging and traction items Used to express the pulling force that causes the current state to move closer to the counterfactual endpoint state; diffusion intensity function and Brownian motion increment Used to express random disturbances in the spread of public opinion, such as sudden comments, statements from key stakeholders, changes in platform recommendations, delays in fact-checking, or fluctuations in status caused by the disclosure of new information.

[0082] In its specific generation process, DBGM uses the current starting state as the initial state, the corresponding counterfactual ending state as the endpoint target, and the scenario conditions as input conditions, gradually generating candidate situation trajectories according to a preset time step. Each candidate situation trajectory consists of multiple consecutive time-stamped public opinion state vectors, and each public opinion state vector includes sentiment intensity, dissemination heat, subject influence, topic shift, factual clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factor.

[0083] For the same current starting state and the same counterfactual ending state, DBGM can perform multiple samplings to generate multiple candidate situational trajectories. For different scenario conditions, DBGM generates candidate situational trajectories under natural evolution, reinforced propagation, issue reversal, and intervention / suppression scenarios, respectively. Thus, the system can form multiple possible continuous evolutionary paths under the same current public opinion state.

[0084] In some implementations, DBGM can also introduce local bridging anchors during the candidate situation trajectory generation process. Local bridging anchors are intermediate constraint states located between the current starting state and the counterfactual ending state, used to define the state areas that the candidate situation trajectory must pass through or approach at key stages. Local bridging anchors can include peak dissemination heat anchors, fact clarification effectiveness anchors, topic shift inflection point anchors, cross-platform backflow anchors, and risk decline anchors. Peak dissemination heat anchors indicate the time and state at which public opinion dissemination may reach its peak; fact clarification effectiveness anchors indicate the state where fact clarification content begins to influence sentiment intensity after being absorbed by the main dissemination path; topic shift inflection point anchors indicate the state where the topic shifts from the original event to a derivative topic; cross-platform backflow anchors indicate the state where the same source topic on other platforms re-influences the original platform; and risk decline anchors indicate the state where dissemination heat and sentiment intensity begin to steadily decline.

[0085] By using local bridging anchors, the candidate situational trajectory generated by DBGM no longer simply moves directly from the current starting state to the counterfactual endpoint state, but is constrained by multiple key states in the intermediate stages. This approach avoids the model directly generating a smooth descent trajectory while ignoring intermediate risk peaks, issue reversals, or resurgence processes.

[0086] In some implementations, DBGM can also adjust the change in intervention response volume based on the action cooling factor. When similar actions occur consecutively within adjacent time windows and their dissemination intensity does not decrease significantly, the action cooling factor increases. In subsequent time steps, DBGM reduces the gain of the intervention response volume corresponding to the same action, or shifts to alternative action paths such as generating topic diversion, authoritative forwarding, and enhanced coverage of fact clarification. In this way, the situation where repeated clarifications become less effective or excessive intervention triggers negative feedback in real-world public opinion management can be simulated.

[0087] In some implementations, DBGM can also use state evidence packages as trajectory sampling constraints. When a key inflection point in a candidate situation trajectory lacks state evidence package support, DBGM reduces the weight of such trajectories in subsequent sampling; when a peak in popularity, issue shift inflection point, or fact clarification effectiveness node in a candidate situation trajectory has corresponding state evidence package support, DBGM increases the sampling stability of such trajectories under similar scenarios. This approach allows for a closer correspondence between the generated trajectory and the actual data source.

[0088] When training the model, DBGM's training samples consist of multiple historical public opinion events. Each historical public opinion event includes historical multi-source time-series data, historical public opinion state sequences, historical dissemination subject structure, historical platform diffusion relationships, historical topic migration directions, historical intervention action data, and historical event convergence results. During training, an early time segment is selected from the historical public opinion state sequence as the historical starting state, and the event convergence stage, the stage peak stage, or the state constructed by the intervention target is selected as the historical ending state. The historical dissemination subject structure, historical platform diffusion relationships, historical topic migration directions, historical cross-platform backflow intensity, and historical intervention action intensity are encoded as historical scenario conditions. DBGM receives the historical starting state, historical ending state, and historical scenario conditions to generate historical candidate situation trajectories.

[0089] During training, historical candidate situation trajectories are compared with real historical public opinion state sequences to calculate state reconstruction error; the endpoint states of historical candidate situation trajectories are compared with historical endpoint states to calculate endpoint constraint error; the direction of change of adjacent states in historical candidate situation trajectories is compared with the direction of change of adjacent states in real historical public opinion state sequences to calculate drift consistency error; and changes in subject propagation, topic migration direction, factual clarification, and intervention response in historical candidate situation trajectories are compared with historical causal dependencies to calculate causal reachability error. By jointly reducing state reconstruction error, endpoint constraint error, drift consistency error, and causal reachability error, the DBGM model parameters are updated. .

[0090] In some implementations, DBGM training also includes local bridging anchor point training. Specifically, peak moments of dissemination heat, inflection points of topic shift, moments when factual clarification takes effect, moments when cross-platform backflow occurs, and moments when risk declines are extracted from historical public opinion events and used as local bridging anchor point samples. During training, DBGM learns the probability and order in which candidate situational trajectories pass through the corresponding local bridging anchor points under different scenario conditions. Through local bridging anchor point training, the model's ability to express risk changes in intermediate stages can be improved.

[0091] In some implementations, DBGM training also includes training with a lag in fact-clarification. Specifically, the temporal relationship between the release time of fact-clarification content, the time of citation by major subjects, the absorption time of the topic cluster, and the time of decline in sentiment intensity is extracted from historical public opinion events. This temporal relationship is used as a training constraint, enabling DBGM to learn the delay pattern between the release and effectiveness of fact-clarification information. This training can reduce the problem of the model prematurely generating sentiment decline trajectories in intervention and suppression scenarios.

[0092] In some implementations, DBGM training also includes cross-platform backflow training. Specifically, it identifies the propagation order, backflow path, and backflow foldback edges of the same topic cluster across different platforms from historical public opinion events, and uses samples where the popularity of one platform declines and then rises on another platform, re-influencing the original platform, as backflow training samples. Through this training, DBGM can generate candidate trend trajectories of a local cooling down followed by a resurgence due to external platform resurgence.

[0093] In some implementations, DBGM training also includes action cooldown training. Specifically, it extracts the number of consecutive executions of similar actions, execution intervals, user feedback text, and subsequent dissemination trends from historical action records, using samples where the effectiveness of consecutive actions decreased or triggered negative feedback as action cooldown training samples. Through this training, DBGM can avoid indefinitely improving the effect estimation of similar intervention actions in subsequent simulations.

[0094] In this embodiment, the current starting point state defines the starting position of the candidate situation trajectory; the counterfactual endpoint state provides the endpoint traction direction of the candidate situation trajectory; the scenario conditions adjust the sentiment drift function and diffusion intensity function; the local bridging anchor point constrains the intermediate key stages of the candidate situation trajectory; the action cooling factor adjusts the way the intervention response changes; the credible propagation delay and clarification absorption gating affect the duration of the effect of factual clarification on sentiment intensity; the cross-platform backflow intensity and backflow foldback edge affect the propagation heat and re-ignition trigger degree; the state evidence package affects the sampling weight of the key inflection points of the candidate situation trajectory; and the generated candidate situation trajectory is used as the object for unreachable trajectory elimination, endpoint constraint satisfaction calculation, risk evolution cost calculation, and trajectory fingerprint similarity screening in step S5.

[0095] This step generates a continuous candidate situation trajectory connecting the current starting state and the counterfactual ending state. This allows the generated trajectory to be subject to the joint constraints of the endpoint target, scenario conditions, propagation mechanism, local bridging anchors, and supporting evidence, while preserving the random perturbation characteristics of public opinion propagation. This improves the continuity, target constraint, stage completeness, and scenario coverage of multi-scenario public opinion simulation.

[0096] S5. Based on the directed dependencies between the communication subject structure, topic migration direction, factual clarification degree, and intervention response quantity, unreachable trajectories are eliminated, and a multi-scenario situational trajectory set is obtained by filtering according to the endpoint constraint satisfaction degree, risk evolution cost, and trajectory fingerprint similarity.

[0097] In this step, the candidate situation trajectories generated in step S4 are filtered to obtain a multi-scenario situation trajectory set. The filtering process includes unreachable trajectory elimination, endpoint constraint satisfaction filtering, risk evolution cost filtering, and trajectory fingerprint similarity filtering.

[0098] First, unreachable trajectories are eliminated based on the directed dependencies between the dissemination subject structure, the direction of topic migration, the degree of factual clarification, and the amount of intervention response. Directed dependencies are used to express the causal constraints between changes in public opinion status. For example, factual clarification usually needs to be published, disseminated, cited by major subjects, or absorbed by major topic clusters before it has an effect on reducing the intensity of emotions; topic shifts usually need to be supported by the dissemination subject path or the semantic evolution path of the text; and a decrease in dissemination heat usually needs to be matched with changes in platform diffusion relationships, natural heat decay, or records of intervention actions.

[0099] Specifically, when the decrease in emotional intensity in a candidate trajectory precedes the increase in factual clarification and is not supported by changes in intervention response, the candidate trajectory is determined to be an unreachable trajectory. This judgment is used to exclude unreasonable trajectories where factual clarification has not yet taken effect and no intervention action has been taken, but negative emotions have decreased significantly in advance.

[0100] When the topic shift changes in a candidate trajectory but there is no reachable propagation path between the corresponding entities in the propagation structure, the candidate trajectory is determined to be an unreachable trajectory. This judgment is used to exclude unreasonable trajectories that show obvious topic shifts but lack propagation by key entities, platform reposting links, or textual semantic continuity.

[0101] When the decrease in dissemination heat in a candidate trajectory is inconsistent with the relationship between platform diffusion and the recorded response actions, the candidate trajectory is determined to be an unreachable trajectory. This judgment is used to exclude unreasonable trajectories where platform diffusion is still increasing, cross-platform backflow is still occurring, and response actions have not produced corresponding feedback, but the dissemination heat has decreased abnormally.

[0102] In some implementations, the system can also generate trajectory interpretation breakpoints based on the state evidence package. A trajectory interpretation breakpoint is a point in time where there is a significant change in the spread intensity, sentiment intensity, topic shift, factual clarification, or intervention response volume within the candidate situation trajectory. For each trajectory interpretation breakpoint, the system extracts the corresponding subject's voice record, platform dissemination record, fact-checking record, topic shift text fragment, or action record from the state evidence package and determines whether the breakpoint is supported by evidence. When a key breakpoint lacks supporting evidence and cannot be explained by adjacent state changes, the system lowers the causal accessibility score of that candidate situation trajectory. Through trajectory interpretation breakpoints, key changes in the candidate situation trajectory can be reviewed point by point.

[0103] After eliminating unreachable trajectories, the endpoint constraint satisfaction rate is calculated for each candidate situation trajectory. The endpoint constraint satisfaction rate is used to determine whether the candidate situation trajectory is close to the corresponding counterfactual endpoint state at the endpoint time, and whether it meets the requirements for propagation constraints, sentiment constraints, credible propagation delay boundaries, cross-platform backflow boundaries, and resurgence retention. Candidate situation trajectories whose endpoint state deviates too much from the counterfactual endpoint state, or that violate propagation constraints, sentiment constraints, and backflow boundaries, are eliminated or have their priority reduced.

[0104] Subsequently, the risk evolution cost is calculated. The risk evolution cost is determined as follows: , in, For the first The risk evolution cost of candidate situation trajectories under similar scenarios For the first Time under similar scenario conditions The intensity of emotions, For the first Time under similar scenario conditions The popularity of the spread For the first Time under similar scenario conditions The degree of topic shift, For the first Time under similar scenario conditions Clarity of facts For the first Time under similar scenario conditions Intervention response volume For the first Endpoint constraint violation under similar scenario conditions For the first End point under similar scenario conditions The degree of re-ignition triggering, For the first End point under similar scenario conditions Action cooling factor, The weighting coefficient for the intensity of emotion. As a weighting coefficient for the spread of popularity, The weighting coefficient for the topic offset. This is the weighting coefficient for the degree of clarification of facts. The weighting coefficients for the intervention response. The weighting coefficient for the amount of violation of the endpoint constraint. This is the weighting coefficient for the re-ignition triggering degree. This is the weighting coefficient for the action cooling factor. For the purpose of public opinion observation or prediction, The final moment, Assign a scenario number.

[0105] Among the aforementioned costs of risk evolution, increased emotional intensity, dissemination intensity, and topic deviation generally indicate an increase in public opinion risk, and are therefore included in the cost of risk evolution with a positive weight; increased factual clarification generally indicates a decrease in public opinion risk, and are therefore included in the cost of risk evolution with a negative weight; increased effective intervention response generally indicates that the handling actions contribute to the reduction of risk, and are therefore included in the cost of risk evolution with a negative weight; increased endpoint constraint violation, resurgence triggering degree, and action cooling factor indicate a decrease in endpoint deviation, subsequent resurgence risk, and the sustained effect of handling actions, respectively, and are therefore included in the cost of risk evolution with a positive weight.

[0106] In some implementations, the system can also generate secondary resurgence check results in addition to the risk evolution cost. The secondary resurgence check results are used to determine whether, after the candidate situation trajectory meets the endpoint constraints, there are still situations such as enhanced backflow and foldback edges, re-participation of high-impact entities, re-propagation of old content, or a renewed increase in topic shift. If the candidate situation trajectory meets the constraints at the endpoint, but the secondary resurgence check results show a significant risk of resurgence, the system retains the trajectory as a risk warning trajectory, but reduces its priority as the preferred handling path.

[0107] After calculating the cost of risk evolution, a trajectory fingerprint is further generated. The trajectory fingerprint is generated from the peak time of dissemination heat, the sequence of inflection points in emotional intensity, the direction of topic shift, the effective time of fact clarification, and the pattern of change in intervention response volume in the candidate situation trajectory. The peak time of dissemination heat indicates the time when public opinion heat is highest; the sequence of inflection points in emotional intensity indicates the order in which negative emotions rise, stabilize, or decline; the direction of topic shift indicates the direction of the shift in discussion topics; the effective time of fact clarification indicates when the content of fact clarification begins to affect emotional intensity or dissemination heat; and the pattern of change in intervention response volume indicates the pattern of change in the intensity of intervention actions over time.

[0108] When the trajectory fingerprint similarity of multiple candidate situation trajectories exceeds a preset threshold, the candidate situation trajectories with higher endpoint constraint satisfaction, lower risk evolution cost, and stronger causal reachability are retained. This method forms a multi-scenario situation trajectory set with differentiated evolution paths, avoiding the output of a large number of highly similar trajectories.

[0109] In some implementations, the system can also generate representative paths for trajectory clusters. A representative path for a trajectory cluster is a representative path selected from a set of candidate situational trajectories with similar trajectory fingerprints. The system selects representative paths for trajectory clusters based on endpoint constraint satisfaction, risk evolution cost, causal reachability score, trajectory interpretation breakpoint completeness, and secondary re-ignition check results. By using representative paths for trajectory clusters, the system can reduce result redundancy while preserving diversity.

[0110] When it comes to model training, this step can train a causal accessibility discrimination model and a trajectory fingerprint similarity evaluation model. The causal accessibility discrimination model uses historical public opinion trajectories, historical dissemination subject structures, historical topic migration directions, changes in historical fact clarification, and changes in historical intervention response volume as training samples. It inputs key change segments from candidate trajectories and outputs a judgment result on whether the trajectory is accessible. During training, supervised training is performed using historically accessible trajectories and artificially constructed inaccessible trajectories, enabling the model to learn the directed dependencies between fact clarification, intervention response, topic migration, and subject dissemination paths. The trajectory fingerprint similarity evaluation model uses historical trajectory fingerprint pairs and their similarity labels as training samples, learning the degree of difference between different trajectories in peak time, inflection point order, topic shift direction, clarification effectiveness time, and intervention response pattern.

[0111] In some implementations, the system can also train a trajectory interpretation breakpoint assessment model. This model uses real key breakpoints from historical public opinion events, corresponding state evidence packages, and manually annotated interpretation results as training samples to learn whether a particular trajectory breakpoint can be explained by subject propagation, platform dissemination, factual clarification, topic shifting, or action. This model can improve the accuracy of unreachable trajectory elimination and trajectory interpretation.

[0112] In this embodiment, the elimination of unreachable trajectories uses the propagation subject structure, platform diffusion relationship, state evidence package, and backflow foldback edge generated in step S1, the topic migration direction determined in step S3, and the candidate situation trajectory generated in step S4; the endpoint constraint satisfaction uses the counterfactual endpoint state generated in step S2 and the endpoint of the candidate situation trajectory generated in step S4; the risk evolution cost is calculated using the public opinion state vector at each moment in the candidate situation trajectory, combined with the endpoint constraint violation amount, re-ignition trigger degree, and action cooling factor; the trajectory fingerprint similarity is obtained using the propagation heat peak, emotion intensity inflection point, topic offset direction, fact clarification effective time, and intervention response amount change pattern in the candidate situation trajectory; the trajectory interpretation breakpoint is supported by the state evidence package; and the trajectory cluster representative path is selected from the set of similar trajectories.

[0113] This step selects trajectories from a large pool of candidate situational trajectories that conform to propagation logic, meet endpoint constraints, have assessable risk costs, and exhibit significant path differences. This avoids the occurrence of inferences from random trajectories generated by DBGM that do not conform to the laws of public opinion dissemination, fail to meet response objectives, or have low response value, thereby improving the rationality, diversity, interpretability, and resilience assessment capabilities of multi-scenario situational trajectory sets.

[0114] S6. Extract the intervention response quantities at each moment from the multi-scenario situation trajectory set, map them to the execution time, target, and intensity of the action, and output the multi-scenario situation deduction results.

[0115] In this step, the system outputs results based on the multi-scenario situation trajectory set obtained in step S5. For each scenario, the system outputs the corresponding situation trajectory, risk evolution cost, critical risk moments, main driving factors, and response action information.

[0116] For natural evolution scenarios, the output shows the natural trends in the spread of the target public opinion event, emotional intensity, topic shift, factual clarification, cross-platform backflow intensity, and resurgence triggering without significantly increasing intervention actions.

[0117] For scenarios involving enhanced dissemination, the output identifies the time window, dissemination subject, dissemination platform, and return path when the amplification factor of the subject, the diffusion factor of the platform, the emotional disturbance factor, or the return trigger factor increases.

[0118] In the case of a topic reversal, output the direction of the topic shift, the main dissemination path that led to the topic reversal, the impact of the topic shift on the dissemination popularity and emotional intensity, and the coverage of the fact-clarification content on the new topic.

[0119] For intervention and suppression scenarios, output candidate situation trajectories that can meet the disposal objectives and have low risk evolution costs, and extract the intervention response quantities at each time step from the trajectory.

[0120] The inverse solution of the intervention response quantity at each moment includes: extracting the sequence of changes in intervention response quantity from the situation trajectory that satisfies the endpoint constraint and has the lowest risk evolution cost or is lower than a preset threshold; determining the intensity of handling at each moment based on the sequence of changes in intervention response quantity; and combining the emotional intensity, dissemination heat, subject influence, topic deviation, and factual clarification at the corresponding moment to map the intensity of handling into clarification release, authoritative forwarding, heat reduction guidance, interception of false information, topic diversion, or platform traffic restriction actions.

[0121] Specifically, when the clarification of facts is low and the intensity of emotions continues to rise at a certain moment, the system maps the intervention response at that moment to a clarification release or authoritative forwarding action. The target is the platform, issue cluster, or high-influence subject where negative emotions are concentrated. The effectiveness of the clarification release action is jointly determined by the gap in clarification of facts, the magnitude of the increase in emotion intensity, and the credible dissemination lag; the effectiveness of the authoritative forwarding action is jointly determined by the subject's influence, the credibility of the clarification content, and the reachability path in the dissemination subject's structure.

[0122] When the spread of information rapidly increases at a certain moment and the influence is concentrated on a few high-influence entities, the system maps the intervention response at that moment into a cooling-down guidance or platform flow-limiting action. The targets are platforms with rapidly increasing spread information, key spread nodes, or high-frequency diffusion links. The strength of the cooling-down guidance action is jointly determined by the spread information growth rate, the platform diffusion factor, and the entity amplification factor; the strength of the platform flow-limiting action is jointly determined by the peak spread information, the cross-platform backflow intensity, and the platform diffusion relationship.

[0123] When the topic shift continues to increase at a certain moment and diffuse topics that deviate significantly from the original event emerge, the system maps the intervention response at that moment into a topic diversion action. The target of this action is the topic cluster, dissemination subject, or platform entry point where the topic shift has occurred. The strength of the topic diversion action is determined by the topic shift, the direction of topic migration, and the emotional disturbance factor.

[0124] When fact-checking data indicates the spread of misinformation, and the corresponding issue cluster continues to proliferate within the dissemination structure, the system maps the intervention response at that moment to a misinformation interception action. The action targets the false content, the dissemination node, or the dissemination chain. The effectiveness of the misinformation interception action is determined by the scope of the misinformation's spread, its popularity, the degree of factual clarification, and the impact on the disseminating entity.

[0125] When the action cooling factor exceeds a preset threshold, the system replaces repeatedly performed actions with alternative action combinations that achieve the same endpoint constraint. For example, when the action cooling factor for continuous clarification releases is high, the system can replace repeated clarification releases with a combination of authoritative forwarding, topic diversion, and visual release of fact-verification content; when the action cooling factor for continuous cooling-down guidance is high, the system can replace repeated cooling-down guidance with a combination of targeted communication with the main body, fact-clarification on key platforms, and topic diversion. The selection criteria for alternative action combinations are: the candidate situation trajectory generated by the alternative action combination still satisfies the endpoint constraint corresponding to the counterfactual endpoint state, and the risk evolution cost does not exceed the preset allowable range.

[0126] In some implementations, the system can also construct a library of equivalent alternative actions. This library stores the effects of different combinations of actions on historical public opinion events. Each set of equivalent alternative actions corresponds to a target effect, such as reducing the intensity of dissemination, reducing emotional intensity, reducing topic deviation, increasing factual clarification, or reducing the trigger for resurgence. When the currently preferred action is affected by action cooling factors, platform execution limitations, insufficient subject response, or delayed clarification absorption, the system selects an alternative action combination from the library that can achieve the same or near-final constraint. This library of equivalent alternative actions improves the flexibility of action output.

[0127] In some implementations, the system can also generate a safety interval window for handling actions. The safety interval window limits the minimum interval or maximum frequency between consecutive executions of the same type of action. When the action cooling factor increases, or when user negative feedback, excessive intervention feedback, or a reverse increase in dissemination heat appears in the status evidence package, the system extends the safety interval window for the corresponding action and prioritizes other types of alternative actions. By using the safety interval window, the system can avoid the decline in handling effectiveness or the emergence of new public opinion risks caused by the continuous execution of similar actions within a short period.

[0128] In some implementations, the results output module can also generate a trajectory interpretation chain. The trajectory interpretation chain explains the reasons for the increase in key risks, decrease in risks, shift in topics, effectiveness of fact clarification, or changes in intervention responses within a candidate situation trajectory. The trajectory interpretation chain is generated jointly by trajectory interpretation breakpoints, state evidence packages, the structure of the disseminating entities, platform diffusion relationships, and records of handling actions. For example, the trajectory interpretation chain can be represented as: a high-influence entity's statement leads to an increase in the platform diffusion factor; the increase in the platform diffusion factor leads to an increase in dissemination popularity; fact clarification content increases in fact clarification degree after being forwarded by an authoritative entity; and the increase in fact clarification degree leads to a decrease in emotional intensity. Through the trajectory interpretation chain, the situation simulation results can not only be represented as numerical curves but also be understood and verified by public opinion analysts.

[0129] When model training is involved, this step can train an intervention action mapping model. The intervention action mapping model uses historical public opinion trajectories as training samples, including sentiment intensity, dissemination popularity, subject influence, topic shift, factual clarification, intervention response change sequences, historical action types, action execution times, target audiences, impact strength, and subsequent dissemination effects. During training, historical state vectors and intervention response changes are input into the model, which outputs action type, execution time, target audience, and impact strength. The output is then compared with historical records and subsequent effects to calculate errors in action type, execution time, target audience, and impact strength. By jointly reducing these errors, the intervention action mapping model can generate executable actions based on changes in intervention response quantities in multi-scenario situational trajectories.

[0130] In some implementations, the equivalent alternative action library can be generated through training on historical action samples. Training samples include the original action combination, alternative action combinations, action execution time, target, action intensity, changes in the action cooldown factor, and subsequent situational changes. The system categorizes different action combinations into the equivalent alternative action library based on whether they can achieve the same or similar endpoint constraints. Through this training, the equivalent alternative action library can output alternative solutions when the action cooldown factor increases or when an action becomes unenforceable.

[0131] In this embodiment, the multi-scenario situation trajectory set is derived from step S5; the sequence of intervention response changes is derived from the public opinion state vector at each moment in the multi-scenario situation trajectory; emotional intensity, dissemination heat, subject influence, topic shift, and factual clarification are used to determine the type of action; the dissemination subject structure and platform diffusion relationship are used to determine the target of the action; risk evolution cost, endpoint constraint satisfaction, re-ignition triggering degree, and action cooling factor are used to determine the intensity of the action and alternative action combinations; the equivalent alternative action library is used to generate alternative action combinations that can achieve the same endpoint constraint; the action safety interval window is used to control the execution frequency of similar actions; the trajectory interpretation chain is used to explain the correspondence between the action and the situation change; the final output of the multi-scenario situation deduction result includes the situation trajectory, key risk moments, risk evolution cost, action execution time, target, and intensity.

[0132] This step transforms multi-scenario situational trajectories into demonstrable, comparable, and actionable projections. This enables the system to not only output trends in public opinion risk but also provide action recommendations corresponding to specific times, targets, and intensities. Furthermore, it allows the system to generate alternative action combinations when the action cooling factor is high or actions are limited, thereby improving the feasibility, sustained effectiveness, and interpretability of public opinion management plans.

[0133] Second Embodiment

[0134] like Figure 2 As shown, the second embodiment provides a public opinion multi-scenario situational analysis system based on SDE and DBGM. This system includes a state construction module, an endpoint generation module, a scenario coding module, a diffusion bridge analysis module, a trajectory filtering module, and a result output module.

[0135] The state construction module aligns multi-source time-series data of the target public opinion event within a unified time window, extracting sentiment intensity, dissemination heat, subject influence, topic shift, factual clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factor to form a public opinion state sequence and generate the dissemination subject structure and platform diffusion relationships. The state construction module's inputs include text content data, comment and repost data, interactive dissemination data, subject relationship data, platform diffusion data, fact verification data, and handling action data; the outputs are the public opinion state sequence, dissemination subject structure, and platform diffusion relationships. This output is sent to the endpoint generation module, scenario coding module, diffusion bridge inference module, and trajectory filtering module. In some implementations, the state construction module is also used to generate state evidence packages, backflow foldback edges, and clarification absorption gating results.

[0136] The endpoint generation module determines the current starting state from the public opinion state sequence and generates multiple counterfactual endpoint states based on the handling objective, propagation constraints, sentiment constraints, credible propagation delay, cross-platform backflow intensity, and resurgence triggering degree. The inputs to the endpoint generation module are the public opinion state sequence, handling objective, propagation constraints, sentiment constraints, credible propagation delay, cross-platform backflow intensity, and resurgence triggering degree; the outputs are the current starting state and multiple counterfactual endpoint states. The current starting state and the counterfactual endpoint states are sent to the diffusion bridge inference module. In some implementations, the endpoint generation module is also used to perform endpoint credibility stratification processing and evidence consistency verification.

[0137] The scenario encoding module encodes scenario conditions based on the dissemination subject structure, platform diffusion relationships, topic migration direction, cross-platform backflow intensity, and intervention action intensity. The module generates subject amplification factors, platform diffusion factors, emotional disturbance factors, topic migration factors, intervention action factors, backflow triggering factors, clarification delay factors, and action cooling factors, and combines these factors into scenario conditions. These scenario conditions are sent to the diffusion bridge deduction module and used to adjust the DBGM trajectory generation process. In some implementations, the scenario encoding module is also used to generate a scenario conflict matrix and a scenario evidence chain.

[0138] The diffusion bridge deduction module couples the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM, and the DBGM generates candidate situation trajectories using the SDE (Situational Deduction) which includes the ending bridging traction term. The inputs to the diffusion bridge deduction module are the current starting state, the counterfactual ending state, and the scenario conditions; the outputs are candidate situation trajectories corresponding to scenarios such as natural evolution, reinforcement propagation, issue reversal, and intervention suppression. The candidate situation trajectories are sent to the trajectory filtering module. In some implementations, the diffusion bridge deduction module is also used to adjust the sampling weights of the candidate situation trajectories based on local bridging anchor points and state evidence packets.

[0139] The trajectory filtering module is used to eliminate unreachable trajectories based on the directed dependencies between the propagation subject structure, topic migration direction, factual clarification level, and intervention response volume, and to obtain a multi-scenario situational trajectory set by filtering according to endpoint constraint satisfaction, risk evolution cost, and trajectory fingerprint similarity. The inputs to the trajectory filtering module are candidate situational trajectories, propagation subject structure, platform diffusion relationship, topic migration direction, counterfactual endpoint state, and constraints; the output is a multi-scenario situational trajectory set. This multi-scenario situational trajectory set is sent to the result output module. In some implementations, the trajectory filtering module is also used to generate trajectory interpretation breakpoints, secondary re-ignition check results, and representative paths for trajectory clusters.

[0140] The results output module is used to inversely extract the intervention response quantities at each moment from the multi-scenario situation trajectory set, mapping them to the execution time, target, and intensity of the action, and outputting the multi-scenario situation inference results. Based on the sequence of intervention response quantity changes and the corresponding moment's emotional intensity, dissemination heat, subject influence, topic shift, and factual clarification, the results output module maps the handling intensity to actions such as clarification release, authoritative forwarding, cooling-down guidance, interception of false information, topic diversion, or platform traffic restriction. When the action cooling factor exceeds a preset threshold, continuously repeated handling actions are replaced with alternative action combinations that can achieve the same endpoint constraint. In some implementations, the results output module is also used to call an equivalent alternative action library to generate a handling action safety interval window and trajectory interpretation chain.

[0141] In this system, the public opinion state sequence output by the state construction module serves as the data foundation for the endpoint generation module, scenario coding module, diffusion bridge inference module, and trajectory selection module. The counterfactual endpoint state output by the endpoint generation module provides endpoint guidance for the diffusion bridge inference module. The scenario conditions output by the scenario coding module provide scenario control for the diffusion bridge inference module. The candidate situation trajectories output by the diffusion bridge inference module provide selection targets for the trajectory selection module. The multi-scenario situation trajectory set output by the trajectory selection module provides the data foundation for the result output module to reverse-engineer the intervention response and output the situation inference results. Thus, a continuous data processing chain is formed among the modules.

[0142] In a specific application example, the target public opinion event is an online controversy triggered by a public event. The system first collects news reports, user comments, forwarding relationships, statements from key stakeholders, fact-checking information, and handling records from multiple platforms after the event occurs, and forms multi-source time-series data in ten-minute time windows. The state construction module extracts sentiment intensity, dissemination heat, subject influence, topic shift, fact clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, resurgence triggering degree, and action cooling factor from each time window to form a public opinion state sequence and generate the dissemination subject structure and platform diffusion relationships.

[0143] Subsequently, the endpoint generation module uses the current public opinion state vector within the current time window as the current starting state. Based on the objectives of reducing dissemination intensity, decreasing negative sentiment intensity, and increasing factual clarification, and considering factors such as the upper bound of dissemination intensity, the upper bound of sentiment intensity, credible dissemination lag, cross-platform backflow intensity, and resurgence trigger rate, it generates counterfactual endpoint states under intervention and suppression scenarios. Simultaneously, the system also generates counterfactual endpoint states under natural evolution, intensified dissemination, and issue reversal scenarios.

[0144] The scenario coding module identifies multiple high-influence entities based on the dissemination subject structure, identifies cross-platform dissemination paths based on platform diffusion relationships, identifies the direction of topic migration based on changes in topic clusters within a continuous time window, and determines the intensity of intervention actions based on existing clarification releases and authoritative reposts. Then, the scenario coding module generates scenario conditions for natural evolution, reinforced dissemination, topic reversal, and intervention suppression, respectively.

[0145] The diffusion bridge simulation module inputs the current starting state, the states of each counterfactual endpoint, and the conditions of each scenario into DBGM, generating multiple candidate situational trajectories. The trajectory filtering module eliminates unreachable trajectories based on the directed dependencies between the propagation subject structure, topic migration direction, factual clarification level, and intervention response volume. It retains trajectories that meet the conditions and have differentiated evolutionary paths based on endpoint constraint satisfaction, risk evolution cost, and trajectory fingerprint similarity. The results output module finally outputs the multi-scenario situational simulation results and, under the intervention suppression scenario, provides corresponding actions for different time windows, including clarification releases, authoritative reposts, de-escalation guidance, misinformation interception, topic diversion, or platform traffic restriction.

[0146] In a further application example, if the spread of a target public opinion event on one platform has decreased, but the same topic cluster begins to rise on another short video platform, the system identifies the risk of backflow by analyzing the cross-platform backflow intensity and backflow foldback edge, and retains the amount of resurgence in the counterfactual endpoint state. At this time, DBGM can generate candidate trend trajectories of cooling down on the main platform, heating up on the short video platform, and resurgence on the main platform. The trajectory filtering module determines whether the trajectory has causal reachability based on the structure of the dissemination subject, the platform diffusion relationship, and the state evidence package. The result output module generates a corresponding resurgence risk warning and provides authoritative forwarding, topic diversion, or fact clarification actions for the dissemination subject on the short video platform.

[0147] In another application example, if the system identifies that the same clarification content has been published multiple times consecutively, but the level of factual clarification is increasing slowly and user feedback text contains negative expressions, the action cooling factor increases. The diffusion bridge inference module reduces the marginal effect of repeated clarification publication actions in subsequent candidate situation trajectories. The output module replaces consecutively repeated clarification publication actions with alternative action combinations such as authoritative reposting, visual publication of fact-verification content, or topic diversion, and adjusts the execution time of the next similar action based on the action safety interval window. Thus, the system can avoid mechanically repeating the same action output.

[0148] Through the above implementation methods, the system can construct a complete sequence of public opinion states based on multi-source time-series data of the target public opinion event, and generate continuous candidate situational trajectories under different scenarios between the current starting state and multiple counterfactual endpoint states. Unreachable trajectories are eliminated by the directed dependencies between the dissemination subject structure, topic migration direction, factual clarification degree, and intervention response quantity. A multi-scenario situational trajectory set is formed through endpoint constraint satisfaction, risk evolution cost, and trajectory fingerprint similarity. Finally, the intervention response quantity is deduced and mapped to specific handling actions. This process improves the target constraint, dissemination rationality, trajectory diversity, handling feasibility, and result interpretability of multi-scenario public opinion extrapolation.

[0149] The above description is merely a specific embodiment of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make equivalent substitutions or adaptive adjustments to the collection methods of multi-source time-series data, the dimensions of the public opinion state vector, the generation method of the counterfactual endpoint state, the encoding method of the scenario conditions, the training method of DBGM, the screening method of candidate situation trajectories, and the action mapping method of intervention response quantities, based on the specific public opinion data source, platform type, time granularity, handling target, scenario category, and handling action type. They can also combine or selectively use credible propagation delay, cross-platform backflow intensity, resurgence triggering degree, action cooling factor, trajectory fingerprint, state evidence package, clarification absorption gating, backflow foldback edge, endpoint credibility layering, scenario conflict matrix, local bridging anchor point, trajectory interpretation breakpoint, equivalent alternative action library, handling action safety interval window, and trajectory interpretation chain, all of which fall within the protection scope of the present invention.

Claims

1. An SDE and DBGM-based public opinion multi-scenario situation deduction method, characterized in that, include: S1. Align the multi-source time-series data of the target public opinion event with a unified time window, extract the intensity of emotion, the heat of dissemination, the degree of influence of the subject, the degree of topic shift, the degree of clarification of facts, the amount of intervention response, the credible dissemination time lag, the intensity of cross-platform backflow, the degree of re-ignition triggering and the action cooling factor, form a public opinion state sequence, and generate the dissemination subject structure and platform diffusion relationship; S2. Determine the current starting state from the public opinion state sequence, and generate multiple counterfactual endpoint states based on the handling target, propagation constraints, emotional constraints, credible propagation time delay, cross-platform backflow intensity, and resurgence trigger degree; S3. The scenario conditions are encoded by the structure of the propagation subject, the platform diffusion relationship, the direction of topic migration, the intensity of cross-platform backflow, and the intensity of intervention actions. S4. Couple the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM, and the DBGM will generate candidate situation trajectories according to the SDE containing the ending bridging traction term. S5. Based on the directed dependency relationship between the dissemination subject structure, topic migration direction, fact clarification degree and intervention response quantity, unreachable trajectories are eliminated, and a multi-scenario situation trajectory set is obtained by filtering according to the endpoint constraint satisfaction degree, risk evolution cost and trajectory fingerprint similarity. S6. Extract the intervention response quantities at each moment from the set of multi-scenario situation trajectories, map them to the execution time, target and intensity of the action, and output the multi-scenario situation simulation results.

2. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The public opinion state vector in the public opinion state sequence is represented as follows: ， in, For a moment The public opinion state vector, For a moment The intensity of emotions, For a moment The popularity of the spread For a moment The degree of influence of the main body For a moment The degree of topic shift, For a moment Clarity of facts For a moment Intervention response volume For a moment The reliable propagation time delay, For a moment Cross-platform backflow strength, For a moment The degree of re-ignition triggering, For a moment Action cooling factor, This is a time for observing or predicting public opinion.

3. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The counterfactual endpoint state is generated as follows: ， in, For the first End point under similar scenario conditions The counterfactual endpoint state, This is the candidate endpoint state. For the first Under similar scenario conditions, the current starting state and disposal targets The target state obtained by mapping, This is the current starting state. For the first Targets for handling similar scenarios Candidate endpoint status Penalties for violating communication and emotional constraints. Candidate endpoint status Penalty for violating the trusted propagation time delay boundary. Candidate endpoint status Penalties for violating cross-platform backflow boundaries Candidate endpoint status The penalty for underestimating the amount of reignition retention. To determine the weights of the propagation constraints and emotional constraints as punishment items, The weight of the reliable propagation time delay boundary penalty term, The weight of the cross-platform backflow boundary penalty term, The weight of the penalty term for the amount of re-ignition retention. Assign scenario numbers, The final moment, It is the vector norm.

4. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The scenario conditions include subject amplification factor, platform diffusion factor, emotional disturbance factor, issue migration factor, intervention action factor, backflow trigger factor, clarification delay factor, and action cooling factor. The subject amplification factor is obtained from the subject's influence and the structure of the dissemination subject; the platform diffusion factor is obtained from the dissemination heat and platform diffusion relationship; the emotional disturbance factor is obtained from the emotional intensity; the issue migration factor is obtained from the issue offset and issue migration direction; the intervention action factor is obtained from the intervention response volume and the degree of factual clarification; the backflow trigger factor is obtained from the cross-platform backflow intensity; the clarification delay factor is obtained from the credible dissemination time lag; and the action cooling factor is obtained from the frequency of action execution and dissemination feedback.

5. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The DBGM generates candidate situational trajectories according to the following SDE: ， in, For the first Time under similar scenario conditions The increase in public opinion status For the first Time under similar scenario conditions The public opinion state vector, For model parameters The following corresponding scenario conditions The public opinion drift function, These are the model parameters for DBGM. For the first Scenario conditions, For the first Time under similar scenario conditions The bridging traction coefficient, For the first End point under similar scenario conditions The counterfactual endpoint state, For time infinitesimal elements, For model parameters The following corresponding scenario conditions The diffusion intensity function, For a moment Brownian motion increment, Assign scenario numbers, The endpoint.

6. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The elimination of unreachable trajectories includes: when the decrease in emotional intensity in a candidate trajectory precedes the increase in factual clarification and there is no supporting change in intervention response, the candidate trajectory is determined to be an unreachable trajectory; when the topic shift in a candidate trajectory changes but there is no reachable dissemination path between the corresponding subjects in the dissemination subject structure, the candidate trajectory is determined to be an unreachable trajectory; when the decrease in dissemination heat in a candidate trajectory is inconsistent with the platform's diffusion relationship and handling action records, the candidate trajectory is determined to be an unreachable trajectory.

7. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The cost of risk evolution is determined as follows: ， in, For the first The risk evolution cost of candidate situation trajectories under similar scenarios For the first Time under similar scenario conditions The intensity of emotions, For the first Time under similar scenario conditions The popularity of the spread For the first Time under similar scenario conditions The degree of topic shift, For the first Time under similar scenario conditions Clarity of facts For the first Time under similar scenario conditions Intervention response volume For the first Endpoint constraint violation under similar scenario conditions For the first End point under similar scenario conditions The degree of re-ignition triggering, For the first End point under similar scenario conditions Action cooling factor, The weighting coefficient for the intensity of emotion. As a weighting coefficient for the spread of popularity, The weighting coefficient for the topic offset. This is the weighting coefficient for the degree of clarification of facts. The weighting coefficients for the intervention response. The weighting coefficient for the amount of violation of the endpoint constraint. This is the weighting coefficient for the re-ignition triggering degree. This is the weighting coefficient for the action cooling factor. For the purpose of public opinion observation or prediction, The final moment, Number the scenarios, Indicates time to Summing up the terms of the expression.

8. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The trajectory fingerprint is generated from the peak time of the spread heat, the order of the inflection point of the emotion intensity, the direction of the topic shift, the time when the fact clarification takes effect, and the change pattern of the intervention response in the candidate situation trajectory. When the similarity of the trajectory fingerprints of multiple candidate situation trajectories is higher than a preset threshold, the candidate situation trajectory with higher endpoint constraint satisfaction, lower risk evolution cost and stronger causal accessibility is retained to form a multi-scenario situation trajectory set with differentiated evolution paths.

9. The method for multi-scenario public opinion situation inference based on SDE and DBGM according to claim 1, characterized in that, The reverse solution of intervention response quantities at each moment includes: extracting the sequence of intervention response quantity changes from the situation trajectory that satisfies the endpoint constraint and has the lowest risk evolution cost or is lower than a preset threshold; determining the intensity of handling at each moment based on the sequence of intervention response quantity changes; and mapping the intensity of handling to actions such as clarification release, authoritative forwarding, cooling guidance, interception of false information, topic diversion, or platform traffic restriction, in conjunction with the emotional intensity, dissemination heat, subject influence, topic deviation, and factual clarification at the corresponding moment. When the action cooling factor is higher than a preset threshold, the continuously repeated handling actions are replaced with alternative action combinations that can achieve the same endpoint constraint.

10. A public opinion multi-scenario situational prediction system based on SDE and DBGM, characterized in that, include: The state construction module is used to align multi-source time-series data of target public opinion events according to a unified time window, extract emotion intensity, dissemination heat, subject influence, topic shift, fact clarification, intervention response volume, credible dissemination lag, cross-platform backflow intensity, re-ignition trigger degree, and action cooling factor, form a public opinion state sequence, and generate the dissemination subject structure and platform diffusion relationship; The endpoint generation module is used to determine the current starting state from the public opinion state sequence, and generate multiple counterfactual endpoint states based on the handling objectives, dissemination constraints, emotional constraints, credible dissemination time delay, cross-platform backflow intensity, and resurgence trigger degree. The scenario coding module is used to encode scenario conditions based on the structure of the dissemination subject, the platform diffusion relationship, the direction of topic migration, the intensity of cross-platform backflow, and the intensity of intervention actions. The diffusion bridge simulation module is used to couple the current starting state, the corresponding counterfactual ending state, and the scenario conditions into the DBGM, and the DBGM generates candidate situation trajectories by SDE containing the ending bridging traction term. The trajectory filtering module is used to eliminate unreachable trajectories based on the directed dependency relationship between the structure of the dissemination subject, the direction of topic migration, the degree of factual clarification, and the amount of intervention response, and to filter and obtain a set of multi-scenario situational trajectories based on the degree of end-point constraint satisfaction, the cost of risk evolution, and the similarity of trajectory fingerprints. The results output module is used to inversely solve the intervention response quantities at each moment from the multi-scenario situation trajectory set, map them into the execution time, target and intensity of the action, and output the multi-scenario situation inference results.

Images