Network group behavior regulation method and system based on emotional entropy evolution

Abstract

This invention relates to a method and system for regulating network group behavior based on emotional entropy evolution. The method constructs a time-varying qualitative interaction graph based on multi-source behavioral data from a target network platform, calculates node emotional entropy, group emotional entropy, cross-group emotional entropy gradients, and edge-level emotional entropy flow, generating an emotional entropy evolution field. Based on this emotional entropy evolution field, combined with cumulative edge-level emotional entropy flow, behavioral offset differences, emotional entropy differences, cross-group propagation connectivity changes, and temporal propagation constraints, it identifies key bridging edges and searches for entropy-increasing trigger chains, outputting the contribution of trigger chains and the risk level of group behavior instability. Under the constraints of risk suppression and normal propagation preservation, it solves and executes the minimum necessary intervention combination. By comparing the counterfactual evolution trajectory under no-intervention conditions with the actual feedback trajectory, it determines the net intervention contribution and performs joint updates accordingly, forming a closed-loop regulation. This invention can improve the accuracy, real-time performance, and controllability of network group behavior regulation in complex propagation scenarios.

Description

Technical Field

[0001] This invention relates to the fields of Internet information processing, group behavior analysis, artificial intelligence, complex network modeling and intelligent regulation, and in particular to a method and system for regulating network group behavior based on emotional entropy evolution. Background Technology

[0002] With the rapid development of social media platforms, content platforms, community forums, and instant messaging platforms, user expression, information dissemination, and group interaction in cyberspace are characterized by high frequency, cross-segmentation, strong coupling, and rapid evolution. Around specific events, topics, or content objects, a large number of users can form intensive interactions in a short period, thus exhibiting obvious phenomena of evolving online group behavior. This type of group behavior is not only influenced by content semantics and individual emotions, but also by multiple factors such as propagation structure, interaction rhythm, group boundaries, and cross-group coupling channels, thus possessing significant dynamic, nonlinear, and sudden diffusion characteristics.

[0003] In existing technologies, the analysis and governance of online group behavior typically focus on sentiment classification, public opinion heat monitoring, keyword detection, key node identification, propagation path tracing, or simple rule intervention. For example, one type of method uses sentiment analysis models to identify the positive, negative, or neutral tendencies of text to determine the overall trend of public opinion; another type of method identifies hot users or hot content by statistically analyzing the number of reposts, comments, popularity values, or centrality indicators; and yet another type of method uses fixed thresholds or manual rules to limit the flow, reduce the ranking, or issue warnings for specific content.

[0004] However, the aforementioned prior art has at least the following shortcomings:

[0005] First, most existing technologies use emotional information as static labels or simple proportional statistics, lacking a unified model of the uncertainty and disorder of group emotions and their propagation and migration process in network structures, making it difficult to characterize the instability mechanism of network groups in the continuous evolution process.

[0006] Secondly, existing technologies mostly focus on identifying key nodes, hot topics, or local paths, making it difficult to identify the key bridging edges and entropy increase triggering chains that truly trigger cross-group emotional entropy leaps and behavioral instability spread. Therefore, it is difficult to accurately locate the key channels for the abnormal spread of group behavior.

[0007] Third, existing intervention methods mostly employ coarse-grained regulation driven by fixed rules, uniform flow restriction, or single indicators. They lack a minimum necessary intervention mechanism to balance risk suppression and preservation of normal transmission, which can easily lead to excessive intervention, too large a scope, or accidental damage to normal transmission.

[0008] Fourth, existing technologies typically adjust parameters directly based on the surface results after intervention, lacking a technical mechanism to compare the counterfactual evolutionary trajectory under no-intervention conditions with the actual feedback trajectory, in order to isolate natural evolutionary factors and quantify the net intervention contribution. This results in insufficient targeting, accuracy, and closed-loop adaptive capability of model updates.

[0009] Therefore, there is an urgent need to provide a new method and system for regulating online group behavior to solve the technical problems of existing technologies, such as insufficient modeling of disordered migration of group emotions, inaccurate identification of key cross-group propagation channels, coarse intervention granularity, and inaccurate feedback updates. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for regulating network group behavior based on emotional entropy evolution, so as to construct an emotional entropy evolution field, uniformly model the emotional disorder migration process in network group behavior, and identify key bridging edges by combining edge-level emotional entropy flow, behavioral offset difference, emotional entropy difference, cross-group propagation connectivity change amount and time sequence propagation constraints, and search for entropy increase triggering chains.

[0011] This invention also improves the accuracy, real-time performance, interpretability, and controllability of network group behavior regulation in complex propagation scenarios by solving the minimum necessary intervention combination under risk suppression constraints and normal propagation damage upper limit constraints, and determining the net intervention contribution based on the difference between the counterfactual evolution trajectory and the actual feedback trajectory. It also jointly updates the key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters.

[0012] To achieve the above objectives, the present invention provides a method for regulating network group behavior based on emotional entropy evolution, executed by a processor, comprising:

[0013] Based on multi-source behavioral data from the target network platform, a time-varying qualitative interaction graph is constructed, and node emotional entropy, group emotional entropy, cross-group emotional entropy gradient, and edge-level emotional entropy flow are calculated to generate an emotional entropy evolution field.

[0014] Based on the emotional entropy evolution field, edge features are extracted from cross-group bridging candidate edges. The edge features include at least the cumulative edge-level emotional entropy flow within the time window, the behavioral offset difference between the nodes at both ends of the edge, the emotional entropy difference between the nodes at both ends of the edge or the group, the cross-group propagation connectivity change after deleting the edge, and the continuous active time of the edge. Under the conditions of cross-group connection and time propagation constraints, key bridging edges are identified.

[0015] Using the key bridging edge as the anchor point, search for node edge sequences along the propagation direction and time sequence, and identify the node edge sequences that satisfy the cumulative edge-level emotional entropy flow enhancement, the continuous increase of node behavior offset and contain at least one key bridging edge as entropy increase trigger chains, and output the contribution of the trigger chain and the risk level of group behavior instability.

[0016] Based on the key bridging edge, the entropy increase triggering chain, the contribution of the triggering link, the risk level of group behavior instability, and the emotional entropy evolution field, solve for and execute the minimum necessary intervention combination that satisfies the risk suppression constraint and the upper limit constraint of normal propagation damage;

[0017] Before executing the minimum necessary intervention combination, a counterfactual evolutionary trajectory under no-intervention conditions is generated. After executing the minimum necessary intervention combination, the actual feedback trajectory is collected. The net intervention contribution is determined based on the difference between the counterfactual evolutionary trajectory and the actual feedback trajectory to remove natural evolutionary factors. The key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters are jointly updated based on the net intervention contribution to form a closed-loop regulation.

[0018] To achieve the above objectives, another aspect of the present invention provides a network group behavior regulation system based on emotional entropy evolution, comprising:

[0019] The Emotional Entropy Evolution Field Construction Module is used to construct a time-varying qualitative interaction graph based on multi-source behavioral data of the target network platform, calculate node emotional entropy, group emotional entropy, cross-group emotional entropy gradient and edge-level emotional entropy flow, and generate an emotional entropy evolution field.

[0020] The bridging edge and trigger chain identification module is used to extract edge features from cross-group bridging candidate edges based on the emotional entropy evolution field, including at least the cumulative edge-level emotional entropy flow, behavioral offset difference, emotional entropy difference, cross-group propagation connectivity change and continuous active time. Under the conditions of satisfying cross-group connection and time propagation constraints, it identifies key bridging edges, searches for entropy increase trigger chains along the propagation direction and time sequence, and outputs the contribution of trigger chains and the risk level of group behavior instability.

[0021] The minimum necessary intervention module is used to solve and execute the minimum necessary intervention combination that satisfies the risk suppression constraint and the normal propagation damage upper limit constraint based on the key bridging edge, the entropy increase triggering chain, the contribution degree of the triggering link, the risk level of group behavior instability and the emotional entropy evolution field.

[0022] The counterfactual correction module is used to generate a counterfactual evolutionary trajectory under no-intervention conditions before executing the minimum necessary intervention combination, collect the actual feedback trajectory after executing the minimum necessary intervention combination, determine the net intervention contribution based on the difference between the counterfactual evolutionary trajectory and the actual feedback trajectory to remove natural evolution factors, and jointly update the key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters based on the net intervention contribution.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects:

[0024] First, by constructing an emotional entropy evolution field, this invention incorporates node emotional entropy, group emotional entropy, cross-group emotional entropy gradient, and edge-level emotional entropy flow into a unified framework, thereby characterizing the migration process of network group emotional disorder in complex propagation structures and improving the ability to depict the process of group behavioral instability.

[0025] Second, this invention identifies key bridging edges by combining cumulative edge-level emotional entropy flow, behavioral offset differences, emotional entropy differences, cross-group propagation connectivity changes, and temporal propagation constraints. It then uses these key bridging edges as anchor points to search for entropy increase triggering chains. This allows for more accurate location of key propagation channels that trigger cross-group diffusion and behavioral instability, improving the accuracy and targeting of abnormal propagation identification.

[0026] Third, by solving the minimum necessary intervention combination under the constraints of risk suppression and normal propagation damage limit, this invention can reduce at least some of the following indicators, such as the number of intervened nodes, the number of intervened edges, the duration of intervention, the intensity of intervention, the affected propagation radius, and the number of affected groups, while achieving the preset adjustment target, thereby reducing the impact on the normal propagation order and improving the precision and controllability of intervention.

[0027] Fourth, this invention determines the net intervention contribution by comparing the counterfactual evolution trajectory under no intervention conditions with the actual feedback trajectory, and accordingly updates the key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters. This allows the interference of natural evolution factors on the feedback results to be removed, improving the accuracy of model updates, closed-loop adaptive capability, and long-term regulation stability.

[0028] Fifth, this invention integrates the design of emotion entropy evolution modeling, key bridging edge identification, entropy increase trigger chain search, minimum necessary intervention solution, and counterfactual closed-loop update, thereby improving the accuracy, real-time performance, interpretability, and engineering deployability of network group behavior regulation in complex propagation scenarios. Attached Figure Description

[0029] Figure 1This is a flowchart of the network group behavior regulation method based on emotional entropy evolution provided by the present invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the present invention, the same terms represent the same technical features, the same symbols represent the same meaning, and different symbols represent different meanings.

[0031] Figure 1 This is a flowchart of the network group behavior regulation method based on emotional entropy evolution provided by the present invention, such as... Figure 1 As shown, the network group behavior regulation method based on emotional entropy evolution provided in this embodiment can be executed by a server, cloud computing platform, edge computing node or other electronic device with data processing capabilities. The method preferably includes steps S1 to S5.

[0032] S1. Acquisition, alignment, annotation, and event-driven coding of multi-source behavioral data.

[0033] Specifically, it includes:

[0034] Obtain raw multi-source behavioral data related to the target event, target topic, or target interaction domain from the target network platform, as the input dataset for step S1. The raw multi-source behavioral data includes at least: user-posted content data, forwarding data, comment data, reply data, like data, user relationship data, group affiliation data, topic affiliation data, and timestamp data; in some embodiments, it may also include collection data, citation data, follow relationship data, geographic area tag data, and device-side context data.

[0035] The original multi-source behavioral data is subjected to unified time standardization processing, which specifically includes: converting data timestamps from different sources into a unified standard time zone; normalizing the precision of data with inconsistent precision at the second, minute, or hour level; and completing data with missing timestamps by using context event time interpolation, log write time backfilling, or window alignment, thereby forming a time-aligned dataset.

[0036] The time-aligned dataset undergoes semantic cleaning and normalization, specifically including: text segmentation, noise reduction, stop word filtering, emoji mapping, merging of network abbreviations, merging of synonyms, and semantic vector encoding; unified text conversion of image and text descriptions, short video titles, link titles, and tag text; and standardized mapping of behavior types, mapping posting, forwarding, commenting, replying, liking, quoting, joining a group, leaving a group, following, and unfollowing to a preset set of behavior types.

[0037] Map the normalized behavior records to event quadruples:

[0038]

[0039] in, For the first 1 event record, Identifier of the initiator of the action. Identify the object associated with the behavior. For behavior type, The time when the action occurs.

[0040] Furthermore, the event quadruple is expanded into an extended event representation:

[0041]

[0042] in, For the first An extended event indicates that, For the first The semantic representation vector of an event, For the first The sentiment probability vector of each event. This refers to the group affiliation tag corresponding to the initiator or object of the event. For propagation context indexing.

[0043] All extended event representations are aggregated to form a heterogeneous time-series observation dataset:

[0044]

[0045] in, This is a heterogeneous time-series observation dataset. To expand the total number of events.

[0046] Wherein, the emotion probability vector The emotion recognition model is used to obtain the emotion data. This model takes text content, emoticons, hashtags, contextual comments, and recent user interactions as input, and outputs a probability distribution over a predefined set of emotion categories.

[0047]

[0048] in, For the first The sentiment probability vector of each event. For emotion recognition models, For the parameters of the emotion recognition model, For the first The input content corresponding to each event, For the first Contextual information for each event.

[0049] The emotion recognition model can be implemented using any of the following: a pre-trained language model, a recurrent neural network, a convolutional neural network, or a graph-text fusion model; preferably, it is implemented using a text classification network pre-trained on a general Chinese corpus and fine-tuned on historical labeled data from the platform. Its training set can be represented as:

[0050]

[0051] in, For the training set of the emotion recognition model, For the first Input of training samples, For the first The sentiment labels of the training samples This represents the number of training samples.

[0052] Training is performed using the cross-entropy loss function:

[0053]

[0054] in, For emotion recognition loss, For the number of emotion categories, For indicator functions, For the model to predict the first The sample belongs to the first The probability of similar emotions.

[0055] This step achieves unified time alignment, semantic normalization, and event-based encoding of multi-source behavioral data, enabling data from heterogeneous sources, with heterogeneous granularities and semantic forms to be fused and expressed within a unified temporal observation framework. This provides a consistent data foundation for subsequent construction of time-varying qualitative interaction graphs, calculation of sentiment entropy, and search of trigger chains, thus improving data availability, graph integrity, and the accuracy of subsequent analysis.

[0056] S2, Construction of time-varying qualitative interaction graph, calculation of sentiment entropy and generation of sentiment entropy evolution field.

[0057] Specifically, it includes:

[0058] The heterogeneous time-series observation dataset output in step S1 As input to step S2, within the preset sliding time window The following is a construction-time variable interaction graph:

[0059]

[0060] in, For a moment The time-varying meta-interaction graph, For a moment The set of nodes, For a moment The set of edges, A type mapping function for nodes or edges. A function for mapping attributes of nodes or edges. The time window length is defined as . The node set includes at least user nodes, content nodes, topic nodes, and group nodes, and the edge set includes at least post edges, forward edges, comment edges, reply edges, co-occurrence edges, membership edges, and cross-group bridging candidate edges.

[0061] Based on the sentiment probability vector of the extended event representation in step S1, events belonging to the same node are weighted and aggregated within a time window to obtain the node at time [time value missing]. Emotion probability distribution:

[0062]

[0063] in, For nodes At any moment Belongs to the The aggregation probability of sentiment-like factors For nodes within the time window The set of event indexes For the first The event at the moment Normalized weights, For the first The event in the The probability of similar emotions.

[0064] Calculate node sentiment entropy based on node sentiment probability distribution:

[0065]

[0066] in, For nodes At any moment Emotional entropy For the number of emotion categories, For nodes At any moment Belongs to the The aggregation probability of emotion-like categories.

[0067] Based on group affiliation tags, the group sentiment entropy is obtained by weighted aggregation of the sentiment entropy of nodes within the group:

[0068]

[0069] in, For the group At any moment The group emotional entropy For the group The corresponding set of nodes, For nodes In the group The weights in the equation. The weights... It can be determined by combining node activity, propagation centrality, and frequency of interaction within the group.

[0070] Calculate the cross-group emotional entropy gradient based on the difference in group emotional entropy and structural distance between different groups:

[0071]

[0072] in, For the group with the group At any moment The cross-group emotional entropy gradient For the group with the group At any moment Structural distance, For smoothing terms. The structural distance. The calculation can be performed using the shortest path length, average number of hops, or the inverse of the sparsity of cross-group connection edges; the preferred method is to use the weighted structural distance constructed by combining the shortest path length and the cross-group edge density.

[0073] Calculate the edge-level sentiment entropy flow based on the sentiment entropy difference, interaction strength, and propagation direction between adjacent nodes:

[0074]

[0075] in, For a moment From node Pointing to node The edge-level emotional entropy flow, For nodes With nodes At any moment The connection indicator quantity, For nodes With nodes Interaction strength weight, and They are nodes With nodes At any moment Emotional entropy This is the propagation direction factor. The interaction strength weight... It can be obtained by weighted normalization based on forwarding frequency, commenting frequency, replying frequency, co-occurrence frequency, and the reciprocal of interaction latency.

[0076] The emotional entropy of nodes, the emotional entropy of groups, the cross-group emotional entropy gradient, and the edge-level emotional entropy flow are accumulated, smoothed, and encapsulated to form the state of the emotional entropy evolution field:

[0077]

[0078] in, For a moment The state of the emotional entropy evolution field.

[0079] To enhance the representation capabilities of node states, edge states, and group states, a time-varying prime interaction graph can be input into the graph representation network to obtain node representation vectors:

[0080]

[0081] in, For nodes At any moment The graph representation vector, To represent networks using graphs, The parameters are for the graph representation network. The graph representation network can be any of a graph neural network, a heterogeneous graph neural network, a graph attention network, or a temporal graph network; a heterogeneous temporal graph neural network is preferred. Its training objectives may include at least one or more of the following: next-time connection prediction, group affiliation prediction, node activity prediction, or auxiliary prediction of the risk level of group behavior instability.

[0082] This step constructs a time-varying qualitative interaction graph and calculates node sentiment entropy, group sentiment entropy, cross-group sentiment entropy gradient, and edge-level sentiment entropy flow. This transforms the sentiment expression in the network group from static labels to a continuous evolutionary field expression that changes with time and network structure. This enhances the ability to perceive the evolutionary trend of group behavior and cross-group diffusion trends, and has the beneficial effect of improving the accuracy of identifying abnormal diffusion precursors.

[0083] S3, Key Bridge Edge Identification, Entropy Increase Trigger Chain Search, and Instability Risk Assessment

[0084] Specifically, it includes:

[0085] The emotional entropy evolution field state output in step S2 Time-varying meta-interaction graph and graphical representation vectors As input to step S3, calculate node behavior offset, edge-level cumulative entropy flow, cross-population propagation connectivity change, and local clustering enhancement.

[0086] The node behavior offset is defined as follows:

[0087]

[0088] in, For nodes At any moment The behavior offset, For nodes At any moment The current behavior response vector, For nodes Historical steady-state behavior baseline vector It is a norm 2. The current behavior response vector The historical steady-state behavior baseline vector is composed of at least one of the following statistical measures within a time window: frequency of speaking, frequency of forwarding, comment density, reply latency, frequency of cross-group interaction, and frequency of topic switching; It can be obtained from the mean statistic, exponential moving average, or cluster center within the historical normal range.

[0089] For each bridge candidate edge Calculate the bridging effect score:

[0090]

[0091] in, For the edge At any moment The bridging affects the score. For the edge The cumulative edge-level sentiment entropy flow within the preset time window To delete edges Changes in cross-group connectivity after transmission For the edge The difference in emotional entropy between two nodes or two groups For the edge The difference in behavioral offset between the two endpoints. These are the weighting coefficients. The weighting coefficients can be set manually, obtained through grid search, Bayesian optimization, or gradient descent; preferably, they are obtained through supervised training based on bridge edge labeled samples.

[0092] A bridging candidate edge that meets at least two of the following conditions is identified as a key bridging edge: the two ends of the edge belong to different groups or different local communities; the cumulative edge-level sentiment entropy flow exceeds a preset threshold; the change in cross-group propagation connectivity after deleting the edge exceeds a preset threshold; the difference in behavioral offset or sentiment entropy between the nodes at the two ends of the edge exceeds a preset threshold; the edge remains active within a preset observation window, and its adjacent propagation events satisfy the temporal propagation constraint. Therefore, the key bridging edges in this invention are not ordinary cross-community connection edges, but are jointly defined by the cumulative edge-level sentiment entropy flow, behavioral offset difference, sentiment entropy difference, change in cross-group propagation connectivity, and temporal propagation constraint.

[0093] Based on the set of key bridging edges, the node edge sequence is searched along the temporal sequence and propagation direction to obtain the entropy increase triggering chain. The criteria for determining the entropy increase trigger chain include at least the following: the cumulative edge-level emotional entropy flow along the link direction shows a monotonically increasing or phased increasing trend; the node behavior offset along the link direction continuously increases; the link contains at least one key bridging edge; and the link propagation contribution exceeds a preset threshold. The parameters corresponding to the criteria can be collectively referred to as the entropy increase trigger chain discrimination parameters.

[0094] The contribution of the triggering link can be defined as:

[0095]

[0096] in, For the first The contribution of each entropy-increasing trigger chain to the triggering chain. For the first The cumulative entropy flow contribution of each entropy-increasing trigger chain. For the first The propagation depth of the chain triggered by entropy increase. For the first The cross-population span of a chain triggered by increasing entropy. For the first The duration of the entropy increase triggering chain. These are the weighting coefficients.

[0097] Based on the number of key bridging edges, the maximum bridging impact score, the maximum triggering link contribution, the peak value of the cross-group emotional entropy gradient, and the local abnormal clustering degree, the risk level of group behavior instability is determined and output. The identification of key bridging edges can be completed by a classification model, with the input being the edge feature vector and the output being the bridging edge probability; the risk level determination can be completed by a multi-classification network, with the input being the global risk feature vector and the output being the risk level distribution.

[0098] The key bridging edge identification model can be represented as follows:

[0099]

[0100] in, For the edge This represents the predicted probability of the critical bridge edge. This is a key bridging edge identification model. Key bridging edge identification model parameters, For the edge The edge feature vector. The edge feature vector includes at least the edge feature vector. , , , And the characteristics of sustained active time.

[0101] This step, by identifying key bridging edges from cross-group connections and further searching for entropy increase triggering chains, can locate key cross-group propagation channels and chain-like triggering paths before group behavior becomes completely unstable. This improves the ability to warn of abnormal spread, the accuracy of locating key intervention targets, and the reliability of risk level discrimination, demonstrating significant technical effectiveness.

[0102] Furthermore, the intervention triggering condition can be determined based on the group behavior instability risk level, the number of key bridging edges, the maximum bridging impact score, and the maximum triggering link contribution. When the group behavior instability risk level is higher than the preset threshold, or at least one of the key bridging edges, the maximum bridging impact score, and the maximum triggering link contribution is higher than the corresponding preset threshold, the intervention triggering condition is determined to be met, and the process proceeds to step S4. Otherwise, steps S2 and S3 are continued for rolling monitoring and updating.

[0103] S4. Solving for the minimum necessary intervention combination, constraint optimization, and strategy execution.

[0104] Specifically, it includes:

[0105] The key bridging edge set, entropy increase trigger chain set, trigger link contribution set, group behavior instability risk level, and emotional entropy evolution field state output in step S3 are used as inputs for step S4 to construct a minimum necessary intervention solution model.

[0106] Define an intervention strategy combination vector:

[0107]

[0108] in, This is a vector of candidate intervention strategy combinations. For the first The intensity or on / off variable of the intervention action. The number of intervention action categories. The intervention actions include at least one or more of the following: target content diffusion radius adjustment, target content ranking weight adjustment, target node interaction threshold adjustment, cross-group bridging channel suppression, topic coupling strength reduction, rhythm delay injection, rational prompt injection, and fact verification prompt injection.

[0109] Establish a constrained multi-objective optimization problem:

[0110]

[0111] in, To determine the feasible intervention domain that satisfies the constraints, To comprehensively optimize the objective function.

[0112] The comprehensive optimization objective function can be expressed as:

[0113]

[0114] in, This is a risk item for the spread of abnormal emotional entropy. To bridge the risk of cross-group transmission, For behavioral deviation risk items, For local anomalous clustering terms, The cost of damage during normal transmission. For intervention cost items, These are the weighting coefficients.

[0115] The constraints must include at least:

[0116]

[0117]

[0118]

[0119] in, The total risk residual after implementing the intervention portfolio, As the risk suppression threshold, This represents the upper limit of damage during normal transmission. and The first The lower and upper limits of intervention actions.

[0120] In this invention, "minimum necessary intervention" means minimizing at least two of the following: the number of intervened nodes, the number of intervened edges, the duration of intervention, the intensity of intervention, the affected propagation radius, and the number of affected groups, while meeting the preset risk suppression target. This avoids the "one-size-fits-all" over-intervention common in existing technologies.

[0121] The minimum necessary intervention solution model can be implemented using an analytical optimizer, a reinforcement learning policy network, or an offline policy regression model. Preferably, a policy network is used to generate candidate intervention combinations.

[0122]

[0123] in, For a moment intervention combination, For policy networks, For policy network parameters, The system state vector is obtained by concatenating the emotional entropy evolution field state, the key bridging edge state, the entropy increase triggering chain state, and the risk level.

[0124] The policy network can be trained using historical intervention logs and a simulation environment, and the reward function can be defined as:

[0125]

[0126] in, For a moment Controlling rewards, The reward weights are used. During training, policy gradient, Actor-Critic, or PPO methods can be employed to improve the solution efficiency and stability under complex constraints.

[0127] After obtaining the optimal intervention combination, it is mapped to the platform action interface to perform corresponding interventions on the target content, target node, target bridging edge, or target group.

[0128] This step solves for the minimum necessary intervention combination under multiple constraints such as risk suppression, preservation of normal dissemination, platform rules, and user experience. It can suppress the spread of abnormal group behavior while minimizing the collateral impact on normal discussion and dissemination, thereby improving the precision, targeting, and rationality of the intervention strategy.

[0129] S5, Counterfactual Evolution Prediction, Net Intervention Contribution Estimation, and Joint Update

[0130] Specifically, it includes:

[0131] The optimal intervention combination, intervention execution log, graph state, emotional entropy evolution field state, and group behavior instability risk level output from step S4 are used as inputs for step S5.

[0132] Before implementing the optimal intervention combination, based on the current system state Construct a counterfactual evolution prediction model to generate future state trajectories under conditions of no intervention:

[0133]

[0134] in, As a counterfactual evolutionary trajectory, For counterfactual evolution prediction models, For the parameters of the counterfactual evolution prediction model, For a moment The system state vector, This indicates that no intervention was imposed. To predict the step size.

[0135] After implementing the optimal intervention combination, continue to collect future data. The actual feedback trajectory within each time step:

[0136]

[0137] in, The actual feedback trajectory is defined as follows. The counterfactual evolution trajectory and the actual feedback trajectory include at least: the group sentiment entropy sequence, the cross-group sentiment entropy gradient sequence, the key bridging edge activity sequence, the entropy increase trigger chain propagation depth sequence, and the risk level sequence.

[0138] Calculate the net intervention contribution based on the difference between the counterfactual evolution trajectory and the actual feedback trajectory:

[0139]

[0140] in, For net intervention contribution, The difference in peak group emotional entropy The peak difference of the cross-group emotional entropy gradient. The critical bridging edge has poor activity. The difference in chain propagation depth is caused by entropy increase. For the duration of instability, The weighting coefficients are as follows. The peak difference and activity difference are both determined by the difference between the counterfactual evolutionary trajectory under no intervention conditions and the actual feedback trajectory, thus enabling the isolation of apparent changes caused by natural evolutionary factors.

[0141] Further construct the joint update loss function:

[0142]

[0143] in, To jointly update the loss function, To compensate for the loss of evolutionary prediction error, The net intervention contributes to the estimation error loss. This is the balance coefficient.

[0144] Simultaneously update the parameters of the key bridging edge identification model using a joint update loss function. Entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters. and parameters of the counterfactual evolution prediction model In some embodiments, the parameters of the emotion recognition model can be further updated. Graph representation of network parameters and risk discrimination model parameters.

[0145] The training data for the counterfactual evolution prediction model comes from historical evolution samples and historical intervention samples of the platform, forming a sequence training set. During training, the historical real trajectory is used as the supervision signal to minimize the difference between the predicted trajectory and the real trajectory. For samples with real intervention logs, counterfactual replay estimation can be performed on the non-intervention branch to improve the model's ability to fit the natural evolution trajectory. The counterfactual evolution prediction model can be implemented using any of the following: temporal graphical neural network, sequence-to-sequence network, Transformer temporal prediction network, or state-space model.

[0146] This step generates counterfactual evolutionary trajectories under no-intervention conditions before intervention and collects actual feedback trajectories after intervention for comparison. This allows for the differentiation between the parts of group behavior changes caused by natural evolutionary factors and those caused by intervention measures, thereby improving the accuracy of intervention effect attribution, the relevance of parameter updates, and the robustness of the overall regulatory loop.

[0147] Furthermore, the adjustment effect can be judged based on the comprehensive risk residual amount, the normal transmission damage cost item, the group behavior instability risk level, and the net intervention contribution. When the comprehensive risk residual amount is lower than the preset threshold, the normal transmission damage cost item is not higher than the preset upper limit, and the group behavior instability risk level drops to the preset level range, the adjustment effect is judged to meet the requirements. Otherwise, the updated model parameters are reused in steps S2 to S5 for the next round of rolling identification, intervention solution, and closed-loop update.

[0148] Corresponding to the above methods, this invention also provides a network group behavior regulation system based on emotional entropy evolution, comprising: an emotional entropy evolution field construction module, a bridging edge and trigger chain identification module, a minimum necessary intervention module, and a counterfactual correction module. Each module can be implemented through software, hardware, or a combination of both.

[0149] The system comprises three modules: an emotional entropy evolution field construction module for data acquisition, preprocessing, construction of time-varying qualitative interaction graphs, and generation of emotional entropy evolution fields in steps S1 and S2; a bridging edge and trigger chain identification module for key bridging edge identification, entropy increase trigger chain search, and risk level judgment in step S3; a minimum necessary intervention module for constrained multi-objective optimization and intervention execution in step S4; and a counterfactual correction module for counterfactual evolution prediction, net intervention contribution estimation, and joint update in step S5. This enables the system to perform real-time perception, fine-grained intervention, and closed-loop adaptive updates for complex propagation scenarios.

[0150] It should be noted that the above embodiments are merely preferred embodiments of the present invention, used to illustrate the technical solution of the present invention, and not to limit the scope of protection of the present invention. Those skilled in the art can make various equivalent substitutions, modifications, or improvements to the technical solution of the present invention without departing from the spirit and substance of the present invention. For example, adjustments or substitutions can be made to the data source format, the sentiment category classification method, the time-varying qualitative interaction graph construction method, the sentiment entropy calculation granularity, the key bridging edge identification condition, the entropy increase trigger chain discrimination method, the minimum necessary intervention solution method, the counterfactual evolution prediction model structure, and the joint update mechanism. All equivalent changes, modifications, and improvements made using the technical concepts described in the present invention specification and claims should fall within the protection scope of the claims of the present invention.

Claims

1. A method for regulating network group behavior based on emotional entropy evolution, characterized in that, Executed by the processor, the process includes: constructing a time-varying qualitative interaction graph based on multi-source behavioral data from the target network platform; calculating node emotional entropy, group emotional entropy, cross-group emotional entropy gradient, and edge-level emotional entropy flow; and generating an emotional entropy evolution field. Based on the emotional entropy evolution field, edge features are extracted from cross-group bridging candidate edges. The edge features include at least the cumulative edge-level emotional entropy flow within the time window, the behavioral offset difference between the nodes at both ends of the edge, the emotional entropy difference between the nodes at both ends of the edge or the group, the cross-group propagation connectivity change after deleting the edge, and the continuous active time of the edge. Under the conditions of cross-group connection and time propagation constraints, key bridging edges are identified. Using the key bridging edge as the anchor point, search for node edge sequences along the propagation direction and time sequence, and identify the node edge sequences that satisfy the cumulative edge-level emotional entropy flow enhancement, the continuous increase of node behavior offset and contain at least one key bridging edge as entropy increase trigger chains, and output the contribution of the trigger chain and the risk level of group behavior instability. Based on the key bridging edge, the entropy increase triggering chain, the contribution of the triggering link, the risk level of group behavior instability, and the emotional entropy evolution field, solve for and execute the minimum necessary intervention combination that satisfies the risk suppression constraint and the upper limit constraint of normal propagation damage; Before executing the minimum necessary intervention combination, a counterfactual evolutionary trajectory under no-intervention conditions is generated. After executing the minimum necessary intervention combination, the actual feedback trajectory is collected. The net intervention contribution is determined based on the difference between the counterfactual evolutionary trajectory and the actual feedback trajectory to remove natural evolutionary factors. The key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters are jointly updated based on the net intervention contribution to form a closed-loop regulation.

2. The method according to claim 1, characterized in that: The time-varying qualitative interaction graph is represented as follows: in, For a moment The time-varying meta-interaction graph, For a moment The set of nodes, For a moment The set of edges, A type mapping function for nodes or edges. The attribute mapping function is for nodes or edges; the set of nodes includes at least user nodes, content nodes, topic nodes and group nodes, and the set of edges includes at least post edges, forward edges, comment edges, reply edges, co-occurrence edges, membership edges and cross-group bridging candidate edges.

3. The method according to claim 1, characterized in that: The node sentiment entropy is represented as: in, For nodes At any moment Emotional entropy For the number of emotion categories, For nodes At any moment Belongs to the The probability of similar emotions; The cross-group emotional entropy gradient is represented as: in, For the group with the group At any moment The cross-group emotional entropy gradient and Groups with the group At any moment The group emotional entropy For the group with the group At any moment Structural distance, This is a smoothing term.

4. The method according to claim 1, characterized in that: The edge-level sentiment entropy flow is represented as: in, For a moment From node Pointing to node The edge-level emotional entropy flow, For nodes With nodes At any moment The connection indicator quantity, For nodes With nodes Interaction strength weight, and They are nodes With nodes At any moment Emotional entropy This is the propagation direction factor.

5. The method according to claim 1, characterized in that: The node behavior offset is represented as follows: in, For nodes At any moment The behavior offset, For nodes At any moment The current behavior response vector, For nodes Historical steady-state behavior baseline vector It is a norm 2. The current behavior response vector It consists of at least one of the following statistical measures within a preset time window: frequency of speaking, frequency of forwarding, comment density, reply delay, frequency of cross-group interaction, and frequency of topic switching.

6. The method according to claim 1, characterized in that: Calculate the bridging impact score for each bridging candidate edge: in, For the edge At any moment The bridging affects the score. For the edge The cumulative edge-level sentiment entropy flow within the preset time window To delete edges Changes in cross-group connectivity after transmission For the edge The difference in emotional entropy between two nodes or two groups For the edge The difference in behavioral offset between the two endpoints. These are the weighting coefficients; Bridge candidate edges that meet at least two of the following conditions are identified as key bridge edges: the two ends of the edge belong to different groups or different local communities; the cumulative edge-level sentiment entropy flow is higher than a preset threshold; the decrease in cross-group propagation connectivity after deleting the edge is higher than a preset threshold; the difference in behavioral offset or sentiment entropy between the nodes at the two ends of the edge is higher than a preset threshold; and the edge remains active for a longer period of time within the observation window than a preset threshold.

7. The method according to claim 1, characterized in that: The entropy increase triggering chain is a sequence of nodes and edges obtained by searching along the temporal order and propagation direction. And it satisfies the following conditions: the cumulative edge-level emotional entropy flow along the link direction monotonically increases or is phased in increase; the node behavior offset continuously increases; the link contains at least one key bridging edge; and the total link contribution exceeds a preset threshold. The trigger link contribution is expressed as: in, For the first The contribution of each entropy-increasing trigger chain to the triggering chain. For the first The cumulative entropy flow contribution of each entropy-increasing trigger chain. For the first The propagation depth of the chain triggered by entropy increase. For the first The cross-population span of a chain triggered by increasing entropy. For the first The duration of the entropy increase triggering chain. These are the weighting coefficients.

8. The method according to claim 1, characterized in that: The minimum necessary intervention combination is solved using a constrained multi-objective optimization model, and the optimization objective function is expressed as: in, For the intervention strategy combination vector, To determine the feasible intervention domain that satisfies the constraints, To comprehensively optimize the objective function; The comprehensive optimization objective function is expressed as follows: = in, This is a risk item for the spread of abnormal emotional entropy. To bridge the risk of cross-group transmission, For behavioral deviation risk items, For local anomalous clustering terms, The cost of damage during normal transmission. These are the weighting coefficients.

9. The method according to claim 1, characterized in that: The net intervention contribution is expressed as: in, For net intervention contribution, The difference in peak group emotional entropy The peak difference of the cross-group emotional entropy gradient. The critical bridging edge has poor activity. The difference in chain propagation depth is caused by entropy increase. For the duration of instability, The weighting coefficients are used; the joint update is based on the net intervention contribution, and simultaneously updates the key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters.

10. A network group behavior regulation system based on emotional entropy evolution, characterized in that, include: The Emotional Entropy Evolution Field Construction Module is used to construct a time-varying qualitative interaction graph based on multi-source behavioral data of the target network platform, calculate node emotional entropy, group emotional entropy, cross-group emotional entropy gradient and edge-level emotional entropy flow, and generate an emotional entropy evolution field. The bridging edge and trigger chain identification module is used to extract edge features from cross-group bridging candidate edges based on the emotional entropy evolution field, including at least the cumulative edge-level emotional entropy flow, behavioral offset difference, emotional entropy difference, cross-group propagation connectivity change and continuous active time. Under the conditions of satisfying cross-group connection and time propagation constraints, it identifies key bridging edges, searches for entropy increase trigger chains along the propagation direction and time sequence, and outputs the contribution of trigger chains and the risk level of group behavior instability. The minimum necessary intervention module is used to solve and execute the minimum necessary intervention combination that satisfies the risk suppression constraint and the normal propagation damage upper limit constraint based on the key bridging edge, the entropy increase triggering chain, the contribution degree of the triggering link, the risk level of group behavior instability and the emotional entropy evolution field. The counterfactual correction module is used to generate a counterfactual evolutionary trajectory under no-intervention conditions before executing the minimum necessary intervention combination, collect the actual feedback trajectory after executing the minimum necessary intervention combination, determine the net intervention contribution based on the difference between the counterfactual evolutionary trajectory and the actual feedback trajectory to remove natural evolution factors, and jointly update the key bridging edge identification parameters, entropy increase trigger chain discrimination parameters, and minimum necessary intervention solution model parameters based on the net intervention contribution.

Images