Cognitive influence counterfactual effect evaluation method and system for network harmful information dissemination

Abstract

This invention discloses a method and system for assessing the counterfactual effectiveness of cognitive impacts on the spread of harmful online information, belonging to the fields of network information content security, causal inference, and cognitive computing. The method constructs a propagation-cognition coupling time-varying graph, extracts affected subgraphs and local repair sets, calculates node cognitive perturbation inputs, and recursively infers the true cognitive state. It performs deletion, weight reduction, or replacement of baseline paths on the same topic only within the local repair set, while maintaining statistical background equivalence outside the affected subgraph, to obtain the counterfactual cognitive state. Furthermore, it generates robust effectiveness scores based on the true-counterfactual difference, bridging effectiveness contribution, persistent impact term, vulnerability weight, and uncertainty standard deviation, outputting the effectiveness level, key affected groups, and key propagation links. This invention can improve the accuracy, interpretability, and stability of assessing the cognitive impact of harmful online information.

Description

Technical Field

[0001] This invention relates to the fields of network information content security, social network propagation analysis, causal inference, cognitive computing and intelligent assessment, and in particular to a method and system for assessing the counterfactual effect of cognitive impact on the spread of harmful information on the internet. Background Technology

[0002] With the widespread use of social media, short video platforms, online communities, and instant messaging platforms, harmful online information exhibits characteristics such as rapid spread, wide reach, strong cross-group penetration, and long-lasting impact. Existing technologies typically focus on harmful content identification, dissemination popularity statistics, sentiment analysis, dissemination link tracing, and risk classification and early warning. While these technologies can identify harmful information and its scale of spread to a certain extent, they still lack a refined, stable, and causally explanatory assessment mechanism for the actual impact of harmful information on users' and groups' cognitive states.

[0003] In existing technologies, one type of method focuses on multimodal content recognition and propagation structure analysis, capable of identifying harmful content in text, images, or audio / video, and constructing a propagation graph based on forwarding, commenting, and following relationships for risk assessment. Another type of method introduces causal inference or counterfactual analysis to compare differences before and after content exposure, or between real and alternative scenarios. However, these methods typically suffer from the following shortcomings: First, they often lack explicit recursive modeling of node cognitive states, making it difficult to continuously characterize the dynamic evolution of cognitive variables such as cognitive acceptance, emotional shift, behavioral tendencies, and memory retention. Second, counterfactual construction often employs full-graph deletion, global replacement, or overall weight reduction, easily perturbing background structures unrelated to the target harmful information, leading to significant causal attribution bias. Third, the impact of cross-group bridging diffusion is usually calculated on the global path, making it difficult to distinguish which bridging effects are directly caused by the exposure of the target harmful information and which belong to natural interactions between groups. Fourth, existing assessment results often remain at the level of mean or single-point scores, insufficiently considering the uncertainties caused by model errors, incomplete structural observations, and counterfactual construction biases, resulting in low stability and credibility of the output results.

[0004] Therefore, there is an urgent need for a cognitive impact assessment method, system, and program product that can perform local counterfactual repair of affected areas, limited recursion of node cognitive states, targeted attribution of bridging diffusion effects, and robust calibration of assessment results on the propagation-cognition coupling graph, so as to improve the accuracy, interpretability, and engineering applicability of cognitive impact assessment of harmful online information. Summary of the Invention

[0005] The purpose of this invention is to overcome the following shortcomings in existing technologies: lack of explicit modeling of the dynamic evolution of cognitive states, excessive perturbation of non-target background structures by counterfactual scenario construction, inaccurate attribution of cross-group bridging effects, and lack of uncertainty constraints in evaluation results. To address these shortcomings, this invention provides a method and system for evaluating the counterfactual effectiveness of cognitive impacts on the spread of harmful online information. By constructing a propagation-cognition coupling time-varying graph, extracting affected subgraphs and determining local repair sets, counterfactual repair is implemented only within these local repair sets while maintaining statistical background equivalence outside the affected subgraphs, thereby obtaining a counterfactual cognitive state with greater causal explanatory power. Furthermore, by combining node cognitive effectiveness increments, bridging effectiveness contributions, persistent impact terms, vulnerability weights, and uncertainty standard deviations, robust effectiveness scores are generated to achieve a refined evaluation of the cognitive impact of harmful information spread, identification of key impact groups, and output of key propagation links.

[0006] To achieve the aforementioned objective, this invention provides, in one aspect, a method for assessing the counterfactual effect of cognitive impact on the spread of harmful online information, which is executed by a processor and includes:

[0007] A propagation-cognition coupling time-varying graph is constructed based on target harmful content, user interaction, group affiliation, and handling events, and propagation states are configured for nodes. Cognitive state and vulnerability weight Based on the temporal reach and cross-group propagation links of the target harmful content, the affected subgraph is extracted. And determine the local repair set Based on the set of exposed content, the cognitive perturbation input of the nodes is calculated and combined with the group mean cognitive state and memory residue. By performing restricted recursion, the true cognitive state can be obtained; only when... Internally, for targets exposed at certain edges, perform deletion, downgrading, or replacement of baseline information paths for the same topic, while maintaining... Equivalent external statistical background to generate counterfactual diagrams And obtain a counterfactual cognitive state ;according to and The difference calculation node cognitive effect increment And only when crossing Calculating the bridge acceptance effect contribution on cross-group paths ;based on , Items with lasting impact , and standard deviation of uncertainty Generate robust effectiveness scores Based on this, the effectiveness level, key impact groups, and key transmission links are output.

[0008] To achieve the aforementioned objective, another aspect of the present invention provides a counterfactual effect assessment system for the cognitive impact of the spread of harmful online information, comprising:

[0009] Includes: a propagation-cognition coupling graph construction module, used to construct a propagation-cognition coupling time-varying graph based on target harmful content, user interaction, group affiliation, and handling events, and to generate the propagation state of nodes. Cognitive state and vulnerability weight The affected subgraph extraction module is used to extract affected subgraphs based on time-series reachable exposure links and cross-group propagation links. And determine the local repair set The real-world cognitive recursion module is used to calculate the cognitive perturbation input of nodes based on the exposed content set and to update it by combining the group mean cognitive state and memory residue. Counterfactual constructs are used only when... Internally, the target exposed edge is deleted, downgraded, or the baseline information path of the same topic is replaced while maintaining... External statistical background equivalence to generate counterfactual diagrams and counterfactual cognitive state Differential bridging computation module, used by... and The difference calculation node cognitive effect increment and only when crossing Calculating the bridge acceptance effect contribution on cross-group paths Robust output module, used for... , Items with lasting impact , and standard deviation of uncertainty Generate robust effectiveness scores It also outputs the effectiveness level, key impact groups, and key transmission links.

[0010] (v) Beneficial effects of existing technologies

[0011] Compared with the prior art, the present invention has at least the following beneficial effects:

[0012] 1. More accurate local counterfactual construction. This invention extracts the affected subgraph and determines the local repair set. It only performs deletion, weight reduction, or replacement of baseline information paths of the same topic on the exposed edges of the target within the local repair set, while maintaining the statistical background equivalence outside the affected subgraph. This avoids the additional perturbation to the non-target background structure caused by full-graph replacement or global edge deletion in the prior art, and improves the causal attribution accuracy of the cognitive impact of harmful information of the target.

[0013] 2. More continuous cognitive state modeling. This invention incorporates the set of exposed content, node cognitive perturbation input, group mean cognitive state, and memory residue into the restricted recursive process of node cognitive state. This enables continuous characterization of the dynamic evolution of cognitive acceptance, emotional shift, behavioral tendency, and memory residue, thereby overcoming the shortcomings of existing technologies that only assess the strength of influence based on the popularity of dissemination or static indicators.

[0014] 3. Clearer attribution of bridging diffusion. This invention calculates bridging effectiveness contributions only on cross-group paths traversing local repair sets, thereby distinguishing cross-group bridging effects directly triggered by exposure to harmful target information from natural inter-group interactions, improving the accuracy of identifying key bridging paths and key impact groups.

[0015] 4. More robust evaluation output. This invention combines the node cognitive effectiveness increment, bridge acceptance effectiveness contribution, persistent influence term, vulnerability weight, and uncertainty standard deviation to generate a robust effectiveness score. This can suppress high uncertainty results and improve the stability and reliability of the output results, even when the propagation structure changes rapidly, observations are incomplete, and the model fluctuates.

[0016] 5. Enhanced engineering applicability. This invention can not only output the effectiveness level, but also the key impact groups and key transmission links, facilitating differentiated interventions, targeted measures, and trend warnings on the platform side. Therefore, it has good engineering deployability and application value. Attached Figure Description

[0017] Figure 1 This is a flowchart of the counterfactual effect assessment method for the cognitive impact of the spread of harmful information on the Internet provided in the first embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to preferred 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 following description, harmful content is uniformly referred to as... The time window is uniformly denoted as The nodes are uniformly denoted as or The content is uniformly recorded as The transmission path is uniformly recorded as The group is uniformly denoted as The same letters indicate the same meaning, and different letters indicate different meanings; the same terms indicate the same technical features, and different terms indicate different technical features. For commonly used technical terms involving English abbreviations or terms, a Chinese annotation is provided upon their first appearance. The method described in this paper can be deployed on platform risk control servers, public opinion analysis servers, cloud-edge collaborative nodes, or other computing devices with graph computing capabilities, and is executed by the processor calling program instructions from memory.

[0019] In this embodiment, the evaluation object is the target harmful content. The system describes the propagation and cognitive impact process across multiple discrete time windows. First, it collects content, user, interaction, group, and handling data from multiple platforms to construct a propagation-cognition coupling time-varying graph. Then, it identifies the affected subgraphs and local remediation sets of the target harmful content on this graph. Next, it calculates content-level harmful risk, node-level cognitive perturbation input, and the true cognitive state. Then, it performs counterfactual remediation only within the local remediation set to obtain the counterfactual cognitive state. Finally, it generates a robust effectiveness score based on the true-counterfact difference, bridging effectiveness contribution, persistent impact term, and uncertainty standard deviation, and outputs the effectiveness level, key affected groups, and key propagation links. These steps are interdependent and sequentially coupled, forming a complete closed loop around "accurately assessing the cognitive impact of target harmful content."

[0020] First Embodiment

[0021] Figure 1 This is a flowchart of the counterfactual effect assessment method for the cognitive impact of the spread of harmful information on the Internet provided in the first embodiment of the present invention, as shown below. Figure 1 As shown in Example 1, the method for assessing the counterfactual effect of cognitive impact on the spread of harmful information on the Internet includes the following steps.

[0022] S1. Multi-source data acquisition, unified time standardization, and basic sample construction

[0023] In this step, harmful content related to the target is collected. The relevant text, images, audio and video, external link summaries, posting behavior, forwarding behavior, commenting behavior, dwell time, click-to-expand, follow relationships, group relationships, handling feedback, and debunking feedback data are all aligned according to a unified time window to form a time window-level observation set:

[0024]

[0025] in, For the first A set of observations within a time window; For the first Content representation of the item; The identifier of the source node that published this content; The behavioral attributes are characterized by at least one or more of the following: publication time, number of interactions, dwell time, and click-to-expand behavior. For relationship records related to this content, at least one or more of the following must be included: forwarding relationship, commenting relationship, citation relationship, and following relationship; To address feedback, at least one or more of the following measures should be taken: deletion, traffic restriction, labeling, debunking, and manual review. For time windows The amount of content within.

[0026] In the data preprocessing stage, the raw data undergoes deduplication, anonymization, outlier removal, missing field completion, cross-platform entity alignment, and event clustering. For cross-platform entity alignment, a binary classification matching model is preferred to determine whether account pairs belong to the same entity. The training samples consist of positive and negative account pairs. Input features include at least username similarity, device fingerprint similarity, historical posting time distribution similarity, overlap of follow relationships, and semantic similarity of content. The training objective is to minimize the binary cross-entropy loss, thereby outputting a unified node identifier at the entity level. For group segmentation, group discovery is preferably performed based on follow relationships, interaction co-occurrence relationships, and content interest similarity to obtain the group label to which each node belongs. .

[0027] The output of this step is the time-stamped set of observations. Candidate node sets and group labels and content harmful to the target Relevant event topic tags. The above outputs serve as inputs for graph construction in step S2, propagation path extraction in step S3, and risk identification in step S4, respectively. By performing unified time-stamping and subject alignment on multi-source data, data fragmentation caused by different platform granularities, multiple accounts for the same subject, and multiple topic branches for the same event can be avoided, thereby improving the consistency and stability of subsequent graph modeling and cognitive assessment.

[0028] S2. Construction of variational graphs during propagation-cognitive coupling, generation of propagation states, and calculation of vulnerability weights.

[0029] Based on the observation set output in step S1 , construct the first The propagation of a time window—a time-varying construction of cognitive coupling:

[0030]

[0031] in, For time windows The propagation-cognitive coupling variation mapping; A set of nodes; Let it be the set of edges; This is a type mapping function used to map nodes and edges to predefined types; This is an attribute mapping function used to load content, behavior, relationship, and disposal features onto nodes and edges.

[0032] In this embodiment, the node set It includes at least content nodes, user nodes, group nodes, and event nodes; the edge set It includes at least the posting edge, forwarding edge, commenting edge, quoting edge, following edge, group affiliation edge, and action edge. For any node... Construct its original node features And the propagation state vector is extracted using HGNN (Heterogeneous Graph Neural Network). :

[0033]

[0034] in, For nodes In the time window The propagation state vector; It is a non-linear activation function; and The parameter matrix to be learned; For nodes In the time window The set of neighboring nodes; For nodes For nodes The right of the border; For the edge The type.

[0035] The edge weight It is determined by interaction strength, semantic similarity, historical influence residue, and latency decay:

[0036]

[0037] in, Interaction strength; For semantic similarity; Residual items from historical impacts; For time delay; , , and These are the weighting coefficients; Indicates a node Normalize all incoming edges.

[0038] After obtaining the propagation state vector Then, the node vulnerability weights are further calculated. :

[0039]

[0040] in, For nodes Vulnerability weights; This is the normalization function; This is an indicator of the extent of the impact on history. Indicators for the degree of slowness of recovery; For group-sensitive attribute indicators; For the sensitivity index of propagation location; , , and These are the weighting coefficients.

[0041] The HGNN preferably adopts a multi-task training method. The training samples consist of historical propagation graphs, node attributes, edge attributes, and supervision labels. The supervision labels include at least two of the following: key propagation node labels, next-time activation labels, and edge existence labels. The training objective is composed of a weighted average of node classification loss, link prediction loss, and next-time propagation state prediction loss, so that the model can both express the propagation structure and retain the ability to evolve over time.

[0042] The output of this step is a propagation-cognitive coupling time-varying model. Propagation state vector , border rights and vulnerability weight .in, and The path extraction and local repair set determination are used in step S3. and The cognitive perturbation input is used for calculation in step S4 and the robust effectiveness score is used for calculation in step S7. By jointly expressing the propagation structure and node vulnerability on the same graph, a unified and coupled structural basis can be provided for subsequent cognitive impact analysis.

[0043] S3. Determining the affected subgraph extraction and local repair set.

[0044] Variational mapping when propagation-cognition coupling is achieved Then, targeting harmful content. The affected subgraphs are extracted according to the temporal reachability of the exposure links and the cross-group propagation links. Preferably, the target harmful content is first recovered based on the content publication time, edge latency, and user contact records. The set of propagation paths Then for all nodes Determine whether it meets the following conditions:

[0045]

[0046] in, For time windows The set of nodes corresponding to the affected subgraph; Targeting harmful content nodes To the node The timing hop count is achievable; This is the upper limit for the number of jumps; For nodes Harmful content to the target The exposure indicator value is 1 if there is exposure, and 0 otherwise. To reach the node The set of propagation paths; Harmful content The set of propagation paths.

[0047] In determining the affected subgraph Then, further from Extracting local repair sets :

[0048]

[0049] in, For local repair collection; Indicates that by node Pointing to node The edge; Representing an edge For cross-group transmission edge; Represents the endpoint node of the edge. There is exposure to content harmful to the target.

[0050] When historical propagation logs are incomplete, a time-series link completion model is preferred to complete the missing edges. This model takes historical interaction sequences, topic consistency, and adjacency relationships as inputs and uses real next-hop propagation nodes as supervision signals for training, thereby obtaining a more complete set of propagation paths. .

[0051] This step outputs the affected subgraph. and local repair collection .in, Used to constrain the scope of subsequent counterfactual repairs. This is used to limit the target edges for deletion, demotion, or path replacement performed in step S5. By restricting counterfactual repair to... This avoids unnecessary disturbance to the background structure that is unrelated to the target harmful content, thus providing a more accurate local boundary for subsequent causal attribution.

[0052] S4, Content-level harmful risk identification, Node-level cognitive perturbation input calculation and real cognitive state deduction

[0053] This step first builds upon the content representation of step S1. Based on the propagation context of step S2, calculate the content-level harmful risk probability. and semantic perturbation vector Preferably, a multimodal fusion network is used, wherein the text sub-encoder preferably uses a Transformer, the image sub-encoder preferably uses a convolutional network, and the audio / video sub-encoder preferably uses a temporal convolutional network. The outputs of each modality are aggregated into a unified content representation at the fusion layer. :

[0054]

[0055] in, For the first The fusion representation of the content; The number of modes; For the first The fusion weights of each modality; For the first Encoding features of each modality; For content The probability of harmful risks; Use the Sigmoid activation function; and For risk output parameters; This is a semantic perturbation vector; and These are the output parameters for the disturbance.

[0056] The training samples of the multimodal fusion network consist of harmful content samples and non-harmful content samples. A weighted combination of cross-entropy loss and hard sample focusing loss is preferably used as the training objective. Furthermore, a metric constraint between positive and negative samples on the same topic can be added to improve the semantic perturbation vector. Distinctiveness.

[0057] Obtain the probability of harmful content. and semantic perturbation vector Then, for any node Calculate its time window The collection of exposed content within And further calculate the contact weight. With node-level cognitive perturbation input :

[0058]

[0059] in, For content For nodes Contact weight; For exposure frequency; Duration of contact; For interaction depth; , , and These are the weighting coefficients; For nodes Cognitive perturbation input; This is a matching function for semantic perturbations and propagation states; For nodes The propagation state vector.

[0060] In this embodiment, the matching function A bilinear mapping network is preferred, and its parameters are jointly trained with the cognitive state recursive model in an end-to-end manner.

[0061] Furthermore, define the node cognitive state vector:

[0062]

[0063] in, For cognitive acceptance; For the emotional shift component; As a behavioral tendency component; This is the residual amount of memory.

[0064] Preferably, the feasible region of cognitive state is defined as:

[0065]

[0066] in, The cognitive state is the feasible region, used to constrain the recursive results to remain within the interpretable interval.

[0067] For any node First, calculate the average cognitive state of the group to which it belongs:

[0068]

[0069] in, For nodes The group belongs to the time window The average cognitive state; For nodes Belonging to the same group The set of nodes.

[0070] Then, the node's cognitive state is updated using a GRU (Gated Recurrent Unit) type cognitive recursive function. Preferably, the update gating coefficient is calculated first. :

[0071]

[0072] in, To update the gating coefficients; and These are the gating parameters; This indicates vector concatenation.

[0073] Then calculate the true cognitive state using the following formula:

[0074]

[0075] in, This is a projection operator used to project intermediate results onto the feasible region. superior; It is a recursive function for cognitive states.

[0076] In this step, the data coupling relationship is: content-level harmful risk probability. and semantic perturbation vector First, contact weights The weighted aggregation forms node-level cognitive perturbation input. Node-level cognitive perturbation input Then, the state of propagation Group mean cognitive status and memory residue Input the cognitive state recursion function together to obtain the actual cognitive state. .

[0077] The training samples of the cognitive state recursive model consist of continuous time window samples. The supervision labels are preferably constructed based on changes in stance, emotional expression, whether similar content continues to be disseminated, and the duration of historical impact of subsequent content releases. A multi-task loss function is used to train each component. Through this hierarchical coupling process from content level to node level and from node level to group level, the true cognitive state can more accurately reflect the dynamic cognitive stimulation and sustained impact of harmful content on different nodes and groups.

[0078] S5. Construction of Local Counterfactual Diagrams, Background Equivalence Constraints, and Recursion of Counterfactual Cognitive States

[0079] After obtaining the affected subgraph Local repair collection After the recursive mechanism of the actual cognitive state, only Internal counterfactual repair was performed to obtain a counterfactual diagram. In this embodiment, it is preferable to first target content that is harmful to the target. Select a set of baseline information paths on the same topic from those paths that are of similar theme, low risk, have similar time distribution, and do not contain paths that spread harmful opinions. Then, use the local repair function. Generate a counterfactual graph:

[0080]

[0081] in, For time windows Counterfactual diagram; This is a local repair function; This is a set of baseline information paths for the same topic.

[0082] Specifically, local repair function right Edges in the process are subject to deletion, weight reduction, or path replacement. For any edge to be repaired... Its counterfactual rights It can be determined by the following formula:

[0083]

[0084] in, For the edge Edge weights in a counterfactual graph; For the set of edges to be deleted; For the set of reduced-weighted edges; For the replacement edge set; For the weighting coefficients, satisfying ; To be with the edge The edge weight of the corresponding baseline edge with the same theme.

[0085] To ensure that counterfactual repair does not significantly perturb the background structure outside the affected subgraph, a background equivalence constraint is applied to the counterfactual graph:

[0086]

[0087] in, Background equivalence bias; For the first One background statistic; For background equivalent tolerance; This represents the background portion after removing the affected subimage from the actual image; This represents the background portion after removing the affected subgraph from the counterfactual graph.

[0088] In this embodiment, the background statistics It should include at least two of the following: non-target node degree distribution, non-target edge delay distribution, and non-target group proportion. If the background equivalence is biased... Exceeding the threshold Then adjust the weighting coefficient. Or reselect the baseline path set Until the background equivalence constraint is satisfied.

[0089] In obtaining the counterfactual diagram Then, using the same propagation state generation model and cognitive state recursion model as in steps S2 and S4, the counterfactual propagation state is recalculated. Counterfactual disturbance input and counterfactual cognitive state Since real-world and counterfactual scenarios share the same propagation state model and the same cognitive recursion function, the differences between them arise only from local repair sets. Changes in the exposure pathway within the body can satisfy the basic requirements of causal control.

[0090] The parameters of the local repair rules are preferably calibrated through historical handling events. The training samples consist of "the actual propagation trajectory and cognitive results before and after actual deletion, traffic restriction, or refutation." The optimization objective is to minimize the deviation between the counterfactual inference results and the actual results after handling, while simultaneously minimizing the background equivalence bias. Through the combined effect of local counterfactual construction and background equivalence constraints, the propagation and cognitive impact directly caused by the target harmful content can be separated without destroying the background structure.

[0091] S6, Node Cognition Effectiveness Increment, Bridge Acceptance Effectiveness Contribution, and Critical Link Calculation

[0092] In obtaining the true state of cognition and counterfactual cognitive state Then, first calculate the incremental effect of node cognition:

[0093]

[0094] in, For nodes In the time window The increase in cognitive effectiveness; Mapping weight vectors to cognitive components; This is an operator that takes the non-negative part of a component.

[0095] In this embodiment, the Each element in the model corresponds to the mapping weights of the cognitive acceptance component, emotional bias component, behavioral tendency component, and memory residue component. The training samples can be composed of historical posterior influence strength labels, with manual evaluation of influence scores, subsequent propagation inhibition degree, and behavioral regression degree as supervision signals, and obtained through regression training.

[0096] Furthermore, only when traversing the local repair set Cross-group path set Calculate the bridge's acceptance efficiency contribution. For any path... Its admission coefficient Defined as: when path Time travel When the first and last nodes of the path belong to different groups and the path length does not exceed the preset upper limit. ;otherwise, Based on this, the bridge's acceptance efficiency contribution can be obtained. :

[0097]

[0098] in, To contribute to the bridge's effectiveness; For path Above The right of the border; For path The set of affected nodes on the endpoint side; The number of nodes in the set.

[0099] For any path Furthermore, the path effectiveness contribution value can be further defined:

[0100]

[0101] in, For path The contribution value of the effective path. For all candidate paths. Sort in descending order and take the first few. This path serves as a key propagation link.

[0102] In this step, the data coupling relationship is: actual cognitive state Counterfactual cognitive state First, couple into node-level components. ; Nodal grade components Then along the traversal of the local repair collection Cross-group paths, based on edge weights Perform path-level aggregation to obtain the bridge acceptance efficiency contribution. Path effect contribution value By calculating the bridging effectiveness contribution only on cross-group paths that have a causal relationship with the local repair set, it is possible to more accurately distinguish between "targeted harmful content-driven bridging effects" and "propagation effects caused by natural interactions between groups".

[0103] S7, Persistent Impact Item, Uncertainty Standard Deviation, Robust Effectiveness Score, and Evaluation Output

[0104] In order to obtain the incremental effect of node cognition Heqiao accepts the contribution of effectiveness Then, first calculate the persistent impact item. :

[0105]

[0106] in, For time windows The ongoing impact; The length of the backtracking window; Let be the time decay coefficient, satisfying ; For nodes In the historical time window Vulnerability weights; For nodes In the historical time window The cognitive effect increment.

[0107] Then, calculate the mean effective value. :

[0108]

[0109] in, The mean of the effect; The weighting coefficients for the bridge's effectiveness; The weighting coefficients for the terms with lasting impact.

[0110] To reflect the uncertainties caused by model errors, incomplete structural observations, and counterfactual construction biases, a combination of Monte Carlo Dropout and deep ensemble is preferred for repeated computation. The mean of the second effective value, denoted as the first The effective mean of the samples was Then the standard deviation of uncertainty The calculation is as follows:

[0111]

[0112] in, This is the sample mean; The standard deviation of uncertainty; The number of samples.

[0113] Furthermore, calculate the robust effectiveness score. :

[0114]

[0115] in, For robust and effective scores; This represents the uncertainty penalty coefficient.

[0116] In order to identify key influence groups, for any group Calculate the group influence score:

[0117]

[0118] in, For the group Impact score; For the group In the time window The set of nodes. For the entire group Sort the groups in descending order and select the top few groups as key impact groups.

[0119] For efficacy levels, robust efficacy scores can be used. Divide the data according to the relationship with the threshold interval, for example:

[0120]

[0121] in, For time windows The level of effectiveness; , and The threshold for classifying levels, and satisfying... .

[0122] The output of this step is the robust effectiveness score. Effectiveness level The ranking results of key impact groups and key propagation links are presented. By incorporating node differential components, bridging diffusion, persistent impact, and uncertainty penalties into a unified scoring framework, misjudgments caused by highly volatile assessment results can be suppressed, improving the robustness, interpretability, and engineering applicability of the output results.

[0123] S8, Access-based Online Updates

[0124] To adapt to changes in network propagation scenarios, handling strategies, and user behavior, this embodiment also establishes an admission-based online update mechanism. Model updates are triggered only when the counterfactual construction satisfies the background equivalence constraint and the deviation between the robust effectiveness score and the posterior observation effectiveness label exceeds a threshold. The update indicator is defined as:

[0125]

[0126] in, To update the indicator value; Background equivalence bias; For background equivalent tolerance; For posterior observation effectiveness labels; For robust and effective scores; To update the threshold.

[0127] when At the same time, using newly added handling feedback, rumor debunking feedback, dissemination regression data, and manually verified labels, the harmful risk identification model, cognitive state recursive function, cognitive component mapping weights, and uncertainty penalty coefficient are jointly updated. Preferably, the joint loss function can be expressed as:

[0128]

[0129] in, For the joint loss function; Identify losses for harmful risks; The loss is calculated by recursion of cognitive states. To construct a calibration loss for counterfactual purposes; To ensure robust effectiveness, the score regresses to the loss. , , and This is the loss weighting coefficient.

[0130] The purpose of this step is to update parameters only when the counterfactual scenario is reasonably constructed and there is a significant deviation between the existing model and the posterior results. This avoids introducing abnormal data or unreliable counterfactual samples into the model, thereby reducing the risk of online drift and improving the stability and reliability of the model in the long run.

[0131] Second Embodiment

[0132] The second embodiment of the present invention only describes the contents that are different from those of the first embodiment; the contents that are the same will not be described again.

[0133] Based on the method in Embodiment 1, this invention also provides a counterfactual effectiveness assessment system for the cognitive impact of the spread of harmful information online. This system includes a propagation-cognition coupling graph construction module, an affected subgraph extraction module, a true cognition recursion module, a counterfactual construction module, a differential bridging calculation module, and a robust output module.

[0134] The propagation-cognition coupling graph construction module is used to execute steps S1 and S2, completing multi-source data acquisition, unified time-stamping processing, propagation-cognition coupling time-varying graph construction, propagation state vector generation, and vulnerability weight calculation; the affected subgraph extraction module is used to execute step S3 to extract the affected subgraph. And determine the local repair set The Real Cognition Recursion Module is used to execute step S4, completing content-level harmful risk identification, node-level cognitive perturbation input calculation, and real cognitive state recursion; the Counterfactual Construction Module is used to execute step S5, based on the local repair set. and the same topic baseline information path set Generate counterfactual graphs And calculate the counterfactual cognitive state The differential bridging computation module is used to execute step S6, calculating the cognitive effect increment of the node. Bridge acceptance effect contribution and contribution value of key propagation links The robust output module is used to execute steps S7 and S8, calculating the persistent impact item. Uncertainty Standard Deviation Steady and effective score It outputs the effectiveness level, key impact groups, and key propagation links, and performs online model updates when update admission conditions are met.

[0135] The modules described above can be physically deployed on the same server or distributed across multiple servers or container nodes, and interact with each other through shared memory, message queues, or remote procedure call interfaces. The data coupling relationship between the modules is consistent with the step coupling relationship in Example 1: the data output by the previous module serves as the input of the next module, thereby ensuring a one-to-one correspondence between the system implementation and the method steps.

[0136] Third Embodiment

[0137] The third embodiment of the present invention only describes the content that is different from the first embodiment; the same content will not be described again.

[0138] A third embodiment of the present invention also provides a computer program product. The computer program product includes program instructions stored in a non-transitory computer-readable storage medium. When the program instructions are loaded and executed by a processor, the processor performs steps S1 to S8 as described in Embodiment 1. The non-transitory computer-readable storage medium may be a hard disk, a solid-state drive, a read-only memory, a random access memory, flash memory, a disk array, a network-attached storage device, or other storage media capable of storing program code.

[0139] In a preferred embodiment, the program product is deployed in the content security analysis platform in the form of a container image, service package, or platform plugin. The program instructions invoke the graph computing engine, deep learning inference engine, and rule scheduling engine to collaboratively complete propagation-cognitive coupling modeling, local counterfactual construction, differential effectiveness calculation, and robust output. The execution process of the program product is consistent with the method steps in Embodiment 1, and therefore will not be described again.

[0140] The specific embodiments of the present invention have been described in detail above with reference to preferred embodiments. Those skilled in the art should understand that, without departing from the concept and essence of the present invention, various equivalent substitutions, combinations, deletions, or modifications can be made to the technical features in the above embodiments. For example, changing the network structure of part of the model, replacing part of the loss function form, adjusting part of the threshold, or adopting other equivalent local counterfactual repair rules; however, as long as it still uses the affected subgraph and the local repair set to limit the counterfactual scope, uses the difference between real and counterfactual cognitive states to calculate the node cognitive effectiveness increment, and combines the bridge acceptance effectiveness contribution, the persistent influence term, and the uncertainty standard deviation to generate a robust effectiveness score, all should fall within the protection scope defined by the claims of the present invention.

Claims

1. A robust counterfactual effectiveness assessment method for the cognitive impact of harmful online information dissemination, characterized in that, include: A propagation-cognition coupling time-varying graph is constructed based on target harmful content, user interaction, group affiliation, and handling events, and propagation states are configured for nodes. Cognitive state and vulnerability weight ; Based on the temporal reach and cross-group propagation links of the target harmful content, the affected subgraph is extracted. And determine the local repair set ; Based on the set of exposed content, the cognitive perturbation input of the calculated nodes is combined with the group mean cognitive state and memory residue. By performing restricted recursion, the true cognitive state can be obtained; only when... Internally, for targets exposed at certain edges, perform deletion, downgrading, or replacement of baseline information paths for the same topic, while maintaining... Equivalent external statistical background to generate counterfactual diagrams And obtain a counterfactual cognitive state ;according to and The difference calculation node cognitive effect increment And only when crossing Calculating the bridge acceptance effect contribution on cross-group paths ;based on , Items with lasting impact , and standard deviation of uncertainty Generate robust effectiveness scores Based on this, the effectiveness level, key impact groups, and key transmission links are output.

2. The method according to claim 1, characterized in that, The affected subgraph satisfy: in, Targeting harmful content nodes To node The timing hop count is achievable; This is the upper limit for the number of jumps; For nodes For nodes Exposure indicator; To reach the node The set of propagation paths; The set of propagation paths corresponding to the target harmful content; the set of local repairs Located within the affected subgraph Furthermore, the endpoint is composed of edges that cross group bridging nodes or exposed user nodes.

3. The method according to claim 1, characterized in that, The node cognitive perturbation input satisfy: in, For content For nodes Contact weight; This is the normalization function; For exposure frequency; Duration of contact; For interaction depth; For nodes The collection of exposed content; For content The probability of harmful risks; For content The semantic perturbation vector; This is a matching function for semantic perturbations and propagation states; , , and These are the weighting coefficients.

4. The method according to claim 3, characterized in that, The cognitive state Including cognitive acceptance component Emotional shift component Behavioral tendency component and residual memory And satisfy: in, For feasible regions Projection operator on; To update the gating coefficients; For cognitive state recursion function; For nodes Group The average cognitive state.

5. The method according to claim 4, characterized in that, The counterfactual diagram Through local repair function Generate, and satisfy: in, A set of baseline information paths for the same topic; The equivalent deviation of the background outside the affected sub-image; For the first One background statistic; For equivalent tolerance; the background statistics include at least two of the following: non-target node degree distribution, non-target edge delay distribution, and non-target group proportion.

6. The method according to claim 5, characterized in that, The incremental effect of node cognition satisfy: in, Mapping weight vectors to cognitive components; The operator for extracting the non-negative part by component; Cognitive Acceptance Component Emotional shift component Behavioral tendency component and residual memory Assign weights separately.

7. The method according to claim 6, characterized in that, The bridge receives performance contribution Only when traversing the local repair set Cross-group path set The above calculations are performed, and the following conditions are met: in, This refers to the path admission coefficient; For the edge The right to the side; For path The set of affected nodes on the endpoint side.

8. The method according to claim 7, characterized in that, The persistent impact item and robust effectiveness score They respectively satisfy: in, The length of the backtracking window; This is the time decay coefficient; This represents the uncertainty penalty coefficient. The mean of the effect; and These are the weighting coefficients.

9. The method according to claim 8, characterized in that, This also includes access-based online updates: only when... and Only then is the cognitive state recursive function updated using the handling feedback, debunking feedback, and behavioral regression results. The matching function The mapping weight vector and the uncertainty penalty coefficient ,in: ,in, To update the indicator value; For posterior observation effectiveness labels; To update the threshold.

10. A robust counterfactual effectiveness assessment system for the cognitive impact of harmful online information dissemination, characterized in that, include: The propagation-cognition coupling graph construction module is used to construct a propagation-cognition coupling time-varying graph based on target harmful content, user interaction, group affiliation, and handling events, and to generate the propagation state of nodes. Cognitive state and vulnerability weight The affected subgraph extraction module is used to extract affected subgraphs based on time-series reachable exposure links and cross-group propagation links. And determine the local repair set ; The real-world cognitive recursion module is used to calculate the cognitive perturbation input of nodes based on the exposed content set and update it by combining the group mean cognitive state and memory residue. ; Counterfactual constructors are used only when... Internally, the target exposed edge is deleted, downgraded, or the baseline information path of the same topic is replaced while maintaining... External statistical background equivalence to generate counterfactual diagrams and counterfactual cognitive state ; Differential bridging computation module, used by and The difference calculation node cognitive effect increment and only when crossing Calculating the bridge acceptance effect contribution on cross-group paths Robust output module, used for... , Items with lasting impact , and standard deviation of uncertainty Generate robust effectiveness scores It also outputs the effectiveness level, key impact groups, and key transmission links.

Images