A rumor prediction method based on cognitive collaboration across domains

Abstract

The present application belongs to the technical field of social network information propagation prediction, and particularly relates to a cross-domain rumor prediction method based on cognitive synergy, comprising: acquiring cross-domain rumor propagation dynamic graphs, user behavior characteristics and propagation domain information within a historical time window; generating an initial cognitive portrait of a user based on community discovery, cognitive depth, cognitive breadth and information credibility, and performing cognitive representation alignment through cross-domain comparative learning; aggregating neighbor cognitive states according to a current propagation topology, and updating the dynamic cognitive state of the user in combination with anti-rumor exposure signals; generating a continuous time-varying game inhibition matrix according to propagation heat, bridging centrality and node influence; updating the edge weight of the propagation topology based on the dynamic cognitive state of the user and the game inhibition matrix, and outputting a future propagation trend prediction result through a dynamic graph neural network and a time convolution network. The present application can depict the closed-loop feedback relationship between user cognition, game inhibition and propagation topology, and improve the stability and interpretability of cross-domain rumor propagation prediction.

Description

Technical Field

[0001] This invention belongs to the field of social network information dissemination prediction technology, specifically involving a cross-domain rumor prediction method based on cognitive collaboration, and more particularly involving a cross-domain rumor dissemination trend prediction method that integrates user dynamic cognitive updates, domain-selective game inhibition, and dissemination topology collaborative evolution. Background Technology

[0002] With the rapid development of social networking platforms, rumors can spread rapidly in a short period of time through interactive behaviors such as forwarding, commenting, replying, and following, migrating between different topics, communities, and communication scenarios. Compared with the spread of rumors within a single communication domain, cross-domain rumor spread not only exhibits significant temporal dynamics but is also influenced by multiple factors, including differences in user cognition, changes in the spread topology, and intervention by anti-rumor information. Therefore, accurately predicting the spread of cross-domain rumors has become a crucial technical issue in social network governance, public opinion monitoring, and information security prevention and control.

[0003] Existing methods for predicting rumor spread typically model the spread based on user behavior characteristics, propagation structure characteristics, or time series characteristics. One type of method uses user profiles, historical interaction behavior, and interest communities to characterize users' acceptance tendencies towards rumor information; another type utilizes graph neural networks, temporal convolutional networks, or recurrent neural networks to model the dynamic changes in the spread structure and popularity of rumors; yet another type focuses on the adversarial relationship between rumor information and debunking information, reducing the scope of rumor spread through intervention strategies or influence control methods. Despite the progress made in rumor detection and prediction, the following significant technical bottlenecks remain:

[0004] 1. The Staticity and Cross-Domain Heterogeneity of User Cognition Modeling. Existing user cognition modeling largely relies on static profiles or simple time-series statistics, neglecting the dynamic evolution of user cognition during rumor propagation. Specifically, users are continuously influenced by neighboring propagation behaviors, exposure to anti-rumor information, and changes in the propagation environment during rumor propagation; their cognitive state is not fixed. Furthermore, existing technologies often fail to effectively address the problem of cognitive heterogeneity in cross-domain scenarios, making it difficult to accurately characterize the dynamic evolution of user cognition during propagation.

[0005] 2. The uniformity and game complexity of the suppression mechanism. The impact of debunking information on the propagation path is non-uniform. In actual social networks, debunking information is usually more likely to concentrate its effect on key propagation areas where high-popularity, high-bridging centrality, or high-influence nodes are located, rather than acting uniformly on all propagation edges. Existing methods often adopt uniform intervention or overall suppression approaches, lacking a quantitative assessment of the value of the propagation domain, and making it difficult to characterize the differentiated game-theoretic suppression effects in different propagation domains.

[0006] 3. Fragmentation and Error Accumulation Between Modules. Existing prediction frameworks are mostly "pipeline" structures, performing feature extraction, structural modeling, and prediction output sequentially. This architecture results in a lack of closed-loop feedback modeling between user cognition, game-theoretic inhibition, and propagation topology. The lack of feedback constraints between modules easily leads to the propagation and accumulation of cognitive estimation bias, topological evolution bias, and prediction errors across different modules, thereby reducing the stability and interpretability of the prediction. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a cross-domain rumor prediction method based on cognitive collaboration, which sequentially executes the following eight steps in the same cross-domain rumor propagation prediction process, including:

[0008] S1: Obtain the dynamic graph sequence of cross-domain rumor propagation, user behavior feature sequence, propagation domain label, historical propagation popularity, bridging centrality and node influence within the historical time window;

[0009] S2: Perform community discovery based on the propagation subgraph of each propagation domain obtained in step S1, and generate a multi-granular initial cognitive profile based on the user's historical interaction intensity, topic coverage, and historical content credibility.

[0010] S3: Use cross-domain contrastive learning to align the representation of the multi-granularity initial cognitive profile generated in step S2 to obtain a cross-domain consistent user initial cognitive representation;

[0011] S4: Based on the current propagation topology aggregated neighbor cognitive state obtained in step S1, and combined with the neighbor cognitive state in the cross-domain consistent user initial cognitive representation obtained in step S3, and combined with the anti-rumor exposure signal, update the user dynamic cognitive state through the gated loop unit.

[0012] S5: Calculate the propagation domain value based on the propagation heat, bridging centrality and node influence obtained in step S1, and generate a continuous time-varying game inhibition matrix by combining the node representations at both ends of the edge.

[0013] S6: Update the propagation topology edge weights based on the user dynamic cognitive state obtained in step S4 and the game inhibition matrix generated in step S5 to obtain the evolutionary propagation topology;

[0014] S7: Input the evolutionary propagation topology obtained in step S6 and the user dynamic cognitive state obtained in step S4 into a dynamic graph neural network to obtain a spatial propagation representation, and input the spatial propagation representation of continuous time steps into a temporal convolutional network to obtain a spatiotemporal propagation representation;

[0015] S8: Based on the spatiotemporal propagation representation obtained in step S7, output the prediction result of the future propagation situation, and use the propagation prediction loss, cross-domain cognitive alignment loss, game-theoretic inhibition sparsity constraint and closed-loop consistency constraint to jointly train the model.

[0016] The beneficial effects of this invention are:

[0017] This invention constructs an initial cognitive profile of a user through multi-granularity cognitive factors (cognitive depth, cognitive breadth, and information credibility), and dynamically updates the user's cognitive state by combining neighbor propagation behavior and anti-rumor exposure signals. At the same time, it generates a continuous time-varying game inhibition matrix based on propagation heat, bridging centrality, and node influence. Unlike existing technologies that only use static cross-domain feature alignment and only use meta-learning domain adaptation, this invention, through cross-domain contrastive learning alignment guided by cognitive profile and a continuous time-varying game inhibition matrix, can characterize the synergistic relationship between changes in user cognition, anti-rumor inhibition, and propagation topology evolution.

[0018] This invention drives the propagation topology edge weight update jointly by the user's dynamic cognitive state and the game inhibition matrix, and combines dynamic graph neural networks, temporal convolutional networks and closed-loop consistency constraints for joint prediction. Unlike existing technologies that use dynamic graph convolutional networks without topology edge weight evolution or static game payoff matrices without deep learning components, this invention reduces the cascading error between modules and improves the stability and interpretability of cross-domain rumor propagation prediction through a cognitive-game collaborative topology evolution mechanism and spatiotemporal propagation representation learning jointly using DGNN and TCN. Attached Figure Description

[0019] Figure 1 This is a flowchart of a cross-domain rumor prediction method based on cognitive collaboration according to the present invention.

[0020] Figure 2 A flowchart for user cognition alignment and representation;

[0021] Figure 3 This is a flowchart of the spatiotemporal joint prediction process. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] This invention proposes a cross-domain rumor prediction method based on cognitive collaboration, such as... Figure 1 As shown, it includes:

[0024] S1: Obtain the dynamic graph sequence of cross-domain rumor propagation, user behavior feature sequence, propagation domain label, historical propagation popularity, bridging centrality and node influence within the historical time window;

[0025] S2: Perform community discovery based on the propagation subgraph of each propagation domain, and generate multi-granular initial cognitive profiles based on the user's historical interaction intensity, topic coverage, and historical content credibility.

[0026] S3: Use cross-domain contrastive learning to align the representations of the multi-granularity initial cognitive profiles to obtain cross-domain consistent user initial cognitive representations;

[0027] S4: Aggregate the cognitive status of neighbors based on the current propagation topology, and combine it with the anti-rumor exposure signal to update the user's dynamic cognitive status through the gated loop unit;

[0028] S5: Calculate the value of the propagation domain based on the propagation heat, bridging centrality, and node influence, and generate a continuous time-varying game inhibition matrix by combining the node representations at both ends of the edge;

[0029] S6: Update the propagation topology edge weights based on the user's dynamic cognitive state and the game inhibition matrix to obtain the evolutionary propagation topology;

[0030] S7: Input the evolutionary propagation topology and the user's dynamic cognitive state into a dynamic graph neural network to obtain a spatial propagation representation, and input the spatial propagation representation of continuous time steps into a temporal convolutional network to obtain a spatiotemporal propagation representation;

[0031] S8: Output the prediction results of future propagation trends based on the spatiotemporal propagation representation, and jointly train the model using propagation prediction loss, cross-domain cognitive alignment loss, game-theoretic inhibition sparsity constraint and closed-loop consistency constraint.

[0032] Methods for obtaining cross-domain rumor propagation data include using social media platform interfaces, web crawling technology, publicly available rumor propagation dataset platforms, or third-party social media data services to obtain data such as rumor events, user interaction relationships, propagation timestamps, user behavior information, and propagation domain tags.

[0033] In this embodiment, step S1: Obtaining the cross-domain rumor propagation dynamic graph sequence, user behavior feature sequence, propagation domain label, historical propagation popularity, bridging centrality, and node influence within a historical time window, including:

[0034] S11: Given a set of time steps At any moment Constructing a dynamic graph of cross-domain rumor propagation ;

[0035] S12: The cross-domain rumor propagation dynamic graph is defined as follows:

[0036]

[0037] in, For a set of user nodes, For a moment The set of propagation edges, For a weighted adjacency matrix, For user behavior feature matrix, For the set of propagation domains, where ; ; ;

[0038] S13: If This indicates that the user With users At any moment The presence of forwarding, commenting, following, or interaction relationships indicates the intensity of dissemination.

[0039] S14: Form a historical propagation dynamic graph sequence by using the propagation dynamic graph of continuous time steps. And simultaneously acquire user behavior feature sequences Propagation domain set Historical popularity Bridge centrality and node influence This serves as the input for subsequent dynamic cognitive updates, game inhibition matrix generation, and propagation topological evolution.

[0040] In this embodiment, step S2: Community discovery is performed based on the propagation subgraph of each propagation domain, and a multi-granularity initial cognitive profile is generated according to the user's historical interaction intensity, topic coverage, and historical content credibility, including:

[0041] S21: Based on the propagation domain label From time Animated graph of cross-domain rumor spread Extraction from the middle belongs to the propagation domain Construct a propagation domain subgraph by identifying user nodes, propagation edges, and their weights. Its definition is:

[0042]

[0043] in, Indicates time Belongs to the propagation domain The set of user nodes, Indicates time Belongs to the propagation domain The set of propagation edges, Represents the propagation domain subgraph The corresponding weighted adjacency matrix;

[0044] S22: In the propagation domain subgraph The above uses a community discovery method based on modularity optimization to divide user interest communities. The modularity function is defined as follows:

[0045]

[0046] in, Represents the propagation domain subgraph Chinese users With users The right to the border, and Representing users respectively With users In the propagation domain subgraph The degree of nodes in Represents the propagation domain subgraph The sum of the edge weights in the middle, and Representing users respectively With users The community to which it belongs A function indicating community affiliation;

[0047] S23: After completing the community segmentation within the propagation domain, based on users Historical interaction intensity calculation of cognitive depth According to the user The breadth of knowledge is calculated based on the range of topics or communities covered. According to the user The credibility of information is calculated based on the credibility of historically published or forwarded content. ;

[0048] S24: User The initial cognitive profile is defined as:

[0049]

[0050] in, Indicates user The depth of their understanding reflects the intensity of their historical interaction with related topics; It indicates the breadth of knowledge, reflecting the range of topics or communities covered by the participant; It indicates the credibility of information and reflects the degree of credibility of its historically published or forwarded content; , and These are the weight parameters.

[0051] In this embodiment, step S3: aligning the representation of the multi-granularity initial cognitive profile using cross-domain contrastive learning includes:

[0052] S31: For users Cross-domain positive samples and negative sample set A cross-domain cognitive alignment loss is constructed. The cross-domain cognitive alignment loss is defined as follows:

[0053]

[0054] in, Represents cosine similarity. Temperature coefficient;

[0055] S32: By minimizing the cross-domain cognitive alignment loss, the cognitive representations of the same user or similar users in different propagation domains are brought closer together, while the cognitive representations of unrelated users are spaced further apart, resulting in a consistent initial cognitive representation of the user across domains.

[0056] In this embodiment, step S4: Based on the current propagation topology, aggregate the neighbor's cognitive state and combine it with the anti-rumor exposure signal, update the user's dynamic cognitive state through the gating loop unit, such as... Figure 2 As shown, it specifically includes:

[0057] S41: Let Indicates time The user cognitive state matrix, where Indicates user The cognitive state vector. The general update form of the user's cognitive state is:

[0058] ),

[0059] in, The results of neighbor perception aggregation For users At any moment Received debunking or exposure suppression signals, For the cognitive state update function, .

[0060] S42: The results of the neighbor cognition aggregation and the influence weight of neighbor perception Defined as:

[0061]

[0062] in, and For learnable projection matrices, Indicates user For users Neighbor perception influences weight;

[0063] S43: The aforementioned anti-rumor exposure signal Defined as:

[0064]

[0065] in, For the edge On the inhibition strength, Indicates user At any moment The intensity of anti-rumor exposure or the weight of the role in debunking rumors;

[0066] S44: Will , and Input gated loop unit, user Cognitive state through Gating mechanism update:

[0067]

[0068]

[0069]

[0070]

[0071] in, , and For learnable weight matrix, , and For bias terms, To update the door, To reset the door, This represents element-wise multiplication. Through the above process, static cognitive factors are extended to time-varying user cognitive states that change with the propagation process.

[0072] Through the above process, static cognitive factors are extended to time-varying user cognitive states that change with the propagation process. Unlike the existing technology where user cognitive states are modeled as static profiles, this invention achieves dynamic updates of user cognitive states through neighbor cognitive aggregation, anti-rumor exposure signals, and GRU gating mechanisms.

[0073] In this embodiment, step S5: Calculate the propagation domain value based on propagation heat, bridging centrality, and node influence, and generate a continuous time-varying game inhibition matrix by combining the node representations at both ends of the edge, including:

[0074] S51: Let Indicates time The inhibition matrix of continuous game, where Indicates the propagation edge At any moment The intensity of suppression. For the propagation domain. Its communicative value is defined as:

[0075]

[0076] in, Indicates the propagation domain The popularity of the spread Indicates bridging centrality. This represents the average influence of key nodes within the domain. , and For learnable weight parameters, ;

[0077] S52: Generate continuous suppression strength based on propagation value and edge node representation:

[0078]

[0079] in, and The node representation is obtained from the dynamic graph encoder. This represents vector concatenation. For the Sigmoid function, The larger the value, the stronger the inhibition of the corresponding propagation path;

[0080] S53: If it is necessary to explicitly represent the mapping from the policy to the inhibition matrix, then it is denoted as:

[0081]

[0082] in, This represents the distribution of domain-level regulation strategies;

[0083] S54: After obtaining the propagation domain value, further generate the domain-level suppression strength:

[0084]

[0085] in, and For learnable parameters, Indicates the propagation domain The overall inhibition strength;

[0086] S55: To avoid the model imposing excessive suppression on all propagation edges, a game-theoretic suppression sparsity constraint is introduced, concentrating the suppression signal more on high-value propagation paths. Unlike existing technologies that employ uniform intervention or overall suppression, this invention's domain-selective game-theoretic suppression can adjust the suppression strength of each propagation edge according to the value differences within the propagation domain. The sparsity constraint is as follows:

[0087]

[0088] In this embodiment, step S6: updating the propagation topology edge weights according to the user's dynamic cognitive state and the game inhibition matrix includes:

[0089] S61: Given time propagation topology User cognitive status Suppression Matrix of Game Theory The edge weight is defined as follows:

[0090]

[0091] in, This indicates the weakening effect of game-theoretic inhibition on the propagation edge. This indicates the impact of user cognitive similarity on the probability of propagation.

[0092] S62: Based on the edge weights of all propagated edges in the next time step The evolved propagation topology is obtained. The evolved propagation topology is then used for subsequent dynamic graph neural network structure encoding.

[0093] In this embodiment, step S7: The evolutionary propagation topology and the user's dynamic cognitive state are input into a dynamic graph neural network to obtain a spatial propagation representation, and the spatial propagation representation of consecutive time steps is input into a temporal convolutional network to obtain a spatiotemporal propagation representation, such as... Figure 3 As shown, it specifically includes:

[0094] S71: The adjacency matrix after adding self-connections and Correspondence Matrix Input a dynamic graph neural network and extract the dynamic graph structure representation according to the following formula:

[0095]

[0096] in, This represents the adjacency matrix after adding self-joins. This is the degree matrix. This indicates that the dynamic user cognitive state is used as the initial node feature. For the first Layered graph neural network parameters, It is a non-linear activation function;

[0097] S72: In the time-dependent modeling stage, the graph representation of continuous time steps is input into the temporal convolutional network to extract the trend of the spread popularity over time.

[0098]

[0099] in, The time window length, The number of layers in the graph neural network is specified. A temporal convolutional network is used to capture propagation changes at different time scales, resulting in a spatiotemporal propagation representation for predicting future propagation patterns. .

[0100] In this embodiment, step S8: Based on the spatiotemporal propagation representation, output the prediction result of the future propagation situation, and jointly train the model using propagation prediction loss, cross-domain cognitive alignment loss, game-theoretic suppression sparsity constraint, and closed-loop consistency constraint, including:

[0101] S81: Given a history window The model aims to analyze the propagation topology, user cognitive state, and game inhibition matrix within the context of future events. Predicting the spread trend at each time step:

[0102]

[0103] in, It can indicate future popularity, scale of forwarding, scope of influence of rumors, or cross-domain migration trends. Let be the propagation prediction function to be learned;

[0104] S82: Learn parameters based on predicted and actual propagation trends. Its optimization objective is:

[0105]

[0106] S83: The dynamic cognitive update, domain-selective game regulation, propagation topology evolution, and propagation situation prediction in the model are represented by the following closed-loop coupled equations:

[0107]

[0108] S84: Input the spatiotemporal representation into the multilayer perceptron, and output the prediction result of the future propagation trend:

[0109]

[0110] S85: To reduce the cascading errors between modules, a closed-loop consistency constraint is introduced. This differs from the lack of feedback constraints between modules in traditional pipeline-style prediction methods in existing technologies. It is defined as follows:

[0111]

[0112] in, , and For closed-loop consistency weights;

[0113] S86: The final optimization objective during model training consists of the propagation situation prediction loss, cross-domain cognitive alignment loss, game-theoretic inhibition sparsity constraint, and closed-loop consistency constraint.

[0114]

[0115] in, To predict losses in the spread of the virus, For cross-domain cognitive alignment loss, To suppress sparsity constraints in game theory, This is a closed-loop consistency constraint. , and This is the balance coefficient;

[0116] S87: Update the dynamic cognitive update function via backpropagation Domain-selective game adjustment function Propagation topological evolution function and propagation trend prediction function The model parameters are used to obtain the trained cross-domain rumor propagation prediction model.

[0117] In this embodiment, the future dissemination trend prediction results include at least one of the following: future dissemination popularity prediction results, future dissemination scale prediction results, future rumor impact range prediction results, and cross-domain migration trend prediction results. The method also outputs a dynamic user cognitive state matrix. Domain-selective game inhibition matrix and the evolved propagation topology It is used to characterize the closed-loop feedback relationship between changes in user cognition, the role of anti-rumor suppression, and the evolution of propagation paths.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cross-domain rumor prediction method based on cognitive collaboration, characterized in that, The following steps are executed sequentially within the same cross-domain rumor propagation prediction process: S1: Obtain the dynamic graph sequence of cross-domain rumor propagation, user behavior feature sequence, propagation domain label, historical propagation popularity, bridging centrality and node influence within the historical time window; S2: Perform community discovery based on the propagation subgraph of each propagation domain obtained in step S1, and generate a multi-granular initial cognitive profile based on the user's historical interaction intensity, topic coverage, and historical content credibility. S3: Use cross-domain contrastive learning to align the representation of the multi-granularity initial cognitive profile generated in step S2 to obtain a cross-domain consistent user initial cognitive representation; S4: Based on the current propagation topology aggregated neighbor cognitive state obtained in step S1, and combined with the neighbor cognitive state in the cross-domain consistent user initial cognitive representation obtained in step S3, and combined with the anti-rumor exposure signal, update the user dynamic cognitive state through the gated loop unit. S5: Calculate the propagation domain value based on the propagation heat, bridging centrality and node influence obtained in step S1, and generate a continuous time-varying game inhibition matrix by combining the node representations at both ends of the edge. S6: Update the propagation topology edge weights based on the user dynamic cognitive state obtained in step S4 and the game inhibition matrix generated in step S5 to obtain the evolutionary propagation topology; S7: Input the evolutionary propagation topology obtained in step S6 and the user dynamic cognitive state obtained in step S4 into a dynamic graph neural network to obtain a spatial propagation representation, and input the spatial propagation representation of continuous time steps into a temporal convolutional network to obtain a spatiotemporal propagation representation; S8: Based on the spatiotemporal propagation representation obtained in step S7, output the prediction result of the future propagation situation, and use the propagation prediction loss, cross-domain cognitive alignment loss, game-theoretic inhibition sparsity constraint and closed-loop consistency constraint to jointly train the model.

2. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S1 obtains cross-domain rumor propagation data, including: (a) Divide the historical dissemination process into multiple consecutive time steps according to the dissemination timestamp, take the users who participated in the dissemination as nodes, and take the forwarding, commenting, replying, following or interaction relationships as dissemination edges to construct a cross-domain rumor dissemination dynamic graph corresponding to each time step; (b) Extract user posting, forwarding, commenting, replying and topic participation behaviors from the dissemination data at each time step to form a user behavior feature sequence, and determine the dissemination domain label based on the topic, platform source or dissemination scenario of the rumor event; (c) Statistically analyze the historical propagation activity, cross-community or cross-domain connectivity, and key user influence of each propagation domain to obtain propagation popularity, bridging centrality, and node influence, which serve as inputs for subsequent cognitive updates, game inhibition, and topological evolution.

3. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S2 generates a multi-granularity initial cognitive profile, including: (a) Construct a propagation subgraph for each propagation domain and use a community discovery method based on modularity optimization to divide user interest communities; (b) Determine the depth of cognition based on the intensity of users’ historical posts, reposts, comments and replies around relevant topics; determine the breadth of cognition based on the coverage of topics, communities or dissemination domains in which users participate; and determine the credibility of information based on the credibility of content posted or reposted by users in the past. (c) Weighted fusion of cognitive depth, cognitive breadth and information credibility is performed to obtain multi-granular initial cognitive profiles for each user.

4. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S3 performs cross-domain cognitive representation alignment, including: (a) Construct cross-domain positive sample pairs by constructing the cognitive profiles of the same user in different communication domains or the cognitive profiles of users with similar historical behaviors, and construct cross-domain negative sample pairs by constructing the cognitive profiles of users whose historical behavior differences meet the preset conditions. (b) The cognitive profiles of users in different propagation domains are mapped to the same representation space by a cognitive encoder, and the similarity of cognitive representations between sample pairs is calculated. (c) By adopting cross-domain contrastive learning objectives, the representation distance between positive sample pairs is reduced and the representation distance between negative sample pairs is increased, resulting in cross-domain consistent initial cognitive representations of users.

5. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S4 updates the user's dynamic cognitive state, including: (a) Determine the set of neighboring users of the target user based on the current propagation topology, and calculate the neighboring cognitive influence weight based on the propagation relationship and the user's cognitive state; (b) Aggregate the cognitive states of neighboring users according to the neighbor cognitive influence weight to obtain the neighbor cognitive aggregation result of the target user; (c) Determine the anti-rumor exposure information of the target user based on the debunking information, clarification information, anti-rumor interaction information received by the target user and the suppression strength of adjacent propagation edges; (d) Input the target user’s previous time step cognitive state, neighbor cognitive aggregation results and anti-rumor exposure information into the gating loop unit to obtain the user’s dynamic cognitive state at the current time step; (e) Summarize the dynamic cognitive states of each user to form a set of user dynamic cognitive states at the current time step.

6. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S5 generates the inhibition matrix for continuous time-varying games, including: (a) Calculate the propagation heat, bridging centrality and average influence of key nodes in each propagation domain at the current time step; (b) The dissemination heat, bridging centrality and average influence of key nodes within the domain are weighted and integrated to obtain the dissemination domain value for measuring the intervention priority of different dissemination domains; (c) Input the current propagation topology and the user's dynamic cognitive state into the dynamic graph encoder to obtain the node representations at both ends of the propagation edge; (d) Input the node representations at both ends of the propagation edge and the propagation domain value of their respective propagation domains into the multilayer perceptron, and obtain the continuous time-varying inhibition strength of each propagation edge through normalization mapping, and combine them to form a game inhibition matrix. (e) Apply sparsity constraints to the game inhibition matrix so that the inhibition signal is concentrated on the critical propagation path with high value in the propagation domain.

7. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S6 updates the propagation topology edge weights, including: (a) Calculate the degree of cognitive similarity between the dynamic cognitive states of users at both ends of the propagation edge; (b) Weaken the original propagation edge weights by using the game inhibition strength, and adjust the propagation probability of the corresponding propagation edge by using the cognitive similarity to obtain the propagation edge weights of the next time step; (c) An evolutionary propagation topology is formed based on all the updated propagation edge weights, and the evolutionary propagation topology is fed back to the user cognition update process and dynamic graph structure encoding process at the next time step.

8. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S7 extracts the spatiotemporal propagation representation, including: (a) Add node self-connections to the evolutionary propagation topology and normalize the propagation edge weights; (b) Using the user's dynamic cognitive state as the initial feature of the node, the information of the user node and its neighboring nodes is aggregated through a dynamic graph neural network to obtain the spatial propagation representation of each time step; (c) Arrange the spatial propagation representations of consecutive time steps within the historical time window in chronological order to form a spatial propagation representation sequence; (d) Input the spatial propagation representation sequence into a temporal convolutional network to extract short-term burst propagation features and long-term propagation evolution features to obtain a spatiotemporal propagation representation for prediction.

9. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, Step S8 involves jointly training the model, including: (a) Input the spatiotemporal propagation representation into a multilayer perceptron and output the prediction result of the future propagation situation; (b) Calculate the propagation prediction loss based on the difference between the predicted propagation trend and the actual propagation trend; (c) The propagation prediction loss is weighted and combined with the cross-domain cognitive alignment loss, the game-inhibition sparsity constraint, and the closed-loop consistency constraint to obtain the joint training objective. (d) The model parameters of the cognitive encoder, gated recurrent unit, dynamic graph encoder, temporal convolutional network and multilayer perceptron are updated by backpropagation to obtain the trained cross-domain rumor propagation prediction model.

10. The cross-domain rumor prediction method based on cognitive collaboration according to claim 1, characterized in that, The prediction results and explanatory information include: (a) at least one of the following: future spread intensity, future spread scale, future scope of influence of rumors, and cross-domain migration trend; (b) User dynamic cognitive state, domain selective game inhibition matrix and evolutionary propagation topology are used to characterize the closed-loop feedback relationship between user cognitive changes, anti-rumor suppression effect and propagation path evolution.

Images