A key user discovery method fusing cognition and interaction

Abstract

The application belongs to the field of network rumor propagation and social network analysis, and specifically relates to a key user discovery method fusing cognition and interaction, comprising: acquiring network data and user attributes, using a graph neural network to cross-dimensionally aggregate features and combining a rumor similarity matrix to generate a weighted adjacency matrix, to realize effective compensation of initial data sparsity; extracting deep semantic features through an automatic encoder, introducing external information intensity to construct a dynamic game model, and quantifying nonlinear evolution of user cognition; using a graph attention network to adaptively learn connection weights, introducing an exponential time decay factor to dynamically adjust time-series interaction weights, and aggregating to generate multi-dimensional user interaction features considering timeliness; splicing and fusing the extracted multi-dimensional features into a comprehensive representation vector and inputting the comprehensive representation vector into a classifier to output key user screening results. Through three-stage cooperation of data sparsity compensation, dynamic evolution of cognition and complex time-effective interaction, the precision of initial feature description of the rumor and the fitting efficiency of personal strategy evolution are improved, and the spread of the network rumor is effectively curbed.

Description

Technical Field

[0001] This invention belongs to the field of online rumor propagation and social network analysis, specifically involving a key user discovery method that integrates cognition and interaction. Background Technology

[0002] Existing research has proposed various analytical approaches to address the problem of key user identification. However, traditional methods based on network structure features (such as centrality index and path analysis) and user static behavioral features (such as activity level, number of followers, and historical sentiment tendencies) often fail to accurately capture the evolution of user cognition and non-uniform dynamic interaction relationships in complex environments when faced with highly uncertain rumor propagation scenarios. This is because the models are too static and have only one dimension.

[0003] Existing research has proposed various modeling and analysis methods to address the aforementioned challenges. For example, some techniques employ evolutionary game theory models to measure the mutual influence of rumors and debunking information, design specific algorithms to perform low-rank dense vectorization of the feature space to address data sparsity, and then concatenate and fuse the feature topology matrix and adjacency matrix before inputting it into a graph attention network to predict user behavior. Other studies calculate the mutual influence matrix of rumors and debunking information based on evolutionary game theory, use topic models to calculate domain influence, and process user mutual influence and domain influence to obtain key user identification results. Still other methods obtain the payoff functions of rumor, debunking, and rumor-promoting messages based on three-party cognitive game theory, quantify the driving force of user state transformation, or introduce time decay mechanisms to weaken the popularity of hot topics.

[0004] While the above methods have achieved some success, they are still difficult to fully characterize the dynamic equilibrium of user strategy choices at the individual level, and their ability to accurately capture cognitive evolution in complex environments is limited. At the same time, in the early stages of rumor dissemination, the number of participating users is small and the topological structure is incomplete, resulting in severe data sparsity, and single-dimensional features are insufficient to comprehensively characterize the dissemination trend.

[0005] Specifically, existing technologies still have significant shortcomings in the following three aspects:

[0006] First, regarding the dynamic evolution of user cognition, some works use traditional evolutionary game theory and replicating dynamic equations for macroscopic prediction, but most schemes do not introduce Nash equilibrium solution mechanisms, making it difficult to accurately capture the nonlinear dynamic evolution at the individual cognitive level, which makes it difficult for traditional static models to adapt to complex real-world communication scenarios.

[0007] Secondly, in terms of data sparsity processing, node embedding algorithms or topic models are usually studied separately, lacking topology enhancement designs that explicitly encode the feature similarity between nodes into the edge weights of the adjacency matrix, which limits the topological expressive power of early sparse networks.

[0008] Finally, in terms of key user discovery and temporal multidimensional interaction quantification, existing game theory models are mostly limited to macro-level prediction or dynamic tracing of information dissemination, failing to deeply coordinate with graph attention networks and time decay factors, or only limited to the time decay of macro-level topic popularity while ignoring the temporal non-uniformity of individual micro-level interactions. This leads to a disconnect between user multidimensional interaction characteristics and game evolution results, making it difficult to effectively quantify the non-uniform interaction relationships that decay over time on social platforms, thus failing to accurately discover key users. Summary of the Invention

[0009] The technical problems to be solved by this invention include: first, how to effectively integrate data across dimensions to comprehensively depict the spread of rumors under the condition of sparse data in the early stage of rumor dissemination; second, how to accurately quantify the nonlinear dynamic evolution of user cognition under the dual impact of rumors and debunking information; and third, how to effectively extract and quantify the complex multidimensional interaction relationships of social networks with time decay characteristics, and coordinate them with the aforementioned characteristics to achieve accurate identification of key users.

[0010] To address the aforementioned technical problems, the present invention employs the following technical solution: a key user discovery method integrating cognition and interaction, which sequentially executes the following six steps within the same key user discovery process. This method includes interconnected core collaborative processes: The first stage (corresponding to steps S1-S2) involves acquiring a set of user attributes and network data containing static attribute features and dynamic behavioral features. Features are aggregated based on a graph neural network, and a rumor similarity matrix is ​​designed using cosine similarity to calculate a weighted adjacency matrix, thereby generating fused user node features to alleviate the initial data sparsity problem. The second stage (corresponding to step S3) introduces an autoencoder to extract deep semantic features of users, combines external information intensity to construct a user cognition evolution model based on dynamic game theory, solves strategies and updates states for the dynamic evolution of user cognition, and accurately quantifies the selection of Nash equilibrium strategies. The third stage (corresponding to step S4) utilizes a graph attention network to calculate the non-uniform attention coefficients between users, and introduces an exponential time decay factor to dynamically adjust the interaction weights over time, aggregating and generating multi-dimensional user interaction features that take into account timeliness.

[0011] The features extracted in the above three stages (corresponding to step S5) are spliced ​​and fused to construct a multi-dimensional comprehensive representation vector, which is then input into the classification layer (corresponding to step S6) for normalization processing, and outputs the probability distribution of users as key users, thus completing the accurate screening and identification of the key user set.

[0012] Compared with existing technologies, the advantages of this invention are as follows: By aggregating multimodal cross-dimensional features through graph neural networks and jointly optimizing with an original rumor similarity matrix, it effectively utilizes domain supplementary information to provide data compensation for the initial sparse network, efficiently alleviating data sparsity. Unlike existing technologies that use the CSR2Vec node embedding algorithm to solve sparsity, this invention explicitly encodes the feature similarity between nodes as edge weights in the adjacency matrix, enhancing the expressive power of topological structure information under data sparsity conditions. By introducing an autoencoder and dynamic game theory, it functionalizes the user's decision utility in the rumor-debunking game network, keenly capturing and accurately quantifying the nonlinear evolution of cognitive state over time and with neighboring strategies. This differs from existing technologies that use... This invention, which uses evolutionary game theory and replicating dynamic equations for propagation prediction, introduces a Nash equilibrium solution mechanism to perform dynamic equilibrium analysis on users' strategy choices in the rumor-debunking game, enabling a more accurate characterization of the nonlinear evolution at the individual cognitive level. Through the synergistic mechanism of graph attention network and time decay factor, it not only adaptively learns the non-uniform connection weights between users but also reflects the high sensitivity of recent interaction behavior to information dissemination, adaptively capturing the timeliness of complex interactions. Unlike existing technologies that only calculate the time decay of the popularity of hot topics without collaborating with graph attention network, this invention jointly models the exponential time decay factor and GAT attention coefficient, so that the interaction weights simultaneously reflect the semantic association strength and time decay characteristics. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall process of a key user discovery method that integrates cognition and interaction in the present invention.

[0014] Figure 2 This is a schematic diagram illustrating the process of multimodal data fusion and feature compensation in heterogeneous propagation networks in the method of this invention;

[0015] Figure 3 This is a schematic diagram of the cognitive-driven rumor-debunking dynamic game modeling and decision evolution in the method of this invention. Detailed Implementation

[0016] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The symbols are defined as follows:

[0018] : indicates that the user is at the Hidden representation of layers;

[0019] : This is the initial feature input;

[0020] : Represents the user's set of neighbors;

[0021] : Represents the normalization coefficient;

[0022] :express The weight matrix of layer -1;

[0023] : Represents the activation function;

[0024] : Indicates the number of elements in a user's neighbor set;

[0025] : Represents user and users Feature similarity;

[0026] : Represents the dot product of two user feature vectors;

[0027] , : denote the norm of the vectors respectively;

[0028] : Represents the edge weights between users (or the edge weights of the weighted adjacency matrix);

[0029] : Represents the node features after fusion;

[0030] : Represents the weight matrix of the fusion layer;

[0031] It refers to the user's sensitivity;

[0032] , : respectively represent in The intensity of rumors and debunking information received by users at any given moment;

[0033] : indicates a time interval;

[0034] : Represents the overall utility of a user's chosen strategy under a specific cognitive state;

[0035] : Represents the basic benefit a user receives when adopting a strategy;

[0036] : The neighbor influence weighting coefficient;

[0037] : is an indicator function;

[0038] : is the time decay factor for interaction weights;

[0039] : is the attenuation coefficient;

[0040] : The time step in which the interaction occurs;

[0041] : Represents the total dimension after concatenating multidimensional features;

[0042] : indicates that the model is for the th The original score is output for each class name;

[0043] : Represents the probability output after normalization;

[0044] : The true label of the sample;

[0045] : represents the cross-entropy loss function.

[0046] A key user discovery method that integrates cognition and interaction, such as Figure 1 As shown, it includes:

[0047] S1: Obtain data on rumor propagation networks and user attribute sets.

[0048] S11: Methods for obtaining social network-related data include using social media APIs, public dataset platforms, or subscribing to third-party data services to obtain advanced social media data and directed graphs in rumor propagation networks.

[0049] S12: Extract static attribute features (including age, number of followers, number of followings) and dynamic behavioral features (including number of likes, number of comments, number of reposts) for each user.

[0050] S2: Based on the rumor propagation network data and user attribute set obtained in step S1, construct a heterogeneous propagation network and generate fused user node features.

[0051] S21: Utilize graph neural networks to aggregate heterogeneous user features across dimensions. The formula for calculating the feature aggregation of a layer is:

[0052]

[0053] in, Indicates user In the Hidden representation of layers, As initial feature input; Indicates user The neighborhood group, Represents the normalization coefficient. express The weight matrix of layer -1 This represents the activation function.

[0054] S22: Next, in order to standardize the size of the neighbor set and ensure the comparability of nodes with different degrees during the information aggregation process, the normalization coefficient is... The formula is:

[0055]

[0056] in, Indicates user The number of elements in the neighbor set, and .

[0057] S23: Simultaneously, to further quantify the similarity between nodes, a rumor similarity matrix is ​​constructed. Using cosine similarity as the weight to measure the feature similarity between nodes, the similarity matrix is ​​defined as follows:

[0058]

[0059] in, Indicates user and users The feature similarity is set to [-1, 1], with a larger value indicating greater feature similarity. This represents the dot product of two user feature vectors. and Let represent the norms of the vectors.

[0060] S24: Then, based on the constructed rumor similarity matrix, a weighted adjacency matrix is ​​generated, defined by the following formula:

[0061]

[0062] in, Indicates user and users The edge weights between them, when the user and users When there is a direct relationship between the edges, the edge weight is equal to their feature similarity. If there is no direct relationship, the edge weight is 0.

[0063] S25: Finally, the fused node features are obtained through the above steps. Its formula is defined as:

[0064]

[0065] in, This represents the edge weights between users. This represents the weight matrix of the fusion layer. The fused node features comprehensively consider multiple user characteristics and the relationships between nodes, providing strong support for subsequent user cognition analysis and interaction network modeling.

[0066] S3: Based on the fused user node features obtained in step S2, construct a user cognitive dynamic game model and update its state, including:

[0067] S31: Use the encoding module of the autoencoder to extract deep semantic features from the fused user node features obtained in step S2, and use the deep semantic features as the initial user cognitive state.

[0068] S32: Considering the immediate evolution of user cognition due to the intensity of external rumors and debunking, the dynamic update formula for user cognition state is as follows:

[0069]

[0070] in, It's the user's sensitivity. and They represent the times respectively. The intensity of rumors and debunking information received by users. It represents a time interval.

[0071] S33: Based on this, construct a dynamic game model of user cognition in the rumor-debunking scenario. The comprehensive utility function of a user's choice of strategy under a specific cognitive state is defined as:

[0072]

[0073] in, This indicates that the user selects a strategy under a specific cognitive state. The effect, It's the user's sensitivity. Indicates the user's strategy Basic returns at that time The neighbor influence weighting coefficient, This represents the edge weights in the weighted adjacency matrix. For the set of neighboring nodes, Let be the indicator function. The optimal policy choice for each user under cognitive drive is determined by solving the Nash equilibrium through the derivative of the combined utility function of each user.

[0074] S4: Based on the rumor propagation network data obtained in step S1, a time decay factor is introduced to generate user interaction features.

[0075] S41: In complex social network interactions, to capture non-uniform edge connectivity, graph attention networks are used to calculate the attention coefficients between user nodes. .

[0076] S42: Simultaneously, a time decay factor is introduced. The interaction weights are dynamically adjusted over time, and the time decay factor is defined as follows:

[0077]

[0078] in, The attenuation coefficient is... The time step in which the interaction occurs; the attention coefficients are calculated using a graph attention network. With time decay factor The feature information of neighboring nodes is weighted and aggregated to generate the final user interaction features.

[0079] S5: Based on the outputs of steps S2, S3, and S4, construct the user's comprehensive representation vector, including:

[0080] S51: Concatenate the user fusion features obtained in step S2, the cognitive state updated by the game in step S3, and the time-sensitive interaction features generated in step S4 to construct a comprehensive user representation vector. .in This represents the total dimension after concatenating multidimensional features.

[0081] S6: Input the classification layer to output probability and filter key users.

[0082] S61: Input the comprehensive representation vector into the classification layer, and use the dimensionality reduction matrix of the classifier. After mapping the comprehensive representation, through The function is normalized to obtain the probability output. .

[0083] S62: The probability values ​​of key users are calculated by function mapping as shown below.

[0084]

[0085] in, The model represents the first The original scores output by each category name are subjected to index mapping to widen the score differences between different categories, and the sum of the index results of all categories is used as the normalized denominator; if the output label is 1, it means that the user is a key user, and if the output label is 0, it means that the user is a regular user.

[0086] S63: Finally, to enable the model to learn to distinguish between key users and ordinary users, the cross-entropy loss function is used for model training, and the calculation formula is as follows:

[0087]

[0088] in, Let be the true label of the sample. When the true label is 1, the loss is calculated using the first term; when the true label is 0, the loss is calculated using the second term. A negative sign is added to the entire value to ensure that the loss is non-negative, thus facilitating the assessment of the loss magnitude.

[0089] 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 key user discovery method integrating cognition and interaction, characterized in that, Perform the following steps sequentially within the same key user discovery process: S1: Obtain rumor propagation network data and user attribute set, the user attribute set including user static attribute features and dynamic behavioral features; S2: Based on the rumor propagation network data obtained in step S1, construct a heterogeneous propagation network, aggregate adjacency information using a graph neural network, and generate a weighted adjacency matrix by combining the rumor similarity matrix constructed based on cosine similarity, to obtain the fused user node features. S3: Use an autoencoder to extract deep semantic features from the fused user node features generated in step S2, and construct a dynamic game model of user cognition in the rumor-debunking scenario to solve strategies and update states for the dynamic evolution of user cognition. S4: Based on the rumor propagation network data obtained in step S1, the attention coefficients between user nodes are calculated using a graph attention network, and a time decay factor is introduced to dynamically adjust the interaction weights over time. The weighted aggregation generates user interaction features. S5: Construct a comprehensive representation vector of the user by combining the fused user node features obtained in step S2, the user cognitive state obtained in step S3, and the user interaction features obtained in step S4. S6: Input the comprehensive representation vector constructed in step S5 into the classification layer, use the normalization function to output the probability value of the user belonging to the key user, and filter out the key user set according to the probability value by setting a threshold.

2. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The features obtained in step S1 include: (a) Extract the user's static attribute features, specifically including the user's age, number of followers, and number of followers; (b) Extract the dynamic behavioral features, specifically including the number of likes, comments and reposts of the user.

3. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The process of calculating the rumor similarity matrix and generating the weighted adjacency matrix in step S2 includes: (a) Cosine similarity is used to measure the feature similarity between users and construct a rumor similarity matrix. The specific calculation method is to solve the dot product of the feature vectors of two users and divide it by the product of the norms of the two feature vectors. (b) Generate a weighted adjacency matrix based on the rumor similarity matrix. When there is a rumor propagation or interaction relationship between two users, the edge weight is assigned the feature similarity value between the two users; when there is no direct relationship between two users, the edge weight is assigned zero.

4. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The process of updating the user's cognitive state in step S3 includes: (a) Set the user's cognitive state to change dynamically due to the dual influence of external rumors and debunking information; (b) The update increment of cognitive state is calculated by combining the user's own sensitivity, the intensity of the rumor information received at a specific time and the intensity of the debunking information, and the time interval of the evolution. (c) Combine the user's cognitive state at the previous moment with the update increment to obtain the dynamically updated user cognitive state.

5. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The process of constructing the user cognitive dynamic game model in step S3 includes: (a) Define the combined utility function of a user choosing different strategies under a specific cognitive state; (b) The comprehensive utility function consists of two superimposed utility components: the first component is the basic benefit obtained by the user by adopting the current strategy, and the second component is the benefit generated by the influence of neighbors, specifically the neighbor influence weight coefficient multiplied by the sum of the edge weights of all neighbor nodes that adopt the same strategy. (c) By differentiating the combined utility function of each user in the network and solving for the Nash equilibrium, the state evolution mechanism driven by cognition can be inferred.

6. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The process of generating user interaction features in step S4 includes: (a) An exponential time decay factor is introduced to dynamically adjust the interaction weights over time. This factor is determined by a preset decay coefficient and the time step in which the interaction occurs. (b) Combining the attention coefficients calculated by the graph attention network with the time decay factor, the feature information of neighboring nodes is weighted and aggregated to generate the final user interaction features with temporal dynamic characteristics.

7. The key user discovery method integrating cognition and interaction according to claim 1, characterized in that, The process of outputting and filtering results in step S6 includes: (a) The comprehensive representation vector is mapped and transformed using the dimensionality reduction matrix of the classifier; (b) Obtain the probability output through normalization. If the output label is 1, it means that the user is a key user. If the output label is 0, it means that the user is an ordinary user. (c) The cross-entropy loss function is used to guide model training, and the loss value is calculated and the model is optimized for the true labels of samples of different categories.

Images