An opinion node identification method based on behavioral sentiment game

Abstract

The application discloses a kind of based on behavior emotional game public opinion node identification method, belong to social network public opinion governance technical field.This method includes: obtaining social media rumor propagation dataset and user attribute strategy set;High-dimensional sparse rumor data is reconstructed using enhanced variational autoencoder, and low-dimensional dense feature matrix is generated in combination with UMAP algorithm;Based on BERT model quantification user fine-grained emotion, construct evolutionary game multidimensional benefit function to simulate the propagation strategy evolution under the emotion driving;Fusion network structure, behavior, emotion and strategy features construct graph neural network model, combine Monte Carlo simulation calculation user propagation contribution degree, identify rumor key node and predict propagation trend.The application solves the problem of high-dimensional sparse rumor data and dynamic coupling of user behavior and emotion, through multi-modal fusion modeling, significantly improves the key user identification precision and prediction stability, suitable for network public opinion control and false information tracing.

Description

Technical Field

[0001] This invention belongs to the field of social network public opinion governance and false information detection technology, specifically involving a public opinion node identification method based on behavioral emotion game theory. Background Technology

[0002] The real-time interactive nature of social media greatly accelerates the spread of rumors and easily disrupts online public order. Therefore, accurately identifying key users in the spread of rumors has become the core of public opinion management. However, existing identification methods are often limited to single-dimensional modeling. Whether relying solely on network topology, user behavior data, or simply analyzing text sentiment, they cannot simultaneously take into account the dynamic evolution of user behavior and the interconnected characteristics of group emotional games, resulting in insufficient identification accuracy and poor stability in propagation prediction. Although the academic community has proposed topological centrality algorithms, sentiment detection models, and evolutionary game simulation frameworks, these technologies are mostly used independently. Topological algorithms ignore the influence of subjective emotions on decision-making, sentiment models struggle to depict temporal dynamics, and game-theoretic schemes lack pre-optimization for high-dimensional sparse data, failing to form an integrated modeling system.

[0003] In terms of data processing and feature extraction, existing technologies face multiple challenges. First, the raw data of social rumors is characterized by high dimensionality and sparsity, making it difficult for traditional dimensionality reduction methods to compress the data while simultaneously preserving the global topological structure and local behavioral features. Second, users' forwarding and debunking behaviors continuously and dynamically switch with the development of public opinion, and algorithms based on static node weights cannot effectively lock the core of dynamic propagation. Finally, group emotions can significantly change individuals' propagation strategies, and existing solutions lack quantitative modeling methods for dynamic game theory driven by emotions, making it difficult to capture the deep-seated driving forces in the evolution of public opinion.

[0004] To address the aforementioned issues, some existing studies have attempted fusion solutions, but specific shortcomings remain. For example, some solutions combine evolutionary game theory with node embedding algorithms, inputting the feature matrix into a graph attention network for prediction; however, their sentiment quantification still relies on traditional classifiers, failing to incorporate pre-trained language models for fine-grained analysis and lacking multi-scenario propagation simulation mechanisms. Other solutions discriminate based on game interaction influence matrices and topic models, but do not involve joint dimensionality reduction techniques based on data reconstruction and manifold approximation, nor do they achieve linked modeling of sentiment and game theory. Furthermore, while some studies have used autoencoders or manifold approximation algorithms simultaneously, the former is only used for credibility assessment rather than data reconstruction, and the latter is only used for topic clustering rather than generating low-dimensional dense feature matrices, failing to fully leverage the potential of these algorithms in feature engineering.

[0005] In summary, existing technologies generally suffer from the problem of independent sentiment quantification and evolutionary game modeling, failing to form a closed loop for sentiment-driven strategy evolution. Furthermore, data-sparse solutions often rely on node embedding or topic models, lacking mechanisms for collaboratively designing data reconstruction and manifold approximation, and lacking multi-scenario propagation simulations to quantify differences in user contribution across different scenarios. Because they fail to comprehensively integrate and model the four types of features—network structure, user behavior, text sentiment, and game strategy—existing technologies struggle to accurately characterize users' potential propagation motives, thus limiting the effectiveness of public opinion management. Summary of the Invention

[0006] The technical problems to be solved by this invention include: First, how to complete sample expansion and low-dimensional mapping of high-dimensional sparse rumor data, and fully preserve the propagation topology and user behavior characteristics; Second, how to quantify the fine-grained emotions of user texts and build a multi-dimensional evolutionary game model to simulate the evolution of emotion-driven propagation strategies; Third, how to integrate multi-dimensional features to construct a propagation prediction model, and combine random simulation to accurately identify key rumor users and output long-term propagation trends.

[0007] To address the aforementioned technical problems, the present invention employs the following technical solution: a method for identifying public opinion nodes based on behavioral sentiment game theory, comprising: acquiring original interaction data and network topology of a social network; preprocessing the original data to obtain a user behavior feature matrix; reconstructing and expanding the user behavior feature matrix using an enhanced variational autoencoder, and performing low-dimensional mapping on the reconstructed joint feature matrix using a uniform manifold approximation algorithm to generate a low-dimensional dense user behavior feature matrix; performing fine-grained sentiment quantification on user text using a pre-trained language model to obtain text sentiment scores, and constructing a multi-dimensional game payoff function based on the sentiment scores to simulate the dynamic evolution of user propagation strategies and output user strategy codes; integrating network topology features, the low-dimensional dense user behavior feature matrix, text sentiment scores, and user strategy codes as initial node features, inputting them into a graph neural network for propagation prediction, and outputting user propagation probabilities; based on user propagation probabilities, statistically analyzing the propagation contribution of each user in different scenarios through multi-round Monte Carlo propagation simulations, screening key public opinion nodes, and outputting propagation trend predictions.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0009] By jointly optimizing the variational autoencoder and the uniform manifold approximation algorithm to address the problem of high-dimensional sparsity of data, the reconstructed joint feature matrix is ​​input into the uniform manifold approximation algorithm, which can simultaneously preserve the global propagation correlation of social networks and the similarity of local user behavior. Fine-grained sentiment quantification is performed through a pre-trained language model and linked to the evolutionary game multidimensional payoff function model. The sentiment quantification score is directly mapped to the propagation strategy preference vector and input into the game payoff calculation, which can accurately characterize the driving effect of emotions on user propagation behavior. By integrating four types of features—network structure, behavior, sentiment, and strategy—to construct a graph neural network propagation node model, and combining it with Monte Carlo multi-scenario propagation simulation to calculate the user propagation contribution, key user identification and public opinion propagation scale prediction can be achieved simultaneously. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall process of the public opinion node identification method based on behavioral emotion game theory of the present invention.

[0011] Figure 2 This is a flowchart of the evolutionary game modeling based on fine-grained emotion quantification in this invention.

[0012] Figure 3 This is a flowchart of the multimodal fusion graph neural network propagation prediction process in this invention. Detailed Implementation

[0013] 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.

[0014] The symbols are defined as follows:

[0015] N: The total number of users participating in online public opinion interaction;

[0016] i and k are both indexes of all users on the network, i = 1, 2, ..., N, k = i = 1, 2, ..., N;

[0017] : The total number of behavioral samples corresponding to user i;

[0018] The dataset is a local behavior dataset specific to user i, and the data distribution varies across different user datasets.

[0019] U: Server-side storage of unlabeled public opinion text sets, used for global knowledge distillation to expand training samples;

[0020] : The j-th public opinion text sample in the unlabeled set, j=1,2,…,J;

[0021] M: Total number of rounds executed in the Monte Carlo simulation;

[0022] T: The total number of future time steps in the prediction of public opinion dissemination.

[0023] The encoder outputs the mean of the latent distribution corresponding to the j-th public opinion text sample.

[0024] The latent distribution logarithm and standard deviation of the encoder output;

[0025] Standard normal random variable, used for reparameter sampling;

[0026] z: The latent feature vector obtained by sampling a single text;

[0027] Data reconstruction loss;

[0028] : Latent distribution KL divergence loss;

[0029] : Joint loss of variational autoencoder;

[0030] : Original cleaned high-dimensional sparse public opinion representation matrix;

[0031] The sample reconstruction representation matrix output by the VAE decoder;

[0032] The global joint representation matrix is ​​obtained by concatenating the original matrix and the reconstructed matrix along the feature dimension;

[0033] High-dimensional feature space user similarity adjacency matrix;

[0034] UMAP low-dimensional embedding space user similarity adjacency matrix;

[0035] Q, K, V: Three weight matrices for multi-head attention;

[0036] The h-th attention calculation branch;

[0037] Sentiment-policy mapping weights;

[0038] b: Emotion mapping bias constant;

[0039] User propagation strategy preference vector;

[0040] Emotional influence coefficient;

[0041] , : The one-hot encoded vectors of the strategies for users i and k;

[0042] M: Game payoff matrix;

[0043] The combined game payoff between user i and user k.

[0044] Strategy execution loss cost.

[0045] : Complete initial representation vector of the user node;

[0046] Network topology feature vector;

[0047] Low-dimensional behavioral characteristics of VAE-UMAP;

[0048] : Text sentiment quantification score vector;

[0049] One-hot encoding of user policies;

[0050] : Nodes retain weights based on their own characteristics;

[0051] Neighborhood aggregation weight matrix;

[0052] : The set of all social neighbors of user i;

[0053] : Update node features in round t;

[0054] Overall loss of GNN;

[0055] Cross-entropy classification loss;

[0056] L2 regularization coefficient;

[0057] All trainable parameters of the network;

[0058] : Probability of user-generated rumors spreading.

[0059] The set of users in the s-th round of simulated propagation;

[0060] All interactive users across the entire network;

[0061] : Coverage ratio of a single simulation propagation.

[0062] A method for identifying public opinion nodes based on behavioral emotion game theory, such as Figure 1 As shown, the specific steps include:

[0063] S1: Obtain raw data on the spread of rumors on social media, basic user attributes, and a set of user strategies for dealing with rumors; the raw data includes user behavior sequences, rumor and comment texts, and the topology of the rumor spread network. Specifically, this includes:

[0064] S11: Data source selection: We selected three types of public social rumor datasets: Weibo rumor dataset, Twitter 15 / 16, and PHEME as data sources. The collected content includes user basic attributes, time-series forwarding and commenting behavior, rumor text, user comment text, and the propagation network topology composed of user follow and forwarding.

[0065] S12: Dissemination strategy classification, which divides users’ behavior in the face of rumors into three categories: active dissemination, neutral observation, and proactive debunking, and completes the user strategy set construction through weak labeling;

[0066] S13: Data standardization and cleaning, removing blank text, duplicate user IDs, and zombie accounts with no interaction behavior, unifying the time-series behavior timestamp format, and constructing a standardized high-dimensional sparse feature matrix.

[0067] S2: Based on the high-dimensional sparse rumor raw data obtained in step S1, an enhanced variational autoencoder is used to augment and reconstruct the high-dimensional sparse rumor data, and a low-dimensional dense feature matrix is ​​generated by combining uniform manifold approximation and projection algorithms. Specifically, this includes:

[0068] S21: Input the original high-dimensional sparse representation matrix of the rumor into the VAE encoder. The encoder consists of two fully connected layers, which output the latent distribution mean vector and the logarithmic standard deviation vector, respectively. Introduce a standard normal random variable to complete the reparameterization operation. This operation can solve the defect that VAE cannot directly sample the distribution to calculate the gradient, ensuring that the model can complete iterative optimization through gradient descent. After completing the reparameter sampling, the latent representation vector corresponding to a single text is obtained. The calculation formula is as follows:

[0069]

[0070] In the formula, Let j be the mean of the latent distribution of the j-th text sample. The standard deviation is the logarithm. It is a standard normal random variable.

[0071] S22: Construct a VAE decoding network to receive the output latent representation vector. The reverse mapping restores the original high-dimensional representation dimension, completing the supplementation and reconstruction of missing and sparse samples. The overall optimization objective of the network is constrained by two parts of loss: sample reconstruction loss and latent distribution KL divergence loss. The joint loss calculation formula is as follows:

[0072]

[0073] In the formula, This represents the feature reconstruction error between the original sample and the reconstructed sample, used to constrain the decoding output to fit the distribution of real propagation behavior. This is the difference in divergence between the encoded latent distribution and the standard normal distribution. Its purpose is to standardize the latent spatial distribution and prevent the collapse of the model's latent representation.

[0074] S23: After completing the VAE decoder sample reconstruction operation, in order to simultaneously preserve the representation distribution of the original public opinion dissemination information and the model-generated supplementary samples, the standardized and cleaned original high-dimensional sparse representation matrix and the VAE network output reconstructed representation matrix are vector-concatenated according to the representation dimension direction to obtain the global joint representation matrix. ;

[0075]

[0076] In the formula, the original feature matrix is ​​denoted as , representing the total number of social network user samples, is the original representation dimension; the VAE-reconstructed feature matrix is ​​denoted as . After the concatenation is completed, the joint representation matrix is ​​used as input to calculate the fuzzy similarity weights between all user samples in the high-dimensional representation space. The nearest neighbor hyperparameter is set to control the local neighborhood calculation range. The similarity weight is used to characterize the degree of similarity in the distribution of any two users in terms of public opinion dissemination behavior and text sentiment dimension. The larger the weight value, the higher the similarity of the two users' dissemination behavior and sentiment tendency, which provides a basic measurement basis for the subsequent calculation of UMAP high- and low-dimensional structural loss.

[0077] S24: Construct a high-dimensional adjacency similarity matrix based on the high-dimensional spatial fuzzy similarity weights obtained in S23. Simultaneously set the target low-dimensional dimension, initialize the low-dimensional embedding vector, and generate a low-dimensional spatial similarity matrix. The overall optimization objective is to minimize the structural difference between the high and low similarity matrices. Gradient descent is used to iteratively update the low-dimensional embedding vectors of all users. A fixed learning rate and a maximum number of iterations are set during the iteration process. After the convergence condition is met, a dense low-dimensional representation matrix is ​​output. It fully preserves the global propagation correlations and local user behavior similarities on social networks.

[0078] S3: As Figure 2 As shown, based on the user comment text obtained in step S1, a pre-trained language model is used to complete fine-grained sentiment quantification of user text, and an evolutionary game multidimensional payoff function is used to model the evolution of propagation strategies driven by user sentiment. Specifically, this includes:

[0079] S31: Extract all user-posted, forwarded, and commented text content, and use the Word2Vec tool to perform word embedding encoding, converting discrete words into fixed-length numerical vectors. The encoded text sequence is then input into a pre-trained BERT base model. Leveraging a multi-head self-attention mechanism, the model mines deep semantic connections within the context, building a three-layer fully connected layer to output the predicted probabilities of positive, negative, and neutral sentiment. The multi-head attention calculation formula is as follows:

[0080]

[0081] in, , , These represent the three weight matrices: query, key, and value. The total number of attention heads; the multi-head attention design can simultaneously capture local emotional keywords in short texts and global emotional tendencies in long comment texts, improving the accuracy of fine-grained emotion recognition.

[0082] S32: Based on the sentiment score of each user text output by the BERT model, construct a linear mapping relationship to convert continuous sentiment scores into a user propagation strategy preference vector. The calculation formula is as follows:

[0083]

[0084] in, The learnable weight matrix for the mapping layer. This is a global bias parameter. This is a vector representing user sentiment scores; this mapping mechanism quantifies the degree to which the strength of emotions drives user propagation choices. The higher the sentiment score, the greater the weight of the corresponding propagation strategy in the vector. After completing preference calculation, the simulation proceeds to the user neighborhood game simulation stage, simulating the changes in payoffs after a single user interacts with their surrounding friends.

[0085] S33: Construct an emotion-weighted multidimensional game payoff formula, integrating four factors: emotion influence coefficient, group strategy proportion, social interaction coefficient, and strategy execution loss, to solve for the payoff between any two users; Payoff calculation formula:

[0086]

[0087] in, The coefficient representing the influence of emotions. , For users ,user The corresponding one-hot encoding vector of the strategy, This is the global game payoff matrix. This represents the execution cost of actions such as forwarding and debunking rumors; the higher the benefit value, the more beneficial the current dissemination strategy is to users, and users are more likely to retain this strategy in the next iteration.

[0088] S34: Iterate through all users across the network, update its own propagation selection by combining its own strategy preference vector and neighborhood interaction reward value; record the strategy switching records of all users in each round of evolution iteration, generate a time-series strategy evolution sequence, which is used as feature input for subsequent GNN models to characterize the game law of dynamic changes in sentiment over time.

[0089] S4: As Figure 3 As shown, the network topology features obtained in step S1, the low-dimensional behavioral features generated in step S2, and the BERT sentiment scores and user policy codes output in step S3 are integrated to construct the initial features of the graph neural network nodes, and the user rumor propagation probability is obtained through training. Specifically, this includes:

[0090] S41: Extract four types of heterogeneous feature vectors and perform a concatenation operation to generate the initial representation of a single-user graph node: the social network topology features obtained in step S1, the VAE-UMAP low-dimensional dense behavioral features output in step S2, the BERT sentiment score output in step S3, and the one-hot encoding of the user policy generated in step S3; the calculation formula for the initial node representation is as follows:

[0091]

[0092] The various features correspond to network connection relationships, user's historical dissemination behavior, text sentiment tendencies, and long-term dissemination choices, respectively. The integration of multiple dimensions can comprehensively depict the potential motivations of users in dissemination.

[0093] S42: Construct a three-layer graph convolutional network to complete multiple rounds of neighborhood message passing. Each layer performs node feature update operations, balancing the weight ratio of its own inherent features and the features propagated from its neighbors; Node feature update formula:

[0094]

[0095] in, Retain weights for the node's own characteristics. The neighborhood information aggregation weight matrix, It represents the entire neighborhood of a user's friends; multi-level aggregation can capture the impact of users indirectly spreading over long distances.

[0096] S43: Introducing L2 regularization constraints to alleviate overfitting, using the cross-entropy between the propagation probability of the true label and the predicted value as the basic loss, and combining them to generate the overall optimized loss function of the model; loss function calculation formula:

[0097]

[0098] in, To propagate the probability cross-entropy loss, For regularization hyperparameters, These are all trainable parameters of the network; the regularization term is used to constrain the range of parameter values ​​to prevent the model from overfitting on small datasets.

[0099] S44: After the network iterates until the loss converges, the final node representation vector is mapped to the interval between 0 and 1 using the Sigmoid activation function to obtain the user's rumor propagation probability. The higher the probability, the greater the likelihood that the user is a key figure in spreading rumors.

[0100] S5: Based on the user rumor propagation probability output in step S4, and combined with Monte Carlo multi-scenario propagation simulation statistical propagation indicators, key users of rumors are selected and propagation trend prediction results are output. Specifically, this includes:

[0101] S51: Randomly select a small number of network nodes as the initial source of public opinion propagation. Use the user propagation probability output in step S4 as the node activation threshold. Update the user propagation activation status of the entire network step by step. When the number of newly propagated users is lower than the set threshold for 10 consecutive time steps, it is determined that the propagation process has converged and stabilized, and the single-round simulation ends.

[0102] S52: After a single simulation, calculate the ratio of the total number of activated propagation nodes to the total number of users in the entire network, and then calculate the propagation coverage ratio; Single simulation coverage ratio:

[0103]

[0104] in, For the first The set of nodes participating in propagation within the simulation round. It represents all user nodes on the social network; the coverage ratio is used to measure the breadth of rumor spread in a single round of simulation.

[0105] S53: Independently repeat M unrelated Monte Carlo simulations, calculate the mean and standard deviation of the total simulation coverage ratio, and simultaneously count the total number of times each user is activated and propagated in all simulation rounds; the mean reflects the average scale of propagation, and the standard deviation measures the degree of propagation fluctuation.

[0106] S54: Sort all users in descending order of cumulative dissemination activation frequency, preset a global contribution threshold, and classify accounts with a frequency higher than the threshold into the key user set of rumors to complete the core dissemination node screening.

[0107] S55: Summarize the full-time activation data of M rounds of simulation, calculate the expected value of the number of propagation nodes in the next T consecutive time steps, and output the quantitative prediction results of the long-term spread scale of rumors, providing data support for public opinion management.

[0108] In a specific embodiment, three types of publicly available social media rumor datasets—Weibo, Twitter15 / 16, and PHEME—were selected for validation experiments. The total number of interactive users across the entire network was N=10,000, divided into multiple differentiated public opinion subsets, with each user group corresponding to differentiated rumor propagation time-series samples. 10,000 unlabeled public opinion comment texts were selected as auxiliary modeling data. The hyperparameter settings were: VAE latent feature dimension 256, UMAP target low-dimensional dimension 64, BERT attention head number h=12, evolutionary game iteration rounds 15, Monte Carlo simulation total rounds M=50, and graph neural network convolutional layers L=3. The network structure was layered: the first three layers were basic feature extraction layers, and subsequent layers were high-level networks that fused sentiment and propagation. Experimental results showed that the present invention outperformed the baseline models DDCA-GCN, GMIN, VoteRank++, and T³RD in both key user identification F1 score and public opinion propagation scale prediction MAE, significantly reducing the standard deviation of the simulated propagation coverage ratio fluctuation and significantly improving the stability of public opinion trend prediction.

[0109] It should be noted that, in an optional embodiment, the user feature fusion weight can also be obtained by optimizing the cross-entropy of the local public opinion subset. However, this implementation does not utilize the global behavioral uncertainty features output by VAE dimensionality reduction, does not link the user's local verification loss sensitivity information, and cannot form a synergistic linkage with the evolutionary game strategy evolution module. Therefore, it is not considered as the preferred implementation of this invention. The preferred implementation of this invention mainly uses the above-mentioned method of jointly calculating the fusion weight by combining the global uncertainty features of VAE and the local verification sensitivity to ensure the closed-loop collaboration of each module process.

[0110] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to parameters such as the number of model layers, feature dimensions, number of game iteration rounds, and number of Monte Carlo simulations in 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 method for identifying public opinion nodes based on behavioral-emotional game theory, characterized in that, Includes the following steps: S1: Obtain raw data on the spread of rumors on social media, basic user attributes, and a set of user strategies for dealing with the spread of rumors; the raw data includes user time-series behavior sequences, rumor text and user comment text, and the topology of the rumor spread network; S2: Based on the acquired high-dimensional sparse rumor raw data, a generative reconstruction model is used to perform sample expansion and reconstruction on the high-dimensional sparse rumor data, and a nonlinear dimensionality reduction algorithm is combined to process the reconstructed data to generate a low-dimensional dense feature matrix. S3: Based on the acquired user comment text, a pre-trained language model is used to perform fine-grained sentiment quantification on the user comment text, construct an evolutionary game multidimensional payoff function of sentiment association, and calculate the evolutionary probability distribution of user propagation strategy based on the multidimensional payoff function; S4: The initial node features of the graph representation learning model are constructed by fusing the four types of information: the acquired network topology features, the generated low-dimensional user behavior features, the output text sentiment scores, and the user policy codes. After training the graph representation learning model, the probability value of rumor propagation for the target user is output. S5: Based on the rumor propagation probability value of the target user, a random simulation method is used to perform multiple rounds of rumor propagation simulation, statistical indicators related to the propagation coverage of all users are counted, a set of key rumor users is selected based on the relevant statistical indicators, and the prediction results of the overall rumor propagation trend in the future multiple time steps are output.

2. The method for identifying public opinion nodes based on behavioral-emotional game theory according to claim 1, characterized in that, The process involves using a generative reconstruction model to augment and reconstruct the high-dimensional sparse rumor original data, and then combining this with a nonlinear dimensionality reduction algorithm to process the reconstructed data and generate a low-dimensional dense feature matrix, including: (a) Input the cleaned high-dimensional sparse rumor feature matrix into the variational autoencoder module, output the mean value and log standard deviation value of the corresponding latent distribution, and use the standard normal random value to complete the reparameter sampling operation to obtain a latent feature vector of uniform length. (b) The sampled latent feature vectors are back mapped back to the original high-dimensional feature space through the decoder module to complete the data augmentation and reconstruction of missing sparse samples; at the same time, the reconstruction error loss and the latent distribution divergence loss are used to jointly constrain the iterative optimization of network parameters. (c) The original high-dimensional feature matrix and the reconstructed feature matrix output by the decoder are concatenated along the feature dimension to generate a global joint feature representation. Based on the concatenated joint features, the behavioral similarity weights of all users in the high-dimensional feature space are calculated. (d) Construct two similarity matrices: a high-dimensional original feature space and a target low-dimensional embedding space. With the goal of reducing the structural differences between the high-dimensional and low-dimensional spaces, the gradient descent method is used to iteratively update the low-dimensional embedding vectors of all users. After the model converges, a dense low-dimensional feature matrix is ​​output.

3. The method for identifying public opinion nodes based on behavioral-emotional game theory according to claim 1, characterized in that, The process involves employing a pre-trained language model to perform fine-grained sentiment quantification on the user comment text, constructing an evolutionary game-theoretic multidimensional payoff function for sentiment association, and calculating the evolutionary probability distribution of the user propagation strategy based on the multidimensional payoff function, including: (a) Basic word vector encoding is performed on all user-posted, forwarded, and commented texts on the platform. The encoded text sequence is input into a pre-trained language model. The deep semantics of the text context are mined by multi-head self-attention mechanism. The predicted probability of positive, negative, and neutral emotions corresponding to each text is output through a fully connected layer. (b) Based on the text sentiment quantification score output by the model, a linear mapping rule from emotion to dissemination strategy is constructed, and a dissemination strategy preference vector exclusive to each user is obtained. The strength of user emotion determines their preference for forwarding, debunking, or waiting and seeing. (c) Construct a multi-dimensional game payoff calculation rule that integrates four types of factors: emotional influence weight, group strategy proportion, social interaction coefficient, and strategy execution loss, and solve the comprehensive payoff value corresponding to different propagation strategies when users interact with friends in their social neighborhood. (d) Combine the user's own strategy preference vector and neighborhood interaction benefit value to iteratively update the single user's propagation behavior selection, fully record the entire process of strategy switching of all users in each round of evolution iteration, and generate a time-series strategy evolution record.

4. The method for identifying public opinion nodes based on behavioral-emotional game theory according to claim 1, characterized in that, The fused network topology features, generated low-dimensional user behavior features, output text sentiment scores, and user policy codes are used to construct the initial node features of a graph representation learning model. After training the graph representation learning model, the target user's rumor propagation probability value is output, including: (a) Construct the initial feature vector of a single user graph node by splicing four types of heterogeneous information. The four types of information are social network topology connection features, user low-dimensional behavior features after dimensionality reduction, text sentiment quantization score, and one-hot encoding of user propagation strategy. (b) The graph neural network performs multi-layer neighborhood message passing operations, aggregates the features of neighboring nodes layer by layer and updates its own node features, balancing the weight of the propagation influence brought by the user's inherent behavioral features and the social neighborhood. (c) Construct cross-entropy loss calculation rules with superimposed parameter regularization constraints, and iteratively train the graph neural network until the network parameters converge and stabilize; (d) Apply Sigmoid activation to the final node feature vector output by the network and map it to the interval between 0 and 1 to output the predicted probability value of each user forwarding and spreading rumors.

5. The method for identifying public opinion nodes based on behavioral-emotional game theory according to claim 1, characterized in that, The method employs a random simulation approach to perform multiple rounds of rumor propagation simulation, statistically analyzes relevant indicators of the propagation coverage of all users, selects a key user set for rumors based on these indicators, and outputs predictions of the overall online rumor propagation trend for multiple future time steps, including: (a) Randomly select a small number of users in the network as the initial source of public opinion propagation, and use the user propagation probability output by the graph neural network as the node activation judgment criterion. Update the propagation activation status of users across the network step by step until the number of newly propagating users in multiple consecutive time steps is lower than the preset threshold, and determine that the single-round simulation has converged. (b) After a single simulation, the ratio of the total number of users who actually participated in the spread of rumors in this round to the total number of users across the entire network is calculated to obtain the coverage ratio of the spread of rumors in this round of simulation. (c) Independently repeat multiple rounds of Monte Carlo simulation without interference, calculate the average and standard deviation of the total simulation coverage, and at the same time count the total number of times each user is activated to participate in the propagation in all simulation rounds. (d) Sort all users in the network in descending order of cumulative dissemination activation frequency, preset a global contribution judgment threshold, and classify users whose cumulative frequency is higher than the threshold into the key user set of rumors; (e) Summarize the full-time node activation data from multiple rounds of simulation, calculate the average number of users participating in the spread of rumors in multiple consecutive time steps in the future, and output the long-term rumor spread scale trend prediction results.

Images