Regional public opinion propagation mode mining method based on graph neural network

Abstract

The invention discloses a regional public opinion propagation mode mining method based on a graph neural network. The method comprises the following steps: S1, collecting and preprocessing public opinion data; s2, constructing a public opinion propagation hypergraph model; s3, hypergraph convolutional neural network feature learning; s4, performing quantum particle swarm optimization on the hypergraph convolutional neural network model; s5, propagation mode clustering analysis is carried out; and S6, outputting a public opinion propagation mode result. According to the method, accurate mining and dynamic identification of the regional public opinion propagation mode are realized, and meanwhile, the high-order interaction expression capability of public opinion propagation modeling, the global optimization efficiency of model parameters and the adaptive classification capability of the propagation mode are remarkably improved; the method can be widely applied to multiple fields of public opinion monitoring, media transmission analysis, brand public opinion management and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of graph neural networks, and particularly to a method for mining regional public opinion dissemination patterns based on graph neural networks. Background Art

[0002] In the current social media environment, the dissemination of public opinion has become an important issue that enterprises and the public are concerned about. The analysis of the dissemination pattern of regional public opinion is of great significance for identifying the dissemination path of public opinion, predicting the development trend of public opinion, and formulating public opinion response strategies. Traditional research methods for public opinion dissemination usually rely on statistics, social network analysis, or machine learning models. For example, through text mining, sentiment analysis, or dissemination dynamics modeling, etc., to process and analyze public opinion data. However, these methods often fail to fully describe the complex relationships involved in the process of public opinion dissemination. Especially in the social network environment, the impact of factors such as multi-agent interaction, information flow, and dissemination path on public opinion dissemination is very significant. Existing methods often build models based on traditional graph networks, such as methods based on ordinary graph neural networks. Although they can learn the dissemination relationships between nodes, due to their structural limitations, it is difficult to effectively capture high-order interaction information and comprehensively describe the multi-agent dissemination pattern in regional public opinion dissemination.

[0003] In existing research, some methods use social network analysis means to construct a public opinion dissemination network and analyze the key nodes of public opinion dissemination by calculating indicators such as degree centrality and betweenness centrality of nodes. However, such methods rely on manually defined features, making it difficult to automatically model the public opinion dissemination pattern and difficult to adapt to the dissemination forms of different public opinion events. In addition, machine learning-based public opinion dissemination analysis methods, such as traditional classification methods like random forest and support vector machine, although they can improve the accuracy of dissemination pattern recognition to a certain extent, in the case of increasing data scale and complex dissemination structure, these methods are difficult to effectively learn the spatio-temporal dynamic features in public opinion dissemination. In recent years, with the rapid development of deep learning technology, graph neural networks have been widely applied to tasks such as social network analysis and public opinion dissemination prediction. Compared with traditional methods, graph neural networks have stronger feature extraction capabilities and can aggregate information based on nodes and their adjacency structures to learn the dissemination relationships between nodes. However, most public opinion dissemination analysis methods based on graph neural networks use models such as ordinary graph convolutional networks or graph attention networks. These models mainly perform calculations based on ordinary graph structures. In the case of relatively complex dissemination relationships, it is difficult to make full use of high-order dissemination information and effectively describe the collaborative dissemination pattern between multiple agents.

[0004] In terms of public opinion dissemination modeling, existing methods usually adopt simple graph construction strategies, such as constructing ordinary graph networks based on social relationships or user interaction relationships. However, this approach cannot reflect the high-order interaction characteristics in the process of public opinion dissemination. For example, in the dissemination process of a certain hot event, it may involve the joint dissemination behavior of multiple user groups. Existing ordinary graph neural networks are difficult to model the information flow jointly participated by multiple subjects, while hypergraph modeling technology can better capture high-order interaction relationships. A hypergraph can connect multiple nodes in a hyperedge, thus effectively representing the dissemination connections between multiple subjects and providing more comprehensive structural information for public opinion dissemination pattern mining. However, there are few studies that apply hypergraph modeling to public opinion dissemination analysis and combine it with hypergraph convolutional neural networks for deep feature learning. Traditional graph neural network methods usually only support the aggregation of first-order or second-order neighbor information, while hypergraph structures can be extended to higher-order information fusion, enabling the model to more accurately learn multi-level association information in the process of public opinion dissemination.

[0005] In addition, in the analysis of public opinion dissemination patterns, the parameter optimization of the model has an important impact on the accuracy and robustness of the final result. Most existing studies use traditional methods such as gradient descent and stochastic optimization for parameter adjustment. However, these methods are prone to falling into local optima and are difficult to effectively improve the global optimization ability of the model. Quantum particle swarm optimization, as an improved intelligent optimization algorithm, can efficiently search for the global optimal solution in a high-dimensional parameter space and has strong global convergence ability. However, there are few studies that apply quantum particle swarm optimization to the optimization of hypergraph convolutional neural networks, resulting in limited performance improvement of hypergraph convolutional neural network models in public opinion dissemination analysis. Traditional optimization methods are difficult to effectively adjust the key parameters of hypergraph convolutional neural networks, such as hyperedge weights, node embedding dimensions, and activation function parameters. Quantum particle swarm optimization can achieve dynamic adjustment of the parameters of hypergraph convolutional neural network models under the combination of global search and local optimization, improving the adaptability of the model in complex dissemination environments.

[0006] On the other hand, in the process of mining the public opinion dissemination mode, clustering analysis is one of the key steps. Traditional clustering methods, such as K-means, DBSCAN, etc., mainly rely on Euclidean distance for similarity calculation. However, public opinion dissemination involves complex spatio-temporal interaction relationships, and the geometric distance in a low-dimensional space alone cannot fully characterize the public opinion dissemination mode. Although the clustering method based on graph embedding can utilize the topological information of nodes, it is difficult to accurately depict the public opinion dissemination mode if it does not combine specific features of public opinion dissemination, such as dissemination similarity, interaction intensity, etc. In addition, the method with a fixed number of clusters is difficult to adapt to the complexity of different public opinion events, resulting in unstable clustering results. Therefore, the existing clustering analysis methods have certain limitations in mining the public opinion dissemination mode, and it is necessary to combine the node embedding features of the hypergraph convolutional neural network and adopt an adaptive clustering strategy to dynamically adjust the number of categories according to the actual distribution of the dissemination mode, so as to more accurately mine the public opinion dissemination mode.

[0007] In summary, the existing public opinion dissemination analysis methods have deficiencies in aspects such as high-order dissemination relationship modeling, deep feature learning, model optimization, and dissemination mode clustering. First, ordinary graph neural network methods are difficult to effectively capture the high-order interaction relationships in the public opinion dissemination process, restricting the model's ability to express complex dissemination modes; second, traditional parameter optimization methods are prone to falling into local optima and cannot fully optimize the key parameters of the hypergraph convolutional neural network model; in addition, the existing clustering analysis methods have problems such as insufficient accuracy and fixed number of categories in dissemination mode recognition, and it is difficult to accurately depict the public opinion dissemination mode. Therefore, there is an urgent need for a regional public opinion dissemination mode mining method based on hypergraph modeling, combined with hypergraph convolutional neural network and quantum particle swarm optimization, to better identify the public opinion dissemination mode and improve the accuracy and robustness of public opinion analysis.

[0008] Therefore, how to provide a regional public opinion dissemination mode mining method based on graph neural network is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose a regional public opinion dissemination mode mining method based on graph neural network, which significantly improves the accuracy, robustness, and global optimization ability of regional public opinion dissemination mode mining.

[0010] The regional public opinion dissemination mode mining method based on graph neural network according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect public opinion data in the target area and preprocess the public opinion data;

[0012] S2. Use the hypergraph modeling technology to construct an opinion dissemination hypergraph model. Map the preprocessed opinion data to the nodes in the hypergraph, construct hyperedges according to the characteristics of multi-agent participation in regional opinion events, which reflect the high-order interaction relationships among multiple nodes, and use high-order graph theory technology to assign weights to the hyperedges;

[0013] S3. Based on the opinion dissemination hypergraph model, use the hypergraph convolutional neural network for deep feature learning, which includes multiple hypergraph convolutional layers. Each layer has learnable hyperedge weights, defines the node embedding dimension, and dynamically adjusts the activation function parameters. Through convolution, activation, and pooling operations, generate a high-dimensional node embedding matrix to reflect the dissemination characteristics, and construct a hypergraph convolutional neural network model;

[0014] S4. Use the improved quantum particle swarm optimization algorithm to globally optimize the key parameters of the hypergraph convolutional neural network model. Take the hyperedge weights, node embedding dimensions, and activation function parameters in the hypergraph convolutional neural network model as optimization variables, and use the quantum state update mechanism to iteratively adjust the parameters to make the predefined objective function reach the global optimum, and output the optimized hypergraph convolutional neural network model;

[0015] S5. Use the optimized hypergraph convolutional neural network model for clustering analysis to automatically mine and identify typical opinion dissemination patterns within the region;

[0016] S6. Output the results of the opinion dissemination pattern, and use the results for opinion risk assessment, visualization display, and decision support.

[0017] Optionally, the S2 specifically includes:

[0018] S21. Map the preprocessed opinion data to the nodes in the hypergraph, and set the opinion dissemination hypergraph model G=(V, E, W), where the node set V consists of user nodes, information source nodes, and event nodes. The user nodes represent social media users, the information source nodes represent the opinion dissemination entities, the event nodes represent the topic words of the opinion events, W is the hyperedge weight matrix, and E is the hyperedge set;

[0019] S22. Construct hyperedges and establish multi-agent dissemination relationships. Based on the user interaction behaviors, common topic participation situations, and dissemination time windows in the opinion data, construct the hyperedge set E, where the hyperedge e i consists of multiple nodes and represents the high-order interaction relationships of multiple agents participating together;

[0020] S23. Assign weights to the hyperedges to obtain the hyperedge weight matrix W. The specific steps include:

[0021] S231. Calculate the interaction intensity f int (e i ), which is calculated based on the user interaction behaviors and satisfies:

[0022]

[0023] Among them, R(v, v ′ ), C(v, v ′ ), and L(v, v ′ ) respectively represent the number of forwarding times, the number of comments, and the number of likes between nodes v and v ′ . β1, β2, and β3 are the weighted coefficients of interaction behaviors;

[0024] S232. Calculate the information similarity f sim (e i ) by using the text embedding method, satisfying:

[0025]

[0026] Among them, T(v) and T(v ′ ) represent the text feature representations of nodes v and v ′ , and cos(T(v), T(v ′ )) is the cosine similarity of the text vectors;

[0027] S233. Calculate the time decay factor f time (e i ) satisfying:

[0028]

[0029] Among them, t0 is the initial time of information dissemination, t is the current time, and λ is the time decay coefficient;

[0030] S234. Set the hyperedge weight matrix W. The weight w i of each hyperedge e i is calculated from the interaction intensity f int (e i ), the information similarity f sim (e i ), and the time decay factor f time (e i ), satisfying:

[0031] w i = α1f int (e i ) + α2f sim (e i ) + α3f time (e i );

[0032] Among them, α1, α2, and α3 are normalization coefficients;

[0033] S24. Output the public opinion dissemination hypergraph model, where the hypergraph model includes a node set V, a hyperedge set E, and a hyperedge weight matrix W.

[0034] Optionally, the S3 specifically includes:

[0035] S31. Set the input of the hypergraph convolutional neural network model as the public opinion dissemination hypergraph model G=(V, E, W), and use the hypergraph adjacency matrix and the hypergraph Laplacian matrix as the graph structure representation for convolutional calculation.

[0036] S32. Define the hypergraph convolution operation, calculate the embedding representation of each node based on the hypergraph structure, perform normalization processing using the hypergraph Laplacian matrix, and combine with the learnable weight matrix to extract the local and global propagation features of the nodes through multi-layer hypergraph convolution operations.

[0037] S33. Extract the node embedding representation, use the multi-layer hypergraph convolution calculation of the hypergraph convolutional neural network model to update the node feature representation layer by layer, and output the node features of the final layer as a high-dimensional embedding matrix, which is used to represent the public opinion dissemination pattern.

[0038] S34. After each layer of hypergraph convolution calculation, use a non-linear activation function to enhance the feature expression ability, apply a pooling operation to reduce the feature dimension, and reduce redundant information.

[0039] S35. Output the hypergraph convolutional neural network model, and use the finally calculated high-dimensional node embedding matrix as the output of the hypergraph convolutional neural network model.

[0040] Optionally, the S4 specifically includes:

[0041] S41. Initialize the optimization variables and parameters, set the hyperedge weight matrix W, the node embedding dimension d, and the activation function parameter θ in the hypergraph convolutional neural network model as the optimization variables, initialize the particle swarm size N, the maximum number of iterations T, the inertia weight w, the individual learning factor c1, and the global learning factor c2, and randomly initialize the position matrix X={x1, x2,..., x N} and the particle velocity matrix V={v1, v2,..., v N} of the particle swarm using a uniform distribution, where X represents the current parameter position of the particle swarm, including the hyperedge weight W, the node embedding dimension d, and the activation function parameter θ, V represents the movement speed of each particle in the search space, and N is the total number of particles in the particle swarm.

[0042] S42. Perform global search based on the quantum state update mechanism. In the initial stage of the search, use the quantum state update mechanism to perform global search and calculate the new position of the particle in the global search space:

[0043]

[0044] Among them, represents the new position of the i-th particle after global search, is the optimal solution found by this particle during historical iterations, and g best represents the global optimal solution in the current particle swarm. w is the inertia weight that controls the search step size, which is larger in the initial stage of optimization and gradually decreases in the later stage of optimization. β is the quantum influence factor that adjusts the exploration ability of the search space. ln represents the logarithmic function with base e, and u is a random number between [0, 1];

[0045] S43. Based on the global search results Utilize the quantum state update mechanism for local optimization and calculate the refined search position of the particle:

[0046]

[0047] Among them, is the new position after local optimization. η is the measurement accuracy control factor for local search, which adjusts the step size of local search, depends on the current velocity direction of the particle of, and L is the quantum measurement scaling factor that controls the search range of local optimization;

[0048] S44. Iteratively update the optimization variables and perform parameter update based on to gradually converge the parameters of the optimized hypergraph convolutional neural network:

[0049]

[0050] Among them, is the hyperedge weight after local optimization calculated by local optimization, is the node embedding dimension after local optimization calculated by local optimization, is the activation function parameter after local optimization calculated by local optimization, and α t is the adaptive learning rate;

[0051] S45. Use the finally converged and optimized hyperedge weight node embedding dimension and activation function parameter as the final optimization result and output the optimized hypergraph convolutional neural network model.

[0052] Optionally, the S5 specifically includes:

[0053] S51. Extract the node embedding representation of the optimized hypergraph convolutional neural network model. Based on the optimized hypergraph convolutional neural network model, extract the node embedding matrix Z;

[0054] S52. Based on the node embedding matrix Z, the propagation similarity matrix S is calculated using cosine similarity, where for any nodes i, j:

[0055]

[0056] where ||·|| represents the Euclidean norm, and Z j represents the embedding vector of the j-th public opinion propagation node, and Z i represents the embedding vector of the i-th public opinion propagation node;

[0057] S53. Based on the propagation similarity, clustering analysis is performed. The nodes are clustered using the S matrix, and the adaptive clustering method is used to dynamically determine the number of categories k of the public opinion propagation mode. The classification result C of the public opinion propagation mode is constructed. First, density analysis is performed based on S to determine the core propagation nodes, and the nodes are grouped according to the similarity of the propagation nodes. Then, the central feature of each category is calculated according to the propagation mode category C to obtain the category center vector:

[0058]

[0059] where μ c represents the central feature of category C c , |C c | is the number of nodes included in this category, reflecting the representative feature of the public opinion propagation mode of this category;

[0060] S54. Evaluate the clustering quality and output the public opinion propagation mode. Calculate the clustering compactness D intra and the separation degree D inter between categories, and output the analysis result of the public opinion propagation mode, including the propagation mode category, the central feature of the category, and the association degree between categories, and use it for public opinion risk assessment, visualization display, and decision support.

[0061] The beneficial effects of the present invention are:

[0062] (1) The present invention adopts the hypergraph modeling technology, maps the users, information sources, and events in the public opinion propagation process into a hypergraph structure, and constructs a high-order propagation relationship of multiple subjects. Compared with the traditional ordinary graph neural network method that can only learn the first-order or second-order adjacency relationship, the hypergraph modeling method of the present invention can effectively depict the complex information propagation mode between multiple nodes, and completely express the multi-subject collaborative propagation behavior in the public opinion propagation process. By endowing the hyperedges with weights through high-order graph theory technology, the present invention can more precisely measure the propagation influence, ensure the accuracy of the public opinion propagation modeling, fully excavate the high-dimensional propagation relationship, and overcome the limitation that the traditional graph neural network model is difficult to handle the propagation participated by multiple parties.

[0063] (2) During the training process of the hypergraph convolutional neural network of the present invention, an improved quantum particle swarm optimization algorithm is introduced to globally optimize the hyperedge weights, node embedding dimensions, and activation function parameters. Compared with the traditional gradient descent optimization method, quantum particle swarm optimization performs global search through the quantum state update mechanism and combines local optimization strategies to dynamically adjust the key parameters of the hypergraph convolutional neural network, thereby avoiding falling into local optima during the optimization process and improving the convergence speed and stability of the model. The introduction of quantum particle swarm optimization enables the hypergraph convolutional neural network model to quickly adapt to different public opinion dissemination scenarios, improving the robustness and generalization ability of the model in complex dissemination environments.

[0064] (3) The present invention adopts an adaptive clustering analysis method based on the embedded features of the hypergraph convolutional neural network. By calculating the propagation similarity matrix, the number of propagation pattern categories is automatically determined, overcoming the problems of traditional clustering methods being sensitive to the number of categories and having poor stability. Compared with the K-means or DBSCAN methods with fixed number of categories, the adaptive clustering strategy of the present invention can dynamically adjust the clustering parameters according to the actual distribution of propagation patterns, thereby accurately identifying different types of public opinion dissemination patterns. In addition, the present invention calculates the central feature vector of the propagation pattern and evaluates the clustering quality through compactness and separation to ensure the rationality of the clustering results. Finally, the proposed method can more accurately identify public opinion dissemination patterns, providing more accurate data basis for public opinion risk assessment, propagation path prediction, and decision-making support. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the accompanying drawings:

[0066] Figure 1 is the overall flowchart of the regional public opinion dissemination pattern mining method based on the graph neural network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0068] Refer to Figure 1 , the regional public opinion dissemination pattern mining method based on the graph neural network includes the following steps:

[0069] S1. Collect public opinion data in the target area and preprocess the public opinion data;

[0070] S2. Use the hypergraph modeling technique to construct a public opinion dissemination hypergraph model. Map the preprocessed public opinion data to the nodes in the hypergraph. According to the characteristics of multi-agent participation in regional public opinion events, construct hyperedges to reflect the high-order interaction relationships among multiple nodes, and use high-order graph theory techniques to assign weights to the hyperedges.

[0071] S3. Based on the public opinion dissemination hypergraph model, use the hypergraph convolutional neural network for deep feature learning, which includes multiple hypergraph convolutional layers. Each layer has learnable hyperedge weights, defines the node embedding dimension, and dynamically adjusts the activation function parameters. Through convolution, activation, and pooling operations, generate a high-dimensional node embedding matrix to reflect the dissemination characteristics and construct a hypergraph convolutional neural network model.

[0072] S4. Use the improved quantum particle swarm optimization algorithm to globally optimize the key parameters of the hypergraph convolutional neural network model. Take the hyperedge weights, node embedding dimensions, and activation function parameters in the hypergraph convolutional neural network model as optimization variables, and use the quantum state update mechanism to iteratively adjust the parameters to make the predefined objective function reach the global optimum, and output the optimized hypergraph convolutional neural network model.

[0073] S5. Use the optimized hypergraph convolutional neural network model for clustering analysis to automatically mine and identify typical public opinion dissemination patterns within the region.

[0074] S6. Output the results of the public opinion dissemination pattern, and use the results for public opinion risk assessment, visualization display, and decision support.

[0075] In this embodiment, the S2 specifically includes:

[0076] S21. Map the preprocessed public opinion data to the nodes in the hypergraph. Set the public opinion dissemination hypergraph model G=(V, E, W), where the node set V consists of user nodes, information source nodes, and event nodes. The user nodes represent social media users, the information source nodes represent the public opinion dissemination entities, the event nodes represent the topic words of public opinion events, W is the hyperedge weight matrix, and E is the hyperedge set.

[0077] S22. Construct hyperedges and establish multi-agent dissemination relationships. Based on the user interaction behavior, common topic participation, and dissemination time window in the public opinion data, construct the hyperedge set E, where the hyperedge e i is composed of multiple nodes, representing the high-order interaction relationships of multiple agents participating together.

[0078] S23. Assign weights to the hyperedges to obtain the hyperedge weight matrix W. The specific steps include:

[0079] S231. Calculate the interaction intensity f int (e i ), which is calculated based on the user's interaction behavior and satisfies:

[0080]

[0081] Among them, R(v, v ′ ), C(v, v ′ ), and L(v, v ′ ) respectively represent the number of forwarding times, the number of comments, and the number of likes between nodes v and v ′ . β1, β2, and β3 are the weighting coefficients of interaction behaviors;

[0082] S232. Calculate the information similarity f sim (e i ) using the text embedding method, satisfying:

[0083]

[0084] Among them, T(v) and T(v ′ ) represent the text feature representations of nodes v and v ′ , and cos(T(v), T(v ′ )) is the cosine similarity of the text vectors;

[0085] S233. Calculate the time decay factor f time (e i ) satisfying:

[0086]

[0087] Among them, t0 is the initial time of information dissemination, t is the current time, and λ is the time decay coefficient;

[0088] S234. Set the hyperedge weight matrix W, and the weight w i of each hyperedge e i is calculated from the interaction intensity f int (e i ), the information similarity f sim (e i ), and the time decay factor f time (e i ), satisfying:

[0089] w i = α1f int (e i ) + α2f sim (e i ) + α3f time (e i );

[0090] Among them, α1, α2, and α3 are normalization coefficients;

[0091] S24. Output the public opinion dissemination hypergraph model, where the hypergraph model includes a node set V, a hyperedge set E, and a hyperedge weight matrix W.

[0092] In this embodiment, the S3 specifically includes:

[0093] S31. Set the input of the hypergraph convolutional neural network model as the public opinion dissemination hypergraph model G=(V, E, W), and use the hypergraph adjacency matrix and the hypergraph Laplacian matrix as the graph structure representation for convolutional calculation.

[0094] S32. Define the hypergraph convolution operation, calculate the embedding representation of each node based on the hypergraph structure, perform normalization processing using the hypergraph Laplacian matrix, combine with the learnable weight matrix, and extract the local and global propagation features of the nodes through multi-layer hypergraph convolution operations.

[0095] S33. Extract the node embedding representation, use the multi-layer hypergraph convolution calculation of the hypergraph convolutional neural network model to update the node feature representation layer by layer, and output the node features of the final layer as a high-dimensional embedding matrix, which is used to represent the public opinion dissemination pattern.

[0096] S34. After each layer of hypergraph convolution calculation, use a non-linear activation function to enhance the feature expression ability, apply a pooling operation to reduce the feature dimension, and reduce redundant information.

[0097] S35. Output the hypergraph convolutional neural network model, and use the finally calculated high-dimensional node embedding matrix as the output of the hypergraph convolutional neural network model.

[0098] In this embodiment, the S4 specifically includes:

[0099] S41. Initialize the optimization variables and parameters, set the hyperedge weight matrix W, the node embedding dimension d, and the activation function parameter θ in the hypergraph convolutional neural network model as the optimization variables, initialize the particle swarm size N, the maximum number of iterations T, the inertia weight w, the individual learning factor c1, and the global learning factor c2, and randomly initialize the position matrix X={x1, x2,..., x N} and the particle velocity matrix V={v1, v2,..., v N} of the particle swarm using a uniform distribution, where X represents the current parameter position of the particle swarm, including the hyperedge weight W, the node embedding dimension d, and the activation function parameter θ, V represents the moving speed of each particle in the search space, and N is the total number of particles in the particle swarm.

[0100] S42. Perform global search based on the quantum state update mechanism. In the initial stage of the search, use the quantum state update mechanism to perform global search and calculate the new position of the particle in the global search space:

[0101]

[0102] Among them, represents the new position of the i-th particle after global search, is the optimal solution found by the particle during the historical iteration process, and g best represents the global optimal solution in the current particle swarm. w is the inertia weight, which controls the search step size. It is larger in the initial stage of optimization and gradually decreases in the later stage of optimization. β is the quantum influence factor, which adjusts the exploration ability of the search space. ln represents the logarithmic function with base e, and u is a random number between [0, 1];

[0103] S43. Based on the global search results Use the quantum state update mechanism for local optimization to calculate the refined search position of the particle:

[0104]

[0105] Among them, is the new position after local optimization. η is the measurement accuracy control factor for local search, which adjusts the step size of local search, depends on the current velocity of the particle in the direction, and L is the quantum measurement scaling factor, which controls the search range of local optimization;

[0106] S44. Iteratively update the optimization variables, and based on perform parameter update to gradually converge the optimized parameters of the hypergraph convolutional neural network:

[0107]

[0108] Among them, is the hyperedge weight after local optimization calculated by local optimization, is the node embedding dimension after local optimization calculated by local optimization, is the activation function parameter after local optimization calculated by local optimization, and α t is the adaptive learning rate;

[0109] S45. Take the finally converged and optimized hyperedge weight node embedding dimension and activation function parameter as the final optimization result, and output the optimized hypergraph convolutional neural network model.

[0110] In this embodiment, the S5 specifically includes:

[0111] S51. Extract the node embedding representation of the optimized hypergraph convolutional neural network model. Based on the optimized hypergraph convolutional neural network model, extract the node embedding matrix Z;

[0112] S52. Based on the node embedding matrix Z, the propagation similarity matrix S is calculated using cosine similarity, where for any node i, j:

[0113]

[0114] Among them, ||·|| represents the normalized Euclidean distance, Z j represents the embedding vector of the jth public opinion propagation node, Z i Represents the embedding vector of the i-th public opinion propagation node;

[0115] S53, cluster analysis is performed based on the similarity of propagation, nodes are clustered using the S matrix, and the number of categories k of the public opinion propagation mode is dynamically determined using an adaptive clustering method. The classification result C of the public opinion propagation mode is constructed. First, density analysis is performed based on S to determine the core propagation nodes, and the propagation nodes are grouped according to their similarity. Then, the central feature of each category is calculated according to the propagation mode category C, and the category center vector is obtained:

[0116]

[0117] Among them, μ c Indicates category C c The central feature of |C c | is the number of nodes included in this category, reflecting the representative characteristics of this type of public opinion propagation model;

[0118] S54. Evaluate clustering quality and output public opinion propagation model, calculate clustering density D intra and the separation between categories D inter , output the analysis results of public opinion propagation patterns, including propagation pattern categories, category center characteristics and the degree of correlation between categories, and are used for public opinion risk assessment, visualization and decision support.

[0119] Embodiment 1:

[0120] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a social media public opinion dissemination analysis system in a provincial capital city. The system is used in media monitoring platforms and corporate brand management departments to monitor and analyze regional public opinion dissemination and assist in decision-making. The goal of this experiment is to use the method of the present invention to explore the public opinion dissemination pattern of the region in public emergencies and verify its effectiveness in terms of accuracy, dissemination pattern recognition ability, optimization ability, etc.

[0121] In this experiment, the public opinion data of a certain hot event on social media in the provincial capital city from May 10, 2023 to June 10, 2023 was selected. This event involves social security issues and has attracted wide attention and discussion. The data sources include Weibo, WeChat public platforms, news portals and short video platforms. A total of about 2.6 million original data were obtained. The data includes information such as user posts, forwards, comments, likes, as well as article content, release time, interaction records, etc. Since the spread of public opinion has regionality, only the interaction data of users within the city was retained. After screening, 1.32 million valid data were obtained.

[0122] First, these data were preprocessed, including removing irrelevant information, de-duplication, text cleaning, sentiment classification, etc. Subsequently, based on the hypergraph modeling technology of the present invention, a public opinion propagation hypergraph was constructed. Users, information sources (news, official accounts, etc.) and event keywords were used as hypergraph nodes, and hyperedges were constructed based on user interaction behaviors (comments, forwards, likes), information similarity (text content similarity) and time decay weights. Finally, a hypergraph structure was formed. The hypergraph constructed in this experiment contains 68,420 nodes and 153,827 hyperedges. Among them, each hyperedge is connected to an average of 4.2 nodes, indicating that the high-order interaction relationship between multiple entities is strong. Compared with the traditional ordinary graph model, the method of the present invention can more accurately depict the multi-agent propagation mode in the process of public opinion propagation, avoiding the defect that the ordinary graph structure cannot effectively represent group propagation.

[0123] In the process of feature learning of the public opinion propagation mode, a hypergraph convolutional neural network (HGCN) was used to extract propagation features. The training data set was trained with the first 80% of the data, and the last 20% was used as the test set. HGCN underwent 5 layers of hypergraph convolutional operations and used the ReLU activation function for non-linear mapping, and finally generated a 256-dimensional high-dimensional node embedding matrix. Since a large number of parameter adjustments are involved in the HGCN training process, an improved quantum particle swarm optimization (QPSO) algorithm was introduced in this experiment to globally optimize the key parameters of HGCN. The experimental results show that QPSO converges to the optimal solution after 50 iterations, and the hyperedge weights, node embedding dimensions and activation function parameters all reach the global optimal state. Compared with the traditional Adam and SGD training methods, the optimization method of the present invention reduces the HGCN training time by 37.4% on the same data set, and the final prediction accuracy is increased by 5.2%.

[0124] Subsequently, based on the high-dimensional node embedding representation calculated by HGCN, the adaptive clustering analysis method of the present invention is adopted to calculate the propagation similarity matrix and classify the public opinion propagation patterns. In this experiment, the similarity degree of different public opinion propagation subjects is first calculated using the propagation similarity matrix, and the density analysis method is used to determine the core propagation group. Subsequently, through the adaptive clustering algorithm, the optimal number of public opinion propagation categories is automatically calculated, and a total of 7 main propagation patterns are identified, including institution-guided type, media-driven type, big V opinion leader type, hot discussion type, rumor diffusion type, emotional catharsis type, and silent attention type. Compared with the traditional K-means method, the clustering method of the present invention can dynamically adjust the number of categories, improve the stability of pattern recognition, and maintain good clustering performance on different data sets.

[0125] To further verify the effectiveness of the present invention in public opinion propagation analysis, the performance of the method of the present invention is compared with existing methods such as ordinary graph neural networks (GCN) and long short-term memory networks (LSTM) in public opinion propagation pattern recognition. The following table shows the comparison data of the public opinion propagation pattern recognition accuracy, calculation time, and convergence speed of this experiment under different methods:

[0126] Table 1: Performance comparison between the method of the present invention and existing methods

[0127]

[0128] The experimental results show that the hypergraph modeling, HGCN, and QPSO methods of the present invention perform excellently in public opinion propagation pattern recognition and are superior to existing GCN, LSTM, and traditional clustering methods in multiple key indicators. First, in terms of the accuracy of propagation pattern recognition, the accuracy of the present invention reaches 89.3%, which is 7.8% higher than that of the ordinary graph neural network (GCN) and 10.5% higher than that of LSTM. This improvement is mainly due to the adoption of hypergraph modeling in the present invention, which more comprehensively depicts the high-order interaction relationships of multiple subjects in the public opinion propagation process, making the propagation pattern recognition more accurate. At the same time, the hypergraph convolutional neural network (HGCN) combines multi-layer hypergraph convolutional operations, which can more fully extract the propagation features of nodes during feature learning and improve the discrimination of public opinion propagation patterns. In addition, the improved quantum particle swarm optimization (QPSO) method further enhances the optimization ability of the model, making the finally recognized propagation pattern more stable and reliable.

[0129] In terms of computational efficiency, the training time of the present invention is 4.2 hours, which has obvious advantages compared with 6.8 hours of the GCN method and 7.4 hours of the LSTM method, reducing the computational time by about 35%-45%. This advantage is mainly attributed to the introduction of the QPSO optimization strategy, which makes the parameter convergence speed faster during the HGCN training process and avoids the oscillation problem that may occur in the traditional gradient descent optimization method. Compared with the situation where GCN and LSTM rely on a large number of manually adjusted hyperparameters, the quantum particle swarm optimization method of the present invention can adjust the hypergraph model parameters more automatically, making the training process more efficient. In addition, compared with the traditional K-means clustering method, although the computational time of the present invention is slightly longer, since K-means cannot capture complex propagation patterns, the final recognition accuracy is only 72.4%, which is difficult to meet the requirements of high-precision public opinion propagation analysis.

[0130] In terms of parameter optimization, the optimization convergence time of the present invention only requires 50 iterations, far less than 75 times of GCN and 85 times of LSTM. QPSO combines the quantum state update mechanism and the local search strategy, achieving a better balance between global search and local optimization, enabling the key parameters of HGCN (hyperedge weight, node embedding dimension, and activation function parameters) to quickly converge to the optimal solution. In contrast, traditional stochastic gradient descent (SGD) or Adam optimization methods are prone to falling into local optima in high-dimensional spaces, resulting in the need for more iterations to converge. Through the adaptive search strategy of QPSO, the present invention can find the optimal parameters within fewer iterations, improving the overall optimization efficiency.

[0131] In summary, the method of the present invention performs excellently in terms of the accuracy, computational efficiency, and convergence speed of public opinion propagation pattern recognition, can more accurately model the high-order interaction relationships in the public opinion propagation process, while reducing the computational overhead and improving the stability of propagation pattern analysis. The experimental results further verify the effectiveness of the present invention, demonstrating its practical application value in large-scale social media public opinion analysis tasks, and providing an efficient and accurate public opinion propagation pattern mining tool for institutions, media monitoring platforms, and enterprise brand management.

[0132] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A method for mining regional public opinion dissemination patterns based on graph neural networks, characterized in that, It includes the following steps: S1. Collect the public opinion data in the target area and preprocess the public opinion data; S2. Use the hypergraph modeling technology to construct a public opinion propagation hypergraph model, map the preprocessed public opinion data to the nodes in the hypergraph, construct hyperedges according to the characteristics of multi-agent participation in regional public opinion events to reflect the high-order interaction relationship between multiple nodes, and use the high-order graph theory technology to assign weights to the hyperedges; S3. Based on the public opinion propagation hypergraph model, use the hypergraph convolutional neural network for deep feature learning, which includes multiple hypergraph convolutional layers. Each layer has learnable hyperedge weights, defines the node embedding dimension and dynamically adjustable activation function parameters. Through convolution, activation and pooling operations, generate a high-dimensional node embedding matrix to reflect the propagation characteristics and construct a hypergraph convolutional neural network model; S4. Use the improved quantum particle swarm optimization algorithm to globally optimize the key parameters of the hypergraph convolutional neural network model. Take the hyperedge weights, node embedding dimensions and activation function parameters in the hypergraph convolutional neural network model as optimization variables, and use the quantum state update mechanism to iteratively adjust the parameters to make the predefined objective function reach the global optimum, and output the optimized hypergraph convolutional neural network model; S5. Use the optimized hypergraph convolutional neural network model for clustering analysis to automatically mine and identify the typical public opinion propagation patterns in the region; S6. Output the public opinion propagation pattern results and use the results for public opinion risk assessment, visualization display and decision support.

2. The method for mining the regional public opinion dissemination pattern based on the graph neural network according to claim 1, wherein The specific content of S2 includes: S21. Map the preprocessed public opinion data to the nodes in the hypergraph. Set the public opinion propagation hypergraph model G=(V, E, W), where the node set V consists of user nodes, information source nodes and event nodes. The user nodes represent social media users, the information source nodes represent the public opinion propagation subjects, the event nodes represent the topic words of public opinion events, W is the hyperedge weight matrix, and E is the hyperedge set; S22. Construct hyperedges and establish multi-agent propagation relationships. Based on the user interaction behaviors, common topic participation, and propagation time window in the public opinion data, construct a set of hyperedges E, where a hyperedge e i is composed of multiple nodes and represents a high-order interaction relationship jointly participated by multiple agents; S23. Assign weights to the hyperedges to obtain the hyperedge weight matrix W. The specific steps include: S231. Calculate the interaction intensity f int (e i ), calculated based on the user's interaction behavior, satisfying: Among them, R(v, v ′ ), C(v, v ′ ), and L(v, v ′ ) respectively represent the number of forwarding times, the number of comment times, and the number of like times between nodes v and v ′ , and β1, β2, and β3 are the weighted coefficients of interaction behaviors; S232. Calculate the information similarity f sim (e i ), which is calculated using the text embedding method and satisfies: Among them, T(v) and T(v ′ ) represent the text feature representations of nodes v and v ′ , and cos(T(v), T(v ′ )) is the cosine similarity of the text vectors; S233. Calculate the time decay factor f time (e i ), satisfying: f time (e i ) = e -λ(t-t0) ; where t0 is the initial time of information propagation, t is the current time, and λ is the time decay coefficient; S234. Set the hyperedge weight matrix W, and the weight w i of each hyperedge e i is calculated from the interaction intensity f int (e i ), the information similarity f sim (e i ), and the time decay factor f time (e i ), and satisfies: w i = α1f int (e i ) + α2f sim (e i ) + α3f time (e i ); where α1, α2, α3 are normalization coefficients; S24. Output the public opinion propagation hypergraph model. The hypergraph model includes the node set V, the hyperedge set E and the hyperedge weight matrix W.

3. The method for mining the regional public opinion dissemination pattern based on the graph neural network according to claim 1, characterized in that, The specific content of S3 includes: S31. Set the input of the hypergraph convolutional neural network model as the public opinion propagation hypergraph model G=(V, E, W), and use the hypergraph adjacency matrix and the hypergraph Laplacian matrix as the graph structure representation for convolution calculation; S32. Define the hypergraph convolution operation, calculate the embedding representation of each node based on the hypergraph structure, perform normalization processing using the hypergraph Laplacian matrix, and combine the learnable weight matrix to extract the local and global propagation characteristics of the nodes through multiple hypergraph convolution operations; S33. Extract the node embedding representation, use the multiple hypergraph convolution calculations of the hypergraph convolutional neural network model to update the node feature representation layer by layer, and output the node features of the final layer as a high-dimensional embedding matrix. This embedding matrix is used to represent the public opinion propagation pattern; S34. After each layer of hypergraph convolution calculation, a non-linear activation function is used to enhance the feature expression ability, and a pooling operation is applied to reduce the feature dimension and reduce redundant information; S35. Output the hypergraph convolutional neural network model, and use the finally calculated high-dimensional node embedding matrix as the output of the hypergraph convolutional neural network model.

4. The method for mining the regional public opinion dissemination mode based on the graph neural network according to claim 1, wherein The specific content of S4 includes: S41. Initialize the optimization variables and parameters. Set the hyperedge weight matrix W, the node embedding dimension d, and the activation function parameter θ in the hypergraph convolutional neural network model as optimization variables. Initialize the particle swarm size N, the maximum number of iterations T, the inertia weight w, the individual learning factor c1, and the global learning factor c2. Randomly initialize the position matrix X = {x1, x2,..., x N} and the particle velocity matrix V = {v1, v2,..., v N} using a uniform distribution. Here, X represents the current parameter position of the particle swarm, including the hyperedge weight W, the node embedding dimension d, and the activation function parameter θ. V represents the movement speed of each particle in the search space, and N is the total number of particles in the particle swarm; S42. Perform global search based on the quantum state update mechanism. In the initial stage of the search, use the quantum state update mechanism to perform global search and calculate the new position of the particle in the global search space: Among them, represents the new position of the i-th particle after global search, is the optimal solution found by the particle during the historical iteration process, and g best represents the global optimal solution in the current particle swarm. w is the inertia weight, which controls the search step size and is relatively large in the initial stage of optimization and gradually decreases in the later stage of optimization. β is the quantum influence factor, which adjusts the exploration ability of the search space. ln represents the logarithmic function with base e, and u is a random number between [0, 1]; S43. Based on the global search results Use the quantum state update mechanism to perform local optimization and calculate the refined search position of the particle: Among them, is the new position after local optimization. η is the measurement accuracy control factor for local search, which adjusts the step size of local search. depends on the current velocity of the particle in the direction. L is the quantum measurement scaling factor, which controls the search range of local optimization. S44. Iteratively update and optimize variables, and update parameters based on to gradually converge the parameters of the optimized hypergraph convolutional neural network: Among them, is the locally optimized hyperedge weight obtained by local optimization calculation, is the locally optimized node embedding dimension obtained by local optimization calculation, is the locally optimized activation function parameter α obtained by local optimization calculation, t is the adaptive learning rate; S45. Take the hyperedge weights after the final convergence optimization Node embedding dimension and the activation function parameters As the final optimization result, output the optimized hypergraph convolutional neural network model.

5. The method for mining the regional public opinion dissemination mode based on the graph neural network according to claim 1, wherein The specific content of S5 includes: S51. Extract the node embedding representation of the optimized hypergraph convolutional neural network model. Based on the optimized hypergraph convolutional neural network model, extract the node embedding matrix Z; S52. Based on the node embedding matrix Z, use the cosine similarity to calculate the propagation similarity matrix S, where for any nodes i, j: Among them, ||·|| represents the Euclidean distance of the norm, and Z j represents the embedding vector of the j-th public opinion dissemination node, and Z i represents the embedding vector of the i-th public opinion dissemination node; S53. Perform clustering analysis based on the propagation similarity. Use the S matrix to cluster the nodes, and use an adaptive clustering method to dynamically determine the number of categories k of the public opinion propagation pattern, and construct the public opinion propagation pattern classification result C. First, perform density analysis based on S to determine the core propagation nodes, group them according to the similarity of the propagation nodes, and then calculate the central feature of each category according to the propagation pattern category C to obtain the category center vector: Among them, μ c represents the central feature of category C c , |C c | is the number of nodes included in this category, reflecting the representative feature of the public opinion dissemination mode of this category; S54. Evaluate the clustering quality and output the public opinion dissemination pattern, and calculate the clustering compactness D intra and the separation degree D between categories inter , output the analysis results of the public opinion dissemination pattern, including the dissemination pattern category, the characteristics of the category center, and the degree of association between categories, and use them for public opinion risk assessment, visual display, and decision support.

Images