Network public opinion monitoring and risk early warning method based on AI technology and deep learning

Abstract

This invention discloses a method for monitoring and warning of online public opinion based on AI technology and deep learning, comprising the following steps: S1, collecting and preprocessing online public opinion data to construct a public opinion propagation graph; S2, initializing the graph structure, constructing a node attribute matrix and an adjacency matrix, and dividing it into a dynamic time-series graph set according to time; S3, using a graph convolutional neural network to encode node features and extract structural influence features; S4, constructing a SEIR variant model, setting a set of node states, and introducing a state memory vector for time decay weighting; S5, performing state propagation and updating, and recording the complete state update trajectory; S6, analyzing changes in infected state nodes to form structured public opinion monitoring results; S7, providing real-time warnings based on monitoring results and risk thresholds to achieve automated public opinion monitoring and warning. This invention combines graph neural networks and state modeling to improve the accuracy of public opinion evolution identification and the efficiency of warning.

Description

Technical Field

[0001] This invention relates to the field of big data public opinion analysis and risk management, and in particular to a method for monitoring and warning of online public opinion based on AI technology and deep learning. Background Technology

[0002] In the context of highly developed internet information technology, online public opinion has become a significant factor influencing social stability, government image, corporate reputation, and public sentiment. Especially with the rapid development of social media, self-media platforms, and short video applications, online public opinion possesses the typical characteristics of "rapid dissemination, diversified content, and strong emotional contagion." Once negative information emerges, its spread far exceeds the information dissemination capabilities of the traditional media era. To address this information dissemination trend, how to monitor and analyze online public opinion in a timely and accurate manner has become a pressing practical problem that various institutions urgently need to solve.

[0003] Existing online public opinion monitoring technologies mainly include keyword-matching-based sentiment collection, dictionary-rule-based sentiment recognition, and statistical methods-based dissemination trend analysis. While these methods can provide a rough analysis of online content in the early stages, they have revealed several shortcomings in the face of today's complex online semantics, diverse information sources, and non-linear evolution of content dissemination paths. First, traditional keyword rules suffer from incomplete semantic coverage; their accuracy drops significantly when faced with a large amount of metaphor, irony, puns, or emerging online language. Second, statistical analysis models cannot effectively combine user interaction behavior with the dissemination network structure, making it difficult to depict the true information diffusion path and thus unable to accurately predict potential dissemination risks.

[0004] In recent years, the rise of artificial intelligence, especially deep learning technology, has brought new solutions to online public opinion analysis. Natural language processing technology has made significant breakthroughs in text classification, sentiment analysis, and topic modeling; graph neural networks, Transformer models, and other models have also provided stronger capabilities for modeling complex structured data. However, existing research on applying AI technology to public opinion monitoring still has significant limitations. On the one hand, most models only focus on understanding the text content itself, ignoring the joint modeling of multi-dimensional dynamic information such as dissemination behavior, user interaction, and temporal evolution. On the other hand, although graph neural networks can characterize the propagation structure between nodes, in dealing with public opinion scenarios that have the dual complexity of "emotional evolution + propagation evolution," existing models lack the ability to model state evolution mechanisms and predict risk paths.

[0005] Specifically, while current graph convolutional models can extract network structure features, they lack the native ability to model node state propagation. Traditional epidemic propagation models, such as SEIR, while possessing state transition mechanisms, cannot effectively integrate the contextual features of nodes in the graph structure, propagation paths, and user semantic features. In practical applications, this disconnect results in significant deficiencies in the prediction accuracy, timely warning, and accurate identification of risk nodes in public opinion propagation models.

[0006] Furthermore, in terms of state modeling, existing models generally adopt static classification methods, lacking time decay mechanisms and historical state memory strategies. This makes it impossible to distinguish whether a user is "continuously infected" or experiencing "brief emotional fluctuations," thus affecting the judgment of the overall transmission trend. At the same time, the lack of means to extract structural risk indicators means that most current early warning systems rely on experience to set thresholds, lacking early warning basis driven by both the evolution of transmission mechanisms and structural evolution.

[0007] In terms of visualizing propagation paths, mainstream technologies can only achieve graph structure presentation and high-frequency node marking, making it difficult to dynamically visualize the state evolution trend during propagation, resulting in insufficient ability to trace the source of risks. Regarding hotspot identification, the lack of propagation gradient and spatial clustering modeling methods often fails to effectively identify highly sensitive areas where public opinion is about to erupt, hindering proactive intervention.

[0008] In summary, existing online public opinion monitoring and risk warning technologies have the following main shortcomings: First, the public opinion propagation process lacks a structured state modeling mechanism, making it impossible to achieve multi-stage dynamic evolution modeling; second, the node state evolution process does not combine structural influence and historical propagation status, resulting in low accuracy in propagation trend prediction; third, there is a lack of comprehensive indicators that integrate graph structure and propagation status, making it difficult to identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential diffusion hotspots; fourth, the risk warning mechanism relies on static indicator settings, making it impossible to achieve adaptive risk level identification and dynamic warning response.

[0009] Therefore, how to provide online public opinion monitoring and risk warning methods based on AI technology and deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] One objective of this invention is to propose a method for monitoring and warning of online public opinion based on AI technology and deep learning. This invention fully integrates graph convolutional neural networks, SEIR variant models and time decay memory mechanisms, and describes in detail the entire process of multi-stage public opinion state modeling, trend identification and automated warning on dynamic time-series propagation graphs. It has the advantages of high identification accuracy, strong response timeliness, strong interpretability of propagation paths and high degree of intelligence in warning strategies.

[0011] The method for monitoring and warning of online public opinion based on AI technology and deep learning according to embodiments of the present invention includes the following steps:

[0012] S1. Collect online public opinion data, preprocess it, and construct a public opinion dissemination map;

[0013] S2. Initialize the graph structure of the public opinion propagation graph, construct the node attribute matrix and adjacency matrix, and divide the propagation sequence into batches according to time to form a dynamic time sequence graph set;

[0014] S3. Use graph convolutional neural networks to encode the features of nodes in the dynamic time series graph set, extract the structural influence features of nodes in the public opinion propagation graph, and output the node feature matrix.

[0015] S4. Construct a variant of the SEIR model, set up a set of node states, use the node feature matrix as the state propagation input, and introduce a state memory vector to perform time decay weighting on the historical state distribution of each node.

[0016] S5. Execute the SEIR variant model state update process, and combine it with the node feature matrix to simulate the propagation process of various states on the time sequence graph in the public opinion propagation graph, and record the complete state update trajectory.

[0017] S6. Based on the status update trajectory, statistically analyze the trend of changes in the number of infected status nodes, identify abnormal growth inflection points, high-frequency propagation sub-graphs and potential hotspot diffusion areas in the propagation process, and form structured public opinion monitoring results.

[0018] S7. Based on the public opinion monitoring results, issue real-time warnings according to preset risk thresholds and output risk warning information to realize the entire process of automated online public opinion monitoring and risk warning.

[0019] Optionally, the online public opinion data includes text content, user identifiers, timestamps, and interaction relationships.

[0020] Optionally, the preprocessing includes data cleaning, topic identification, entity disambiguation, node normalization, and edge relation extraction.

[0021] Optionally, the node status includes a propagable state, an exposed state, an infected state, and a dissipated state. The propagable state refers to potential users who have not been reached by negative information. The exposed state refers to users who have been exposed but have not yet developed a clear emotional inclination. The infected state refers to users who have been affected by public opinion content and have shown obvious propagation behavior such as forwarding, commenting, and expressing strong opinions. The dissipated state refers to users who no longer participate in propagation, whose emotions have subsided, or who have deleted the content.

[0022] Optionally, the risk warning information includes a time period, high-risk node identifiers, propagation paths, and predicted propagation impact range.

[0023] Optionally, S2 specifically includes:

[0024] S21. Initialize the graph structure of the public opinion propagation graph, extract the user entity set U and the topic entity set H, and construct the node set V = U∪H. The public opinion propagation graph is composed of node features and edge weights. The nodes represent users or topic entities, and the edges represent the propagation behavior between users.

[0025] S22. Generate an edge set E = {e} based on user interaction behavior. ij}, and calculate the edge weight matrix W = {w ij}, satisfying w ij =α1·f r +α2·f c +α3·f l , where e ij Indicates user u i To user u j The act of disseminating information, w ij Represents node v i to node v j edge weight, f r f represents the frequency of forwarding. c Indicates comment frequency, f l The frequency of likes is represented by α1, α2, α3 ∈ [0, 1], which represent different weighting factors.

[0026] S23. Construct the node attribute matrix X = {x i The node attributes include node type identifier, historical dissemination behavior statistics, sentiment tendency score, time activity score, and topic relevance score, where x i This represents the attribute feature vector of the i-th node;

[0027] S24. Construct the adjacency matrix A = {a} of the graph. ij}, where a ij =1 indicates that there is a slave node v i to node v j The spread of the behavior, a ij =0 indicates that there is no propagation behavior;

[0028] S25. Sort the propagation behavior according to the timestamp t, and let the total time range of the propagation data be T = {t1, t2, ..., t}. k The entire public opinion dissemination graph is divided into K dissemination snapshot graphs according to the time interval Δt, thus forming a dynamic time series graph set.

[0029]

[0030] Among them, G 1 G 2 G K G represents the 1st, 2nd, and Kth propagation snapshots. k Indicates the time interval [t] k-1 ,t k The propagation status within ) This represents the set of active nodes in the k-th time period. Let X represent the set of propagation edges occurring within the k-th time interval. k For the time slice of the corresponding node attribute matrix, A k Let W be the state of the adjacency matrix in the k-th time interval. k Let represent the state of the edge weight matrix during the k-th time interval.

[0031] Optionally, S3 specifically includes:

[0032] S31, For dynamic timing diagram sets Each propagation snapshot graph G k Graph convolution feature encoding operations are performed separately, and a multi-layer graph convolutional neural network is used to extract the structural representation of nodes;

[0033] S32. Let v be the node in the l-th layer graph convolution. i The features are represented as The initial input is That is, X k The corresponding i-th vector in the graph, the graph convolution propagation formula is:

[0034]

[0035] in, Represents node v i The feature vector at layer l+1, Represents node v j The feature vector of the l-th layer, This indicates the propagation snapshot image G. k In and node v i The set of connected neighbor nodes, Represents node v i With node v j The edge weights in the k-th propagation graph. Represents node v i ,v j In the propagation of snapshot image G k The degree value in the middle satisfies W (l) Let B be the trainable weight matrix of the l-th layer. (l) Let σ(·) be the self-connection transformation matrix of the l-th layer, and let σ(·) represent the ReLU activation function.

[0036] S33. After performing a graph convolution operation at layer l, extract the structural influence features of all nodes in the k-th time series graph and output the node feature matrix H. k .

[0037] Optionally, S4 specifically includes:

[0038] S41, in each propagation snapshot graph G k In the middle, for each node v i ∈V k Set the state vector as Let represent the probability of propagation, exposure, infection, and resolution of a node at time step k, respectively, and satisfy the constraints.

[0039] S42, convert the node feature matrix H k As input for state propagation, a state memory vector is introduced. Represents node v i State decay history at time step k:

[0040]

[0041] Where λ∈(0,1) is the time decay coefficient. For node v i The state vector at time step τ;

[0042] S43. Construct a variant model of the fusion graph propagation mechanism and the SEIR state transition mechanism, and define node v. i The state update formula at time step k+1 is:

[0043]

[0044] in, Represents node v i In the state vector at time step k+1, σ(·) represents the Softmax function, which normalizes the state distribution. Represents node v i The characteristic matrix at time step k Represents node v i Historical state memory vector, Represents a propagation snapshot diagram G k In and node v i The set of connected neighbor nodes, Represents neighbor node v j The state distribution at time step k Represents an edge (v) i ,v jIn the propagation graph, || represents the vector concatenation operation, W0 represents the weight matrix of the SEIR state mapping, and b is the bias term.

[0045] Optionally, S5 specifically includes:

[0046] S51. Based on the SEIR variant model, dynamic time series graph set Snapshot of the spread in China G k All nodes v i ∈V k The execution of the state propagation process simulates the propagation of various states on the time sequence graph in the public opinion propagation graph, and obtains the state distribution vector of each node at time step k+1. The results are then written into the node state sequence record table.

[0047] S52, Combining the node feature matrix H k For each node state vector in the propagation snapshot graph Auxiliary adjustments are made to enhance the sensitivity of the state update process to the potential influence of the graph structure, thereby obtaining the node feature-aware state distribution.

[0048] S53. Perform state propagation sequentially from k=1 to k=K, recording the state distribution of all nodes at each time step, and generating a set of state update trajectories. Where S k Let represent the complete state matrix of all nodes at the k-th time step, satisfying k = 1, 2, ..., K.

[0049] Optionally, S6 specifically includes:

[0050] S61, Based on state-updated trajectory sets Extract the infection state probability of all nodes at each time step k. Calculate the number of infected nodes The infection status node satisfies Threshold θ∈(0,1);

[0051] S62, Sequence of the number of infected state nodes Differential analysis is performed to identify time points where the growth rate exceeds the slope threshold δ as abnormal growth inflection points, and the corresponding time step index set is recorded.

[0052] S63, Based on Propagation Snapshot G k The adjacency matrix state A in k Extract the set of the largest connected subgraphs between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify high-frequency propagation subgraphs that meet the propagation index threshold conditions;

[0053] S64, Time slice X combined with node attribute matrix k With the state vector of the node at each time step The propagation sensitivity score and spatial propagation gradient of the nodes are calculated. Region clustering is performed on the edge weight-dense regions in the high-frequency propagation subgraph, and potential hotspot diffusion regions are marked.

[0054] S65. The infection status node change trend, abnormal growth inflection point index, high-frequency propagation sub-graph identifier and potential hotspot diffusion area number are stored in a structured manner to generate a structured public opinion monitoring result data table and provide a basis for risk warning.

[0055] The beneficial effects of this invention are:

[0056] First, this invention introduces a graph convolutional neural network to perform deep feature encoding on nodes in a dynamic public opinion propagation graph. This fully leverages the structural influence of users within the propagation network, overcoming the shortcomings of traditional methods that cannot effectively model the coupling relationship between node semantics and structural information. This enhances the ability to identify key nodes and propagation paths in public opinion evolution. Furthermore, the evolution of a node's state during propagation is driven not only by the propagation influence of neighboring nodes but also by incorporating the decaying memory of its historical propagation behavior, making state updates more contextually continuous and logically consistent.

[0057] Secondly, this invention innovatively extends the SEIR variant model in a structured way and introduces a deep feature-guided and graph structure-aware mechanism. This makes the state propagation process no longer an isolated probability transfer, but rather driven by a combination of node semantic expression, adjacency propagation influence, and historical state trajectories, thereby achieving dynamic and refined modeling of public opinion propagation status on the time axis. Through this mechanism, this invention can effectively identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential hotspot areas in the propagation process, solving the problem of insufficient ability to identify risk outbreak thresholds and spatially sensitive propagation areas in existing technologies.

[0058] Finally, based on the above propagation modeling and trajectory analysis results, this invention achieves structured and quantitative public opinion monitoring output and risk classification early warning, significantly improving the automation and intelligence level of public opinion management. Compared with traditional public opinion analysis methods that rely on rule matching and experience setting, this invention can achieve real-time, high-precision risk node identification and early warning, possessing strong practical deployment value and promotion prospects, and significantly improving the timeliness, accuracy, and scientific nature of public opinion response. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of the online public opinion monitoring and risk warning method based on AI technology and deep learning proposed in this invention;

[0061] Figure 2 This is a flowchart for constructing a dynamic time-series graph set for the network public opinion monitoring and risk warning method based on AI technology and deep learning proposed in this invention;

[0062] Figure 3 This is a flowchart illustrating the structured monitoring result generation process of the network public opinion monitoring and risk warning method based on AI technology and deep learning proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1-3 A method for monitoring and warning of online public opinion based on AI technology and deep learning includes the following steps:

[0065] S1. Collect online public opinion data, preprocess it, and construct a public opinion dissemination map;

[0066] S2. Initialize the graph structure of the public opinion propagation graph, construct the node attribute matrix and adjacency matrix, and divide the propagation sequence into batches according to time to form a dynamic time sequence graph set;

[0067] S3. Use graph convolutional neural networks to encode the features of nodes in the dynamic time series graph set, extract the structural influence features of nodes in the public opinion propagation graph, and output the node feature matrix.

[0068] S4. Construct a variant of the SEIR model, set up a set of node states, use the node feature matrix as the state propagation input, and introduce a state memory vector to perform time decay weighting on the historical state distribution of each node.

[0069] S5. Execute the SEIR variant model state update process, and combine it with the node feature matrix to simulate the propagation process of various states on the time sequence graph in the public opinion propagation graph, and record the complete state update trajectory.

[0070] S6. Based on the status update trajectory, statistically analyze the trend of changes in the number of infected status nodes, identify abnormal growth inflection points, high-frequency propagation sub-graphs and potential hotspot diffusion areas in the propagation process, and form structured public opinion monitoring results.

[0071] S7. Based on the public opinion monitoring results, issue real-time warnings according to preset risk thresholds and output risk warning information to realize the entire process of automated online public opinion monitoring and risk warning.

[0072] This invention provides an automated and intelligent method for monitoring and warning of online public opinion by constructing a complete online public opinion data processing chain and combining graph structure modeling, feature extraction, state propagation and risk warning mechanisms. This significantly improves the accuracy of public opinion evolution trend modeling and the real-time performance of warning response.

[0073] In this embodiment, the online public opinion data includes text content, user identifiers, timestamps, and interaction relationships.

[0074] This invention ensures the multidimensional integrity of data sources by collecting text content, user identifiers, timestamps, and interaction relationships from online public opinion data, which is conducive to a more comprehensive reconstruction of user behavior and time evolution patterns in the process of public opinion dissemination.

[0075] In this embodiment, the preprocessing includes data cleaning, topic identification, entity disambiguation, node normalization, and edge relation extraction.

[0076] This invention introduces specific operations such as data cleaning, topic identification, entity disambiguation, node normalization, and edge relationship extraction in the preprocessing stage, which improves the quality of data structuring and enhances the accuracy and robustness of subsequent graph model processing.

[0077] In this embodiment, the node status includes a propagable state, an exposed state, an infected state, and a dissipated state. The propagable state refers to potential users who have not been reached by negative information. The exposed state refers to users who have been exposed but have not yet developed a clear emotional tendency. The infected state refers to users who have been affected by public opinion content and have shown obvious propagation behavior such as forwarding, commenting, and expressing strong opinions. The dissipated state refers to users who no longer participate in propagation, whose emotions have subsided, or who have deleted the content.

[0078] This invention refines the classification of node behaviors in the process of public opinion dissemination by introducing four distinct states: transmissible, exposed, infected, and resolved. This helps to accurately depict the evolution path of emotions and the dynamics of the dissemination chain.

[0079] In this embodiment, the risk warning information includes a time period, high-risk node identifiers, propagation paths, and predicted propagation impact range.

[0080] This invention specifically defines risk warning information as time period, high-risk node identifier, propagation path and predicted propagation impact range, making the warning results quantifiable, traceable and visualized, which facilitates actual decision-making, deployment and response operations.

[0081] In this embodiment, S2 specifically includes:

[0082] S21. Initialize the graph structure of the public opinion propagation graph, extract the user entity set U and the topic entity set H, and construct the node set V = U∪H. The public opinion propagation graph is composed of node features and edge weights. The nodes represent users or topic entities, and the edges represent the propagation behavior between users.

[0083] S22. Generate an edge set E = {e} based on user interaction behavior. ij}, and calculate the edge weight matrix W = {w ij}, satisfying w ij =α1·f r +α2·f c +α3·f l , where e ij Indicates user u i To user u j The act of disseminating information, w ij Represents node v i to node v j edge weight, f r f represents the frequency of forwarding. c Indicates comment frequency, f l The frequency of likes is represented by α1, α2, α3 ∈ [0, 1], which represent different weighting factors.

[0084] S23. Construct the node attribute matrix X = {x i The node attributes include node type identifier, historical dissemination behavior statistics, sentiment tendency score, time activity score, and topic relevance score, where x i This represents the attribute feature vector of the i-th node;

[0085] S24. Construct the adjacency matrix A = {a} of the graph. ij}, where a ij =1 indicates that there is a slave node v i to node v j The spread of the behavior, a ij =0 indicates that there is no propagation behavior;

[0086] S25. Sort the propagation behavior according to the timestamp t, and let the total time range of the propagation data be T = {t1, t2, ..., t}. k The entire public opinion dissemination graph is divided into K dissemination snapshot graphs according to the time interval Δt, thus forming a dynamic time series graph set.

[0087]

[0088] Among them, G 1 G 2 GK G represents the 1st, 2nd, and Kth propagation snapshots. k Indicates the time interval [t] k-1 ,t k The propagation status within ) This represents the set of active nodes in the k-th time period. Let X represent the set of propagation edges occurring within the k-th time interval. k For the time slice of the corresponding node attribute matrix, A k Let W be the state of the adjacency matrix in the k-th time interval. k Let represent the state of the edge weight matrix during the k-th time interval.

[0089] This invention constructs a time-series-based public opinion propagation graph set by constructing a node set, attribute matrix, edge weight matrix, adjacency matrix, and dynamic time sequence graph in stages. This set can accurately restore the propagation status of network information in different time periods and provide a dynamic structural foundation for subsequent state modeling.

[0090] In this embodiment, S3 specifically includes:

[0091] S31, For dynamic timing diagram sets Each propagation snapshot graph G k Graph convolution feature encoding operations are performed separately, and a multi-layer graph convolutional neural network is used to extract the structural representation of nodes;

[0092] S32. Let v be the node in the l-th layer graph convolution. i The features are represented as The initial input is That is, X k The corresponding i-th vector in the graph, the graph convolution propagation formula is:

[0093]

[0094] in, Represents node v i The feature vector at layer l+1, Represents node v j The feature vector of the l-th layer, This indicates the propagation snapshot image G. k In and node v i The set of connected neighbor nodes, Represents node v i With node v j The edge weights in the k-th propagation graph. Represents node v i ,v j In the propagation of snapshot image G k The degree value in the middle satisfies W(l) Let B be the trainable weight matrix of the l-th layer. (l) Let σ(·) be the self-connection transformation matrix of the l-th layer, and let σ(·) represent the ReLU activation function.

[0095] S33. After performing a graph convolution operation at layer l, extract the structural influence features of all nodes in the k-th time series graph and output the node feature matrix H. k .

[0096] This invention employs graph convolutional neural networks to perform multi-layer feature encoding on dynamic propagation graphs, extracting the structural influence features of nodes in the propagation network. This enables the role of nodes in state evolution to be accurately modeled, significantly enhancing the representational and generalization capabilities of the overall model.

[0097] In this embodiment, S4 specifically includes:

[0098] S41, in each propagation snapshot graph G k In the middle, for each node v i ∈V k Set the state vector as Let represent the probability of propagation, exposure, infection, and resolution of a node at time step k, respectively, and satisfy the constraints.

[0099] S42, convert the node feature matrix H k As input for state propagation, a state memory vector is introduced. Represents node v i State decay history at time step k:

[0100]

[0101] Where λ∈(0,1) is the time decay coefficient. For node v i The state vector at time step τ;

[0102] S43. Construct a variant model of the fusion graph propagation mechanism and the SEIR state transition mechanism, and define node v. i The state update formula at time step k+1 is:

[0103]

[0104] in, Represents node v i In the state vector at time step k+1, σ(·) represents the Softmax function, which normalizes the state distribution. Represents node v i The characteristic matrix at time step k Represents node v i Historical state memory vector, Represents a propagation snapshot diagram G k In and node v i The set of connected neighbor nodes, Represents neighbor node v j The state distribution at time step k Represents an edge (v) i ,v j In the propagation graph, || represents the vector concatenation operation, W0 represents the weight matrix of the SEIR state mapping, and b is the bias term.

[0105] This invention constructs a SEIR variant state propagation formula that integrates graph structure, historical states, and adjacency propagation, thereby achieving continuous and responsive modeling of node states in the time dimension and enhancing the predictive ability for sudden propagation events and public opinion diffusion nodes.

[0106] In this embodiment, S5 specifically includes:

[0107] S51. Based on the SEIR variant model, dynamic time series graph set Snapshot of the spread in China G k All nodes v i ∈V k The execution of the state propagation process simulates the propagation of various states on the time sequence graph in the public opinion propagation graph, and obtains the state distribution vector of each node at time step k+1. The results are then written into the node state sequence record table.

[0108] S52, Combining the node feature matrix H k For each node state vector in the propagation snapshot graph Auxiliary adjustments are made to enhance the sensitivity of the state update process to the potential influence of the graph structure, thereby obtaining the node feature-aware state distribution.

[0109] S53. Perform state propagation sequentially from k=1 to k=K, recording the state distribution of all nodes at each time step, and generating a set of state update trajectories. Where S k Let represent the complete state matrix of all nodes at the k-th time step, satisfying k = 1, 2, ..., K.

[0110] This invention executes the SEIR variant state update process, combines node characteristics to adjust the state propagation trajectory, and fully records the node state evolution path during the propagation process, making the dynamic evolution process of the entire public opinion diffusion chain visible, calculable, and traceable.

[0111] In this embodiment, S6 specifically includes:

[0112] S61, Based on state-updated trajectory sets Extract the infection state probability of all nodes at each time step k. Calculate the number of infected nodes The infection status node satisfies Threshold θ∈(0,1);

[0113] S62, Sequence of the number of infected state nodes Differential analysis is performed to identify time points where the growth rate exceeds the slope threshold δ as abnormal growth inflection points, and the corresponding time step index set is recorded.

[0114] S63, Based on Propagation Snapshot G k The adjacency matrix state A in k Extract the set of the largest connected subgraphs between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify high-frequency propagation subgraphs that meet the propagation index threshold conditions;

[0115] S64, Time slice X combined with node attribute matrix k With the state vector of the node at each time step The propagation sensitivity score and spatial propagation gradient of the nodes are calculated. Region clustering is performed on the edge weight-dense regions in the high-frequency propagation subgraph, and potential hotspot diffusion regions are marked.

[0116] S65. The infection status node change trend, abnormal growth inflection point index, high-frequency propagation sub-graph identifier and potential hotspot diffusion area number are stored in a structured manner to generate a structured public opinion monitoring result data table and provide a basis for risk warning.

[0117] This invention analyzes the changing trends of infection status nodes, identifies abnormal growth inflection points, high-frequency propagation subgraphs, and potential hotspot areas, and generates structured monitoring results, thereby achieving early identification of inflection points in public opinion outbreaks and precise positioning of risk diffusion trends.

[0118] Example 1:

[0119] To verify the feasibility of this invention in practice, it was applied to the dissemination of online public opinion during a trending internet event. This event triggered large-scale user participation, forwarding, commenting, and emotional expression within a short period, exhibiting widespread dissemination paths and rapid evolution, displaying typical characteristics of sudden online public opinion events. In this scenario, a large number of internet users voiced their opinions on the same topic, some platforms experienced information peaks within a short time, and some users' dissemination behaviors exhibited characteristics such as aggregation, high frequency, and strong emotionality, posing a high risk of dissemination.

[0120] In the method of this invention, firstly, comments, microblogs, short video comments, and other content related to the hot topic on the network platform are automatically collected, covering text content, user identification, timestamps, and interactive behaviors (such as comments, reposts, and likes). After cleaning, deduplication, entity recognition, and structuring, a public opinion propagation graph with nodes and edge weights is constructed. Nodes represent users or topic entities, edges represent propagation behaviors, and edge weights are calculated by comprehensively considering the interaction frequency between users. Then, the entire propagation graph is segmented along the time dimension, dividing it into multiple propagation snapshot graphs to form a dynamic time series graph set for subsequent modeling of graph neural networks and SEIR models.

[0121] Next, a graph convolutional neural network is applied to extract features from the propagation graph within each time period, exploring the propagation influence of each node in the structure. A variant of the SEIR model is used to model the propagation state of each node, introducing a state memory vector. Combining the graph structure and historical propagation behavior, the propagation state of each node at different propagation stages is dynamically predicted. The system simulates and records the state change trajectories of all nodes at different time steps.

[0122] Subsequently, based on the status trajectory, the system statistically analyzes the changing trends in the number of infected status nodes, identifies multiple abnormal inflection points and peak periods in the propagation process, and analyzes the connectivity structure between infected nodes in the propagation graph to identify multiple high-frequency propagation subgraphs. Furthermore, it calculates potential hotspot diffusion areas by combining node characteristics and propagation activity. Finally, the system generates a public opinion risk warning report based on a custom risk threshold and issues a Level 3 warning 6 hours before the propagation reaches a preset risk level.

[0123] In this application, compared with traditional online public opinion monitoring methods based on keyword monitoring and sentiment analysis, this invention demonstrates significant advantages in several key indicators. Specifically, in terms of the accuracy of high-risk node identification, this invention improves by an average of 22.4%; in terms of the lead time for predicting abnormal growth points, the average prediction time reaches 5.7 hours; in terms of the accuracy of propagation path reconstruction, the modeling method based on graph models and state evolution accurately reconstructs 87.3% of the actual propagation path chains; and in terms of hotspot area identification, compared with traditional methods based on sentiment hot word analysis for hotspot focusing, this invention improves the accuracy of structural density analysis by approximately 26.8%. The following is a summary table of core results data generated during the implementation process:

[0124] Table 1 Comparison of Public Opinion Dissemination Monitoring and Risk Early Warning Effects of the Invention

[0125]

[0126]

[0127] Table 1 verifies the practicality, accuracy, and early warning effect of this invention in real-world public opinion dissemination processes. Particularly in complex dissemination structures, high-concurrency user behavior, and nonlinear evolution scenarios, the multi-dimensional joint modeling and deep state perception mechanism of this invention demonstrates good robustness and adaptability. The modeling approach, combining graph structure and dissemination state as dual-mode inputs, effectively solves key technical challenges in traditional public opinion analysis systems, such as incomplete dissemination chain mining, lagging state recognition, and subjective reliance on risk assessment. It possesses high engineering application value and promotion potential.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring and warning of online public opinion based on AI technology and deep learning, characterized in that, Includes the following steps: S1. Collect online public opinion data, preprocess it, and construct a public opinion dissemination map; S2. Initialize the graph structure of the public opinion propagation graph, construct the node attribute matrix and adjacency matrix, and divide the propagation sequence into batches according to time to form a dynamic time sequence graph set; S3. Use graph convolutional neural networks to encode the features of nodes in the dynamic time series graph set, extract the structural influence features of nodes in the public opinion propagation graph, and output the node feature matrix. S4. Construct a variant of the SEIR model, set up a set of node states, use the node feature matrix as the state propagation input, and introduce a state memory vector to perform time decay weighting on the historical state distribution of each node. S4 specifically includes: S41, in each propagation snapshot diagram In the middle, for each node Set the state vector as , respectively representing the node at time step The probability of transmission, the probability of exposure, the probability of infection, and the probability of resolution, and satisfying the constraints. ; S42, Convert the node feature matrix As input for state propagation, a state memory vector is introduced. , representing a node At time step State decay history: ； in, The time decay coefficient, For nodes At time step The state vector; S43. Construct a variant model of the fusion graph propagation mechanism and the SEIR state transition mechanism, and define the nodes. At time step The state update formula is: ； in, Represents a node At time step The state vector, This represents the Softmax function, which normalizes the state distribution. Represents a node At time step eigenmatrix Represents a node Historical state memory vector, Represents a propagation snapshot diagram In and nodes The set of connected neighbor nodes, Representing neighboring nodes At time step The state distribution, Representing an edge Weights in the propagation graph This represents a vector concatenation operation. The weight matrix represents the SEIR state mapping. For bias terms; S5. Execute the SEIR variant model state update process, and combine it with the node feature matrix to simulate the propagation process of various states on the time sequence graph in the public opinion propagation graph, and record the complete state update trajectory. S6. Based on the status update trajectory, statistically analyze the trend of changes in the number of infected status nodes, identify abnormal growth inflection points, high-frequency propagation sub-graphs and potential hotspot diffusion areas in the propagation process, and form structured public opinion monitoring results. S7. Based on the public opinion monitoring results, issue real-time warnings according to preset risk thresholds and output risk warning information to realize the entire process of automated online public opinion monitoring and risk warning.

2. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, The online public opinion data includes text content, user identifiers, timestamps, and interaction relationships.

3. The method for monitoring and risk warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, The preprocessing includes data cleaning, topic identification, entity disambiguation, node normalization, and edge relation extraction.

4. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, The node status includes a propagable state, an exposed state, an infected state, and a deflated state. The propagable state refers to potential users who have not been reached by negative information. The exposed state refers to users who have been exposed but have not yet developed a clear emotional inclination. The infected state refers to users who have been influenced by public opinion content and have shown obvious propagation behavior such as forwarding, commenting, and expressing strong opinions. The deflated state refers to users who no longer participate in propagation, whose emotions have subsided, or who have deleted the content.

5. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, The risk warning information includes the time period, high-risk node identifiers, propagation path, and predicted propagation impact range.

6. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Initialize the graph structure of the public opinion dissemination graph and extract the user entity set. With topic entity set Build a node set The public opinion propagation graph is composed of node features and edge weights, where nodes represent users or topic entities, and edges represent propagation behaviors between users; S22. Generate an edge set based on user interactions. And calculate the edge weight matrix. ,satisfy ,in, Indicates user To users The act of disseminating information Represents a node To the node edge weights, Indicates the frequency of forwarding. Indicates the frequency of comments. Indicates the frequency of likes. , representing different weighting factors; S23. Construct the node attribute matrix The node attributes include node type identifier, historical dissemination behavior statistics, sentiment tendency score, time activity score, and topic relevance score, among which... Indicates the first The attribute feature vector of each node; S24. Constructing the adjacency matrix of the graph. ,in This indicates the existence of slave nodes. To the node The spread of the behavior This indicates that there is no transmission activity; S25. Dissemination activities are time-stamped. Sort the data, assuming the total time range of the propagated data is... The entire public opinion dissemination map is divided into time intervals. Divided into Each propagation snapshot diagram forms a dynamic time series diagram set. : ； in, Indicates the first A snapshot of the spread, Indicates the time interval The propagation state within, Indicates the first A set of active nodes within a time period Indicates the first The set of propagation edges occurring within a time period For the time slice of the corresponding node attribute matrix, For the adjacency matrix at the th The status within a time period, The edge weight matrix is ​​at the th The state within a time period.

7. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, S3 specifically includes: S31, For dynamic timing diagram sets Each propagation snapshot diagram Graph convolution feature encoding operations are performed separately, and a multi-layer graph convolutional neural network is used to extract the structural representation of nodes; S32, Let the first Nodes in layer graph convolution The features are represented as The initial input is ,Right now The corresponding number in The graph convolution propagation formula is as follows: (The graph contains vectors.) ； in, Represents a node In the The feature vector of the layer, Represents a node In the The feature vector of the layer, Indicates a snapshot of the propagation process. In and nodes The set of connected neighbor nodes, Represents a node With nodes In the Edge weights in a propagation graph, Represents a node In the spread of snapshot images The degree value in the middle satisfies , For the first The trainable weight matrix of the layer, For the first The self-connection transformation matrix of the layer, express Activation function; S33, Through After the layer graph convolution operation, the first layer is extracted. The structural influence characteristics of all nodes in the time series graph are analyzed, and the node feature matrix is ​​output. .

8. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Based on the SEIR variant model, dynamic time series graph set Snapshot of the spread in China All nodes The execution state propagation process simulates the propagation of various states on a time sequence graph within the public opinion propagation diagram, obtaining the state propagation process of each node at time step. State distribution vector The results are then written to the node state sequence record table. S52, Combining node feature matrix For each node state vector in the propagation snapshot graph Auxiliary adjustments are made to enhance the sensitivity of the state update process to the potential influence of the graph structure, thereby obtaining the node feature-aware state distribution. S53, Follow the time step arrive Perform state propagation sequentially, record the state distribution of all nodes at each time step, and generate a set of state update trajectories. ,in Indicates the first The complete state matrix of all nodes at each time step satisfies .

9. The method for monitoring and warning of online public opinion based on AI technology and deep learning according to claim 1, characterized in that, S6 specifically includes: S61, Based on state-updated trajectory sets Extract each time step Probability of infection status of all nodes Calculate the number of infected nodes. The infection state node satisfies threshold ; S62, Sequence of the number of infected state nodes Perform differential analysis to identify conditions where the growth rate exceeds the slope threshold. The time point is used as the inflection point of abnormal growth, and the corresponding time step index set is recorded. ; S63, Based on Propagation Snapshot Diagram The adjacency matrix state Extract the set of the largest connected subgraphs between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify high-frequency propagation subgraphs that meet the propagation index threshold conditions; S64. Time slices combined with node attribute matrix With the state vector of the node at each time step The propagation sensitivity score and spatial propagation gradient of the nodes are calculated, and the dense edge weight regions in the high-frequency propagation subgraph are clustered and potential hotspot diffusion regions are marked. S65. The infection status node change trend, abnormal growth inflection point index, high-frequency propagation sub-graph identifier and potential hotspot diffusion area number are stored in a structured manner to generate a structured public opinion monitoring result data table and provide a basis for risk warning.

Images