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

Abstract

The invention discloses a network public opinion monitoring and risk early warning method based on an AI technology and deep learning, and the method comprises the following steps: S1, collecting and preprocessing network public opinion data, and constructing a public opinion propagation graph; s2, initializing a graph structure, constructing a node attribute matrix and an adjacent matrix, and dividing into a dynamic time sequence graph set according to time; s3, node features are coded by using a graph convolutional neural network, and structural influence features are extracted; s4, constructing an SEIR variation model, setting a node state set, and introducing a state memory vector to perform time decay weighting; s5, executing state propagation and updating, and recording a complete state updating track; s6, analyzing infection state node changes, and forming a structured public opinion monitoring result; and S7, performing real-time early warning based on a monitoring result and a risk threshold, and realizing public opinion monitoring and early warning automation. According to the method, the graph neural network and state modeling are combined, and the public opinion evolution recognition precision and early warning efficiency are improved.

Description

Technical Field

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

[0002] In the context of a highly developed internet, online public opinion has become a significant factor impacting 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 has acquired the hallmarks of rapid dissemination, diverse content, and strong emotional appeal. Once negative information emerges, its explosive growth far surpasses the information dissemination capabilities of traditional media. To address this information dissemination dynamic, timely and accurate monitoring and analysis of online public opinion has become a pressing challenge for various organizations.

[0003] Existing online public opinion monitoring technologies primarily include keyword-matching-based opinion collection, dictionary-based sentiment recognition, and statistical-based trend analysis. While these methods were able to provide a rough analysis of online content in the early stages, they have become increasingly deficient in the current landscape of complex online semantics, diverse information sources, and nonlinear content dissemination pathways. First, traditional keyword-based rules suffer from incomplete semantic coverage, significantly reducing recognition accuracy when faced with a large number of metaphors, sarcasm, puns, or emerging online lingo. Second, statistical analysis models fail to effectively integrate user interactions with the structure of dissemination networks, making it difficult to accurately depict the true path of information diffusion and, consequently, unable to accurately predict potential dissemination risks.

[0004] In recent years, the rise of artificial intelligence, particularly deep learning, has brought new solutions to online public opinion analysis. Natural language processing technology has achieved significant breakthroughs in text classification, sentiment analysis, and topic modeling; models such as graph neural networks and Transformers have also provided enhanced capabilities for modeling complex structured data. However, existing research on the application of 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 time evolution. On the other hand, although graph neural networks can depict the dissemination structure between nodes, in the context of public opinion, which has the dual complexity of "emotional evolution + dissemination 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 structural features, they lack the native ability to model node state propagation. Traditional epidemic transmission models, such as SEIR, while equipped with state transition mechanisms, are unable to effectively incorporate the contextual characteristics of nodes, propagation paths, and user semantic features within the graph structure. In practical applications, this disconnect results in significant deficiencies in public opinion transmission models in terms of prediction accuracy, timely warnings, and accurate identification of risky nodes.

[0006] Furthermore, in terms of state modeling, existing models generally employ static classification methods, lacking time decay mechanisms and historical state memory strategies. This makes it impossible to distinguish whether a user is "persistently infected" or experiencing "temporary emotional fluctuations," which in turn affects the overall assessment of the spread of the virus. Furthermore, the lack of a method for extracting structural risk indicators means that most current early warning systems rely on empirical thresholds, lacking a basis for early warning driven by both the evolution of transmission mechanisms and structural evolution.

[0007] When it comes to visualizing communication paths, mainstream technologies are limited to graph structure presentation and high-frequency node labeling. Dynamic visualization, combined with state evolution trends during the communication process, is difficult to integrate, resulting in insufficient traceability of risk points. When it comes to identifying hotspots, the lack of propagation gradient and spatial clustering modeling methods often prevents the effective identification of highly sensitive areas where public opinion is about to erupt, making proactive intervention difficult.

[0008] To sum up, the existing network public opinion monitoring and risk warning technologies have the following major defects: First, the public opinion propagation process lacks a structured state modeling mechanism, and multi-stage dynamic evolution modeling cannot be realized; 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 and cannot achieve adaptive risk level identification and dynamic warning response.

[0009] Therefore, how to provide a network public opinion monitoring and risk warning method based on AI technology and deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

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

[0011] The network public opinion monitoring and risk warning method based on AI technology and deep learning according to an embodiment of the present invention includes the following steps:

[0012] S1. Collect online public opinion data, pre-process it, and construct a public opinion propagation 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 based on time to form a dynamic time series 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 SEIR variant model, set the node state set, use the node feature matrix as the state propagation input, introduce the state memory vector, and perform time decay weighting on the historical state distribution of each node;

[0016] S5. Execute the state update process of the SEIR variant model and, combined with the node feature matrix, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram and record the complete state update trajectory;

[0017] S6. Count the number of infected nodes based on the status update trajectory, identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential hotspot diffusion areas during the propagation process, and form structured public opinion monitoring results.

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

[0019] Optionally, the online public opinion data includes text content, user identification, timestamp and interactive relationship.

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

[0021] Optionally, the node status includes a transmissible state, an exposed state, an infected state, and a resolved state. The transmissible state refers to potential users who have not been exposed to negative information. The exposed state refers to wait-and-see users who have been exposed but have not yet developed obvious emotional tendencies. The infected state refers to users who have been affected by public opinion content and have shown obvious dissemination behaviors such as forwarding, commenting, and intense expression. The resolved state refers to users who no longer participate in dissemination, whose emotions have subsided, or whose content has been deleted.

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

[0023] Optionally, the 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 a 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 propagation behaviors between users.

[0025] S22. Generate an edge set E based on the interaction behavior between users. 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 Represents user u i To user u j The act of spreading information, ij Represents node v i To node v j The edge weight, f r Indicates the forwarding frequency, f c Indicates the frequency of comments, f l represents the frequency of likes, α1, α2, α3∈[0,1], representing different weight factors;

[0026] S23, construct node attribute matrix X = {x i}, the node attributes include node type identification, historical communication behavior statistics, sentiment tendency score, time activity score and topic relevance score, where x i Represents the attribute feature vector of the i-th node;

[0027] S24, construct the adjacency matrix A of the graph = {a ij}, where a ij =1 indicates that there is a slave node v i To node v j The dissemination behavior of ij =0 means there is no spreading behavior;

[0028] S25. Sort the propagation behaviors according to the timestamp t. Let the total time range of the propagation data be T = {t1, t2, ..., t k}, the entire public opinion propagation graph is divided into K propagation snapshot graphs according to the time interval Δt, that is, a dynamic time series graph set is formed

[0029]

[0030] Among them, G 1 ,G 2 ,G K Represents the 1st, 2nd, Kth propagation snapshot graph, G k Indicates that in the time interval [t k-1 ,t k ) within the propagation state, represents the set of active nodes in the kth time period, represents the set of propagation edges occurring in the kth time period, X k is the time slice of the corresponding node attribute matrix, A k is the state of the adjacency matrix in the kth time period, W k is the state of the edge weight matrix in the kth time period.

[0031] Optionally, the S3 specifically includes:

[0032] S31. Dynamic timing diagram collection Each propagation snapshot graph G in k Perform graph convolution feature encoding operations separately and use a multi-layer graph convolutional neural network to extract the structural representation of the nodes;

[0033] S32, let node v in the l-th layer graph convolution i The characteristic representation of The initial input is That is X k The corresponding i-th vector in the graph convolution propagation formula is:

[0034]

[0035] in, Represents node v i The feature vector at the l+1th layer, Represents node v j The feature vector at layer l, Represents the propagation snapshot graph G k 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 snapshot graph G k The degree value in satisfies W (l) is the trainable weight matrix of layer l, B (l) is the self-connection transformation matrix of layer l, σ(·) represents the ReLU activation function;

[0036] S33, after the l-layer graph convolution operation, 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, the S4 specifically includes:

[0038] S41. In each propagation snapshot graph G k For each node v i ∈V k Set the state vector to They represent the propagation probability, exposure probability, infection probability and elimination probability of the node at time step k, respectively, and satisfy the constraints

[0039] S42, the node feature matrix H k As state propagation input, and introduce state memory vector Represents node v i The state decay history at time step k is:

[0040]

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

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

[0043]

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

[0045] Optionally, the S5 specifically includes:

[0046] S51, based on the SEIR variant model, dynamic time series diagram set Medium propagation snapshot graph G k All nodes v i ∈V k Execute the state propagation process, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram, and obtain the state distribution vector of each node at time step k+1 And write the result into the node status sequence record table;

[0047] S52, combined with the node feature matrix H k , for each node state vector in the propagation snapshot graph Perform auxiliary adjustments to enhance the sensitivity of the state update process to the potential influence of the graph structure and obtain the node feature-aware state distribution;

[0048] S53, execute state propagation in sequence from k=1 to k=K according to the time step, record the state distribution of all nodes in each time step, and generate a state update trajectory set Among them S k represents the complete state matrix of all nodes at the kth time step, satisfying k=1,2,…,K.

[0049] Optionally, the S6 specifically includes:

[0050] S61. Update trajectory set based on state Extract the infection status probability of all nodes in each time step k Calculate the number of infected nodes The infected state node satisfies Threshold θ∈(0,1);

[0051] S62, sequence of number of infected nodes Perform differential analysis to identify the time point where the growth rate exceeds the slope threshold δ as the abnormal growth inflection point, and record the corresponding time step index set

[0052] S63, based on the propagation snapshot graph G k The adjacency matrix state A in k , extract the maximum connected subgraph set between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify the high-frequency propagation subgraph that meets the propagation index threshold condition;

[0053] S64, combined with the time slice X of the node attribute matrix k and the state vector of the node at each time step Calculate the node's propagation sensitivity score and spatial propagation gradient, perform regional clustering on the edge weight-dense areas in the high-frequency propagation subgraph, and mark potential hotspot diffusion areas;

[0054] S65. Structured storage is performed on the infection status node change trend, abnormal growth inflection point index, high-frequency propagation subgraph identifier, and potential hotspot diffusion area number to generate a structured public opinion monitoring result data table and provide a judgment basis for risk warning.

[0055] The beneficial effects of the present invention are:

[0056] First, this paper introduces a graph convolutional neural network to perform deep feature encoding on nodes in the dynamic public opinion diffusion graph, fully exploring the structural influence of users in the diffusion network. This overcomes the limitation of traditional methods that cannot effectively model the coupling relationship between node semantics and structural information, thereby improving the ability to identify key nodes and diffusion paths in public opinion evolution. Furthermore, the evolution of node states during the diffusion process is not only driven by the influence of neighboring nodes, but also incorporates the decaying memory of their historical diffusion behavior, making state updates more contextually continuous and realistically logical.

[0057] Secondly, the present invention innovatively expands the SEIR variant model into a structured form and introduces deep feature guidance and graph structure perception mechanisms, so that the state propagation process is no longer an isolated probability transfer, but is driven by a combination of node semantic expression, adjacent propagation influence, and historical state trajectory, thereby achieving dynamic and refined modeling of the public opinion propagation state on the timeline. Through this mechanism, the present invention can effectively identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential hot spots in the propagation process, solving the problem of insufficient ability to identify risk outbreak critical points and spatial propagation sensitive areas in existing technologies.

[0058] Finally, based on the results of the aforementioned propagation modeling and trajectory analysis, this invention achieves structured, quantified public opinion monitoring output and risk-graded early warning, significantly enhancing the automation and intelligence level of public opinion management. Compared to traditional public opinion analysis methods that rely on rule matching and empirical settings, this invention enables real-time, high-precision identification of risk nodes and proactive early warnings. It possesses strong practical deployment value and widespread adoption, significantly improving the timeliness, accuracy, and scientific nature of public opinion response. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

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

[0061] Figure 2 Construct a flow chart for the dynamic time sequence diagram set of the network public opinion monitoring and risk warning method based on AI technology and deep learning proposed in the present invention;

[0062] Figure 3 This is a flowchart for generating structured monitoring results for the network public opinion monitoring and risk warning method based on AI technology and deep learning proposed in this invention. DETAILED DESCRIPTION

[0063] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0064] refer to Figure 1-3 , a network public opinion monitoring and risk warning method based on AI technology and deep learning includes the following steps:

[0065] S1. Collect online public opinion data, pre-process it, and construct a public opinion propagation 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 based on time to form a dynamic time series 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 SEIR variant model, set the node state set, use the node feature matrix as the state propagation input, introduce the state memory vector, and perform time decay weighting on the historical state distribution of each node;

[0069] S5. Execute the state update process of the SEIR variant model and, combined with the node feature matrix, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram and record the complete state update trajectory;

[0070] S6. Count the number of infected nodes based on the status update trajectory, identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential hotspot diffusion areas during the propagation process, and form structured public opinion monitoring results.

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

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

[0073] In this embodiment, the network public opinion data includes text content, user identification, timestamp and interactive relationship.

[0074] The present invention ensures the multi-dimensional integrity of data sources by collecting text content, user identification, timestamps and interactive relationships in online public opinion data, which is conducive to more comprehensive restoration of user behavior and time evolution laws in the process of public opinion dissemination.

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

[0076] The present 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 implementation, the node states include a transmissible state, an exposed state, an infected state, and a resolved state. The transmissible state refers to potential users who have not been exposed to negative information. The exposed state refers to wait-and-see users who have been exposed but have not yet developed obvious emotional tendencies. The infected state refers to users who have been affected by public opinion content and have shown obvious dissemination behaviors such as forwarding, commenting, and intense expression. The resolved state refers to users who no longer participate in dissemination, whose emotions have subsided, or whose content has been deleted.

[0078] The present invention refines the node behavior classification in the process of public opinion dissemination by introducing four clear states: transmissible, exposed, infected, and dissolved, which helps to accurately depict the emotion evolution path and the dynamics of the dissemination chain.

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

[0080] The present invention specifically defines risk warning information as time period, high-risk node identification, propagation path and predicted propagation impact range, so that the warning results are quantifiable, traceable and visual, 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 a 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 propagation behaviors between users.

[0083] S22. Generate an edge set E based on the interaction behavior between users. 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 Represents user u i To user u j The act of spreading information, ij Represents node v i To node v j The edge weight, f r Indicates the forwarding frequency, f c Indicates the frequency of comments, f l represents the frequency of likes, α1, α2, α3∈[0,1], representing different weight factors;

[0084] S23, construct node attribute matrix X = {x i}, the node attributes include node type identification, historical communication behavior statistics, sentiment tendency score, time activity score and topic relevance score, where x i Represents the attribute feature vector of the i-th node;

[0085] S24, construct the adjacency matrix A of the graph = {a ij}, where a ij =1 indicates that there is a slave node v i To node v j The dissemination behavior of ij =0 means there is no spreading behavior;

[0086] S25. Sort the propagation behaviors according to the timestamp t. Let the total time range of the propagation data be T = {t1, t2, ..., t k}, the entire public opinion propagation graph is divided into K propagation snapshot graphs according to the time interval Δt, that is, a dynamic time series graph set is formed

[0087]

[0088] Among them, G 1 ,G 2 ,GK Represents the 1st, 2nd, Kth propagation snapshot graph, G k Indicates that in the time interval [t k-1 ,t k ) within the propagation state, represents the set of active nodes in the kth time period, represents the set of propagation edges occurring in the kth time period, X k is the time slice of the corresponding node attribute matrix, A k is the state of the adjacency matrix in the kth time period, W k is the state of the edge weight matrix in the kth time period.

[0089] The present invention forms a set of time-series public opinion propagation graphs by constructing node sets, attribute matrices, edge weight matrices, adjacency matrices and dynamic time-series graphs in stages. It 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. Dynamic timing diagram collection Each propagation snapshot graph G in k Perform graph convolution feature encoding operations separately and use a multi-layer graph convolutional neural network to extract the structural representation of the nodes;

[0092] S32, let node v in the l-th layer graph convolution i The characteristic representation of The initial input is That is X k The corresponding i-th vector in the graph convolution propagation formula is:

[0093]

[0094] in, Represents node v i The feature vector at the l+1th layer, Represents node v j The feature vector at layer l, Represents the propagation snapshot graph G k 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 snapshot graph G k The degree value in satisfies W(l) is the trainable weight matrix of layer l, B (l) is the self-connection transformation matrix of layer l, σ(·) represents the ReLU activation function;

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

[0096] The present invention uses a graph convolutional neural network to perform multi-layer feature encoding on the dynamic propagation graph, extracting the structural influence characteristics of the nodes in the propagation network, so that the role of the nodes in the state evolution can be accurately modeled, significantly enhancing the representation and generalization capabilities of the overall model.

[0097] In this embodiment, the S4 specifically includes:

[0098] S41. In each propagation snapshot graph G k For each node v i ∈V k Set the state vector to They represent the propagation probability, exposure probability, infection probability and elimination probability of the node at time step k, respectively, and satisfy the constraints

[0099] S42, the node feature matrix H k As state propagation input, and introduce state memory vector Represents node v i The state decay history at time step k is:

[0100]

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

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

[0103]

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

[0105] The present invention realizes the continuity and responsiveness modeling of node states in the time dimension by constructing a SEIR variant state propagation formula that integrates graph structure, historical state and adjacent propagation, thereby enhancing the prediction ability of sudden propagation events and public opinion diffusion nodes.

[0106] In this embodiment, the S5 specifically includes:

[0107] S51, based on the SEIR variant model, dynamic time series diagram set Medium propagation snapshot graph G k All nodes v i ∈V k Execute the state propagation process, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram, and obtain the state distribution vector of each node at time step k+1 And write the result into the node status sequence record table;

[0108] S52, combined with the node feature matrix H k , for each node state vector in the propagation snapshot graph Perform auxiliary adjustments to enhance the sensitivity of the state update process to the potential influence of the graph structure and obtain the node feature-aware state distribution;

[0109] S53, execute state propagation in sequence from k=1 to k=K according to the time step, record the state distribution of all nodes in each time step, and generate a state update trajectory set Among them S k represents the complete state matrix of all nodes at the kth time step, satisfying k=1,2,…,K.

[0110] The present invention executes the SEIR variant state update process, adjusts the state propagation trajectory in combination with node characteristics, and completely 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. Update trajectory set based on state Extract the infection status probability of all nodes in each time step k Calculate the number of infected nodes The infected state node satisfies Threshold θ∈(0,1);

[0113] S62, sequence of number of infected nodes Perform differential analysis to identify the time point where the growth rate exceeds the slope threshold δ as the abnormal growth inflection point, and record the corresponding time step index set

[0114] S63, based on the propagation snapshot graph G k The adjacency matrix state A in k , extract the maximum connected subgraph set between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify the high-frequency propagation subgraph that meets the propagation index threshold condition;

[0115] S64, combined with the time slice X of the node attribute matrix k and the state vector of the node at each time step Calculate the node's propagation sensitivity score and spatial propagation gradient, perform regional clustering on the edge weight-dense areas in the high-frequency propagation subgraph, and mark potential hotspot diffusion areas;

[0116] S65. Structured storage is performed on the infection status node change trend, abnormal growth inflection point index, high-frequency propagation subgraph identifier, and potential hotspot diffusion area number to generate a structured public opinion monitoring result data table and provide a judgment basis for risk warning.

[0117] The present invention analyzes the changing trends of infection status nodes, identifies abnormal growth inflection points, high-frequency propagation subgraphs and potential hot spots, generates structured monitoring results, and realizes the early identification of public opinion outbreak inflection points and the precise positioning of risk diffusion trends.

[0118] Example 1:

[0119] To verify the feasibility of this invention, we applied it to the spread of online public opinion surrounding a hot internet event. This event triggered widespread user participation, forwarding, commentary, and emotional expression within a short period of time. The event also exhibited broad dissemination paths and rapid evolution, demonstrating the typical characteristics of sudden online public opinion. In this scenario, a large number of online users collectively voiced opinions around the same topic, with some platforms experiencing short-term information spikes. Some users exhibited clustered, high-frequency, and highly emotional dissemination behaviors, posing a high risk of dissemination.

[0120] In the method of the present invention, comments, microblogs, short video comments and other content related to the hot event on the network platform are first automatically collected, covering text content, user identity identification, timestamps and interactive behaviors (such as comments, reposts, likes). After cleaning, deduplication, entity recognition and structured processing, a public opinion propagation graph with node and edge weights is constructed. The nodes represent users or topic entities, the edges represent propagation behaviors, and the edge weights are calculated comprehensively based on the interaction frequency between users. Then, the entire propagation graph is segmented based on time as the dimension, and a plurality of propagation snapshot graphs are divided into a dynamic time series graph set for subsequent graph neural network and SEIR model modeling.

[0121] Next, a graph convolutional neural network was applied to extract features from the propagation graph within each time period, exploring the propagation influence of each node within the structure. A SEIR variant model was used to model the propagation state of each node. A state memory vector was introduced, combining the graph structure and historical propagation behavior to dynamically predict the propagation state of each node at different propagation stages. The system simulated and recorded the state change trajectories of all nodes at different time steps.

[0122] Subsequently, based on the status trajectory, the system calculates the changing trend of the number of infected nodes, identifying multiple abnormal inflection points and peak periods during the spread process. It then analyzes the connectivity structure between infected nodes in the spread graph, identifying multiple high-frequency spread subgraphs. It then further calculates potential hotspots of spread based on node characteristics and spread activity. Finally, the system generates a public opinion risk warning report based on custom risk thresholds and issues a Level 3 warning six hours in advance, before the spread reaches the preset risk level.

[0123] In this application, the present invention has shown significant advantages in many key indicators compared to traditional network public opinion monitoring methods based on keyword monitoring and sentiment analysis. Specifically, in terms of the accuracy of high-risk node identification, the present invention has increased by an average of 22.4%; in terms of the advance amount of abnormal growth point prediction, the average advance prediction time has reached 5.7 hours; in terms of the accuracy of propagation path restoration, the modeling method based on the combination of graph model and state evolution accurately restored 87.3% of the real propagation path chain; in terms of hot spot area identification, compared with the traditional method based on the hot spot focus method of emotional hot word analysis, the present invention has increased the accuracy of the structural density analysis dimension by about 26.8%. The following is a summary table of the core result data generated during the implementation process:

[0124] Table 1 Comparative data of public opinion dissemination monitoring and risk warning effects of the present invention

[0125]

[0126]

[0127] Table 1 demonstrates the practicality, accuracy, and early warning effectiveness of this method in real-world public opinion dissemination. The multi-dimensional joint modeling and deep state perception mechanism of this method demonstrates excellent robustness and adaptability, particularly in complex dissemination structures, highly concurrent user behavior, and nonlinear evolution scenarios. This modeling approach, combining graph structure and dissemination state dual-mode input, effectively addresses key technical challenges faced by traditional public opinion analysis systems, such as incomplete dissemination of dissemination chains, delayed state identification, and subjective reliance on risk assessment. This approach offers significant engineering value and potential for widespread adoption.

[0128] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A network public opinion monitoring and risk warning method based on AI technology and deep learning, characterized by: The steps include: S1. Collect online public opinion data, pre-process it, and construct a public opinion propagation 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 based on time to form a dynamic time series 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 SEIR variant model, set the node state set, use the node feature matrix as the state propagation input, introduce the state memory vector, and perform time decay weighting on the historical state distribution of each node; S5. Execute the state update process of the SEIR variant model and, combined with the node feature matrix, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram and record the complete state update trajectory; S6. Count the number of infected nodes based on the status update trajectory, identify abnormal growth inflection points, high-frequency propagation subgraphs, and potential hotspot diffusion areas during the propagation process, and form structured public opinion monitoring results. S7. Based on the results of public opinion monitoring, real-time warnings are issued according to the preset risk thresholds, and risk warning information is output to realize the automation of the entire process of online public opinion monitoring and risk warning.

2. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The network public opinion data includes text content, user identification, timestamp and interactive relationship.

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

4. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The node states include the spreadable state, exposed state, infected state and resolved state. The spreadable state refers to potential users who have not been exposed to negative information. The exposed state refers to wait-and-see users who have been exposed but have not yet developed obvious emotional tendencies. The infected state refers to users who have been affected by public opinion content and have shown obvious spread behaviors such as forwarding, commenting and intense expression. The resolved state refers to users who no longer participate in the spread, whose emotions have subsided or whose content has been deleted.

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

6. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The S2 specifically includes: 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 a 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 propagation behaviors between users. S22. Generate an edge set E based on the interaction behavior between users. 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 Represents user u i To user u j The act of spreading information, ij Represents node v i To node v j The edge weight, f r Indicates the forwarding frequency, f c Indicates the frequency of comments, f l represents the frequency of likes, α1, α2, α3∈[0,1], representing different weight factors; S23, construct node attribute matrix X = {x i }, the node attributes include node type identification, historical communication behavior statistics, sentiment tendency score, time activity score and topic relevance score, where x i Represents the attribute feature vector of the i-th node; S24, construct the adjacency matrix A of the graph = {a ij }, where a ij =1 indicates that there is a slave node v i To node v j The dissemination behavior of ij =0 means there is no spreading behavior; S25. Sort the propagation behaviors according to the timestamp t. Let the total time range of the propagation data be T = {t1, t2, ..., t k }, the entire public opinion propagation graph is divided into K propagation snapshot graphs according to the time interval Δt, that is, a dynamic time series graph set is formed Among them, G 1 ,G 2 ,G K Represents the 1st, 2nd, Kth propagation snapshot graph, G k Indicates that in the time interval [t k-1 ,t k ) within the propagation state, represents the set of active nodes in the kth time period, represents the set of propagation edges occurring in the kth time period, X k is the time slice of the corresponding node attribute matrix, A k is the state of the adjacency matrix in the kth time period, W k is the state of the edge weight matrix in the kth time period.

7. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Dynamic timing diagram collection Each propagation snapshot graph G in k Perform graph convolution feature encoding operations separately and use a multi-layer graph convolutional neural network to extract the structural representation of the nodes; S32, let node v in the l-th layer graph convolution i The characteristic representation of The initial input is That is X k The corresponding i-th vector in the graph convolution propagation formula is: in, Represents node v i The feature vector at the l+1th layer, Represents node v j The feature vector at layer l, Represents the propagation snapshot graph G k In 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 snapshot graph G k The degree value in satisfies W (l) is the trainable weight matrix of layer l, B (l) is the self-connection transformation matrix of layer l, σ(·) represents the ReLU activation function; S33, after the l-layer graph convolution operation, extract the structural influence features of all nodes in the k-th time series graph and output the node feature matrix H k .

8. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41. In each propagation snapshot graph G k For each node v i ∈V k Set the state vector to They represent the propagation probability, exposure probability, infection probability and elimination probability of the node at time step k, respectively, and satisfy the constraints S42, the node feature matrix H k As state propagation input, and introduce state memory vector Represents node v i The state decay history at time step k is: Among them, λ∈(0,1) is the time attenuation coefficient, For node v i The state vector at time step τ; S43. Construct a variant model that integrates the graph propagation mechanism and the SEIR state transition mechanism, and define the node v i The state update formula at time step k+1 is: in, Represents node v i At the state vector of time step k+1, σ(·) represents the Softmax function, which normalizes the state distribution. Represents node v i The feature matrix at time step k, Represents node v i The historical state memory vector, Represents the propagation snapshot graph G k In node v i The set of connected neighbor nodes, Represents the neighbor node v j The state distribution at time step k, Represents the edge (v i ,v j ) is the weight in the propagation graph, || represents the vector concatenation operation, W0 represents the weight matrix of SEIR state mapping, and b is the bias term.

9. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The S5 specifically includes: S51, based on the SEIR variant model, dynamic time series diagram set Medium propagation snapshot graph G k All nodes v i ∈V k Execute the state propagation process, simulate the propagation process of various states on the time series diagram in the public opinion propagation diagram, and obtain the state distribution vector of each node at time step k+1 And write the result into the node status sequence record table; S52, combined with the node feature matrix H k , for each node state vector in the propagation snapshot graph Perform auxiliary adjustments to enhance the sensitivity of the state update process to the potential influence of the graph structure and obtain the node feature-aware state distribution; S53, execute state propagation in sequence from k=1 to k=K according to the time step, record the state distribution of all nodes in each time step, and generate a state update trajectory set Among them S k represents the complete state matrix of all nodes at the kth time step, satisfying k=1,2,…,K.

10. The network public opinion monitoring and risk warning method based on AI technology and deep learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Update trajectory set based on state Extract the infection status probability of all nodes in each time step k Calculate the number of infected nodes The infected state node satisfies Threshold θ∈(0,1); S62, sequence of number of infected nodes Perform differential analysis to identify the time point where the growth rate exceeds the slope threshold δ as the abnormal growth inflection point, and record the corresponding time step index set S63, based on the propagation snapshot graph G k The adjacency matrix state A in k , extract the maximum connected subgraph set between infected nodes, count the propagation radius, node density and cumulative propagation intensity of each subgraph, and identify the high-frequency propagation subgraph that meets the propagation index threshold condition; S64, combined with the time slice X of the node attribute matrix k and the state vector of the node at each time step Calculate the node's propagation sensitivity score and spatial propagation gradient, perform regional clustering on the edge weight-dense areas in the high-frequency propagation subgraph, and mark potential hotspot diffusion areas; S65. Structured storage is performed on the infection status node change trend, abnormal growth inflection point index, high-frequency propagation subgraph identifier, and potential hotspot diffusion area number to generate a structured public opinion monitoring result data table and provide a judgment basis for risk warning.

Images