An opinion analysis method, system, device and medium based on a propagation graph

Abstract

A method, system, device, and medium for public opinion analysis based on a propagation graph, relating to the field of data processing technology, includes: acquiring public opinion data; extracting participating accounts as nodes and extracting basic attribute data for each node; constructing directed propagation edges based on interaction records, content similarity, and timestamps to generate a public opinion propagation graph; extracting topological and sentiment features of each node in the graph, including forwarding contribution and negative sentiment amplification, and calculating node influence values ​​by considering follower count, interaction volume, media level, forwarding contribution, and negative sentiment amplification; identifying initial node and first cross-platform node, and identifying key propagation nodes based on influence value, negative sentiment amplification, and out-degree change rate; extracting global propagation features of the graph and calculating risk levels; outputting structured propagation paths and generating priority prompts for handling. This application improves the accuracy and reliability of public opinion analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a public opinion analysis method, system, device, and medium based on propagation graphs. Background Technology

[0002] With the widespread use of social media, the speed and scope of online public opinion events are increasing daily. Effective analysis and guidance of these events have become crucial issues for businesses and organizations. Currently, some public opinion analysis techniques typically assess event impact based on single or a few isolated indicators. For example, they might judge the importance of a node solely by counting the number of reposts or comments, or simply by relying on the number of followers of an account. However, this approach often overlooks the actual propagation path of information within complex network structures, the amplification effect of cross-platform dissemination, and the dynamic changes and true influence of each node during the propagation process. Therefore, relying solely on these static and localized indicators makes it difficult to comprehensively and dynamically assess the overall risk of a public opinion event, nor can it accurately identify the truly key propagation nodes. Clearly, the public opinion analysis methods in these technologies have significant shortcomings in terms of the accuracy and reliability of assessing the overall risk of an event and identifying key propagation nodes. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides a method, system, device, and medium for public opinion analysis based on propagation graphs.

[0004] Firstly, this application provides a public opinion analysis method based on a propagation graph, comprising: acquiring multi-source public opinion data associated with a target public opinion event; extracting participating accounts as nodes in a propagation network; and extracting basic attribute data for each node, including posting timestamp, platform affiliation, number of followers, historical interaction volume, and media level; extracting content features from the multi-source public opinion data; constructing directed propagation edges based on interaction records between nodes, content similarity, and posting timestamps to generate a public opinion propagation graph; extracting network topology features of each node in the public opinion propagation graph, including out-degree, downstream connected node size, and forwarding contribution, as well as sentiment features including the emotional polarity and negative sentiment amplification of the published content; using the number of followers, historical interaction volume, media level, forwarding contribution, and negative sentiment amplification as influencing factors, performing feature normalization and weighted calculation to obtain the influence value of each node; and traversing the public opinion propagation graph to select the node with the earliest posting timestamp. The system identifies the first-time cross-platform node as the node whose platform changes and whose direct downstream platform changes the earliest. Based on the influence value, negative sentiment amplification, and out-degree change rate of each node within a preset time window, key propagation nodes are identified according to preset rules. Global propagation features of the public opinion propagation graph are extracted, including at least: propagation speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. Propagation speed is the number of new nodes added per unit time, and high-influence nodes are key propagation nodes with influence values ​​greater than a preset influence threshold. The global propagation features are quantified based on preset risk assessment logic to obtain the overall propagation risk level of the target public opinion event. Based on the first-time node, the first-time cross-platform node, and key propagation nodes, a structured propagation path of the target public opinion event is output, and a priority response suggestion is generated in conjunction with the overall propagation risk level.

[0005] By adopting the above technical solutions, multi-source public opinion data is acquired and node and basic attribute data are extracted, enabling a comprehensive understanding of the accounts involved in the dissemination. A directed propagation edge is constructed based on interaction records, content similarity, and timestamps to generate a public opinion propagation graph, clearly showing the direction of information flow. Network topology and sentiment features of nodes are extracted, and feature normalization and weighted calculations are performed to obtain influence values, enabling a comprehensive assessment of node influence. Identifying initial nodes, first cross-platform nodes, and key propagation nodes helps to accurately locate important propagation nodes. Global propagation features are extracted and quantitatively calculated to obtain the overall propagation risk level, enabling a comprehensive assessment of public opinion event risks. A structured propagation path is output, and combined with the risk level, a priority prompt for handling is generated, providing clear guidance for public opinion handling. This achieves the effect of improving the accuracy and reliability of public opinion assessment.

[0006] Optionally, a directed propagation edge is constructed based on the interaction behavior records between nodes, content similarity, and publication timestamps to generate a public opinion propagation graph. This includes: parsing multi-source public opinion data, extracting explicit interaction behavior records between nodes (including forwarding, citation, or reprinting behaviors), and constructing explicit directed propagation edges between corresponding nodes based on the order of explicit interaction behavior records and publication timestamps; for target nodes without explicit interaction behavior records, using a pre-trained semantic feature extraction model to convert the content published by the target node into text vectors, and selecting at least one candidate upstream node whose publication time is within a preset source tracing time window and whose cosine similarity with the target node's text vector is greater than a preset similarity threshold; based on the cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node; selecting the candidate upstream node corresponding to the maximum association weight, and constructing a unique implicit directed propagation edge between the candidate upstream node corresponding to the maximum value and the target node; and fusing the explicit and implicit directed propagation edges to generate a public opinion propagation graph.

[0007] By adopting the above technical solutions, explicit directed propagation edges are constructed using explicit interaction behavior records, which can intuitively reflect the information flow between nodes that exhibit forwarding, citation, or reprinting behaviors. For nodes without explicit interaction, candidate upstream nodes are screened using a pre-trained model and cosine similarity, and the association weights are calculated by combining the time decay effect to construct implicit directed propagation edges, which can uncover potential information propagation paths. Finally, the two types of directed propagation edges are fused to generate a public opinion propagation map, which can more comprehensively and accurately reflect the actual flow direction and propagation relationship of public opinion information in the propagation network, providing a more reliable data foundation for subsequent public opinion analysis.

[0008] Optionally, based on cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node, including: calculating the association weight between the candidate upstream node and the target node according to the following comprehensive weight formula: W （u,v） =sim(u,v)·e -λΔt Among them, W （u,v） is the association weight between candidate upstream node u and target node v; sim(u,v) is the cosine similarity between the text vectors of the content published by candidate upstream node u and target node v; Δt is the time difference between the publication timestamps of target node v and candidate upstream node u; λ is the preset time decay coefficient; the time decay coefficient λ is dynamically assigned according to the platform characteristics of the target public opinion event or the propagation decay law of similar public opinion events in the past.

[0009] By adopting the above technical solution, when constructing the public opinion dissemination graph, the association weight between the target node and the candidate upstream node is calculated using a comprehensive weight formula. This takes into account content similarity and time decay effect. Furthermore, the time decay coefficient can be dynamically assigned according to platform characteristics or the dissemination decay law of similar public opinion events in the past. This can more accurately construct implicit directed propagation edges, making the public opinion dissemination graph more accurately reflect the information dissemination path, thereby improving the accuracy of subsequent calculation of node influence values, identification of key dissemination nodes, and assessment of the overall dissemination risk level.

[0010] Optionally, the number of account followers, historical interaction volume, media level, forwarding contribution, and negative sentiment amplification are used as influencing factors. Feature normalization and weighted calculation are performed to obtain the influence value of each node. This includes: performing logarithmic normalization on the number of account followers and historical interaction volume to obtain the normalized values ​​of the number of account followers and historical interaction volume; mapping the media level to [0,1] according to the preset level-score mapping table to obtain the normalized value of the media level; and calculating the influence value of each node according to the following formula: Influence value = W1 × Normalized value of number of account followers + W2 × Normalized value of historical interaction volume + W3 × Normalized value of media level + W4 × Forwarding contribution + W5 × Negative sentiment amplification; where W1, W2, W3, W4, and W5 are the weight coefficients of each influencing factor, and their sum is 1.

[0011] By adopting the above technical solutions, logarithmic normalization is performed on the number of account followers and historical interaction volume, and media level is normalized according to the mapping table. This makes the influence factors of different levels comparable, which is convenient for subsequent calculations. By comprehensively considering the influence factors such as the number of account followers, historical interaction volume, media level, repost contribution, and negative sentiment amplification, and performing weighted calculations according to the formula, the influence value of each node can be obtained more comprehensively and accurately. This helps to more accurately identify key dissemination nodes and assess the overall dissemination risk in the future.

[0012] Optionally, based on the influence value, negative sentiment amplification, and out-degree change rate of each node within a preset time window, key dissemination nodes are identified according to preset rules, including: for any node in the public opinion dissemination graph, obtaining the out-degree change rate of the node within a preset time window; if the out-degree change rate of the node within the preset time window is greater than a preset first threshold, the node is identified as a node with increased dissemination speed; if the influence value of the node is greater than a preset second threshold, the node is identified as a media amplification node; if the negative sentiment amplification of the node is greater than a preset third threshold, the node is identified as a node with concentrated negative sentiment; the nodes with increased dissemination speed, media amplification, and concentrated negative sentiment are determined as key dissemination nodes.

[0013] By adopting the above technical solution, key dissemination nodes of a target public opinion event can be classified and identified based on the node's influence value, negative sentiment amplification degree, and out-degree change rate. Specifically, nodes that increase dissemination speed, nodes that amplify media, and nodes where negative sentiment is concentrated can be identified. This allows for a more accurate determination of key dissemination nodes in the target public opinion event, which is conducive to more effective analysis and response to public opinion events.

[0014] Optionally, the global dissemination characteristics also include: the number of media reposts and the proportion of complaints. The number of media reposts represents the number of nodes with a media level greater than a preset threshold, and the proportion of complaints represents the ratio of nodes whose published content contains complaints or rights protection information to the total number of nodes. Based on a preset risk assessment logic, the global dissemination characteristics are quantitatively calculated to obtain the overall dissemination risk level of the target public opinion event. This includes: normalizing the dissemination speed after truncating it to an upper limit to obtain a normalized dissemination speed value; normalizing the number of platforms involved, the number of high-influence nodes, and the number of media reposts to obtain normalized platform numbers, normalized high-influence node numbers, and normalized media repost numbers, respectively; weighting and summing the normalized dissemination speed value, normalized platform number value, proportion of negative sentiment nodes, normalized high-influence node numbers, normalized media repost numbers, proportion of complaints, and the preset risk weights corresponding to each global dissemination characteristic to obtain an overall risk score; and determining the overall dissemination risk level of the target public opinion event based on the preset risk score range in which the overall risk score falls.

[0015] By adopting the above technical solutions, the number of media reposts and the proportion of complaints are incorporated into the overall dissemination characteristics. This allows the assessment to take into account the number of nodes with significant media influence and the proportion of complaint and rights protection content, thus more comprehensively reflecting the characteristics of public opinion events. The dissemination speed, the number of platforms involved, the number of high-influence nodes, and the number of media reposts are normalized for easier subsequent unified calculations. The overall risk score is obtained by weighting and summing the normalized values, the proportion of nodes with negative sentiment, the proportion of complaints, and preset risk weights. The overall dissemination risk level is determined based on the range in which the score falls, enabling a more accurate and comprehensive assessment of the overall dissemination risk of the target public opinion event.

[0016] Optionally, based on the initial node, the first cross-platform node, and key propagation nodes, a structured propagation path for the target public opinion event is output, and a priority prompt for handling is generated in conjunction with the overall propagation risk level. This includes: starting from the initial node, traversing step by step to the terminal node of the public opinion propagation graph according to the information flow direction indicated by the directed propagation edges between nodes in the public opinion propagation graph, generating multiple candidate propagation paths from the initial node to the terminal node; filtering the multiple candidate propagation paths, retaining candidate propagation paths that pass through at least one key propagation node or the first cross-platform node, to obtain a set of filtered propagation paths; when the number of paths in the set of filtered propagation paths exceeds the preset number of paths... At the upper limit, the sum of the influence values ​​of all nodes along each filtered propagation path is calculated as the path influence score. The paths are then sorted from highest to lowest influence score, and a predetermined number of propagation paths are retained to obtain the target propagation path set. Key propagation nodes and first-time cross-platform nodes are marked on each propagation path in the target propagation path set to generate a structured propagation path for the target public opinion event. The response level of the target public opinion event is determined based on the overall propagation risk level. The nodes are sorted from highest to lowest influence value according to the initial node, first-time cross-platform node, and each key propagation node to generate a node handling order. The response level and handling order are combined to generate a handling priority prompt.

[0017] By adopting the above technical solution, candidate propagation paths are generated starting from the initial node. After screening and calculating the path influence score, a set of target propagation paths is obtained. After marking key propagation nodes and the first cross-platform node, the structured propagation path of the target public opinion event can be clearly presented. The response level is determined according to the overall propagation risk level, and the disposal order is generated by sorting the node influence values. Then, the disposal priority prompts are generated. This helps enterprises and institutions to accurately grasp the public opinion propagation path, determine the disposal priority according to the risk level and the importance of the nodes, and improve the efficiency and pertinence of public opinion disposal.

[0018] Optionally, after generating the public opinion dissemination graph, the above method further includes: determining the dissemination stage of the public opinion dissemination graph to identify the current dissemination stage of the target public opinion event; wherein, the dissemination stage determination includes: dividing the evolution process of the public opinion dissemination graph into multiple consecutive time slices based on the publication timestamps of each node in the public opinion dissemination graph; for each time slice, counting the number of newly added nodes, the number of newly added directed dissemination edges, and the average weight value of the newly added directed dissemination edges to obtain the dissemination activity index corresponding to the time slice; wherein the weight value of each newly added directed dissemination edge is calculated based on the cosine similarity of the text vectors of the content published by the two associated nodes and the publication time difference; comparing the dissemination activity index with a preset dissemination stage threshold range to determine the dissemination stage label corresponding to each time slice, wherein the dissemination stage label includes at least one of the incubation period, outbreak period, diffusion period, and decline period; determining the dissemination stage label corresponding to the latest time slice as the current dissemination stage of the target public opinion event; and outputting the current dissemination stage as additional information of the structured dissemination path when outputting the structured dissemination path of the target public opinion event.

[0019] Optionally, the above method also includes a propagation trend prediction step: obtaining the topological evolution sequence of the public opinion propagation graph within multiple historical time windows, the topological evolution sequence including the number of nodes, the number of directed propagation edges, and the out-degree change data of each node corresponding to each time window; inputting the topological evolution sequence into a preset time series prediction model, the time series prediction model predicts the number of new nodes and the number of new directed propagation edges in the public opinion propagation graph in the next time window based on the topological evolution sequence; based on the number of new nodes and the number of new directed propagation edges, combined with the influence value of each node in the current public opinion propagation graph, identifying candidate nodes that may become key propagation nodes in the next time window; outputting the prediction results of the candidate nodes, and using the prediction results as supplementary information for the priority indication of handling.

[0020] Optionally, after obtaining the number of new nodes predicted by the time-series prediction model in the next time window, the method further includes: calculating a propagation trend factor based on the predicted number of new nodes in the next time window and the actual number of new nodes in the current time window; wherein, the propagation trend factor = (predicted number of new nodes) / (actual number of new nodes + σ), where σ is a very small positive value; when the propagation trend factor is greater than 1, multiplying the overall risk score corresponding to the overall propagation risk level by the propagation trend factor to obtain a corrected risk score; when the propagation trend factor is less than or equal to 1, keeping the overall risk score unchanged; redetermining the overall propagation risk level of the target public opinion event based on the preset risk score range in which the corrected risk score is located, and updating the handling priority prompt based on the redetermined overall propagation risk level.

[0021] Optionally, in calculating the influence value, the weight coefficients of the influence factors are dynamically determined based on the current dissemination stage of the target public opinion event. The current dissemination stage includes the incubation period, outbreak period, diffusion period, and decline period. A corresponding preset weight coefficient set is configured for each of the incubation period, outbreak period, diffusion period, and decline period. Specifically: in the preset weight coefficient set corresponding to the incubation period, the weight coefficients of media level and negative sentiment amplification are greater than the weight coefficients of other influence factors; in the preset weight coefficient set corresponding to the outbreak period, the weight coefficients of forwarding contribution and historical interaction volume are greater than the weight coefficients of other influence factors; in the preset weight coefficient set corresponding to the diffusion period, the weight coefficient of account follower count is greater than the weight coefficients of other influence factors; and in the preset weight coefficient set corresponding to the decline period, the weight coefficient of negative sentiment amplification is greater than the weight coefficients of other influence factors.

[0022] Optionally, after identifying nodes that increase the speed of dissemination, nodes that amplify media, and nodes that concentrate negative emotions, the method also includes: determining whether there are overlapping nodes among the key dissemination nodes that simultaneously satisfy at least two classification rules; if there are overlapping nodes, upgrading the overlapping nodes to core high-risk nodes; when generating priority prompts, assigning core high-risk nodes the highest level of priority, and highlighting core high-risk nodes in a differentiated manner in the structured dissemination path.

[0023] In a second aspect of this application, a public opinion analysis system based on a propagation graph is also provided, used to execute any of the aforementioned methods for public opinion analysis based on a propagation graph, comprising: an acquisition module, used to acquire multi-source public opinion data associated with a target public opinion event, extract accounts participating in the propagation as nodes in the propagation network, and extract basic attribute data for each node including posting timestamp, platform affiliation, number of followers, historical interaction volume, and media level; a generation module, used to extract content features from the multi-source public opinion data, construct directed propagation edges based on interaction records between nodes, content similarity, and posting timestamps, and generate a public opinion propagation graph; a calculation module, used to extract network topology features of each node in the public opinion propagation graph including out-degree, downstream connected node size, and forwarding contribution, as well as sentiment features including the sentiment polarity of the published content and the amplification of negative emotions; using the number of followers, historical interaction volume, media level, forwarding contribution, and amplification of negative emotions as influencing factors, performing feature normalization and weighted calculation to obtain the influence value of each node; and an identification module, This module is used to traverse the public opinion dissemination graph, identifying the node with the earliest publication timestamp as the initial node, and the node whose platform and the platform of its direct downstream nodes have changed, and whose publication time is the earliest, as the first cross-platform node. Based on the influence value, negative sentiment amplification, and out-degree change rate of each node within a preset time window, it identifies key dissemination nodes according to preset rules. The evaluation module extracts global dissemination features from the public opinion dissemination graph, including at least: dissemination speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. Dissemination speed is the number of new nodes added per unit time, and high-influence nodes are key dissemination nodes with influence values ​​greater than a preset influence threshold. Based on preset risk assessment logic, the global dissemination features are quantified to obtain the overall dissemination risk level of the target public opinion event. The output module, based on the initial node, the first cross-platform node, and key dissemination nodes, outputs the structured dissemination path of the target public opinion event and generates a priority prompt for handling based on the overall dissemination risk level.

[0024] In a third aspect of this application, an electronic device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the program to implement the method steps of any of the above claims.

[0025] In a fourth aspect of this application, a computer-readable storage medium is also provided, which stores instructions that, when executed, perform the method steps of any of the above claims.

[0026] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: 1. By acquiring multi-source public opinion data and extracting node and basic attribute data, it can comprehensively grasp the account information involved in the dissemination; by constructing directed propagation edges based on interaction behavior records, content similarity, and timestamps to generate a public opinion propagation graph, it can clearly present the direction of information flow; by extracting the network topology and sentiment characteristics of nodes and performing feature normalization and weighted calculation to obtain the influence value, it can comprehensively assess the influence of nodes; by identifying the initial node, the first cross-platform node, and key propagation nodes, it helps to accurately locate important propagation nodes; by extracting global propagation characteristics and quantifying the overall propagation risk level, it can comprehensively assess the risk of public opinion events; by outputting structured propagation paths and generating handling priority prompts based on risk levels, it provides clear guidance for public opinion handling; thus improving the accuracy and reliability of public opinion assessment. 2. Starting from the initial node, candidate propagation paths are generated. After screening and calculating the path influence score, a set of target propagation paths is obtained. After marking key propagation nodes and the first cross-platform node, the structured propagation path of the target public opinion event can be clearly presented. The response level is determined according to the overall propagation risk level, and the disposal order is generated by combining the node influence value. Then, the disposal priority prompts are generated, which helps enterprises and institutions to accurately grasp the public opinion propagation path, determine the disposal priority according to the risk level and the importance of the node, and improve the efficiency and pertinence of public opinion disposal. Attached Figure Description

[0027] Figure 1 This is a flowchart of a public opinion analysis method based on propagation graphs provided in an embodiment of this application; Figure 2 This is a system architecture diagram of public opinion dissemination path identification and key node positioning provided in the embodiments of this application; Figure 3 This is an example diagram of public opinion network propagation and key nodes provided in an embodiment of this application; Figure 4 This is a structural block diagram of a public opinion analysis system based on propagation graphs provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0028] Explanation of reference numerals in the attached figures: 500 - Electronic device; 501 - Processor; 502 - Communication bus; 503 - User interface; 504 - Network interface; 505 - Memory. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0030] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0031] In the description of the embodiments of this application, the term "multiple" means two or more. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0032] This application provides a public opinion analysis method based on propagation graphs, referring to... Figure 1 , Figure 1 This application provides a flowchart of a public opinion analysis method based on propagation graphs, which includes the following steps: Step S101: Obtain multi-source public opinion data associated with the target public opinion event, extract the accounts involved in the dissemination as nodes of the dissemination network, and extract basic attribute data of each node including the release timestamp, platform, number of followers, historical interaction volume and media level; Step S102: Extract content features from multi-source public opinion data, construct directed propagation edges based on interaction records between nodes, content similarity, and publication timestamps, and generate a public opinion propagation graph. Step S103: Extract the network topology features of each node in the public opinion propagation graph, including out-degree, downstream connected node scale and forwarding contribution, as well as the sentiment features including the emotional polarity of the published content and the amplification of negative emotions; use the number of account followers, historical interaction volume, media level, forwarding contribution and negative emotion amplification as influencing factors, perform feature normalization and weighted calculation, and obtain the influence value of each node. Step S104: Traverse the public opinion dissemination map, identify the node with the earliest release timestamp as the initial node, and identify the node whose platform changes and whose direct downstream platform changes as the earliest release time as the first cross-platform node; and identify key dissemination nodes according to preset rules based on the influence value, negative sentiment amplification degree and out-degree change rate of each node within a preset time window. Step S105: Extract the global propagation features of the public opinion propagation map. The global propagation features include at least: propagation speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. The propagation speed is the number of new nodes added per unit time, and the high-influence nodes are key propagation nodes with an influence value greater than a preset influence threshold. Based on the preset risk assessment logic, the global propagation features are quantitatively calculated to obtain the overall propagation risk level of the target public opinion event. Step S106: Based on the initial node, the first cross-platform node, and the key dissemination nodes, output the structured dissemination path of the target public opinion event, and generate a handling priority prompt in combination with the overall dissemination risk level.

[0033] Through the above steps, multi-source public opinion data is acquired and node and basic attribute data are extracted, enabling a comprehensive understanding of the accounts involved in the dissemination. A directed propagation edge is constructed based on interaction records, content similarity, and timestamps to generate a public opinion propagation graph, clearly showing the direction of information flow. Network topology and sentiment characteristics of nodes are extracted, and feature normalization and weighted calculations are performed to obtain influence values, enabling a comprehensive assessment of node influence. Identifying initial dissemination nodes, first cross-platform nodes, and key propagation nodes helps to accurately locate important propagation nodes. Global propagation characteristics are extracted and quantitatively calculated to obtain the overall propagation risk level, enabling a comprehensive assessment of public opinion event risks. A structured propagation path is output, and combined with the risk level, a priority prompt for handling is generated, providing clear guidance for public opinion handling. This achieves the effect of improving the accuracy and reliability of public opinion assessment.

[0034] This embodiment constructs a multi-dimensional, graph-based complete public opinion analysis process. The overall principle is to first build a weighted dissemination graph using accounts as nodes and multiple dissemination criteria, then integrate five dimensions—fans, interaction, media, downstream coverage, and negative amplification—to quantify the influence of a single account. It also identifies key nodes such as initial releases, first cross-platform events, and various other types, integrates multiple global indicators to quantify overall public opinion risk, and finally outputs a structured dissemination path and generates priority prompts for handling. In other words, it models the participants and their interactions in a public opinion event as a directed weighted graph (i.e., a dissemination graph). By calculating the multi-dimensional influence of each node in the graph and quantifying the global dissemination characteristics, it ultimately achieves a complete automated analysis process from "data collection" to "structured output." Specifically, this method first extracts the accounts involved in the dissemination as nodes and labels each node with basic attributes such as time, platform, and number of followers, laying the data foundation for subsequent analysis. Second, it constructs directed propagation edges through multi-dimensional clues such as forwarding, similarity, and temporal proximity, linking isolated single pieces of public opinion content into a propagation network reflecting the direction of information flow. Next, it extracts the topological features (such as out-degree and downstream coverage scale) and sentiment features (such as content sentiment polarity and negative amplification) of each node on this network, and calculates the node influence value through weighted normalization. This value comprehensively reflects the node's contribution to dissemination and the content's inflammatory nature. Then, it identifies the initial node and the first cross-platform node based on the graph's timestamp information and network structure, and identifies other key nodes in combination with the influence value. It then extracts global features such as dissemination speed, number of platforms (cross-platform breadth), proportion of negative sentiment nodes, and number of high-influence nodes from the graph, calculates the overall risk score with weights, and maps it to a risk level. Finally, it generates a structured dissemination path based on the graph and generates priority prompts for handling based on the risk level and node influence. Current technologies rely solely on static metrics like followers and reposts to assess account importance, failing to establish a complete dissemination network, identify cross-platform propagation sources, and offer limited risk assessment dimensions. They cannot differentiate between different types of amplification nodes, and analysis of public opinion based on static, localized indicators leads to inaccurate overall risk assessments, ineffective identification of key dissemination nodes, and a lack of tiered response guidelines. This embodiment addresses this by reconstructing the true information flow through a complete data graph, objectively quantifying an account's actual dissemination and amplification capabilities using multi-factor weighting, accurately locating key cross-platform and explosive entities, and dynamically calculating risks using multi-dimensional indicators. This achieves a structured presentation of public opinion propagation paths and automatic generation of response priorities, significantly improving the accuracy and reliability of public opinion tracing and risk assessment, and providing quantitative decision-making support for financial and corporate public opinion management.

[0035] In step S101, multi-source public opinion data refers to public opinion-related content from different social media platforms (such as Weibo, Douyin, Zhihu, Xiaohongshu, etc.). The data formats of different platforms vary and need to be obtained through their respective APIs or web crawlers. A propagation network node is an abstraction of each participating account as a node in graph theory; a node is the basic building block of a propagation network. In graph theory, a network consists of two parts: "nodes" and "edges." Nodes represent entities in the network, and edges represent the relationships between entities. This step completes the "node extraction" work, abstracting each participating account as a node; however, nodes alone are not enough. Each node needs to carry sufficient attribute information to support subsequent analysis and calculations. This step selects five basic attributes: the posting timestamp refers to the precise time, accurate to the second, when an account first posted content related to the target public opinion event, used for subsequent initial node identification and time-series analysis. The platform refers to the name of the social media platform to which the account belongs, used for subsequent cross-platform diffusion detection. Account follower count measures the potential audience size of a node; historical interaction volume measures the historical appeal of the node's content, such as the average interaction volume of content posted by the account within a certain period before the release of content related to the target public opinion event (e.g., 30 days, or other periods), usually the sum of comments, likes, and reposts. Historical interaction volume reflects the historical appeal of the account's content. Media level refers to the authority level of the media to which the account belongs. The system has a preset level-score mapping table, for example, A-level mainstream media = 0.90, B-level local media = 0.70, C-level self-media influencers = 0.50, D-level ordinary personal accounts = 0.15, used to distinguish the dissemination weight of different types of accounts. Graph theory is a mature mathematical tool for processing complex relational networks. The selection of these five attributes is not arbitrary, but based on in-depth analysis of the mechanism of public opinion dissemination. They respectively cover the "time dimension," "spatial dimension," "scale dimension," "quality dimension," and "authority dimension," forming a complete node profile.

[0036] In step S102, "content features" refers to the textual semantic features (such as keywords and topic vectors) extracted after NLP processing of public opinion data. "Interaction behavior records" include traceable behaviors such as forwarding, quoting, reposting, and mentioning. "Content similarity" is used to infer implicit propagation relationships when there are no explicit interaction records. "Directed propagation edges" refer to directional connections from information source nodes to receiving / propagating nodes, representing the direction of information flow. The "public opinion propagation graph" is a directed graph composed of all nodes and directed propagation edges. A dual-track edge construction strategy of "explicit + implicit" can be adopted. Explicit edges are based on interaction behaviors (forwarding, quoting, reposting) explicitly recorded by the platform; these edges are accurate and reliable but have limited coverage. Implicit edges are inferred based on content similarity and temporal proximity. If the content of two nodes is highly similar semantically and their publication times are close, a substantial propagation relationship can be reasonably inferred, even without explicit platform records.

[0037] In step S103, five key influencing factors were selected from two dimensions: "network topology" and "content sentiment." Network topology characteristics assess the propagation contribution of nodes from a graph structure perspective. A larger out-degree indicates that the node directly propagates to more downstream nodes, and a larger downstream connected node scale indicates a wider indirect influence range. Forwarding contribution is determined based on the ratio of the downstream connected node scale to the total number of nodes, condensing this information into a ratio between 0 and 1. Sentiment characteristics assess the "inflammatory" nature of nodes from a content semantic perspective. Sentiment polarity reflects the emotional tendency of the content, and negative sentiment amplification is determined based on the proportion of nodes publishing negative sentiment content among the node and its corresponding direct downstream nodes. Negative sentiment amplification further measures the concentration of negative content in the node and its direct downstream nodes, reflecting the node's emotional amplification effect during propagation. After normalization, the five factors are weighted and summed to form a comprehensive score, i.e., the influence value. The innovation of this influence value lies in the fact that it is not a static indicator dependent on a single source, but a dynamic comprehensive indicator that integrates network location, propagation contribution, and content characteristics.

[0038] In step S104, automated identification is achieved by combining "deterministic rules" and "threshold filtering." Identifying the initial node is the simplest and most definitive step; the earliest published content is the source, determined by its timestamp, leaving no room for ambiguity. Identifying the first cross-platform node is also based on deterministic rules, finding nodes where a platform switch occurs between the original platform and its direct downstream platforms, thus identifying the trigger point for cross-platform diffusion. Identifying key propagation nodes uses a threshold filtering method. For example, nodes with an influence value exceeding a threshold are media amplification nodes; nodes with a negative sentiment amplification exceeding a threshold are nodes with concentrated negative sentiment; and nodes with an out-degree change rate exceeding a threshold are nodes with increased propagation speed. Each of these three types of nodes captures "abnormal" nodes across three different dimensions: "high influence," "high negativity," and "high acceleration," which, when combined, form the key node set.

[0039] In step S105, the overall propagation risk level of the target public opinion event is assessed based on global propagation characteristics. Propagation speed reflects how quickly the public opinion spreads; the faster the speed, the shorter the window for intervention. The number of platforms involved reflects the degree to which the public opinion has "gone beyond its original scope"; the more platforms involved, the wider the impact and the greater the difficulty of control. The proportion of negative sentiment nodes reflects the "negative concentration" of the public opinion; the higher the negative proportion, the greater the damage to brand reputation. The number of high-influence nodes reflects the number of "heavyweight" participants; the more high-influence nodes, the greater the likelihood of the event being amplified. These four characteristics correspond to the four dimensions of "speed," "breadth," "concentration," and "weight," respectively, and together constitute a macro-level description of the overall situation of the public opinion event.

[0040] In step S106, the contribution lies in "information visualization" and "decision directives." Regarding information visualization: starting from the initial node, the entire graph is traversed along directed edges, generating all paths from the source to the end node. Then, paths passing through key propagation nodes (these are the true "risk propagation main lines") are selected, and key propagation node labels are marked on the paths, ultimately outputting structured propagation chains. This way, users see not a complex graph, but clear propagation chains. Regarding decision directives: risk levels are mapped to response levels (solving the question of "what level of force to use"), and key nodes are sorted by influence (solving the question of "who to contact first"). These two are combined into standardized instructions such as "[Level X Response] Please handle in order: ①…②…③…". This provides users with an action plan that can be directly executed, rather than a pile of data requiring secondary processing. Based on the risk level and the influence ranking of key nodes, users (such as corporate public relations teams and government cyberspace administration departments) are provided with priority node order and response suggestions.

[0041] This embodiment achieves a structured presentation of public opinion dissemination paths by constructing a directed propagation graph; it achieves precise positioning of key nodes through multi-dimensional influence assessment; it achieves automatic detection of cross-platform diffusion through cross-platform node identification; it achieves dynamic assessment of risk levels through global feature weighted quantification; and it provides clear and actionable priority guidance for risk management through structured output.

[0042] In an optional embodiment, a directed propagation edge is constructed based on the interaction behavior records between nodes, content similarity, and publication timestamps to generate a public opinion propagation graph. This includes: parsing multi-source public opinion data, extracting explicit interaction behavior records between nodes (including forwarding, citation, or reprinting behaviors), constructing explicit directed propagation edges between corresponding nodes based on the order of explicit interaction behavior records and publication timestamps; for target nodes without explicit interaction behavior records, using a pre-trained semantic feature extraction model to convert the content published by the target node into text vectors, and selecting at least one candidate upstream node whose publication time is within a preset source tracing time window and whose cosine similarity with the target node's text vector is greater than a preset similarity threshold; based on the cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node; selecting the candidate upstream node corresponding to the maximum association weight, constructing a unique implicit directed propagation edge between the candidate upstream node corresponding to the maximum value and the target node; and fusing the explicit and implicit directed propagation edges to generate a public opinion propagation graph.

[0043] In the above embodiments, explicit directed propagation edges are constructed using explicit interaction behavior records, which can intuitively reflect the information flow between nodes that exhibit forwarding, citation, or reprinting behaviors. For nodes without explicit interaction, candidate upstream nodes are screened using a pre-trained model and cosine similarity, and the association weights are calculated by combining the time decay effect to construct implicit directed propagation edges, which can uncover potential information propagation paths. Finally, the two types of directed propagation edges are fused to generate a public opinion propagation map, which can more comprehensively and accurately reflect the actual flow direction and propagation relationship of public opinion information in the propagation network, providing a more reliable data foundation for subsequent public opinion analysis.

[0044] The construction of propagation edges can be divided into two categories: "explicit" and "implicit." Explicit propagation edges are built based on explicit interactive behaviors (forwarding, quoting, reposting) recorded by the platform. These edges have high credibility but limited coverage because many substantial "propagation" behaviors (such as a user seeing a Weibo post and then publishing similar content on other platforms) do not have explicit platform interaction records. Implicit propagation edges, on the other hand, are inferred based on content similarity analysis. If an account publishes content highly similar to an upstream account, and the publication time is after the upstream account, then even without explicit forwarding or quoting records, it is reasonable to infer the existence of a "substantive propagation relationship." Specifically, this method first analyzes all explicit interactive behaviors from multi-source public opinion data, directly constructing explicit directed propagation edges between corresponding nodes. Then, for "isolated" target nodes not covered by any explicit edges, a semantic feature extraction model is used to convert their content into text vectors. Within a preset source tracing time window, candidate upstream nodes with content similarity higher than a threshold are searched. When multiple candidate upstream nodes exist, a time decay effect is introduced to calculate a comprehensive association weight; the more similar the content and the closer the time, the greater the weight. The candidate upstream node with the highest weight is selected to construct a unique implicit directed propagation edge. Finally, the two types of edges are merged to form a complete graph. This method solves the problem that existing technologies cannot identify the propagation relationships between content in scenarios such as cross-platform propagation, secondary creation propagation, and screenshot propagation where there are no clear forwarding or citation records, leading to breaks and omissions in the propagation graph. By introducing an implicit edge inference mechanism based on content similarity and time decay, this method can effectively fill the gaps in propagation relationships that explicit interactive records cannot cover, significantly improving the completeness and accuracy of the propagation graph, and providing a more reliable data foundation for subsequent node influence calculation and key node identification. In related technologies, the construction of propagation relationships typically relies solely on explicit forwarding or citation records. However, in cross-platform propagation scenarios, short video content on Douyin may originate from text and image posts on Weibo, but Douyin's forwarding records do not point to Weibo accounts. In derivative works scenarios, users may rephrase the core arguments of a long article in their own words, without forwarding or citation, but essentially propagating the original content. These situations lead to numerous breaks and omissions in the propagation graph constructed solely based on explicit records, severely impacting the accuracy of subsequent analysis. This embodiment introduces an implicit edge inference mechanism based on content similarity and time decay. By fusing explicit and implicit edges, a more complete propagation network is constructed, providing more comprehensive data support for subsequent node influence calculation and key node identification.

[0045] In an optional embodiment, based on cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node, including: calculating the association weight between the candidate upstream node and the target node according to the following comprehensive weight formula: W（u,v） =sim(u,v)·e -λΔt Among them, W （u,v） is the association weight between candidate upstream node u and target node v; sim(u,v) is the cosine similarity between the text vectors of the content published by candidate upstream node u and target node v; Δt is the time difference between the publication timestamps of target node v and candidate upstream node u; λ is the preset time decay coefficient; the time decay coefficient λ is dynamically assigned according to the platform characteristics of the target public opinion event or the propagation decay law of similar public opinion events in the past.

[0046] In the above embodiments, when constructing the public opinion dissemination graph, the association weight between the target node and the candidate upstream node is calculated using a comprehensive weight formula. This takes into account content similarity and time decay effects. Furthermore, the time decay coefficient can be dynamically assigned based on platform characteristics or the dissemination decay pattern of similar public opinion events in the past. This allows for a more accurate construction of implicit directed propagation edges, enabling the public opinion dissemination graph to more accurately reflect the information dissemination path. This, in turn, improves the accuracy of subsequent calculations of node influence values, identification of key dissemination nodes, and assessment of the overall dissemination risk level.

[0047] The association between candidate upstream nodes and target nodes is modeled as a comprehensive weight determined by two core factors: content similarity and temporal proximity. Higher content similarity indicates a semantically closer relationship between the two pieces of content, increasing the likelihood of a propagation relationship. A smaller time difference indicates that the two pieces of content were published closer together, increasing the likelihood of a direct propagation relationship. More importantly, the impact of the time factor is not linear but exhibits an exponential decay—as the time difference increases, even highly similar content will rapidly decrease in the probability of an actual propagation relationship. This is the essence of the "time decay effect." Specifically, this method uses formula W... （u,v）=sim(u,v)·e^(-λ·Δt) calculates the association weight, with the core innovation being the time decay term e^(-λ·Δt): when the time difference Δt=0, the decay term is 1, indicating that if two pieces of content are published simultaneously, the probability of their propagation relationship is entirely determined by similarity; as Δt increases, the decay term decreases exponentially, indicating that even if the similarity is high, two pieces of content with a large time difference are likely just "coincidental similarities" rather than "propagation relationships". The time decay coefficient λ is dynamically assigned based on platform characteristics or historical data of similar events: on platforms with fast propagation speeds such as Weibo, λ takes a larger value (0.3-0.5), indicating that a time difference of several hours will significantly reduce the probability of propagation; on platforms with slower propagation speeds such as Zhihu, λ takes a smaller value (0.1-0.2), allowing a larger time difference to be accepted as a valid propagation relationship. This solves the problem of "misjudgment" risk in related technologies that rely solely on content similarity to determine propagation relationships. Two pieces of content with high similarity may simply be reporting the same news event simultaneously, rather than having a propagation relationship, especially when a public opinion event triggers simultaneous reporting by multiple media outlets. By introducing a time decay effect, this method effectively suppresses such misjudgments, making the inferred propagation relationships closer to the actual propagation process and significantly improving the accuracy and reliability of implicit propagation edge construction. Through dynamic assignment of the time decay coefficient λ, this method can adapt to the propagation characteristics of different platforms, maintaining high inference accuracy in various application scenarios. By selecting the maximum value through comprehensive weight sorting, it ensures that each target node is associated only with the most likely propagation source, avoiding a large number of unreasonable propagation branches in the graph.

[0048] In an optional embodiment, the number of account followers, historical interaction volume, media level, forwarding contribution, and negative sentiment amplification are used as influencing factors. Feature normalization and weighted calculation are performed to obtain the influence value of each node. This includes: performing logarithmic normalization on the number of account followers and historical interaction volume to obtain normalized values ​​for the number of account followers and historical interaction volume; mapping the media level to [0,1] according to a preset level-score mapping table to obtain a normalized value for the media level; and calculating the influence value of each node according to the following formula: Influence value = W1 × Normalized value of number of account followers + W2 × Normalized value of historical interaction volume + W3 × Normalized value of media level + W4 × Forwarding contribution + W5 × Negative sentiment amplification; where W1, W2, W3, W4, and W5 are the weight coefficients of each influencing factor, and their sum is 1.

[0049] In the above embodiments, logarithmic normalization is performed on the number of account followers and historical interaction volume, and media level is normalized according to the mapping table. This makes the influence factors of different magnitudes within a comparable range, which is convenient for subsequent calculations. By comprehensively considering the influence factors such as the number of account followers, historical interaction volume, media level, forwarding contribution, and negative sentiment amplification, and performing weighted calculations according to the formula, the influence value of each node can be obtained more comprehensively and accurately. This helps to more accurately identify key dissemination nodes and assess the overall dissemination risk in the future.

[0050] The influence of a node is not determined by a single factor, but rather by five major influencing factors: the number of followers, historical interaction volume, media level, repost contribution, and negative sentiment amplification. These five factors characterize the node's impact on public opinion dissemination from different perspectives: the number of followers represents the potential audience size, interaction volume represents the content's attractiveness, media level represents the institution's credibility and authority, repost contribution represents the actual dissemination coverage, and negative sentiment amplification represents the content's inflammatory and emotionally guiding effect. Since these five dimensions have completely different dimensions and numerical ranges (follower counts can range from tens to tens of millions, while repost contribution is a ratio of 0 to 1), direct weighted summation is meaningless; therefore, normalization is necessary. For data such as follower count and interaction volume, which exhibit a long-tail distribution (a few accounts have extremely high values, while most accounts have relatively low values), this method uses logarithmic transformation followed by normalization to compress extreme values ​​to a reasonable range, making the normalized values ​​of most nodes distinguishable. For media ratings, which inherently possess an ordered categorization nature, this method directly maps to [0,1] using a pre-defined rating-score mapping table; for example, A-level mainstream media = 0.90, B-level local media = 0.70, C-level self-media influencers = 0.50, and D-level ordinary personal accounts = 0.15. This is just one example; other values ​​can certainly be used. Forwarding contribution and negative sentiment amplification are naturally within the [0,1] range and do not require additional normalization. After normalizing all factors, the final influence value is obtained through weighted summation. The normalized values ​​of each dimension are multiplied by their respective weight coefficients and then summed to obtain a comprehensive score. The weight coefficients reflect the relative importance of each dimension in judging "node importance." As an optional approach, a set of weight coefficients W1 to W5 can be pre-set as fixed weight coefficients. Alternatively, the aforementioned set of weight coefficients W1 to W5 can be dynamically adjusted according to the stage of public opinion development (i.e., the current stage of public opinion dissemination). In related technologies, the assessment of node importance typically relies on a single indicator (such as the number of followers or reposts), which fails to comprehensively reflect the node's true contribution to the dissemination network. Even when using multiple indicators, the lack of scientific normalization processing leads to the inability to reasonably integrate indicators with different dimensions and distributions. This method, through the systematic selection of five influencing factors, a differentiated normalization strategy, and a weighted summation mechanism, achieves accurate quantification of node influence, providing a reliable data foundation for subsequent identification and prioritization of key nodes.

[0051] In an optional embodiment, key propagation nodes are identified according to preset rules based on the influence value, negative sentiment amplification, and out-degree change rate of each node within a preset time window. This includes: for any node in the public opinion propagation graph, obtaining the out-degree change rate of the node within a preset time window; if the out-degree change rate of the node within the preset time window is greater than a preset first threshold, then the node is identified as a propagation speed-increasing node; if the influence value of the node is greater than a preset second threshold, then the node is identified as a media amplification node; if the negative sentiment amplification of the node is greater than a preset third threshold, then the node is identified as a negative sentiment concentration node; and the propagation speed-increasing node, media amplification node, and negative sentiment concentration node are determined as key propagation nodes.

[0052] In the above embodiments, key propagation nodes of a target public opinion event can be classified and identified based on the node's influence value, negative sentiment amplification degree, and out-degree change rate. Specifically, nodes that increase propagation speed, nodes that amplify media, and nodes that concentrate negative sentiment can be identified, thereby more accurately determining key propagation nodes in the target public opinion event, which is conducive to more effective analysis and response to public opinion events.

[0053] This embodiment provides a method for identifying key propagation nodes. Specifically, this method defines three different types of key propagation nodes: propagation speed-boosting nodes, media amplification nodes, and negative sentiment concentration nodes. Propagation speed-boosting nodes refer to those nodes that trigger a large number of new downstream propagations in a short period of time. These nodes act as "accelerators" in the propagation network; before their appearance, public opinion spreads slowly, but after their appearance, public opinion spreads rapidly. The basis for identifying these nodes is the out-degree change rate of the node within a preset time window: if the node's out-degree (number of direct downstream nodes) increases significantly within a short period (e.g., within 30 minutes), it indicates that the node has just published or forwarded some content, triggering a large number of new propagations. Media amplification nodes refer to high-influence accounts with influence values ​​exceeding the general level. These nodes are usually institutional accounts with a large number of followers or high media ratings. Their participation signifies that public opinion has achieved a "magnification effect," escalating from an ordinary topic to a public issue. The basis for identifying these nodes is whether their influence value exceeds a preset threshold. Negative sentiment concentration nodes refer to nodes where negative content accounts for an extremely high proportion of their own content and that of their direct downstream nodes. These nodes act as "gathering places" and "amplifiers" of negative sentiment, and their existence signifies that public opinion is evolving in a negative direction. The basis for identifying these nodes is whether the amplification of negative sentiment exceeds a preset threshold. For example, if a node has an out-degree of 0 at 10:00 and an out-degree of 8 at 10:30, then the out-degree change rate within 30 minutes is 8 / 0.5 = 16 (nodes / hour). If the first threshold is 10 nodes / hour, then this node is considered a node with increased propagation speed. For example, the preset second threshold is 0.6. This value can be adjusted according to business needs. When it is necessary to more sensitively capture potential major risks, the second threshold can be lowered (e.g., 0.50); ​​when it is necessary to focus on the most core nodes, the second threshold can be raised (e.g., 0.70). It should be noted that this second threshold can be equal to or different from the aforementioned preset influence threshold. For example, the aforementioned preset influence threshold can be 0.6, 0.7, or other values. The two are respectively used to filter media amplification nodes or high-influence nodes based on the magnitude of the influence value. That is to say, high-influence nodes may include media amplification nodes, and media amplification nodes may not necessarily be high-influence nodes. Different thresholds can be configured. For example, the aforementioned third threshold is 0.7 (or other values). The content of such nodes may be negative, or the content itself may be neutral but its comment section is filled with a lot of negative emotions. The definition of "key nodes" in related technologies is vague. They often only focus on "big V" accounts with a large number of followers, ignoring those nodes that suddenly erupt in specific events and actually drive the spread; they also ignore nodes with a high concentration of negative emotions.This method filters key nodes in the propagation graph by setting thresholds for three different dimensions: out-degree change rate, influence value, and negative emotion amplification. This allows for the comprehensive and accurate identification of different types of key nodes, providing complete node information for subsequent structured path output and priority generation.

[0054] In an optional embodiment, the global dissemination features further include: the number of media reposts and the proportion of complaints, wherein the number of media reposts represents the number of nodes with a media level greater than a preset level threshold, and the proportion of complaints represents the ratio of nodes whose published content contains complaints or rights protection content to the total number of nodes; the global dissemination features are quantitatively calculated based on a preset risk assessment logic to obtain the overall dissemination risk level of the target public opinion event, including: normalizing the dissemination speed after truncating the upper limit to obtain a normalized dissemination speed value; normalizing the number of platforms involved, the number of high-influence nodes, and the number of media reposts to obtain normalized platform number values, normalized high-influence node number values, and normalized media repost number values ​​respectively; weighted summation is performed based on the normalized dissemination speed value, normalized platform number value, proportion of negative sentiment nodes, normalized high-influence node number value, normalized media repost number value, proportion of complaints, and the preset risk weights corresponding to each global dissemination feature to obtain an overall risk score; the overall dissemination risk level of the target public opinion event is determined according to the preset risk score range in which the overall risk score is located.

[0055] In the above embodiments, the number of media reposts and the proportion of complaints are incorporated into the overall dissemination characteristics, so that the assessment takes into account the number of nodes with high media influence and the proportion of complaint and rights protection content, thus more comprehensively reflecting the characteristics of public opinion events. The dissemination speed, the number of platforms involved, the number of high-influence nodes, and the number of media reposts are normalized to facilitate subsequent unified calculations. The overall risk score is obtained by weighting and summing the normalized values, the proportion of negative sentiment nodes, the proportion of complaints, and the preset risk weights. The overall dissemination risk level is determined according to the range in which the score falls, which can more accurately and comprehensively assess the overall dissemination risk of the target public opinion event.

[0056] This embodiment provides a method for extracting global dissemination features and assessing the risk of target public opinion events based on these features. The risk level of a public opinion event is not determined by a single factor, but is composed of multiple factors, including dissemination speed, number of platforms (or breadth of spread), concentration of negative sentiment, participation of high-influence nodes, degree of media involvement, and proportion of complaints (or concentration of complaints). Among these, the number of media reposts reflects the mainstream media's attention to the event, which is a unique risk factor of this method. The involvement of mainstream media usually means that the event has risen from "hot online discussion" to "public issue," and the nature of the risk has undergone a qualitative change. The proportion of complaints reflects whether the event involves sensitive areas such as consumer rights and product quality. The public opinion risk of such content is often higher and more persistent. This method extracts the above six global dissemination features, normalizes each feature (normalization is performed for dissemination speed, number of platforms, number of high-influence nodes, and number of media reposts, with dissemination speed being normalized after upper limit truncation; for the proportion of negative sentiment nodes and the proportion of complaints, their proportion values ​​are used directly, since the proportion itself is a normalized value between 0 and 1), and then obtains the overall risk score through weighted summation. The weighted summation coefficients reflect the relative contributions of each risk factor in determining the final risk level. Finally, based on the preset range into which the risk score falls, public opinion events are classified into different risk levels (e.g., general concern, key concern, high risk, major risk). Upper limit truncation normalization involves setting an upper limit value, calculating values ​​exceeding that limit based on the upper limit, and then dividing by that limit. For example, regarding the spread rate: if the upper limit is set at 50 per hour, and the actual peak is 52 per hour, then the truncation is 50, and the normalized value is 50 / 50 = 1.0. Truncation is used to prevent extreme outliers from dominating the overall risk score. Platform number normalization involves dividing the number of platforms involved by a preset upper limit (e.g., 5 platforms). If 3 platforms are actually involved, then 3 / 5 = 0.60. The number of media reposts refers to the number of nodes in the dissemination graph whose media rating score is higher than a preset threshold (or whose media rating normalized value is greater than the preset media rating threshold). For example, there are 4 nodes with a media rating score greater than 0.9 (representing authoritative media) (such as "National XX Weekly" and "XX Daily"). These nodes are usually official accounts of various media outlets, and their participation in reporting signifies that the event has gained "media attention," which is an important signal of escalating public opinion risk. The risk weights corresponding to various global dissemination characteristics can be pre-set. For example, the risk weights corresponding to dissemination speed, number of platforms involved, proportion of nodes with negative sentiment, number of high-influence nodes, number of media reposts, and proportion of complaints are 0.25, 0.15, 0.20, 0.15, 0.15, and 0.10, respectively.The preset risk score ranges can include: general concern (0≤Risk<0.30), key concern (0.30≤Risk<0.50), high risk (0.50≤Risk<0.75), and significant risk (0.75≤Risk≤1.00). The overall risk level of a target public opinion event can be determined by which preset risk score range the overall risk score falls into. Related risk assessment methods are typically based on qualitative judgments or single statistical indicators (such as looking only at the total number of reposts), lacking a systematic multi-factor quantitative model. This leads to strong subjectivity and poor repeatability in risk level determination. Furthermore, these methods lack the introduction of key risk factors such as the number of media reposts and the proportion of complaints, resulting in a lag in the perception of qualitative changes in public opinion. This method, through a quantitative weighted model of six risk factors, achieves objectivity, automation, and dynamism in risk level determination, providing users with clear early warning signals.

[0057] In an optional embodiment, based on the initial node, the first cross-platform node, and key propagation nodes, a structured propagation path for the target public opinion event is output, and a handling priority prompt is generated in conjunction with the overall propagation risk level. This includes: starting from the initial node, traversing step by step to the end node of the public opinion propagation graph according to the information flow direction indicated by the directed propagation edges between nodes in the public opinion propagation graph, generating multiple candidate propagation paths from the initial node to the end node; filtering the multiple candidate propagation paths, retaining candidate propagation paths that pass through at least one key propagation node or the first cross-platform node, to obtain a set of filtered propagation paths; when the number of paths in the set of filtered propagation paths exceeds a preset value... When the maximum number of paths is reached, the sum of the influence values ​​of all nodes traversed on each filtered propagation path is calculated as the path influence score. Paths are sorted from highest to lowest influence score, and a predetermined number of propagation paths at the top of the list are retained to obtain the target propagation path set. Key propagation nodes and first-time cross-platform nodes are marked on each propagation path in the target propagation path set to generate a structured propagation path for the target public opinion event. The response level of the target public opinion event is determined based on the overall propagation risk level. The nodes are sorted from highest to lowest influence value according to the initial node, first-time cross-platform node, and each key propagation node to generate a node handling order. The response level and handling order are combined to generate a handling priority prompt.

[0058] In the above embodiments, candidate propagation paths are generated starting from the initial node. After screening and calculating the path influence score, a set of target propagation paths is obtained. After marking key propagation nodes and the first cross-platform node, the structured propagation path of the target public opinion event can be clearly presented. The response level is determined according to the overall propagation risk level, and the disposal order is generated by combining the node influence value. Then, the disposal priority prompts are generated, which helps enterprises and institutions to accurately grasp the public opinion propagation path, determine the disposal priority according to the risk level and the importance of the nodes, and improve the efficiency and pertinence of public opinion disposal.

[0059] Starting from the initial node, traversing the entire propagation graph along directed edges generates multiple propagation paths from the source to all end nodes. However, not all paths have the same analytical value; some paths do not pass through any key nodes, indicating they are merely "branch propagation" and not the main risk propagation chain. Therefore, this method first retains paths that pass through at least one key propagation node or the first cross-platform node, as these are the truly noteworthy "main lines of risk propagation." When the retained paths are still too numerous (exceeding the preset path count limit), this method calculates the sum of the influence values ​​of all nodes on each path as the path's "influence score," retaining the highest-scoring paths to ensure the output propagation paths are both concise and representative. After obtaining the target propagation path set, key propagation nodes and the first cross-platform node are highlighted on each path using annotations, forming a structured visual output. Regarding handling priorities, this method determines the response level based on the overall risk level and the handling order based on the influence value of each key node, combining the two into a complete handling priority prompt. Related technologies typically only output a simple "propagation chain" or "propagation tree," resulting in redundant information and a lack of focus. Regarding decision-making, they only indicate "the involvement of a certain influential figure," without providing actionable solutions regarding "the appropriate level and order of action." This method addresses the technical issues of redundant output information and insufficient decision support by combining path filtering with influence scoring and ranking, and by combining response level with action order. The action priority prompt includes two dimensions: response level and action order. The response level, determined based on the overall propagation risk level, indicates the level of action and resource allocation for the target public opinion event, including but not limited to response time limits, decision-making levels, and the scope of available resources. The action order, sorted from highest to lowest influence value of key propagation nodes, indicates the order in which key propagation nodes should be addressed. The system outputs a combination of response level and action order, enabling users to allocate resources based on the response level and execute actions on key propagation nodes sequentially according to the action order, achieving precise matching of resources and risk levels.

[0060] In an optional embodiment, after generating the public opinion propagation graph, the method further includes: determining the propagation stage of the public opinion propagation graph to identify the current propagation stage of the target public opinion event; wherein, the propagation stage determination includes: dividing the evolution process of the public opinion propagation graph into multiple consecutive time slices based on the publication timestamps of each node in the public opinion propagation graph; for each time slice, counting the number of newly added nodes, the number of newly added directed propagation edges, and the average weight value of the newly added directed propagation edges to obtain the propagation activity index corresponding to the time slice; wherein the weight value of each newly added directed propagation edge is calculated based on the cosine similarity of the text vectors of the content published by the two associated nodes and the publication time difference; comparing the propagation activity index with a preset propagation stage threshold range to determine the propagation stage label corresponding to each time slice, the propagation stage label including at least one of incubation period, outbreak period, diffusion period, and decline period; determining the propagation stage label corresponding to the latest time slice as the current propagation stage of the target public opinion event; when outputting the structured propagation path of the target public opinion event, the current propagation stage is output as additional information of the structured propagation path.

[0061] In the above embodiments, determining the propagation stage of the public opinion propagation map can identify the current propagation stage of the target public opinion event. By dividing the time slices using the content release timestamps and counting the number of new nodes, the number of new directed propagation edges, and the average weight value, a propagation activity index can be derived, which helps to accurately grasp the dynamic changes in public opinion propagation. By comparing the propagation activity index with the preset propagation stage threshold range to determine the propagation stage label, it is possible to clearly understand whether the public opinion is in the incubation period, outbreak period, diffusion period, or decline period. Using the current propagation stage as additional information output of the structured propagation path can provide a more comprehensive reference for public opinion management, enabling decision-makers to formulate more targeted strategies according to different propagation stages.

[0062] Public opinion events are not static but undergo a complete life cycle from emergence, outbreak, diffusion, to decline. Different stages have different propagation characteristics (e.g., few new nodes during the incubation period, a surge in new nodes during the outbreak period, a steady increase in new nodes during the diffusion period, and a sharp decrease in new nodes during the decline period). Furthermore, the propagation risks and response strategies differ significantly at each stage. Therefore, accurately determining the current propagation stage of an event is crucial for risk assessment and decision-making. Specifically, this method divides the evolution of the public opinion propagation graph into multiple consecutive time slices (e.g., one slice every 30 minutes) based on the content publication timestamps of each node. For each time slice, the number of new nodes, the number of new directed propagation edges, and the average weight of the new edges are counted—these three indicators together constitute the "propagation activity index," reflecting the activity level of public opinion within that time period. Then, the propagation activity index is compared with a preset stage threshold range to determine the propagation stage label (incubation period, outbreak period, diffusion period, decline period) corresponding to each time slice. The stage label of the latest time slice is the current propagation stage. Finally, the current propagation stage is presented as additional information when outputting the structured propagation path. The average weight of newly added edges within a certain time slice is the arithmetic mean of the weights of all newly added edges within that time slice. The calculation method for the weight value is the comprehensive weight formula in the aforementioned embodiment. Assume that the number of newly added nodes ΔN=15, the number of newly added propagation edges ΔE=18, and the average weight of the newly added propagation edges Wav=0.632 in a certain time slice. Let the weighting coefficients be α1=0.4, α2=0.35, and α3=0.25. Given that the maximum number of newly added nodes in all slices is Nmax=60 and the maximum number of newly added edges is Emax=55, then the propagation activity index A=0.4×(15 / 60)+0.35×(18 / 55)+0.25×0.632=0.372. If the preset propagation stage threshold ranges are: incubation period [0,0.15), outbreak period [0.15,0.35), diffusion period [0.35,0.65], and decline period [0.65,1.00], then A=0.372 falls within the diffusion period range, indicating that the slice is in the diffusion period. Public opinion analysis of related technologies is mostly static, failing to perceive the dynamic evolution stage of an event, resulting in a lack of timeliness and specificity in risk assessment and response decisions. This embodiment uses time slicing and quantification of dissemination activity indicators to dynamically track the evolution of public opinion. It achieves automatic and objective determination of dissemination stages through stage threshold comparison, avoiding the subjectivity of manual judgment. The output of the current dissemination stage allows users to adjust their response strategies and resource allocation according to the actual stage of the event, achieving "phased and differentiated" risk management. It also enables automatic and accurate determination of dissemination stages, providing a basis for dynamic risk assessment and the formulation of phased response strategies.

[0063] In an optional embodiment, the method further includes a propagation trend prediction step: obtaining the topological evolution sequence of the public opinion propagation graph within multiple historical time windows, the topological evolution sequence including the number of nodes, the number of directed propagation edges, and the out-degree change data of each node corresponding to each time window; inputting the topological evolution sequence into a preset time series prediction model, the time series prediction model predicts the number of new nodes and the number of new directed propagation edges in the public opinion propagation graph in the next time window based on the topological evolution sequence; identifying candidate nodes that may become key propagation nodes in the next time window based on the number of new nodes and the number of new directed propagation edges, combined with the influence value of each node in the current public opinion propagation graph; outputting the prediction results of the candidate nodes, and using the prediction results as supplementary information for the priority indication of handling.

[0064] In the above embodiments, the topological evolution sequence of the public opinion propagation map within multiple historical time windows is obtained. The number of new nodes and directed propagation edges in the next time window is predicted using a time series prediction model. The candidate nodes that may become key propagation nodes are identified by combining the influence value of the current node. This can predict the public opinion propagation trend in advance, output the candidate node prediction results and use them as supplementary information for handling priority prompts. This helps to more comprehensively and dynamically evaluate public opinion events and provide more sufficient basis for public opinion handling.

[0065] This embodiment provides a method for predicting the spread of public opinion. Public opinion spread is not entirely random or unpredictable; its diffusion process is profoundly influenced by the structure of the spread network and the influence of nodes, exhibiting certain evolutionary patterns. By analyzing the topological evolution sequence of the spread graph within multiple historical time windows (including changes in the number of nodes, edges, and out-degree changes of each node), the spread trend pattern of public opinion can be extracted. After these patterns are input into a time-series prediction model, the model can learn the mapping relationship from historical evolution to future trends, thereby predicting the number of new nodes and edges in the next time window. More importantly, simply predicting "how many new nodes" is insufficient; it is also necessary to determine "who will trigger the new nodes." The time-series prediction model can employ a Long Short-Term Memory (LSTM) network or its variants, using the node increment sequence, edge increment sequence, and node out-degree change sequence within historical time windows as input features, and outputting the predicted value for the next time window through an encoder-decoder structure. During the model training phase, the evolution sequence of the spread graph of historical public opinion events is used as training samples, and the actual observed data for the next window is used as labels. Mean squared error is used as the loss function for parameter optimization. This method combines the predicted number of new nodes / edges with the current influence value of each node to identify candidate nodes that are likely to become key propagation nodes in the next time window. Specifically, nodes with already high current influence and whose out-degree change rate has been continuously increasing in recent windows are most likely to trigger a large number of new propagations in the next time window. The prediction results of these candidate nodes are output as supplementary information to indicate the priority of handling, enabling users to prepare response resources in advance.

[0066] Example: The system has completed the construction of a propagation graph, containing 120 nodes and 188 directed edges. The current time is 12:00 noon, and the event has lasted for 2 hours. The system has obtained the topological evolution sequence of a certain public opinion event from its initial occurrence at 10:00 to the current time of 12:00. In the first window, from 10:00 to 10:30, the graph contains only 8 nodes and 9 directed edges; this is the initial stage of the event, and propagation has not yet expanded on a large scale. In the second window, from 10:30 to 11:00, the total number of nodes in the graph increases to 25, with 17 new nodes added; the total number of directed edges increases to 32, with 23 new directed edges added. In the third window, from 11:00 to 11:30, the graph enters a rapid expansion phase, with the total number of nodes surging to 68, 43 new nodes added; the total number of directed edges increases to 97, with 65 new directed edges added. During the fourth window from 11:30 to 12:00, the graph continued its rapid expansion, reaching a total of 120 nodes (52 new nodes) and a total of 188 directed edges (91 new directed edges). In addition to the incremental data for nodes and edges, the system also recorded the out-degree changes of each node within each window. Specifically, node V2 had an out-degree of 0 in the first window, increased to 3 in the second, rose to 5 in the third, and further increased to 6 in the fourth, showing a continuous upward trend. Node V3 had an out-degree of 0 in the first window, increased to 1 in the second, rose to 2 in the third, and then surged to 8 in the fourth, showing an accelerated upward trend in the later stages. Node V4 had an out-degree of 0 in the first and second windows, increased to 0 in the third, increased to 1 in the fourth, and then increased to 2 in the fourth, with a relatively gradual increase. Node V1 had an out-degree of 2 in the first window, and its out-degree remained at 2 in subsequent windows, without further increase. The system inputs the topological evolution sequences of these four windows into a pre-set time-series prediction model. This model employs a Long Short-Term Memory (LSTM) network structure and has been pre-trained on a large amount of historical public opinion event propagation and evolution data. The model's input features include the sequence of newly added nodes, the sequence of newly added directed propagation edges, and the sequence of out-degree changes of each node in the graph within these four windows. The model's task is to predict the number of newly added nodes and the number of newly added directed propagation edges in the next window (12:00 to 12:30) based on the evolution patterns of these four windows. After calculation, the model outputs the prediction results: during the next window (12:00 to 12:30), the expected number of newly added nodes is 35, and the expected number of newly added directed propagation edges is 48. After obtaining the predicted number of newly added nodes and the number of newly added directed propagation edges, the system begins to identify candidate nodes that may become key propagation nodes in the next window. The system traverses all 120 nodes in the current graph and filters each node.The selection criteria are based on two conditions: first, the node's current influence value must be greater than 0.60; second, the node's out-degree change rate in the most recent two historical windows must show a continuously increasing trend. A node is selected as a candidate node because it simultaneously possesses the characteristics of "currently having a certain influence" and "becoming increasingly active," meaning it is most likely to trigger a large amount of new propagation in the next window. After system traversal, it was found that node V2's current influence value is 0.739, greater than 0.60, and its out-degree increased from 5 to 6 during the third to fourth window, continuing its upward trend. Therefore, this node was identified as a candidate key node, with a high prediction confidence level. Node V3's current influence value is 0.733, greater than 0.60, and its out-degree increased sharply from 2 to 8 during the third to fourth window, showing an accelerating upward trend. Therefore, this node was also identified as a candidate key node, with a medium prediction confidence level. Node V1 has an influence value of 0.35, below the threshold of 0.60. Therefore, although it is the initial node, the prediction model believes it is unlikely to trigger a large amount of new propagation in the next window and is not identified as a candidate node. Node V4 has an influence value of 0.25, also below the threshold, and is not identified as a candidate node. The system ultimately outputs the candidate node prediction results: approximately 35 new nodes and approximately 48 new directed propagation edges are predicted during the next window from 12:00 to 12:30. There are two candidate accounts that may become key propagation nodes: node V2 (high confidence) and node V3 (medium confidence). The system recommends that node V2 be the highest priority prediction target, and suggests that the operations team prepare a response plan for this account in advance, including but not limited to pre-communication, monitoring its subsequent content, and preparing clarification statements. In related technologies, public opinion analysis only traces and analyzes after the public opinion has occurred, and cannot predict the possible future propagation trend, resulting in the handling work always being in a "passive response" state. By the time high-risk nodes are discovered, the public opinion has already spread on a large scale. This method introduces time-series prediction based on graph topology evolution, enabling early prediction of key nodes and providing a technical possibility for proactive public opinion risk prevention and control.

[0067] In an optional embodiment, after obtaining the number of new nodes predicted by the time-series prediction model in the next time window, the method further includes: calculating a propagation trend factor based on the predicted number of new nodes in the next time window and the actual number of new nodes in the current time window; wherein, the propagation trend factor = (predicted number of new nodes) / (actual number of new nodes + σ), where σ is a very small positive value; when the propagation trend factor is greater than 1, multiplying the overall risk score corresponding to the overall propagation risk level by the propagation trend factor to obtain a corrected risk score; when the propagation trend factor is less than or equal to 1, keeping the overall risk score unchanged; redetermining the overall propagation risk level of the target public opinion event based on the preset risk score range in which the corrected risk score is located, and updating the handling priority prompt based on the redetermined overall propagation risk level.

[0068] In the above embodiments, the propagation trend factor is calculated based on the predicted number of new nodes in the next time window and the actual number of new nodes in the current time window, which can dynamically reflect the trend of public opinion propagation. The overall risk score is corrected based on the propagation trend factor, which can more accurately redetermine the overall propagation risk level of the target public opinion event, ensuring the timeliness and accuracy of risk assessment. The updated handling priority prompts based on the redetermined overall propagation risk level can enable the handling strategy to better adapt to the real-time development of public opinion and adjust the response measures in a timely manner.

[0069] If the number of new nodes in the next time window can be predicted, comparing this predicted value with the actual value in the current window yields a quantitative indicator measuring the "acceleration of propagation"—the propagation trend factor. A propagation trend factor greater than 1 indicates that public opinion is spreading rapidly; a propagation trend factor less than or equal to 1 indicates that public opinion is spreading slowly or developing steadily. This propagation trend factor can serve as a "correction multiplier" for risk scoring. If the propagation trend is predicted to accelerate, even if the current risk score is low, the actual risk is likely to rise in the short term; conversely, if the propagation trend is predicted to slow down, the current high-risk score may not be sustainable. Specifically, after obtaining the predicted number of new nodes, this method calculates the propagation trend factor τ = predicted number of new nodes / (actual number of new nodes + σ). The denominator is added with σ to prevent division by zero errors. σ is a very small positive value, for example, σ less than 10. -4(or other values). When τ>1, the overall risk score is multiplied by τ to obtain the corrected risk score; when τ≤1, the risk score remains unchanged. Then, the risk level is re-determined based on the corrected risk score, and the handling priority prompt is updated. For example, if the actual number of newly added nodes in the current time window (11:30-12:00) is 52, while the model predicts that 70 new nodes will be added in the next window, then τ=70 / (52+σ)≈1.35>1. Assuming the original risk score is 0.5834, the corrected score is 0.5834×1.35≈0.787, jumping from the high-risk zone (0.50-0.75) to the major risk zone (0.75-1.00). The system upgrades the risk level to "major risk" and updates the handling priority prompt. The system upgrades the risk level to "major risk" and updates the handling priority. The risk assessment in the aforementioned embodiment is based on current and historical data, reflecting the current spread and failing to reflect the future trend of public opinion. If public opinion is spreading rapidly, even if the current risk score is low, it may quickly escalate to high risk, but related technologies cannot capture this "acceleration" signal. This embodiment uses a propagation trend factor to proactively correct the risk score, enabling the risk assessment to reflect not only the current state but also future trends, achieving an upgrade from "static risk assessment" to "dynamic trend perception."

[0070] In an optional embodiment, the weight coefficients of the influencing factors are dynamically determined based on the current propagation stage of the target public opinion event, which includes the incubation period, outbreak period, diffusion period, and decline period. A set of preset weight coefficients is configured for each of the incubation period, outbreak period, diffusion period, and decline period. Specifically: in the preset weight coefficient set corresponding to the incubation period, the weight coefficients of media rating and negative sentiment amplification are greater than the weight coefficients of other influencing factors; in the preset weight coefficient set corresponding to the outbreak period, the weight coefficients of forwarding contribution and historical interaction volume are greater than the weight coefficients of other influencing factors; in the preset weight coefficient set corresponding to the diffusion period, the weight coefficient of account follower count is greater than the weight coefficients of other influencing factors; and in the preset weight coefficient set corresponding to the decline period, the weight coefficient of negative sentiment amplification is greater than the weight coefficients of other influencing factors.

[0071] In the above embodiments, the weight coefficients of influencing factors are dynamically determined according to the different propagation stages of the target public opinion event (latent period, outbreak period, diffusion period, and decline period). During the latent period, the influence of media level and the amplification of negative emotions is highlighted, which can pay more attention to media and negative emotional factors in the initial stage of public opinion. During the outbreak period, the role of forwarding contribution and historical interaction volume is emphasized, which can grasp the key to the rapid spread of public opinion. During the diffusion period, the weight of account followers is increased, which is conducive to grasping the elements of the widespread spread of public opinion. During the decline period, the amplification of negative emotions is emphasized, which helps to deal with the negative situation in the final stage of public opinion, thereby more accurately calculating the influence value of each node and improving the accuracy and reliability of public opinion analysis.

[0072] This embodiment introduces a dynamic weight adjustment mechanism, which optimizes the method of calculating influence value. At different stages of public opinion dissemination, the contribution of each influencing factor to the "importance of the node" varies; fixed weights can lead to inaccurate assessments at certain stages. During the incubation period, the source of public opinion often comes from ordinary individual users or small self-media outlets. At this time, media level and the amplification of negative emotions are key to judging "potential risk sources." A high media level means that even if the current dissemination volume is not large, its subsequent diffusion ability is stronger. A high amplification of negative emotions means that the content is inflammatory and more likely to trigger a subsequent outbreak. During the outbreak period, the rapid expansion of the dissemination network is mainly driven by forwarding behavior. Forwarding contribution and historical interaction volume become core indicators. Forwarding contribution directly measures the "hub" role of the node in the dissemination network, while historical interaction volume reflects the "eye-catching ability" of the account's content. During the diffusion period, multi-party participation becomes the main characteristic, and the number of account followers becomes a key factor determining the breadth of dissemination. Participants during the diffusion period include a large number of ordinary users, and accounts with a large number of followers are more likely to push the event to the next level of engagement at this stage. During the decline phase, the "residual risk" of public opinion mainly lies in the continued fermentation and repeated spread of negative content, making the amplification of negative emotions the only factor requiring focused attention. This method assigns a specific set of weight coefficients to each stage of dissemination, allowing the calculation of influence values ​​to adapt to different stages of public opinion evolution. For example, the weight configurations of the five major influencing factors (account follower count, historical interaction volume, media level, repost contribution, and amplification of negative emotions) at different dissemination stages are as follows: Latent phase: [0.1, 0.1, 0.4, 0.1, 0.3]; Outbreak phase: [0.1, 0.3, 0.1, 0.4, 0.1]; Diffusion phase: [0.5, 0.1, 0.2, 0.1, 0.1]; Decline phase: [0.1, 0.1, 0.2, 0.1, 0.5]; This is just one example. Fixed-weight schemes in related technologies ignore the differences in the dissemination stages, inevitably leading to inaccurate assessments at certain stages. For example, during the incubation period, an account with a small number of followers but high media rating and highly inflammatory content may be underestimated; similarly, during the explosive growth period, a "key promoter" with a small number of followers but a large contribution to reposting may also be underestimated. This method, through stage-adaptive dynamic weight configuration, makes the calculation of influence values ​​more closely match the dissemination characteristics of each stage, improving the accuracy of node influence assessment and providing a foundation for more accurate identification of key nodes.

[0073] In an optional embodiment, after identifying nodes that increase propagation speed, nodes that amplify media, and nodes that concentrate negative emotions, the method further includes: determining whether there are overlapping nodes among the key propagation nodes that simultaneously satisfy at least two classification rules; if there are overlapping nodes, upgrading the overlapping nodes to core high-risk nodes; when generating a handling priority prompt, assigning the core high-risk nodes the highest level of handling priority, and highlighting the core high-risk nodes in a differentiated manner in the structured propagation path.

[0074] In the above embodiments, overlapping nodes in key propagation nodes can be identified and upgraded and marked as core high-risk nodes, and core high-risk nodes are given the highest level of handling priority, which can ensure that the most risky nodes are dealt with first and improve the efficiency of public opinion handling. Core high-risk nodes are highlighted in a differentiated manner in the structured propagation path, which makes it easy to intuitively locate and pay attention to high-risk nodes, and provides clear guidance for public opinion analysis and handling.

[0075] When a node simultaneously satisfies multiple key classification rules (e.g., it is simultaneously a node that accelerates dissemination, a node that amplifies media impact, and / or a node that concentrates negative emotions), it indicates that the node exhibits "abnormal" or "critical" characteristics across multiple dimensions. It is not only an "accelerator" of dissemination but also an "amplifier" of negative emotions and a "megaphone" of high influence. Such multi-dimensionally overlapping nodes pose a far higher risk than nodes that only satisfy a single rule and should be given greater attention. Specifically, after identifying three types of key nodes, this method determines whether there are overlapping nodes that simultaneously satisfy at least two classification rules. If so, these overlapping nodes are upgraded and marked as "core high-risk nodes." When generating priority handling prompts, core high-risk nodes receive the highest level of handling priority (even if their influence value is not the highest, they should still be handled first) and are differentiated and highlighted in the structured dissemination path (e.g., red borders, flashing markers, special icons, etc.) so that users can easily identify these most dangerous nodes. While existing technologies identify multiple critical nodes, they treat all nodes equally, leaving users unable to distinguish which are the "most dangerous," leading to the dilemma of making secondary judgments during decision-making. This method, by upgrading the marking of overlapping nodes, automatically distinguishes between "ordinary critical nodes" and "core high-risk nodes," providing users with more accurate risk signals.

[0076] The following description, in conjunction with specific embodiments, illustrates this method and system for identifying public opinion dissemination paths and locating key nodes. Figure 2 The system architecture diagram for identifying public opinion dissemination paths and locating key nodes provided in the application embodiment includes a public opinion content and dissemination data acquisition unit, a dissemination node construction unit, a dissemination relationship construction unit, a node influence calculation unit, a key dissemination node identification unit, a dissemination risk level calculation unit, and a dissemination path and handling priority output unit. The method includes the following steps.

[0077] Step 1: Construct propagation nodes The system uses the source accounts, media accounts, comment accounts, forwarding accounts, and publishing platforms of each piece of content under the same public opinion event as nodes in the dissemination network, and records the number of followers, historical interaction volume, media level, platform, and first appearance time of each node.

[0078] Step Two: Establishing Propagation Relationships The system establishes directed propagation edges between nodes based on the following relationship: Forwarding relationship: There are records of forwarding, quoting or reprinting, with the direction from the forwarded party to the forwarding party; Content similarity: The text semantic vector or multimedia feature similarity of the content published by the two nodes is higher than the preset value; Time proximity: The time difference between the publication of content on two nodes falls within a preset time window; Interactive relationships: There are interactions such as comments and likes; The propagation edge is weighted, and the weight is determined by the content similarity and time proximity. The higher the similarity and the closer the time, the greater the weight.

[0079] Step 3: Calculate node influence The system calculates the influence Inf(v) of each node v using the following formula: Inf(v)=γ1·F(v)+γ2·E(v)+γ3·M(v)+γ4·R(v)+γ5·N(v); The factors in the formula are defined as follows: F(v) is the normalized value of the number of followers of the account: the number of followers of the node account is logarithmically transformed and normalized to [0,1].

[0080] E(v) is the normalized value of content interaction: the total number of comments, likes and reposts of node content is logarithmically transformed and normalized.

[0081] M(v) is the media level normalization value: it is mapped to [0,1] according to the media level to which the node belongs, with mainstream media having a higher value than self-media.

[0082] R(v) represents the forwarding contribution: the proportion of the number of nodes reachable downstream in the propagation network from this node to the total number of nodes in the network.

[0083] N(v) represents the amplification of negative emotions: the proportion of negative emotional content in the content of this node and its direct downstream content.

[0084] γ1 to γ5 are preset coefficients and their sum is 1. In one implementation, the values ​​are γ1=0.2, γ2=0.2, γ3=0.2, γ4=0.25, and γ5=0.15.

[0085] Step 4: Identify key propagation nodes The system identifies key nodes based on their influence, propagation level, propagation speed, and cross-platform diffusion. Specifically: the node that first appears is identified as the initial node; nodes whose downstream nodes first connect to nodes on different platforms are identified as the first cross-platform nodes; nodes with a significant increase in the number of new downstream nodes per unit time are identified as nodes with increased propagation speed; nodes with influence exceeding a preset value are identified as high-interaction nodes or media amplification nodes; and nodes with a negative sentiment amplification weight exceeding a preset value are identified as nodes with concentrated negative sentiment.

[0086] Step 5: Calculate the transmission risk level The system calculates the risk value (Risk) of a public opinion event's spread using the following formula: Risk=δ1·V+δ2·P+δ3·Eneg+δ4·H+δ5·Mcnt+δ6·C In the formula: V represents the propagation speed, which is the number of new nodes per unit time truncated to an upper limit and then normalized; P represents the normalized value of the number of platforms involved; Eneg represents the proportion of negative sentiment content; H represents the normalized value of the number of high-influence nodes; Mcnt represents the normalized value of the number of media reposts; C represents the concentration of complaints, which is the proportion of complaint-related content; δ1 to δ6 are preset coefficients, and their sum is 1. Based on the preset range in which Risk falls, the system classifies public opinion events into four levels: general concern, key concern, high risk, and major risk.

[0087] Step Six: Output Propagation Path and Handling Priority The system outputs the propagation path of the public opinion event, including the initial source, main diffusion path, key propagation nodes, cross-platform propagation nodes, current propagation stage, propagation risk level, and handling priority. The handling priority is determined based on the propagation risk level and the influence of key nodes.

[0088] Figure 3 This is an example diagram of public opinion network propagation and key nodes provided in this application embodiment, showing the complete public opinion propagation topology link, which facilitates users to quickly locate key roles in the propagation network. The link starts from the initial account and flows to two paths to forwarding nodes A and B respectively; forwarding A connects to cross-platform nodes (green fill, representing the first cross-platform node of the event); forwarding B connects to high-influence media (red border, representing the media amplification node); the cross-platform nodes and high-influence media extend outward to comment aggregation and secondary diffusion nodes respectively.

[0089] Specific Implementation Example: Identification and Risk Assessment of the Propagation Path of a Negative Public Opinion Scenario: A bank's credit card business is accused of improper debt collection practices. The relevant content was first published on a social media account, and after being forwarded multiple times, it was reported by a high-influence media outlet and spread across platforms to short video platforms.

[0090] Steps 1 and 2: The system uses 120 accounts and platforms under this event as nodes, and establishes directed weighted propagation edges based on forwarding records, content similarity (semantic similarity greater than 0.8), time proximity (within a 24-hour window), and account association to form a propagation network.

[0091] Step 3: Taking the high-influence media node v as an example, its fan count is log-normalized as F(v)=0.9, its interaction volume is normalized as E(v)=0.85, its media level weight is M(v)=0.95, its downstream reachable node ratio is R(v)=0.6, and its negative sentiment amplification weight is N(v)=0.8. Using γ1=0.2, γ2=0.2, γ3=0.2, γ4=0.25, and γ5=0.15, we calculate Inf(v)=0.2×0.9+0.2×0.85+0.2×0.95+0.25×0.6+0.15×0.8=0.81, thus identifying it as a media amplification node.

[0092] Step 4: The system identifies the initial social media account as the initial node, the forwarding account connected to the short video platform as the first cross-platform node, and the above media nodes as media amplification nodes.

[0093] Step 5: This event saw a large number of new nodes added per unit time, with a normalized propagation speed of V=0.8, involving 3 platforms (P=0.6), a negative sentiment percentage (Eneg=0.75), 2 high-influence nodes (H=0.5), 1 media repost (Mcnt=0.4), and a complaint concentration (C=0.5). Using δ1=0.25, δ2=0.15, δ3=0.2, δ4=0.15, δ5=0.15, and δ6=0.1, the Risk is calculated as 0.25×0.8+0.15×0.6+0.2×0.75+0.15×0.5+0.15×0.4+0.1×0.5=0.625, falling into the high-risk range.

[0094] Step Six: The system outputs the propagation path: initial social media account → multiple forwarding accounts → cross-platform nodes → high-influence media → comment aggregation. Key nodes and high-risk levels are marked, and the handling priority is determined to be high. It is recommended to initiate special handling.

[0095] Parameter configurability notes: In this embodiment, the weight coefficients γ1–γ5, δ1–δ6, risk level interval threshold, similarity threshold, and time window are all configurable parameters. They can be adjusted by operators according to business characteristics or determined based on historical event regression, and do not constitute a limitation on the scope of protection of this invention.

[0096] This application also provides a public opinion analysis system based on propagation graphs, used to execute the public opinion analysis method based on propagation graphs in any of the foregoing embodiments, such as... Figure 4 As shown, the system includes: The acquisition module is used to acquire multi-source public opinion data associated with the target public opinion event, extract the accounts involved in the dissemination as nodes in the dissemination network, and extract basic attribute data of each node, including the release timestamp, platform, number of followers, historical interaction volume, and media level. The generation module is used to extract content features from multi-source public opinion data, and construct directed propagation edges based on the interaction behavior records between nodes, content similarity, and publication timestamps to generate a public opinion propagation graph. The calculation module is used to extract the network topology features of each node in the public opinion dissemination graph, including out-degree, downstream connected node scale and forwarding contribution, as well as the sentiment features including the emotional polarity of the published content and the degree of negative sentiment amplification. The number of account followers, historical interaction volume, media level, forwarding contribution and negative sentiment amplification are used as influencing factors to perform feature normalization and weighted calculation to obtain the influence value of each node. The identification module is used to traverse the public opinion dissemination map, identify the node with the earliest release timestamp as the initial node, and identify the node whose platform has changed and whose direct downstream platform has changed, and whose release time is the earliest as the first cross-platform node; and based on the influence value, negative sentiment amplification degree and out-degree change rate of each node within a preset time window, identify key dissemination nodes according to preset rules. The assessment module is used to extract global propagation features from the public opinion propagation map. The global propagation features include at least: propagation speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. Among them, propagation speed is the number of new nodes added per unit time, and high-influence nodes are key propagation nodes with an influence value greater than a preset influence threshold. Based on the preset risk assessment logic, the global propagation features are quantitatively calculated to obtain the overall propagation risk level of the target public opinion event. The output module is used to output the structured propagation path of the target public opinion event based on the initial node, the first cross-platform node, and key propagation nodes, and to generate a handling priority prompt in combination with the overall propagation risk level.

[0097] It should be noted that the devices or systems provided in the above embodiments are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept. Other device or system embodiments correspond to the aforementioned method embodiments. Other technical features are described in the previous embodiments and will not be repeated here.

[0098] This application also provides a computer-readable storage medium storing instructions that, when executed, perform the steps of any of the methods described above.

[0099] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0100] This application also discloses an electronic device. For example... Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0101] The communication bus 502 is used to enable communication between these components.

[0102] The user interface 503 may include a display screen and a camera. Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0103] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0104] The processor 501 may include one or more processing cores. The processor 501 connects to various parts of the electronic device (such as a server) using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 505, and by calling data stored in memory 505. Optionally, the processor 501 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 501 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 501 and may be implemented as a separate chip.

[0105] The memory 505 may include random access memory (RAM) or read-only memory. Optionally, the memory 505 may include a non-transitory computer-readable storage medium. The memory 505 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 505 may also be at least one storage device located remotely from the aforementioned processor 501. (Refer to...) Figure 5 The memory 505, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program based on a propagation graph-based public opinion analysis method.

[0106] exist Figure 5In the illustrated electronic device 500, the user interface 503 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 501 can be used to call an application program of a public opinion analysis method based on propagation graph stored in the memory 505. When executed by one or more processors 501, the electronic device 500 performs one or more of the methods described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0107] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0108] In the various embodiments provided in this application, it should be understood that the disclosed apparatus or system can be implemented in other ways. For example, the apparatus or system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0109] The above description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will be readily apparent to those skilled in the art upon consideration of the disclosure herein.

[0110] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art that are not described in this disclosure.

Claims

1. A public opinion analysis method based on propagation graphs, characterized in that, include: Obtain multi-source public opinion data related to the target public opinion event, extract the accounts involved in the dissemination as nodes in the dissemination network, and extract basic attribute data of each node, including the release timestamp, platform, number of followers, historical interaction volume, and media level; Extract the content features of the multi-source public opinion data, and construct directed propagation edges based on the interaction behavior records between nodes, content similarity, and the publication timestamps to generate a public opinion propagation graph; Extract the network topology features of each node in the public opinion propagation map, including out-degree, downstream connected node scale and forwarding contribution, as well as the sentiment features including the sentiment polarity of the published content and the amplification of negative emotions. The influence value of each node is obtained by using the number of followers of the account, the amount of historical interaction, the media level, the repost contribution, and the negative sentiment amplification as influencing factors, and performing feature normalization and weighted calculation. Traverse the aforementioned public opinion dissemination graph, identify the node with the earliest publication timestamp as the initial node, and identify the node whose platform and the platform of its direct downstream node have changed, and whose publication time is the earliest, as the first cross-platform node. Based on the influence value of each node, the amplification of negative emotions, and the out-degree change rate of the node within a preset time window, key propagation nodes are identified according to preset rules. Extract the global propagation features of the public opinion propagation map. The global propagation features include at least: propagation speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. The propagation speed is the number of new nodes added per unit time, and the high-influence nodes are key propagation nodes with an influence value greater than a preset influence threshold. Quantify the global propagation features based on a preset risk assessment logic to obtain the overall propagation risk level of the target public opinion event. Based on the initial node, the first cross-platform node, and the key propagation node, the structured propagation path of the target public opinion event is output, and a handling priority prompt is generated in combination with the overall propagation risk level.

2. The method according to claim 1, characterized in that, Based on the interaction records between nodes, content similarity, and the publication timestamp, a directed propagation edge is constructed to generate a public opinion propagation graph, including: The multi-source public opinion data is analyzed to extract explicit interaction behavior records between nodes. These explicit interaction behavior records include forwarding, citation, or reprinting behaviors. Based on the explicit interaction behavior records and the order of the publication timestamps, explicit directed propagation edges are constructed between the corresponding nodes. For target nodes that do not have explicit interaction behavior records, a pre-trained semantic feature extraction model is used to convert the content published by the target node into a text vector, and at least one candidate upstream node is selected whose publication time is within a preset source tracing time window and whose cosine similarity with the text vector of the target node is greater than a preset similarity threshold. Based on the cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node. Select the candidate upstream node corresponding to the maximum value of the association weight, and construct a unique implicit directed propagation edge between the candidate upstream node corresponding to the maximum value and the target node; The explicit directed propagation edges and the implicit directed propagation edges are merged to generate the public opinion propagation map.

3. The method according to claim 2, characterized in that, Based on the cosine similarity and the time difference between the target node and each candidate upstream node, a time decay effect is introduced to calculate the association weight between the target node and each candidate upstream node, including: The association weight between the candidate upstream node and the target node is calculated using the following comprehensive weighting formula: W （u,v） =sim(u,v)·e -λΔt ; Among them, W （u,v） λ is the association weight between candidate upstream node u and target node v; sim(u,v) is the cosine similarity between the text vectors of the content published by candidate upstream node u and target node v; Δt is the time difference between the publication timestamps of target node v and candidate upstream node u; λ is a preset time decay coefficient; the time decay coefficient λ is dynamically assigned according to the platform characteristics of the target public opinion event or the propagation decay law of similar public opinion events in history.

4. The method according to claim 1, characterized in that, The influence values ​​of each node are obtained by using the account's follower count, historical interaction volume, media level, repost contribution, and negative sentiment amplification as influencing factors, performing feature normalization and weighted calculation, including: Logarithmic normalization is performed on the number of followers of the account and the number of historical interactions to obtain the normalized value of the number of followers of the account and the normalized value of the number of historical interactions. The media level is mapped to [0,1] according to a preset level-score mapping table to obtain the media level normalized value; Calculate the influence value of each node using the following formula: Influence value = W1 × Normalized value of account followers + W2 × Normalized value of historical interaction volume + W3 × Normalized value of media level + W4 × Contribution of reposts + W5 × Amplification of negative emotions; Among them, W1, W2, W3, W4 and W5 are the weight coefficients of each influencing factor, and their sum is 1.

5. The method according to claim 1, characterized in that, The key propagation nodes are identified according to preset rules based on the influence value of each node, the amplification of negative emotions, and the rate of change of the node's out-degree within a preset time window, including: For any node in the public opinion propagation map, obtain the out-degree change rate of the node within a preset time window; If the out-degree change rate of the node within a preset time window is greater than a preset first threshold, then the node is identified as a node with improved propagation speed. If the influence value of the node is greater than the preset second threshold, the node will be identified as a media amplification node. If the amplification of negative emotions at a node exceeds a preset third threshold, the node will be identified as a node with concentrated negative emotions. The nodes that increase the speed of propagation, the nodes that amplify media, and the nodes that concentrate negative emotions are identified as the key propagation nodes.

6. The method according to claim 1, characterized in that, The global dissemination features also include: the number of media reprints and the proportion of complaints, wherein the number of media reprints is used to represent the number of nodes whose media level is greater than a preset level threshold, and the proportion of complaints is used to represent the ratio of nodes whose published content contains complaints or rights protection content to the total number of nodes. The global propagation characteristics are quantified based on a preset risk assessment logic to obtain the overall propagation risk level of the target public opinion event, including: The propagation speed is truncated to an upper limit and then normalized to obtain a normalized propagation speed value; the number of platforms involved, the number of high-influence nodes, and the number of media reposts are normalized to obtain normalized platform number values, normalized high-influence node number values, and normalized media repost number values, respectively. The overall risk score is obtained by weighting and summing the normalized values ​​of the propagation speed, the number of platforms, the proportion of negative sentiment nodes, the number of high-influence nodes, the number of media reprints, the proportion of complaints, and the preset risk weights corresponding to each of the global propagation features. The overall dissemination risk level of the target public opinion event is determined based on the preset risk score range in which the overall risk score falls.

7. The method according to claim 1, characterized in that, Based on the initial node, the first cross-platform node, and the key propagation nodes, a structured propagation path for the target public opinion event is output, and a handling priority prompt is generated in conjunction with the overall propagation risk level, including: Starting from the initial node, following the information flow direction indicated by the directed propagation edge between nodes in the public opinion propagation graph, the process is traversed step by step to the end node of the public opinion propagation graph, generating multiple candidate propagation paths from the initial node to the end node. Multiple candidate propagation paths are filtered, and candidate propagation paths that pass through at least one of the key propagation nodes or the first cross-platform node are retained to obtain a set of filtered propagation paths. When the number of paths in the set of filtered propagation paths exceeds a preset upper limit for the number of paths, the sum of the influence values ​​of all nodes passed through on each filtered propagation path is calculated as the path influence score. The paths are sorted from high to low according to the path influence scores, and a preset number of propagation paths at the top of the sorted list are retained to obtain a set of target propagation paths. The key propagation nodes and the first cross-platform node are marked in each propagation path of the target propagation path set to generate a structured propagation path for the target public opinion event; The response level for the target public opinion event is determined based on the overall dissemination risk level. The nodes are sorted from highest to lowest based on their influence values, and the order in which they are processed is determined. The response level is combined with the processing order to generate the processing priority prompt.

8. A public opinion analysis system based on propagation graphs, characterized in that, The public opinion analysis method based on propagation graphs according to any one of claims 1 to 7 includes: The acquisition module is used to acquire multi-source public opinion data associated with the target public opinion event, extract the accounts involved in the dissemination as nodes in the dissemination network, and extract basic attribute data of each node, including the release timestamp, platform, number of followers, historical interaction volume, and media level. The generation module is used to extract the content features of the multi-source public opinion data, construct directed propagation edges based on the interaction behavior records between nodes, content similarity and the publication timestamp, and generate a public opinion propagation graph. The calculation module is used to extract network topology features of each node in the public opinion propagation map, including out-degree, downstream connected node scale and forwarding contribution, as well as sentiment features including the emotional polarity of the published content and the amplification of negative emotions; the number of account followers, the amount of historical interaction, the media level, the forwarding contribution and the amplification of negative emotions are used as influencing factors, and feature normalization and weighted calculation are performed to obtain the influence value of each node. The identification module is used to traverse the public opinion dissemination map, identify the node with the earliest release timestamp as the initial node, identify the node whose platform changes and whose direct downstream platform changes, and whose release time is the earliest as the first cross-platform node; and identify key dissemination nodes according to preset rules based on the influence value of each node, the amplification of negative emotions, and the out-degree change rate of the node within a preset time window. The evaluation module is used to extract the global propagation features of the public opinion propagation map. The global propagation features include at least: propagation speed, number of platforms involved, proportion of negative sentiment nodes, and number of high-influence nodes. The propagation speed is the number of new nodes added per unit time, and the high-influence nodes are key propagation nodes with an influence value greater than a preset influence threshold. The global propagation features are quantitatively calculated based on a preset risk assessment logic to obtain the overall propagation risk level of the target public opinion event. The output module is used to output the structured propagation path of the target public opinion event based on the initial node, the first cross-platform node, and the key propagation node, and to generate a handling priority prompt in combination with the overall propagation risk level.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images