A knowledge graph-based network public opinion key node discovery method

Abstract

The application belongs to the field of data processing, and particularly relates to a network public opinion key node discovery method based on a knowledge graph. The method aims to solve the problems of traditional public opinion monitoring lag and difficulty in capturing budding hotspots, and the steps include: performing dynamic data collection and initialization preprocessing; constructing a dynamic heterogeneous knowledge graph; performing time decay weight updating; calculating evolution momentum and identifying abnormal nodes; entity association verification and early warning pushing. By introducing a time decay mechanism and a second-order momentum algorithm, the application can accurately capture nodes with explosion potential, significantly improving the forward-looking nature and noise filtering precision of early warning.

Description

Technical Field

[0001] This invention relates to the fields of data processing and online public opinion analysis, and more specifically, to a method for discovering key nodes in online public opinion based on knowledge graphs. Background Technology

[0002] In today's internet environment, social media has become a core arena for the public to express their opinions and emotions, and the resulting online public opinion has a profound impact on social governance and stability. Online public opinion analysis technology aims to extract the inherent patterns of information dissemination from massive amounts of fragmented user interaction data. By modeling complex user interaction behaviors, real-time perception of public opinion trends can be achieved, thereby providing decision support for public opinion guidance and information security.

[0003] Key node discovery based on knowledge graphs is an effective way to improve the accuracy of public opinion monitoring. This method constructs a structured graph containing entities, attributes, and their relationships, aiming to accurately locate key users who can guide public opinion or accelerate information dissemination. Its core idea is to transform discrete user behavior into a computable graph structure and comprehensively consider the topological position of nodes in the network and the semantics of their associated content to quantitatively assess the influence and dissemination effectiveness of nodes in specific events.

[0004] However, existing technologies generally have several limitations. Traditional methods often rely on static graph theory algorithms for topological analysis, neglecting the highly time-sensitive nature of social media information, leading to a significant lag in the identification of potential outbreak nodes. Some knowledge graph-based solutions overemphasize static matching of textual semantics, lacking a characterization of dynamic social relationship chains between nodes and the dynamics of information diffusion, making it difficult to truly reflect the evolution of node influence in the interactive environment. Furthermore, conventional models fail to fully consider the decay effect of user interaction over time and the driving force of sentiment on propagation, making the identification process susceptible to interference from historically high-weight accounts, and making it difficult to identify high-potential abnormal nodes in the critical early stages of public opinion evolution. Summary of the Invention

[0005] To overcome the aforementioned technical deficiencies, this invention proposes a method for discovering key nodes in online public opinion based on knowledge graphs. This method, by integrating dynamic temporal modeling and semantic association analysis, aims to achieve more real-time, accurate, and forward-looking early warning of key public opinion nodes. To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] A method for discovering key nodes in online public opinion based on knowledge graphs includes the following specific steps:

[0007] Step S1: Perform dynamic data acquisition and initialization preprocessing: The system accesses multi-source social media stream data through a real-time data interface. This multi-source social media stream data covers forwarding chains, comment interactions, and user attribute information. The data preprocessing layer performs unstructured text cleaning and standardization logic, transforming the raw data into a user, action, and entity quadruple structure carrying high-precision timestamps. A hash algorithm is used to uniquely identify entities, constructing an initial candidate dataset. Step S2: Construct a dynamic heterogeneous knowledge graph: The graph construction engine abstracts users in the social network as nodes and user interactions such as following, forwarding, commenting, and liking as heterogeneous edges. A heterogeneous network model containing entity attributes and topological relationships is established through a graph database storage engine. Basic weight coefficients are assigned to each edge based on the initial interaction strength, forming an initial dynamic graph with multi-dimensional association characteristics. Step S3: Perform time decay weight update: The core computing engine periodically triggers the time decay operator, using a time decay function to update the weights of nodes. The connection strength is corrected in real time. By calculating the difference between the current system time and the time of the interaction, the influence weight of historical outdated interactions is dynamically reduced to ensure that the graph structure reflects the current propagation trend of the public opinion window in real time. Step S4: Calculate evolution momentum and identify abnormal nodes: The system extracts the influence feature vector of the node within a continuous sliding time window. The node evolution speed is obtained by calculating the difference in influence scores between adjacent windows, and the rate of change of speed is calculated to obtain evolution momentum. When the evolution momentum exceeds the preset dynamic acceleration threshold, the node is marked as a momentum abnormal node in a burst growth trend. Step S5: Entity association verification and early warning push: Combining the semantic layer information in the knowledge graph, the matching degree between the content entities associated with the momentum abnormal node and the sensitive topic library is analyzed. By calculating the semantic relevance weight between the node content and the core sensitive entities, non-public opinion-related natural heat noise is filtered out. When the comprehensive relevance ratio exceeds the preset safety threshold, the target node and its propagation chain are pushed to the early warning queue.

[0008] Further, in step S1, the formula for calculating semantic consistency is:

[0009]

[0010] in, and These represent the first two text feature vectors. One portion, For vector dimensions.

[0011] Furthermore, in step S2, the node Basic Influence Score Determined by the following formula:

[0012]

[0013] in, For nodes The set of incoming neighbors, For the outgoing neighbors, For the node To the node Interaction weights This is the preset weight adjustment factor.

[0014] Furthermore, in step S3, the time decay function is:

[0015]

[0016] in, As the initial interaction weights, This is the time decay coefficient.

[0017] Furthermore, in step S3, the attenuation coefficient The dynamic adjustment logic is based on the topic's popularity entropy value. The calculation formula is:

[0018]

[0019] in, For the first Information distribution frequency within each time slice This represents the total number of fragments.

[0020] Furthermore, in step S4, the evolution speed of node influence By calculating two adjacent time windows and The change in internal influence score is obtained using the following formula:

[0021]

[0022] Evolutionary momentum The second derivative of the evolution rate in the time dimension is an approximation, and the formula is as follows:

[0023]

[0024] Furthermore, in step S4, to improve the robustness of recognition, the system introduces a Laplace smoothing term to correct the evolutionary momentum. The corrected momentum... for:

[0025]

[0026] in, For the change in velocity, For the time span, This is a preset smoothing constant used to prevent calculation errors caused by a denominator of zero.

[0027] Furthermore, in step S5, the formula for calculating the entity association score R is:

[0028]

[0029] in, and The first of the two entities respectively dimensional embedding vector components, The preset dimension of the vector space.

[0030] Furthermore, in step S5, the comprehensive relevance weight also needs to be combined with the node's sentiment polarity features and sentiment score. The sentiment analysis model is used to obtain the sentiment, and its calculation logic is as follows:

[0031]

[0032] in, Frequency of positive sentiment words Frequency of words expressing negative emotions.

[0033] Furthermore, the coverage rate The calculation formula is:

[0034]

[0035] in, The set of user nodes that are directly or indirectly affected by this node. This represents the total number of active users across the entire network.

[0036] Furthermore, the first Layer Node Feature Matrix The updated formula is:

[0037]

[0038] in, To increase the adjacency matrix after self-connection, This is the corresponding degree matrix. The weight parameter matrix is ​​a learnable matrix. It is a non-linear activation function.

[0039] Furthermore, the public opinion prediction model is a logistic regression growth model:

[0040]

[0041] in, This is the upper limit of the capacity to handle public opinion. This is the initial heat value. is the growth rate constant.

[0042] Compared with the prior art, the present invention has the following significant advantages:

[0043] 1. Improved real-time performance and forward-looking capabilities: Through dynamic time decay mechanism and second-order evolution momentum monitoring, the system can not only reflect the current propagation core in real time, but also discover potential critical nodes in their infancy in advance based on "acceleration" rather than "speed", which greatly shortens the early warning response time and gains the initiative in handling the situation.

[0044] 2. High accuracy and strong noise resistance: It adopts a dual filtering mechanism of "dynamic topology analysis + deep semantic verification". Momentum analysis identifies abnormal nodes from the perspective of propagation dynamics, while semantic-sentiment correlation verification identifies real risks at the content level, effectively distinguishing between artificially manipulated public opinion and naturally formed hot topics, and reducing the false alarm rate.

[0045] 3. Excellent system adaptability and scalability: The time decay coefficient can be adaptively adjusted according to the topic's popularity, enabling the model to flexibly cope with different types of topic lifecycles. The overall architecture design supports distributed stream processing and edge computing, meeting the real-time processing needs of massive social media data and possessing good horizontal scalability.

[0046] 4. Comprehensive and in-depth analysis dimensions: This method deeply integrates social network analysis, time series modeling, natural language processing and knowledge graph technology, providing a three-dimensional analysis perspective from micro-node behavior to macro-public opinion trends, and enhancing the scientific and accurate nature of public opinion governance decisions. Attached Figure Description

[0047] Figure 1 A schematic diagram of the system architecture of the overall technical solution of this invention;

[0048] Figure 2 A schematic diagram of the principle framework for anomaly node identification based on evolutionary momentum;

[0049] Figure 3 A schematic diagram illustrating the construction and weight allocation process of a dynamic heterogeneous knowledge graph;

[0050] Figure 4 A schematic diagram illustrating the principle of dynamic correction of edge weights based on time decay operators;

[0051] Figure 5 A flowchart illustrating the integrated early warning logic that combines semantic association and sentiment features;

[0052] Figure 6 A diagram illustrating the data interaction relationships between the multi-level modules within the system; Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, a detailed description is provided below in conjunction with the accompanying drawings and embodiments. The following embodiments are for illustrative purposes only and do not constitute a limitation on the scope of protection.

[0054] Example 1

[0055] In this embodiment, a method for discovering key nodes in online public opinion based on knowledge graphs deploys its core execution logic within a cluster of public opinion monitoring servers equipped with high-performance computing capabilities. This server cluster is connected to global internet backbone nodes via high-speed fiber optic links, ensuring real-time capture of streaming data from various social media platforms, news portals, and forums.

[0056] In the above method, the specific implementation process of step S1, dynamic data acquisition and initialization preprocessing, is as follows:

[0057] The system accesses multi-source social media streaming data via pre-defined RESTful API interfaces or WebSocket long-connection protocols. The data acquisition module boasts high concurrency processing capabilities, employing a combination of a distributed crawler cluster and official data authorization interfaces to obtain comprehensive information covering forwarding chains, comment interactions, like records, and user profiles. Before the data enters the preprocessing layer, a traffic shaping module buffers it, utilizing a Kafka distributed message queue to partition and store massive amounts of data arriving instantaneously, ensuring load balancing in subsequent processing stages.

[0058] The data preprocessing layer performs deep cleaning logic. First, regular expressions are used to clean unstructured text, removing HTML tags, garbled characters, redundant URL links, and meaningless special symbols. For emojis specific to social media, the system establishes a mapping table to convert them into corresponding semantic tags. Then, text deduplication logic is executed. By calculating the SimHash value of the text, completely duplicated or highly similar redundant content is quickly identified and removed from massive amounts of data to reduce subsequent computational overhead.

[0059] During preprocessing, the system transforms the cleaned raw data into a "user-action-entity" quadruple structure carrying high-precision timestamps. The user dimension includes a unique identifier (UID), historical activity level, number of followers, and authentication attributes; the action dimension defines the type of interaction, including forwarding, commenting, liking, and referencing; the entity dimension is extracted from the text using a named entity recognition algorithm, including keywords related to people, organizations, locations, and specific events; the timestamps are accurate to the millisecond level, serving as the baseline coordinates for subsequent dynamic evolution calculations. To ensure efficient data retrieval, the system uses the MurmurHash3 algorithm to uniquely identify entities, mapping variable-length strings to fixed-length integer indices, constructing an initial candidate dataset, and storing it in a Redis distributed cache for rapid access by the subsequent graph construction module.

[0060] In the above method, step S2, the specific implementation process of constructing a dynamic heterogeneous knowledge graph, is as follows:

[0061] The graph construction engine reads four-tuple data from the cache and uses the JanusGraph distributed graph database as its storage platform. During the construction process, the system abstracts each independent user in the social network as a node and the dynamic interaction behavior between users as heterogeneous edges. The attribute information of a node includes, but is not limited to, the user's social influence index, historical posting frequency, geographic location tags, and follower list. The type of edge is defined in detail according to the action dimension. For example, a "forward edge" represents a direct path of information dissemination, and a "comment edge" represents the interaction and collision of opinions.

[0062] When establishing the topological relationships between nodes and edges, the graph construction engine simultaneously performs the association of entity attributes and semantic relationships. Keywords extracted by the entity recognition module are mapped to "topic nodes" in the knowledge graph, and user nodes and topic nodes are connected through "mention edges." This heterogeneous network model not only depicts the social connections between people but also the semantic connections between people and content. In the initial weight allocation phase, the system assigns a basic weight coefficient to each edge based on the intensity of interaction. For example, forwarding behavior is given a higher initial weight because it represents secondary dissemination of information; commenting behavior is next, representing deep interaction; and liking behavior has the lowest weight, representing lightweight recognition. All weight data is stored in matrix form in the adjacency list structure of the graph storage engine, forming an initial dynamic graph with multi-dimensional relational features.

[0063] In the above method, the specific implementation process of step S3, which involves updating the time decay weight, is as follows:

[0064] The core computing engine integrates a high-frequency triggered time decay operator, which is periodically activated by the distributed task scheduling system. Due to the highly time-sensitive nature of social media information, the influence of historical interactions on the current public opinion situation weakens rapidly over time. The system uses an exponential decay function to correct the connection strength between nodes in real time. The specific calculation logic is as follows: by obtaining the current system clock's Tnow value, and retrieving the interaction occurrence time recorded on each edge... Calculate the time difference between the two.

[0065] The mathematical implementation of the time decay operator is as follows:

[0066]

[0067] in, For the corrected dynamic weights, The basic interaction weights assigned in step S2, This is the time decay coefficient. In this embodiment, It's not fixed, but rather dynamically and adaptively adjusted based on the topic's lifecycle characteristics. The system identifies whether a topic is in its explosive, stable, or declining phase by calculating the time-series distribution of its popularity. For sudden, explosive topics in their explosive phase, the adjustment is reduced... The value should be adjusted to retain more recent interaction intensity; for older topics that are entering a decline phase, the value should be increased. This accelerates the elimination of historical information. This real-time correction mechanism ensures that the topological structure of the knowledge graph can dynamically fit the actual dissemination trend of the current public opinion window, effectively eliminating the interference of outdated "zombie" connections on influence assessment.

[0068] In the above method, the specific implementation process of step S4, calculating evolutionary momentum and identifying anomalous nodes, is as follows:

[0069] The system extracts the influence feature vector of nodes within consecutive sliding time windows. The step size and length of the sliding time window are dynamically configured according to the monitoring level; during periods of heightened public opinion sensitivity, the window length is shortened to increase the sampling frequency. At the end of each time window, the calculation module iterates through all dynamically weighted edges of the node in the graph and, using an improved PageRank algorithm or HITS algorithm, calculates the node's influence score within the current window. .

[0070] The evolutionary momentum algorithm detects anomalies by comparing changes in influence between adjacent windows. Let the influence of a node within time window T1 be... In the subsequent time window Internal influence The evolution speed of node influence Defined as the rate of change of influence over time. Furthermore, the system calculates the rate of change of evolutionary velocity, i.e., evolutionary momentum. .

[0071] The formula for calculating the evolutionary momentum is as follows:

[0072]

[0073] in, and These represent the evolutionary rates within adjacent time windows. When the evolutionary momentum... Exceeding the preset dynamic acceleration threshold When this occurs, it indicates that the node's influence is experiencing a non-linear, explosive growth. This growth often foreshadows the node becoming a new core or driving force in the spread of public opinion. Threshold The system is based on statistical analysis of historical big data. By calculating the mean and standard deviation of the momentum of nodes under similar topics, it dynamically defines the boundary of anomalies using the principle of three times the standard deviation. Nodes marked as having "momentum anomalies" will be added to a high-priority monitoring list, and the system will simultaneously record all downstream nodes in their propagation chain, providing data support for subsequent prediction of the spread range.

[0074] In the above method, the specific implementation process of step S5, entity association verification and early warning push, is as follows:

[0075] For the momentum anomaly nodes identified in step S4, the system initiates deep semantic verification logic. Utilizing a pre-built sensitive entity library and topic ontology model, combined with semantic layer information from the knowledge graph, it analyzes whether the content involved in the anomaly nodes poses a public opinion risk. The semantic analysis module employs a deep learning model based on the Transformer architecture to map the text content published by the node to a high-dimensional semantic space and calculates its cosine similarity to the sensitive topic vector.

[0076] The entity association verification process not only focuses on keyword matching but also emphasizes the logical connections between entities. For example, if an anomalous momentum node is simultaneously associated with "specific policies," "negative sentiment terms," ​​and "multiple high-influence unverified accounts," the system will use the path search function of the knowledge graph to calculate the semantic relevance weight between the node's content and the core sensitive entities. This process effectively filters out noise from natural popularity growth caused by entertainment hotspots or everyday topics. When the overall relevance ratio exceeds a preset safety threshold, the system automatically generates an early warning report.

[0077] The early warning report includes a profile of the node, a propagation path map, the expected scale of spread, and a recommended level of response. Early warning information is formally pushed to the early warning queue via an encrypted channel, triggering different notification mechanisms based on the early warning level, including but not limited to system pop-ups, instant messaging push notifications, and email alerts. Simultaneously, the system archives the complete evolution data of the early warning event for subsequent algorithm model iteration and optimization.

[0078] In the specific application scenario of this embodiment, assume the system is monitoring discussions about "new energy policies" on a social media platform. 1. During the data acquisition phase, the system captures 150,000 interaction data points generated within 30 minutes via a streaming interface. The preprocessing layer transforms these into 120,000 standard quadruplets, containing 30,000 users and entities including "subsidies," "electricity prices," and "environmental protection." 2. When constructing the graph, the system identifies several core discussion clusters. The initial weight allocation module establishes edges based on forwarding relationships, with an initial weight set to 1.0. 3. As time progresses, the interaction weight from 30 minutes ago decreases to 0.4 due to the time decay operator, while the interaction weight within the last 5 minutes remains above 0.95. 4. At this point, the system discovers a previously inactive emerging account whose influence score rapidly jumps from 0.5 to 8.0 within the last two 5-minute windows, and the calculated evolutionary momentum... The content significantly exceeded the preset threshold. 5. The semantic verification module intervened and found that the content posted by this account involved misleading interpretations of policies and was associated with multiple known sensitive entities. The system immediately identified it as a key warning node and pushed the warning information to the management terminal within 10 seconds.

[0079] Example 2

[0080] This embodiment focuses on performance optimization within a distributed streaming computing framework. Step S1 employs a two-layer buffer architecture, using LVS load balancing and Kafka message queues to handle high-concurrency data. The preprocessing logic is encapsulated as a microservice, supporting automatic scaling.

[0081] Step S2 introduces graph partitioning technology, using the METIS algorithm to divide the graph and reduce network latency. User reputation factors are incorporated into weight allocation.

[0082] Step S3 employs a lazy update and hotspot triggering strategy, with GPU-accelerated weight updates.

[0083] Step S4 introduces multi-dimensional feature vectors, such as sentiment entropy, and uses the Flink streaming engine for incremental computation. The community detection algorithm defines the scope of influence.

[0084] Step S5 employs enhanced knowledge embedding, incorporates human feedback loops, and optimizes threshold adjustment. Multi-tenant isolation supports personalized rules.

[0085] This embodiment demonstrated extremely high processing efficiency in handling a sudden public security incident. Even with peak traffic reaching 50,000 records per second, the system maintained an end-to-end analysis latency of less than 2 seconds. Through a distributed stream computing architecture, it successfully identified 3 core source nodes and 12 key propagation nodes within 15 minutes of the public opinion crisis escalating, providing precise decision support for relevant departments.

[0086] Example 3

[0087] This embodiment focuses on performance optimization within a distributed streaming computing framework. Step S1 employs a two-layer buffer architecture, using LVS load balancing and Kafka message queues to handle high-concurrency data. The preprocessing logic is encapsulated as a microservice, supporting automatic scaling.

[0088] Step S2 introduces graph partitioning technology, using the METIS algorithm to divide the graph and reduce network latency. User reputation factors are incorporated into weight allocation.

[0089] Step S3 employs a lazy update and hotspot triggering strategy, with GPU-accelerated weight updates.

[0090] Step S4 introduces multi-dimensional feature vectors, such as sentiment entropy, and uses the Flink streaming engine for incremental computation. The community detection algorithm defines the scope of influence.

[0091] Step S5 employs enhanced knowledge embedding, incorporates human feedback loops, and optimizes threshold adjustment. Multi-tenant isolation supports personalized rules.

[0092] In the event of an emergency, the system processes 50,000 records per second with a latency of less than 2 seconds, successfully identifying critical nodes.

[0093] The advantage of this embodiment lies in its strong scalability and fault tolerance. In actual deployment, by deploying 50 edge nodes nationwide, the system achieves refined perception of public opinion trends in various provinces and cities. In a public opinion fluctuation targeting a specific regional policy, the edge nodes completed preliminary node identification within 3 minutes of the event occurring, issuing a local warning approximately 8 minutes earlier than a purely centralized architecture, effectively preventing the cross-regional spillover of public opinion risks.

[0094] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for discovering key nodes in online public opinion based on knowledge graphs, characterized in that, Includes the following steps: Step S1: Perform dynamic data acquisition and initialization preprocessing: The system accesses multi-source social media stream data through a real-time data interface. The multi-source social media stream data covers forwarding chains, comment interactions, and user attribute information. The data preprocessing layer performs unstructured text cleaning and standardization logic, transforming the raw data into a user, action, and entity quadruple structure carrying a high-precision timestamp, and using a hash algorithm to uniquely identify entities, thus constructing an initial candidate dataset. Step S2: Construct a dynamic heterogeneous knowledge graph: The graph construction engine abstracts users in the social network as nodes, and the interaction behaviors of users such as following, forwarding, commenting, and liking as heterogeneous edges. Through the graph database storage engine, a heterogeneous network model containing entity attributes and topological relationships is established, and basic weight coefficients are assigned to each edge according to the initial interaction strength, forming an initial dynamic graph with multi-dimensional association characteristics. Step S3: Perform time decay weight update: The core computing engine periodically triggers the time decay operator, and uses the time decay function to correct the connection strength between nodes in real time. By calculating the difference between the current system time and the time of the interaction, the influence weight of historical outdated interactions is dynamically reduced to ensure that the graph structure reflects the current public opinion window's propagation trend in real time. Step S4: Calculate evolution momentum and identify abnormal nodes: The system extracts the influence feature vector of nodes within a continuous sliding time window, obtains the node evolution speed by calculating the difference in influence scores between adjacent windows, and calculates the rate of change of speed to obtain evolution momentum. When the evolution momentum exceeds a preset dynamic acceleration threshold, the node is marked as an abnormal momentum node in a burst growth state. Step S5: Entity Association Verification and Early Warning Push: Combining semantic layer information in the knowledge graph, analyze the matching degree between the content entities associated with the momentum anomaly nodes and the sensitive topic library. By calculating the semantic relevance weight between the node content and the core sensitive entities, filter out natural heat noise that is not related to public opinion. When the comprehensive relevance ratio exceeds the preset safety threshold, push the target node and its propagation chain to the early warning queue.

2. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, In step S1, the data preprocessing layer performs deep cleaning logic on the input unstructured text: using regular expressions to identify and remove Hypertext Markup Language tags, garbled character sequences, redundant Uniform Resource Locator links, and meaningless special symbols from the original text. For emojis unique to social media, a semantic mapping table is established to convert them into corresponding text tags. Text deduplication logic is executed, using feature hash values ​​to identify and remove completely duplicated or highly similar redundant content from massive data streams. A named entity recognition algorithm based on bidirectional long short-term memory networks and conditional random fields is used to extract keywords, and the semantic consistency between texts is calculated using the cosine similarity formula. This semantic consistency is determined based on the ratio of the product of two text feature vectors to the product of the magnitudes of the two text feature vectors. Duplicate information is merged by setting a first preset similarity threshold. The cleaned raw data is transformed into a user, action, and entity quadruple structure carrying high-precision timestamps. The user dimension includes a unique identifier, historical activity level, number of followers, and authentication attributes; the action dimension defines the type of interaction; and the entity dimension includes keywords related to people, organizations, locations, and specific events. A hash algorithm is used to uniquely identify entities, mapping variable-length strings to fixed-length integer indices, constructing an initial candidate dataset, and storing it in a distributed cache.

3. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, In step S2, the process of constructing a dynamic heterogeneous knowledge graph involves weighted calculation of the in-degree and out-degree of nodes: the graph construction engine reads quadruple data from the distributed cache, uses a distributed graph database as the storage base, abstracts each independent user in the social network as a node, and abstracts the dynamic interaction behavior between users as heterogeneous edges. The node's attribute information includes the user's social influence index, historical posting frequency, geographic location tags, and follower list; the edge types are defined in detail according to the action dimension, including forwarding edges that represent the direct dissemination path of information, commenting edges that represent the interaction and collision of opinions, and liking edges that represent lightweight recognition. Keywords extracted by the entity recognition module are mapped to topic nodes in the knowledge graph, and user nodes and topic nodes are connected by mention edges. In the initial weight allocation stage, a basic weight coefficient is assigned to each edge according to the different interaction strengths. The basic influence score of a node is determined by a linear combination of the weighted sum of the interaction weights of the inbound neighbor set and the weighted sum of the interaction weights of the outbound neighbor set. Passive attention and active dissemination power are balanced by a preset weight adjustment factor. All weight data is stored in matrix form in the adjacency list structure of the graph storage engine.

4. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, In step S3, the specific form of the time decay function adopts an exponential degradation model: the core computing engine periodically triggers the time decay operator, and the corrected dynamic weight is determined based on the product of the initial interaction weight and the base of the natural logarithm, wherein the exponential part is the product of the negative time decay coefficient and the time difference; the time difference is the difference between the current system time and the time when the interaction occurs; the time decay coefficient is dynamically adjusted according to the life cycle characteristics of the topic. The dynamic adjustment logic of the attenuation coefficient is based on the topic's heat entropy value, which is determined by the logarithmic weighted sum of the information distribution frequencies within a preset time segment. When the heat entropy value is in a preset low range, it indicates that the information is highly concentrated, and the system automatically increases the time attenuation coefficient to accelerate the elimination of old information. For breaking news topics, the time attenuation coefficient is set within a preset range to enhance sensitivity to the latest information. A real-time correction mechanism ensures that the topological structure of the knowledge graph dynamically fits the actual dissemination trend of the current public opinion window.

5. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, In step S4, the evolution rate of node influence is obtained by calculating the ratio of the change in influence score within two adjacent time windows to the time span: the system extracts the influence feature vector of the node within a continuous sliding time window, and the step size and length of the sliding time window are dynamically configured according to the monitoring level; at the end of each time window, the calculation module calculates the influence score of the node in the current window by traversing all dynamic weighted edges of the node in the graph and combining the improved web page ranking algorithm. Evolutionary momentum is defined as an approximation of the second derivative of evolutionary velocity in the time dimension, calculated by inter-time intervals.

6. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 5, characterized in that, In step S4, in order to improve the robustness of recognition, the system introduces a Laplace smoothing term to correct the evolutionary momentum: the corrected momentum is determined based on the ratio of the sum of the velocity change and the first preset smoothing constant, and the sum of the time span and the second preset smoothing constant; the first preset smoothing constant and the second preset smoothing constant are used to prevent calculation anomalies caused by the denominator being zero. The system introduces a multi-dimensional feature vector, which includes influence score, node sentiment entropy, domain coverage, and propagation centripetal force. The momentum calculation module uses a sliding window aggregation operator, implemented in the stream computing engine. Incremental calculation is achieved by maintaining historical influence snapshots of nodes in real time through the state backend. When momentum anomalies are detected, the community discovery algorithm is used to define the potential audience range affected by the node in real time and calculate the contagion intensity of public opinion spread. By introducing a Laplace smoothing term in the evolutionary momentum calculation, the computational fluctuations caused by data sparsity or collection jitter are resolved, and the frequency of false warnings caused by data noise is reduced.

7. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, In step S5, the entity association verification process employs knowledge graph embedding technology: entities are mapped to a low-dimensional continuous vector space, and the association score between a node entity and a sensitive entity is measured by the reciprocal of the Euclidean distance between the two entities in the low-dimensional continuous vector space; the semantic analysis module uses a deep learning model to map the text content published by the node to a high-dimensional semantic space and calculates its cosine similarity with the sensitive topic vector; the entity association verification process focuses on the logical association between entities and calculates the semantic relevance weight between the node content and the core sensitive entity through the path search function of the knowledge graph; When the overall relevance ratio exceeds a preset safety threshold, the system automatically generates an early warning report; the early warning report includes the node's profile information, propagation path map, expected spread scale, and recommended handling level; Warning information is pushed to the warning queue through an encrypted channel, and system pop-ups, instant messaging tool pushes, or email alerts are triggered according to the warning level.

8. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 7, characterized in that, In step S5, the relevance weights are combined with the emotional polarity features of the nodes: the emotional score is obtained using a text sentiment analysis model, and the emotional score is determined based on the difference between the frequency of positive sentiment words and the frequency of negative sentiment words, as well as the ratio of the sum of the frequencies of positive sentiment words, the frequencies of negative sentiment words, and a preset constant; when the absolute value of the emotional score is greater than a preset emotional threshold and the association score is greater than a preset association threshold, a warning of the corresponding level is triggered; a layer of manual intervention feedback loop is added to the verification logic, and the system collects feedback results while pushing warnings, and uses reinforcement learning algorithms to automatically adjust the semantic relevance judgment threshold; The early warning push module supports multi-tenant isolation, allowing different management departments to set personalized sensitive entity databases and early warning rules based on their own areas of interest. The system archives the complete evolution data of early warning events for subsequent algorithm model iteration and optimization.

9. The method for discovering key nodes of online public opinion based on knowledge graphs according to claim 1, characterized in that, It also includes a quantitative assessment of the propagation coverage of nodes: the propagation coverage rate is determined based on the ratio of the number of elements in the set of user nodes affected by the node to the total number of active users in the entire network; when the growth rate of the propagation coverage rate exceeds the preset growth threshold within a preset time period, the system automatically upgrades the monitoring priority of the node; the knowledge graph is deeply represented by feature extraction operators based on graph convolutional neural networks, by adding the product of the adjacency matrix after self-connection, the corresponding degree matrix, the node feature matrix, and the learnable weight parameter matrix, and then processing it through a nonlinear activation function to obtain the global structural features of the nodes; After identifying key nodes, a public opinion trend prediction based on an information flow dynamics model is performed; the predicted change in popularity over time follows a logistic regression growth model, which is determined based on the upper limit of public opinion carrying capacity, the initial popularity value, and the growth rate constant; the parameters are fitted in real time using the least squares method to control the prediction error within a preset error range.

Images