An opinion depth analysis and early warning method based on a multi-modal model factory

Abstract

This invention relates to the field of intelligent public opinion analysis technology. Specifically, it relates to a method for in-depth public opinion analysis and early warning based on a multimodal model factory. In this invention, the feature calculation layer extracts user influence, specific risk themes, and dynamic features in parallel through multiple engines. The graph platform synchronously constructs a propagation relationship graph. The graph reasoning engine extracts graph structure features and generates a structured feature set. The feature set is input into the model factory. The scenario recognition module determines the type of public opinion event. The model scheduler and pipeline assembler generate a customized analysis pipeline. After parallel processing of feature slices, the scoring fusion module outputs three types of quantitative scores: source, content, and propagation. The rule reasoning engine combines the scores and event types to determine the risk. High-risk events trigger a causal analysis module to trace core propagation nodes and path mutation points, generating a comprehensive early warning level and basis report, forming a self-evolving closed loop, improving the accuracy of public opinion analysis and the timeliness of early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent public opinion analysis technology, specifically to a method for in-depth public opinion analysis and early warning based on a multimodal model factory. Background Technology

[0002] Intelligent public opinion analysis is an important technology that is specifically applied to the in-depth analysis of various internet public opinion events, the location of core dissemination nodes, and the early warning of risk levels. Its core is to achieve accurate identification and timely response to public opinion risks by integrating multi-dimensional public opinion characteristics and customized analysis processes, and to meet the core needs of rapid public opinion dissemination, multi-source data, and complex scenarios.

[0003] Current public opinion analysis methods struggle to overcome the limitations of single-dimensional analysis when dealing with multi-source, heterogeneous public opinion data. They fail to accurately capture the essential risks and dissemination patterns of public opinion events. Existing public opinion analysis technologies often process user behavior, text content, and dissemination paths in isolation, without constructing a multimodal fusion system that encompasses user influence, content risk themes, dynamic temporal characteristics, and dissemination structure features. This results in incomplete feature coverage and a lack of dynamically customized analysis processes based on public opinion event types. The use of fixed models to handle various events fails to adapt to the risk characteristics of different public opinion events. Furthermore, risk assessment relies solely on surface feature comparison, lacking in-depth tracing of the causal relationships of public opinion outbreaks. It is difficult to locate core dissemination nodes and path abrupt change points, and the analysis model cannot be continuously optimized through incremental samples. This leads to low accuracy in public opinion risk identification, delayed early warnings, and a lack of interpretability in early warning results. Consequently, it is difficult to meet the needs for accurate, efficient, and traceable early warnings in complex public opinion scenarios. To address this technical problem, we propose a public opinion in-depth analysis and early warning method based on a multimodal model factory. Summary of the Invention

[0004] The purpose of this invention is to provide a method for in-depth analysis and early warning of public opinion based on a multimodal model factory, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, one of the objectives of this invention is to provide a method for in-depth analysis and early warning of public opinion based on a multimodal model factory, comprising the following steps: S1. Multi-source public opinion data is processed through the feature calculation layer. Among them, the user entity feature engine extracts user influence features, the text semantic feature engine injects domain knowledge to identify specific risk topics, and the real-time statistical feature engine calculates dynamic features. At the same time, the graph platform constructs a propagation relationship graph in parallel and extracts the graph structure features in user influence features, specific risk topics, and dynamic features through the graph reasoning engine to generate a structured feature set. S2. Input the structured feature set into the standardized model component library pre-set in the model factory. The scenario recognition module of the standardized model component library judges the structured feature set and outputs the public opinion event type. The model scheduler selects components according to the public opinion event type, and the pipeline assembler generates a customized analysis pipeline. The customized analysis pipeline performs multi-model collaborative processing on the structured feature set and outputs a quantitative score through the scoring fusion module. S3. The rule reasoning engine loads the preset judgment rules and combines them with the quantitative score to determine the risk. For high-risk events, the causal analysis module is triggered, and key nodes are traced based on the propagation path of the platform in the figure. The warning level calculator generates a comprehensive warning level and basis report based on the causal conclusions obtained by the causal analysis module through weighted calculation. S4. The comprehensive early warning level and the report are fed back to the feature warehouse as incremental samples to evaluate the effectiveness of each model component in the current pipeline and drive the model factory to perform incremental optimization of the components, forming a self-evolving closed loop.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts user influence, specific risk themes, and dynamic features through a multi-engine parallel feature calculation layer. It combines a graph platform to construct a propagation relationship graph and a graph reasoning engine to extract graph structure features to generate a structured feature set. The model factory customizes the analysis pipeline according to the type of public opinion event. The rule reasoning engine combines quantitative scoring to initially judge the risk. High-risk events trigger a causal analysis module to trace the core propagation nodes and path mutation points. The warning level calculator generates interpretable warning levels through a three-dimensional decision matrix, and the results are fed back to drive incremental model optimization to form a self-evolving closed loop. This achieves comprehensive coverage of multimodal public opinion features, accurate adaptation of analysis process to event type, no omissions or misjudgments in risk judgment, traceability of core propagation nodes, quantitative and interpretable warning levels, and continuous model optimization. It effectively solves the problems of low risk identification accuracy, delayed warnings, and uninterpretable results caused by the isolation of multi-dimensional features, fixed processes, difficulty in tracing causality, and lack of self-evolving capabilities in existing public opinion analysis technologies. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Please see Figure 1 As shown, this embodiment provides a method for in-depth public opinion analysis and early warning based on a multimodal model factory, including the following steps: S1. Multi-source public opinion data is processed through the feature calculation layer. Among them, the user entity feature engine extracts user influence features, the text semantic feature engine injects domain knowledge to identify specific risk topics, and the real-time statistical feature engine calculates dynamic features. At the same time, the graph platform constructs a propagation relationship graph in parallel and extracts the graph structure features in user influence features, specific risk topics, and dynamic features through the graph reasoning engine to generate a structured feature set. S2. Input the structured feature set into the standardized model component library pre-set in the model factory. The scenario recognition module of the standardized model component library judges the structured feature set and outputs the public opinion event type. The model scheduler selects components according to the public opinion event type, and the pipeline assembler generates a customized analysis pipeline. The customized analysis pipeline performs multi-model collaborative processing on the structured feature set and outputs a quantitative score through the scoring fusion module. S3, the rule reasoning engine loads preset judgment rules and combines quantitative scoring to determine risk. Among them, the causal analysis module is triggered for high-risk events, and key nodes are traced based on the propagation path of the platform in the figure. The warning level calculator generates a comprehensive warning level and basis report based on the causal conclusions obtained by the causal analysis module through weighted calculation. S4. The comprehensive early warning level and the report are fed back to the feature warehouse as incremental samples to evaluate the effectiveness of each model component in the current pipeline and drive the model factory to perform incremental optimization of the components, forming a self-evolving closed loop.

[0010] The user entity feature engine constructs a multi-dimensional influence assessment matrix by aggregating user registration information and historical behavior data to extract user influence features. The text semantic feature engine injects industry risk keyword library into a pre-built domain knowledge graph and uses an attention mechanism-enhanced topic matching model to identify specific risk topics. The real-time statistical feature engine dynamically calculates the source sentiment fluctuation variance and posting frequency change rate as dynamic features based on a sliding time window.

[0011] When constructing the propagation relationship graph in parallel, user entities are used as nodes and forward and backward propagation relationships are used as edges. Node attributes inherit user influence characteristics, and edge weights are dynamically calculated based on forwarding latency attenuation coefficients and content similarity. The propagation relationship graph uses a dynamic graph structure storage engine to update the state changes of nodes and edges in real time.

[0012] The graph reasoning engine performs multi-hop neighbor sampling and subgraph aggregation operations on the propagation relationship graph, extracts information propagation depth features through random walk algorithm, calculates propagation scale features using maximum connected subgraph detector, and iteratively solves the centrality index of key nodes based on weighted PageRank algorithm. The aforementioned features are merged with user influence features, risk themes and dynamic features into a structured feature set.

[0013] The context recognition module uses a lightweight cascaded classifier, which includes: firstly, performing a primary event coarse screening based on specific risk theme features, and then performing a secondary fine classification by integrating dissemination scale features, and outputting preset public opinion event tags.

[0014] The model scheduler has built-in event type and component mapping rule tables, and the pipeline assembler generates customized analysis pipelines through topological sorting of directed acyclic graphs, where the component connection order is determined by the preset causal dependencies in the domain knowledge base.

[0015] When the customized analysis pipeline runs, a feature segmentation parallel processing mechanism is adopted, which includes: splitting the structured feature set into user entity segments, text semantic segments and graph structure segments according to modality, inputting them into the selected components for parallel computation, and the scoring fusion module weighting and fusing the component output results through learnable weight tensors to generate three types of quantitative scores: source risk score, content risk score and dissemination risk score.

[0016] The judgment rules loaded by the rule reasoning engine adopt a hierarchical conditional expression storage structure, which includes: the first layer of rules compares the thresholds of three types of quantitative scores; the second layer of rules associates event type labels; and the initial risk judgment result includes two states: low risk and high risk.

[0017] When an event is identified as high-risk, the causal analysis module initiates reverse propagation path tracing. This includes starting from the current event source node based on the propagation relationship graph, visiting key nodes in descending order of centrality index, and finally locating the core propagation nodes and path mutation points that led to the outbreak of public opinion by comparing the differences in dynamic characteristics between nodes.

[0018] The early warning level calculator constructs a three-dimensional decision matrix, specifically including: The matrix input dimensions include rule hit results, number of key nodes, and incremental risk of propagation. It generates a comprehensive early warning level through a preset fuzzy comprehensive evaluation algorithm, automatically associates path mutation point features with event type labels, and generates a report based on the findings.

[0019] Further explanation is needed regarding the multimodal model factory-based deep analysis workflow for public opinion. The feature calculation layer is the core step in obtaining high-quality input features. Through the parallel collaboration of the user entity feature engine, text semantic feature engine, and real-time statistical feature engine, multi-source public opinion data is processed from three dimensions: user influence, content risk themes, and time-series dynamic changes. Simultaneously, in conjunction with graph structure feature extraction from the graph platform, a comprehensive structured feature set is ultimately formed. The specific implementation method is as follows: The user entity feature engine focuses on the "human" dimension, constructing a multi-dimensional influence assessment matrix by aggregating user registration information and historical behavior data. This allows for the precise extraction of user influence characteristics. User registration information includes account level, authentication type, registration duration, and domain tags, which directly reflect the account's initial credibility. Historical behavior data includes the number of posts, reposts, comments, likes, favorites, total number of followers, follower activity, and the number of times content has been cited in the past 90 days, which primarily reflects the user's dissemination ability and industry influence.

[0020] The process of constructing the multidimensional influence assessment matrix is ​​as follows: First, five core evaluation dimensions were identified: fan influence, content dissemination power, interaction activity, account credibility, and domain authority. Each dimension corresponds to 3-4 specific indicators. Then, each indicator was standardized, mapping indicators of different magnitudes to a 0-1 range. Weights were assigned to each dimension based on the industry characteristics of the public opinion analysis scenario. Finally, the standardized indicator values ​​were integrated by dimension to form a three-dimensional matrix of user ID, evaluation dimensions, and indicator values. The comprehensive score for each user in the matrix represents the user's influence characteristic; a higher score indicates a greater potential influence in public opinion dissemination. Simultaneously, textual semantic features... The engine approaches the issue from the "content" dimension, injecting an industry risk keyword library into a pre-built domain knowledge graph. It then uses an attention-enhanced topic matching model to identify specific risk topics. The domain knowledge graph is a pre-built semantic network of industry risks, containing a hierarchical structure of risk topics, core concepts related to topics, and semantic relationships. The industry risk keyword library is generated based on this knowledge graph, covering core risk keywords, synonyms, near-synonyms, and scenario-based variations, and is weighted according to risk level. It also supports dynamic updates. The injection process involves binding the keyword library with the semantic vectors of the domain knowledge graph, giving the keywords hierarchical risk semantic associations.

[0021] The specific recognition process of the attention-enhanced topic matching model is as follows: The public opinion text is converted into word vectors. The model's attention layer then automatically focuses on words in the text that match the industry risk keyword database, assigning higher weights to these high-risk words. Simultaneously, it combines semantic relationships from the domain knowledge graph to uncover implicit risk associations within the text. Finally, the weighted text vectors are compared with preset vectors for specific risk topics. If the similarity exceeds a predefined threshold, the text is determined to belong to the corresponding risk topic, thus completing the identification of specific risk topics. Furthermore, the real-time statistical feature engine captures changing trends from a temporal dynamic dimension. Based on a sliding time window, it dynamically calculates the variance of sentiment fluctuations and the rate of change in posting frequency of information sources as dynamic features reflecting the active state of public opinion. The sliding time window refers to a time interval that continuously scrolls for a fixed duration, its length set according to the typical cycle of public opinion dissemination. The window's sliding step size is consistent with its length, ensuring real-time capture of the latest temporal changes. The calculation process for the variance of sentiment fluctuations is as follows: First, a sentiment analysis model is used to score the sentiment of all texts published by the sources within the window. The score ranges from -1 to 1, where -1 represents extremely negative and 1 represents extremely positive. Then, the variance of all sentiment scores within the window is calculated. The larger the variance, the more intense the fluctuation of the source's sentiment.

[0022] The variance value serves as a characteristic of emotional fluctuations, and the calculation process for the posting frequency change rate is as follows: The total number of posts from the current information source within the sliding window is counted and compared with the total number of posts from the previous adjacent window. The difference between the two is calculated and divided by the total number of posts from the previous window. This rate of change directly reflects the change in the activity level of the information source. These two indicators together constitute dynamic features, reflecting the emotional stability and activity trend of the information source in real time, providing temporal dimension support for subsequent risk assessment. The user influence features, specific risk theme features, and dynamic features extracted by the three engines complement each other with the graph structure features extracted by the graph platform, comprehensively covering the core dimensions of public opinion: "people, content, time sequence, and dissemination structure." Finally, they are integrated to form a standardized structured feature set, providing high-quality input for the customized analysis pipeline of the subsequent model factory.

[0023] The specific implementation method for parallel construction of the propagation relationship graph in the graph platform is as follows: After the feature extraction of user, text, and time series dimensions is completed at the feature calculation layer, the graph platform needs to construct the propagation relationship graph in parallel. This visualizes the propagation structure features of public opinion data as a graph model. By defining users as nodes and propagation relationships as edges, and assigning dynamic attributes to nodes and edges, the information flow path is accurately depicted, laying the foundation for the subsequent extraction of structural features by the graph inference engine. The specific implementation method is as follows: When the platform constructs a graph model using user entities as nodes, the core is to achieve the inheritance of node attributes to user influence characteristics. Node attributes refer to the structured information attached to each user node in the graph. Inheritance means accurately associating the user influence characteristics generated by the user entity feature engine with the corresponding nodes, ensuring that the node information can reflect the potential role of users in public opinion dissemination. The specific process is as follows: First, a unique node ID is assigned to each user entity, and a mapping table between user entities and node IDs is established. The mapping relationship is synchronized to the feature repository in real time to ensure a one-to-one correspondence between user information and nodes. Then, the multi-dimensional influence assessment matrix data of the user is retrieved from the feature repository, and these dimension scores are used as the core attributes of the node and written into the node attribute table. To ensure the timeliness of attributes, an attribute synchronization mechanism is set up every 5 minutes. When the user entity feature engine updates a user's influence characteristics, the feature repository will trigger an update event. After receiving the event, the platform in the graph immediately retrieves the latest score, updates the attribute value of the corresponding node, and refreshes the timestamp to ensure that the node attributes are always consistent with the user's real influence. After determining the nodes, connections are built using forward and backward propagation relationships as edges. Forward and backward propagation relationships refer to the forward path and reverse tracing path of information propagation. A forward propagation relationship is like "User A forwards User B's content", corresponding to the directed edge "A→B" in the graph, representing the direction of information propagation from B to A. A backward propagation relationship (backward edge) is like "User A's content is cited by User B", corresponding to the directed edge "B←A" in the graph, which facilitates subsequent tracing of the source of information.

[0024] The calculation of edge weights needs to be dynamically determined by combining the forwarding delay attenuation coefficient and content similarity. Edge weights are indicators that quantify the strength of propagation relationships; higher weights indicate a greater impact of the propagation path on public opinion dissemination. The forwarding delay attenuation coefficient is an attenuation factor set based on the time difference between the forwarding behavior and the original content publication. The original content publication time is denoted as T0, the forwarding behavior occurrence time is denoted as T1, and the forwarding delay ΔT equals T1-T0. Based on the attenuation law of public opinion propagation, the coefficient is set as follows: when ΔT is less than or equal to 10 minutes, the attenuation coefficient is equal to 1; when 10 minutes is less than ΔT and less than or equal to 30 minutes, the attenuation coefficient is equal to 0.8; when 30 minutes is less than ΔT and less than or equal to 60 minutes, the attenuation coefficient is equal to 0.5; and when ΔT is greater than 60 minutes, the attenuation coefficient is equal to 0.2. Content similarity is an indicator that measures the semantic consistency between forwarded content and original content, and is calculated using the cosine similarity algorithm. First, extract the text vectors of the original content and the forwarded content respectively. Then, calculate the cosine similarity between the two vectors. The similarity value ranges from 0 to 1, with values ​​closer to 1 indicating greater consistency in content. The specific calculation logic for edge weights is as follows: weight equals forwarding delay attenuation coefficient × 0.4 + content similarity × 0.6. For example, if a forwarding behavior ΔT equals 8 minutes and the content similarity is 0.9, then the edge weight equals 1 × 0.4 + 0.9 × 0.6, which equals 0.4 + 0.54, which equals 0.94. This indicates that the propagation path has a high strength. The propagation relationship graph uses a dynamic graph structure storage engine to achieve real-time updates of node and edge state changes. The dynamic graph structure storage engine is a graph database system that supports real-time addition and deletion of nodes / edges and updates of attributes / weights. The specific update process is divided into two categories: node updates and edge updates. When the feature repository pushes an update event for a user's influence dimension score, the storage engine locates the node ID corresponding to that user, directly overwrites the old score field in the attribute table, and updates the timestamp. The entire process is completed within 100ms, ensuring that node attributes are synchronized without delay. Edge updates include three scenarios: First, adding edges. When a new forwarding / referencing behavior is detected, the engine immediately creates a forward edge "C→A" and a backward edge "A←C", calculates the initial edge weights, and writes them to the edge attribute table, including "edge ID, source node ID, target node ID, weight, forwarding latency, content similarity, and edge creation time". The engine employs several mechanisms: first, updating the edge weights; second, recalculating the forwarding latency attenuation coefficient for all edges created more than 10 minutes ago, and updating the edge weights based on changes in content similarity; and third, updating the edge status. When a forwarding action is detected as deleted, the engine marks the corresponding edge's "status" field from "valid" to "invalid" instead of directly deleting the edge, thus preserving the source data for subsequent analysis. All update operations are triggered by an event-driven mechanism, supplemented by scheduled tasks for calibration, to ensure that the propagation relationship graph remains synchronized with the actual state of public opinion propagation.

[0025] After the graph platform completes the dynamic construction of the propagation relationship graph, the graph inference engine needs to perform structured feature extraction operations on the graph. By focusing on key propagation areas through multi-hop neighbor sampling and subgraph aggregation, the engine avoids the waste of resources caused by full-graph computation. Then, by combining random walk, maximum connected subgraph detection, and weighted PageRank algorithm, graph structure features are extracted from three dimensions: propagation path depth, diffusion range, and core node influence. Finally, these features are fused with the features extracted from user entities, text semantics, and the real-time statistics engine to form a structured feature set covering multiple dimensions. The specific implementation method is as follows: The graph reasoning engine first performs multi-hop neighbor sampling and subgraph aggregation operations on the propagation relationship graph. Multi-hop neighbor sampling refers to collecting neighbor nodes layer by layer according to a preset number of hops, starting from the initial associated node of the public opinion event, to ensure focus on key propagation layers. The specific sampling process is as follows: In 1-hop sampling, starting from the initial node, direct neighbor nodes with edge weights greater than or equal to 0.5 are selected, with a maximum of 20 nodes collected to avoid computational overload due to too many neighbors. If fewer than 20 nodes are selected, all nodes are collected. In 2-hop sampling, starting from the node obtained from 1-hop sampling, neighbors with edge weights greater than or equal to 0.3 are selected, and the weight threshold is lowered to expand coverage, but weakly related nodes are filtered out. Each 1-hop node has a maximum of 15 2-hop neighbors collected. In 3-hop sampling, starting from the 2-hop node, neighbors with edge weights greater than or equal to 0.2 are selected, with each 2-hop node having a maximum of 10 3-hop neighbors collected. Subgraph aggregation involves combining the sampled 1-3 hop nodes, the initial node, and all edges (including forward and backward edges) between these nodes to form a local subgraph. The features of this subgraph are then aggregated and calculated. The aggregation dimensions include the average user influence score of all nodes in the subgraph, the average weight of edges in the subgraph, and the proportion of high-weight edges in the subgraph, reflecting the concentration of the core propagation path. These aggregation results serve as the basic structural features at the subgraph level, laying the foundation for subsequent deep feature extraction.

[0026] Subsequently, the information propagation depth features are extracted using a random walk algorithm. The random walk algorithm is an unsupervised learning method that simulates the flow path of information in a propagation graph. By randomly selecting the next hop node, it captures the potential depth and path diversity of information propagation. The information propagation depth features are indicators that quantify the levels and diffusion capabilities of information propagation, including three core dimensions: average propagation depth, maximum propagation depth, and path diversity. The specific implementation process is as follows: Using the initial node obtained from subgraph aggregation as the starting point for the walk, the number of walks is set to 100 to balance computational efficiency and result stability. The maximum number of steps in each walk is 10 to avoid meaningless long-tail paths. During the walk, at each step, the next hop node is randomly selected from all the neighboring nodes of the current node according to the edge weight ratio. At the same time, the complete path of each walk is recorded. After the walk is completed, the average propagation depth, maximum propagation depth, and path diversity are calculated. These three indicators together constitute the information propagation depth feature. While extracting the propagation depth feature, the propagation scale feature is calculated using the maximum connected subgraph detector. The maximum connected subgraph refers to the largest subset of all nodes in the propagation relationship graph that can be reached by edges. The size of this subset directly reflects the actual spread range of the public opinion event. The propagation scale feature includes three indicators: the number of nodes in the connected subgraph, the number of edges in the connected subgraph, and the subgraph density, which respectively quantify the number of users spreading, the number of propagation relationships, and the degree of propagation density. The specific detection process is as follows: A breadth-first search algorithm is used, starting from the initial nodes obtained by subgraph aggregation, and traversing all nodes reachable by edges layer by layer. These nodes and the edges connecting them form a connected subgraph. If there are multiple disconnected subgraphs, the subgraph with the most nodes is selected as the largest connected subgraph. Then, the number of nodes and edges in the subgraph are counted, and the subgraph density is calculated. The density is equal to the number of edges / (number of nodes × (number of nodes - 1)). The higher the density, the stronger the propagation relationship between users, and the more concentrated the public opinion spread, forming a propagation scale characteristic. In addition, the centrality index of key nodes is solved iteratively based on the weighted PageRank algorithm. The weighted PageRank algorithm is based on the traditional PageRank (which calculates importance based on the in-degree of nodes) and introduces edge weight optimization. The higher the edge weight, the more importance score the source node transmits to the target node, which can more accurately reflect the impact of the strength of the propagation relationship on the node's influence. The centrality index is the node importance score output by the algorithm. The higher the score, the more core the node is in the public opinion propagation. The specific iterative process is as follows: First, an initial centrality score is assigned to all nodes within the subgraph, each score being 1 / total number of nodes. Then, iterative calculation begins. In each iteration, each node distributes its current centrality score proportionally to the neighboring nodes pointed to by its outgoing edges, according to the edge weights. A damping coefficient is introduced, set to 0.85 (an industry-standard value) to prevent excessive concentration of scores. Each node's new centrality score equals the damping coefficient × (the sum of scores assigned to neighboring nodes) + (1 - damping coefficient) × initial score. This iteration is repeated until convergence. The iteration stops when the change in centrality score for all nodes is less than 0.001 in two consecutive iterations. At this point, the final score for each node is its centrality. The centrality of the top 10% of nodes is selected as key nodes, and their centrality scores are used as key node features. After extracting the graph structure features, these features need to be merged with user influence features, specific risk topic features, and dynamic features into a structured feature set. The aforementioned features are the propagation depth features, propagation scale features, and key node features extracted by the graph inference engine. The user influence feature is the multi-dimensional influence score generated by the user entity feature engine. The specific risk topic feature is the risk topic tag and topic matching similarity output by the text semantic feature engine. The dynamic feature is the sentiment fluctuation variance and posting frequency change rate calculated by the real-time statistical feature engine. The merging process is as follows: First, all features are associated by user and event dimensions. Using the user node ID as the core key, the influence features and the depth / scale of the public opinion event in which the user participated are associated. If it is a key node, the centrality score is added. At the same time, using the event ID as the auxiliary key, the specific risk topic tags, topic similarity and dynamic features of the event are associated. Then, all features are standardized and the categorical features are one-hot encoded. Finally, according to the field structure of "user ID, event ID, user influence feature group, risk topic feature group, dynamic feature group, dissemination depth feature group, dissemination scale feature group, key node feature group, feature merging timestamp", all features are integrated into a structured feature set in JSON format and stored in the feature repository to provide standardized input for the customized analysis pipeline of the subsequent model factory.

[0027] After constructing the structured feature set, the model factory determines the specific type of public opinion event through the scenario recognition module. This module employs a lightweight cascaded classifier, achieving accurate event type determination through two steps: coarse screening and fine classification. First, it narrows the scope based on risk themes at the content dimension, then combines scale characteristics at the dissemination dimension to pinpoint the event type. Subsequently, the model scheduler and pipeline assembler collaborate to match corresponding components according to the event type and generate a customized analysis pipeline based on component dependencies, ensuring the analysis process is adapted to the event characteristics. The lightweight cascaded classifier uses low-computational-complexity rule-based classification logic, adaptable to real-time public opinion analysis needs. Specifically, it is executed in two steps: primary event coarse screening and secondary fine classification. The first-level event coarse screening is implemented based on specific risk theme characteristics. It extracts risk theme tags and matching similarities of the current public opinion event from the structured feature set, completes the screening according to the preset threshold, retains the candidate event types that meet the requirements, and filters out irrelevant events to reduce the computational pressure of subsequent fine classification. The second-level fine classification integrates the propagation scale characteristics to complete the event type determination. The propagation scale characteristics include the maximum number of nodes in the connected subgraph, the number of edges in the subgraph, and the subgraph density. First, the propagation scale characteristics are defined hierarchically, and then the mapping rules between candidate event types and propagation scale hierarchies are constructed. The candidate event types are matched with the propagation scale hierarchies of the current event to obtain the corresponding preset public opinion event tags, and finally complete the accurate determination of the event type.

[0028] After the context recognition module outputs the public opinion event type, the model scheduler first needs to match the corresponding analysis component. Then, the pipeline assembler generates the execution order through directed acyclic graph topology sorting. The core is to ensure that the components run according to causal dependencies to avoid process chaos. The model scheduler has a built-in event type and component mapping rule table. The event type and component mapping rule table is a pre-configured structured table stored in the model factory's metadata database. The table structure includes three columns: "Event Type", "Core Analysis Component List", and "Component Dependency Hint". The event type column corresponds exactly to the preset public opinion event tags output by the context recognition module. The core analysis component list column stores standardized components adapted to the event type. These components all come from the model factory's standardized model component library and are reusable. The component dependency hint column briefly marks the key dependencies between components, providing a basis for subsequent pipeline assembly. When the event type output by the context recognition module is received, the model scheduler extracts the corresponding core analysis component list and dependency hint by accurately matching the event type column and passes it to the pipeline assembler. The pipeline assembler generates customized analysis pipelines through topological sorting of a directed acyclic graph (DAG). A DAG is an acyclic graph structure consisting of component nodes and directed edges. Component nodes represent core analysis components extracted from the mapping table, and directed edges represent causal dependencies between components. Topological sorting is an algorithm that sorts the DAG nodes to avoid analysis failures caused by reversed execution order. The specific generation process is as follows: Based on the component list and dependency hints provided by the model scheduler, and combined with the pre-defined causal dependencies in the domain knowledge base, the pre-dependent components of each component are determined. A Directed Acyclic Graph (DAG) is drawn, with components as nodes and dependencies as directed edges. For example, node A: user risk level assessment component, node B: fund flow related component, edge A→B; node C: key propagation node location component, node D: risk diffusion prediction component, edge C→D. The Kahn algorithm is used for sorting. First, the in-degree of each component node is calculated, along with the number of directed edges pointing to that node. For example, node A has an in-degree of 0. Node B has an in-degree of 1. Nodes with an in-degree of 0 are added to the queue. A node is taken from the queue and added to the pipeline execution order list. Then, all outgoing edges of that node are deleted, and the in-degree of the dependent nodes is updated. The in-degree of node B changes from 1 to 0, and the in-degree of node D changes from 1 to 0. The above operations are repeated until the queue is empty and all nodes are added to the execution order list. The pipeline execution order list obtained by topological sorting is transformed into an executable process script, clarifying the input, output, and running parameters of each component, and finally forming a customized analysis pipeline adapted to the current event type.

[0029] When a customized analysis pipeline is running, if the entire structured feature set is processed serially, the high feature dimensionality can easily lead to computational delays, failing to meet the needs of real-time public opinion analysis. Therefore, a feature sharding parallel processing mechanism is introduced. By splitting features according to data modality and allocating dedicated components for parallel computation, processing efficiency is significantly improved. At the same time, by leveraging the learnable weight tensor of the scoring fusion module, the output of multiple components is transformed into standardized three-category risk scores, ensuring the objectivity and accuracy of the assessment results. The specific implementation method is as follows: The structured feature set is modally split into user entity slices, text semantic slices, and graph structure slices. Modality refers to the source and type dimension of feature data. Different modal features correspond to different core perspectives of public opinion analysis. After splitting, components adapted to each modality can focus on processing the corresponding data and avoid cross-modal data interference. The user entity slice focuses on the source (user) dimension, including all user-related features extracted from the structured feature set, specifically the core indicators of the user multidimensional influence assessment matrix, the identifier of whether the user is a key node, and the centrality score of the key node. These features directly reflect the potential risk of users as public opinion sources. The text semantic slice focuses on the content dimension, including specific risk topic tags, topic matching similarity (semantic similarity value between text and risk topic), and text sentiment score (basic data for calculating sentiment fluctuations output by the real-time statistical feature engine), which is the core support for content risk assessment. The graph structure slice focuses on the propagation dimension, including propagation depth features (average propagation depth, maximum propagation depth, path diversity), propagation scale features (maximum number of nodes in the connected subgraph, number of edges, subgraph density), and average weight of propagation relationship edges, which are used to quantify the risk level of the public opinion propagation process.

[0030] The specific steps for the splitting process are as follows: A predefined feature mapping rule table is used to clarify the modality to which each structured feature field belongs. The rule table is stored in the metadata management module of the model factory and can be dynamically updated as industry sentiment features are added. The structured feature set of the current sentiment event is retrieved from the feature repository, and the modality to which each field belongs is matched according to the rule table. The fields and their corresponding values ​​are written to three temporary data containers respectively. Numerical features of user entity segments and graph structure segments are retained to two decimal places, and categorical features of text semantic segments are converted to string format, and similarity values ​​are retained to three decimal places to ensure that subsequent components can directly read them. A segment integrity check is performed to compare the feature fields actually contained in each segment with the preset fields in the rule table. If there are any missing fields, they are automatically filled from the structured feature set. If the filling fails, an alarm is triggered and the flow is terminated. To prevent incomplete data processing by components, after successful verification, the three shards are temporarily stored in the pipeline's temporary data cache in JSON format, awaiting component invocation. Subsequently, the three shards are input into selected components for parallel computation. These selected components are standardized components adapted to various modalities, matched by the model scheduler based on public opinion event type matching. The user entity shard corresponds to the source risk assessment component, focusing on analyzing user risk; the text semantic shard corresponds to the content risk identification component, focusing on judging content risk; and the graph structure shard corresponds to the propagation risk analysis component, focusing on assessing propagation risk. All components come from the standardized model component library of the model factory, possessing unified input / output interfaces to ensure data interaction compatibility during parallel processing. Parallel computation is implemented using a multi-threaded processing framework. The specific process is as follows: The pipeline starts three independent threads, each bound to a source risk assessment component and a user entity segment, a content risk identification component and a text semantic segment, and a dissemination risk analysis component and a graph structure segment, respectively. Each thread synchronously reads the corresponding segment data, and the components process independently according to preset logic. The source risk assessment component calculates the user's source risk base score (range 0-100) through weighted summation, with the weights of each influence dimension determined based on industry data training. The content risk identification component multiplies the topic matching similarity by 100, combines it with the sentiment score, positively adding negative sentiment scores and negatively canceling out positive scores, and outputs the content risk base score (0-100) and the dissemination risk score. The analysis components set scoring rules based on the scale and depth of dissemination and output a basic score (0-100) for dissemination risk. After each component completes its processing, it writes its own output basic score, calculation basis, and processing timestamp into the intermediate result library of the pipeline in the format of "component ID-basic score-details-timestamp". During parallel computing, the pipeline monitoring module tracks the progress of each thread in real time. If a thread times out, it is restarted and the fragmented data is reread to ensure that all three components can output valid results. Finally, the scoring fusion module uses a learnable weight tensor to weight and fuse the component output results to generate three types of quantitative scores: source risk score, content risk score, and dissemination risk score.

[0031] The learnable weight tensor is a 3×3 weight matrix, with rows corresponding to three types of basic scores and columns corresponding to three types of final risk scores. Each weight value in the matrix is ​​obtained through training on historical public opinion samples from the model factory. The training data consists of labeled public opinion events from the past 6 months, including the basic scores output by the components and the final risk scores determined by humans. The training objective is to minimize the error between the weighted fusion predicted score and the manually labeled score, so that the weights can adapt to the importance of each basic score under different public opinion scenarios. The weight tensor is stored in the parameter management module of the model factory and is updated monthly based on newly labeled samples. The specific fusion process is as follows: The scoring fusion module reads the base scores of the three components from the intermediate result library and retrieves the latest learnable weight tensor. Assuming the current tensor is [[0.8,0.1,0.1],[0.15,0.7,0.15],[0.05,0.1,0.85]], where the first row [0.8,0.1,0.1] represents the weights of the source base score corresponding to the three final scores; the second row [0.15,0.7,0.15] represents the weights of the content base score; and the third row [0.05,0.1,0.85] represents the weights of the propagation base score. The final risk score is calculated as S1 × corresponding weight + S2 × corresponding weight + S3 × corresponding weight, resulting in the three quantitative scores and the source risk score. The risk score is calculated as S1×0.8+S2×0.15+S3×0.05, focusing on the comprehensive risk of the information source dimension. The content risk score is calculated as S1×0.1+S2×0.7+S3×0.1, focusing on the comprehensive risk of the content dimension. The dissemination risk score is calculated as S1×0.1+S2×0.15+S3×0.85, focusing on the comprehensive risk of the dissemination dimension. The calculation results are normalized to ensure that all three scores fall within the range of 0-100 points, and one decimal place is retained. Finally, the three quantitative scores, their corresponding weight tensor versions, and the fusion timestamp are packaged into a risk score report, which is then passed to the subsequent rule inference engine for risk determination. At the same time, the results are stored in the results warehouse for subsequent retrospective analysis.

[0032] After the scoring fusion module generates three types of quantitative scores—source risk score, content risk score, and dissemination risk score—the rule-based reasoning engine needs to make an initial judgment on public opinion risk based on preset judgment rules. Through a hierarchical conditional expression storage structure, it first completes basic screening using scoring thresholds, and then combines event type tags to achieve accurate judgment, avoiding misjudgments caused by a single rule. When a high-risk event is determined, the causal analysis module immediately initiates reverse propagation path tracing, locating core propagation nodes and path abrupt change points from the propagation relationship graph, providing in-depth causal evidence for subsequent early warnings. The specific implementation method is as follows: The judgment rules loaded by the rule reasoning engine adopt a hierarchical conditional expression storage structure. This structure breaks down the judgment rules into two layers of interconnected conditional logic. Each layer is stored in the rule database as a parsable expression, which facilitates dynamic rule updates based on changes in industry public opinion and supports rapid matching calculations. The first layer of rules compares thresholds for three types of quantitative scores. The core is to filter out obviously low-risk public opinion events through basic scores, reducing the computational load of subsequent in-depth judgments. The three types of quantitative scores are: source risk score (0-100 points, quantifying the risk level of the user's source; the higher the score, the greater the likelihood that the user will cause risk), content risk score (0-100 points, quantifying the risk level of the public opinion content; the higher the score, the higher the probability that the content contains risky information), and dissemination risk score (0-100 points, quantifying the risk intensity of the public opinion dissemination process; the higher the score, the greater the likelihood that the dissemination range is wide and the spread is fast).

[0033] The specific judgment process of the first-level rules is as follows: First, statistical analysis is performed based on historical public opinion data from the past 12 months to screen all event samples marked as "low risk" and "high risk". The distribution characteristics of the three types of scores in each type of sample are calculated, and finally, the thresholds are determined. The source risk score threshold is set to 45 (the source risk score of more than 90% of low-risk events is less than or equal to 45), the content risk score threshold is set to 50 (the content risk score of more than 92% of low-risk events is less than or equal to 50), and the dissemination risk score threshold is set to 40 (the dissemination risk score of more than 88% of low-risk events is less than or equal to 40). These thresholds are stored in the first-level threshold table of the rule database and are recalibrated monthly based on new samples. Then, the rule inference engine retrieves the three types of quantitative scores of the current event from the score fusion module and calculates them according to the expression "(source risk score greater than 45) OR A threshold comparison is performed between "(content risk score greater than 50) OR (dissemination risk score greater than 40)". If none of the three scores exceed the corresponding threshold, it is directly marked as "preliminary low risk" without proceeding to the second-level rule. If at least one score exceeds the threshold, it is judged as "requiring further analysis". The result, along with the three scores, is then passed to the second-level rule to complete the basic screening. The second-level rule associates the event type label with the event type to perform a precise risk assessment on the "requiring further analysis" event, ultimately outputting two states: low risk and high risk. The event type label is a preset public opinion event label output by the scenario recognition module. Different event types have different risk impact ranges and degrees of harm, requiring adaptation to differentiated judgment standards. Therefore, the core of the second-level rule is to establish the association logic of "event type - score threshold adjustment - risk judgment". The specific implementation process is as follows: First, a second-level association rule table is constructed in the rule database. The table structure includes "event type label, source threshold adjustment coefficient, content threshold adjustment coefficient, propagation threshold adjustment coefficient, and high-risk judgment condition." The high-risk judgment condition is: source risk score greater than the adjusted source threshold AND content risk score greater than the adjusted content threshold OR propagation risk score greater than the adjusted propagation threshold AND content risk score greater than the adjusted content threshold. The risk judgment condition is: propagation risk score greater than the adjusted propagation threshold AND content risk score greater than the adjusted content threshold. When the rule inference engine receives a result indicating "further analysis required," it first retrieves the event type label output by the scenario recognition module, matches the corresponding adjustment coefficient and judgment condition in the second-level association rule table, and then compares the three quantitative scores of the current event with the adjusted thresholds. If the high-risk judgment condition is met, a "high-risk" status is output; otherwise, a "low-risk" status is output, completing the initial risk assessment. When the rule inference engine determines it to be a high-risk event, the causal analysis module immediately initiates reverse propagation path tracing. Reverse propagation path tracing refers to starting from the source node of the current public opinion event and tracing along the reverse edge of the propagation relationship graph, i.e., the platform in the graph. The constructed backward propagation relationship traces the information dissemination path, identifying key nodes and abrupt changes that drive the outbreak of public opinion. The core principle is to uncover the root causes of public opinion spread through a logic of finding the reasons behind the results. The specific tracking process revolves around the propagation relationship graph. First, the current event source node is determined. The event source node is the user node that initially published risky content, based on the content publication timestamp and propagation path tracing markers, when the graph platform constructs the propagation relationship graph. This node information is stored in the source node marker table of the graph platform, which can be directly retrieved by the causal analysis module. Then, starting from the source node, the process proceeds along the centrality index... The key nodes are visited in descending order. The centrality index is the node importance score obtained by the graph reasoning engine based on the weighted PageRank algorithm. The descending order means that all nodes in the propagation relationship graph that have direct or indirect propagation relationship with the source node are sorted from high to low according to their centrality scores. Then, the nodes are visited in order according to the sorting results. The propagation path between the node and the source node is recorded each time, and the basic information and propagation relationship of the node are extracted. This process continues until the nodes with the top 20% centrality are visited. This approach balances the coverage of core nodes and computational efficiency. The top 20% of nodes can cover more than 90% of the propagation influence.

[0034] Finally, by comparing the differences in dynamic characteristics between nodes, the core propagation nodes and abrupt changes in the path leading to the outbreak of public opinion were identified. Dynamic characteristic differences refer to the numerical changes in dynamic characteristics of adjacent access nodes. Dynamic characteristics are the source sentiment fluctuation variance and posting frequency change rate calculated by the real-time statistical feature engine. The source sentiment fluctuation variance quantifies the emotional stability of the content published by the node, while the posting frequency change rate quantifies the change in the node's activity level during the spread of public opinion. A higher change rate indicates a sudden increase in node activity. The specific comparison process is as follows: First, dynamic feature data of each node in the access path is retrieved from the feature repository. Then, the feature difference between adjacent nodes is calculated, and a feature difference threshold is preset. If the feature difference between adjacent nodes exceeds the threshold, the next node is determined to be a path mutation point. The dynamic feature mutation of this node is the key to causing public opinion to change from a slow spread to an explosive spread. At the same time, nodes with high centrality and obvious feature mutations are core propagation nodes. These nodes are not only key hubs in the propagation, but also the core triggers for the outbreak of public opinion. Through the above process, the causal analysis module can output a list of core propagation nodes, the location of path mutation points and the corresponding feature mutation details, providing in-depth causal basis for subsequent warning level calculation.

[0035] The specific implementation method for constructing the three-dimensional decision matrix and generating the basis report for the early warning level calculator is as follows: After the causal analysis module locates the core propagation nodes and path abrupt change points of high-risk events, the early warning level calculator needs to construct a three-dimensional decision matrix based on multi-dimensional key information. By integrating three core indicators—risk assessment results, the scale of core propagation nodes, and the growth trend of propagation risk—and using a fuzzy comprehensive evaluation algorithm, it achieves accurate quantification of risk levels. Simultaneously, it correlates key characteristics of causal analysis with event types to generate a traceable basis report, providing clear guidance for public opinion response. The specific implementation method is as follows: The three-dimensional decision matrix constructed by the early warning level calculator includes rule hit results, the number of key nodes, and the incremental risk score. These three dimensions quantify the severity of high-risk events from the perspectives of "matching degree of judgment criteria," "core propagation scale," and "risk growth rate," respectively, ensuring that the evaluation dimensions are comprehensive and each has its own emphasis. First, the definitions and value logic of each dimension are clarified. Rule hit results refer to the degree of matching between the number of high-risk judgment conditions actually met when the rule reasoning engine determines high risk and the number of preset total conditions for this event type. There are two total conditions. If the current event meets two conditions, it is considered a "complete hit," and if it meets one condition, it is considered a "partial hit." This dimension directly reflects the rigor of risk judgment. The number of key nodes is the total number of core propagation nodes that are located in the top 20% of the centrality index and exhibit significant abrupt changes in dynamic characteristics, as identified by the causal analysis module through backpropagation path tracing. These nodes are the key hubs driving the outbreak of public opinion; a larger number indicates a stronger core driving force for the spread of public opinion. The increment of the propagation risk score is the difference between the propagation risk score of the current high-risk event and the average propagation risk score of similar low-risk events in the past 6 months. A larger difference indicates a higher increase in the propagation risk of the current event compared to regular low-risk events, and a faster rate of risk spread. In specific implementation, it is necessary to first standardize and quantify each dimension, converting qualitative or count information into values ​​in the 0-1 range to provide a unified benchmark for matrix calculation. For rule hit results, a preset quantization value of 1.0 corresponds to a complete hit, and a quantization value of 0.6 corresponds to a partial hit. For partially met conditions, the quantization value is directly extracted from the rule reasoning engine's judgment records. The engine simultaneously records the number of conditions met. Regarding the number of key nodes, based on historical data statistics of high-risk events over the past 12 months, the number is divided into three levels: "small number (1-3)," "medium number (4-8)," and "large number (9 or more)," corresponding to quantization values ​​of 0.4, 0.7, and 1.0 respectively. After quantification of each dimension, the three quantization values ​​are combined into a three-dimensional vector according to the rule hit results and the number of key nodes in the order of the risk score increment. This vector serves as the input vector for the three-dimensional decision matrix. After quantification of the input dimensions, a comprehensive warning level is generated using a preset fuzzy comprehensive evaluation algorithm. The fuzzy comprehensive evaluation algorithm is a mathematical method for handling multi-factor fuzzy evaluation problems. It can transform fuzzy information from multiple dimensions into clear evaluation results, adapting to the characteristic of fuzzy boundaries in public opinion risk level determination. The specific implementation steps are as follows: The factor set and weight set are determined. The factor set consists of three input dimensions: rule hit result U1, number of key nodes U2, and incremental risk score U3. The weight set is determined by back-calculation based on expert experience and the response effects of historical high-risk events. A comment set and membership matrix are established. The comment set represents the preset comprehensive warning levels, divided into four levels: "Blue Warning (Low Urgency)," "Yellow Warning (Medium Urgency)," "Orange Warning (High Urgency)," and "Red Warning (Extremely High Urgency)." The membership matrix is ​​calculated by statistically analyzing historical data to determine the degree of membership of each input dimension to different warning levels, thus linking the weight set and membership matrix. The membership matrix is ​​subjected to fuzzy multiplication to obtain the comprehensive membership degree of each warning level. The comprehensive warning level is determined, and the level with the highest comprehensive membership degree is selected as the result. Simultaneously with generating the comprehensive warning level, the warning level calculator automatically associates path mutation point features with event type labels to generate a complete report. Path mutation point features are key information about mutation points located by the causal analysis module, including mutation point node ID, corresponding user account, dynamic feature mutation details, and the timestamp of the mutation. Event type labels are refined labels output by the scenario recognition module. The process of generating the associated report is as follows: First, path abrupt change point feature data is retrieved from the causal analysis module, and event type labels and typical risk characteristics of this type of event are retrieved from the scenario identification module. Then, the input dimension details of the three-dimensional decision matrix (rule hit status, number and list of key nodes, calculation process of propagation risk score increment), key parameters of fuzzy comprehensive evaluation (weight set, membership matrix, membership of each level), path abrupt change point features, event type labels and typical features are integrated into a structured report. The report structure is divided into four parts: "Warning Level Result", "Judgment Basis", "Details of Core Propagation Nodes", and "Path Abrupt Change Point Analysis". The judgment basis must clearly explain how each input dimension affects the level result (e.g., "Because the number of key nodes reaches 5 (medium, quantitative value 0.7), the propagation risk score increment is 35 points (large increment, quantitative value 0.9), and the comprehensive membership degree is the highest red warning, it is judged as a red warning"). Finally, the report is exported as a PDF or HTML format, stored synchronously in the results repository, and pushed to the public opinion response terminal to ensure that staff can clearly understand the origin of the warning level and the core risk points, providing a basis for the formulation of subsequent response measures.

[0036] In this invention, the feature calculation layer extracts user influence, specific risk themes, and dynamic features in parallel through multiple engines. The graph platform synchronously constructs a propagation relationship graph, the graph reasoning engine extracts graph structure features, generates a structured feature set, and inputs the feature set into the model factory. The scenario recognition module determines the type of public opinion event, and the model scheduler and pipeline assembler generate a customized analysis pipeline. After parallel processing of feature fragments, the scoring fusion module outputs three types of quantitative scores: source, content, and propagation. The rule reasoning engine combines the scores with the event type to determine the risk. High-risk events trigger a causal analysis module to trace core propagation nodes and path mutation points, generate a comprehensive early warning level and basis report, forming a self-evolving closed loop, improving the accuracy of public opinion analysis and the timeliness of early warning.

[0037] 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 preferred examples and are not intended to limit 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 the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for in-depth analysis and early warning of public opinion based on a multimodal model factory, characterized in that: Includes the following steps: S1. Multi-source public opinion data is processed through the feature calculation layer. Among them, the user entity feature engine extracts user influence features, the text semantic feature engine injects domain knowledge to identify specific risk topics, and the real-time statistical feature engine calculates dynamic features. At the same time, the graph platform constructs a propagation relationship graph in parallel and extracts the graph structure features in user influence features, specific risk topics, and dynamic features through the graph reasoning engine to generate a structured feature set. S2. Input the structured feature set into the standardized model component library pre-set in the model factory. The scenario recognition module of the standardized model component library judges the structured feature set and outputs the public opinion event type. The model scheduler selects components according to the public opinion event type, and the pipeline assembler generates a customized analysis pipeline. The customized analysis pipeline performs multi-model collaborative processing on the structured feature set and outputs a quantitative score through the scoring fusion module. S3. The rule reasoning engine loads the preset judgment rules and combines them with the quantitative score to determine the risk. For high-risk events, the causal analysis module is triggered, and key nodes are traced based on the propagation path of the platform in the figure. The warning level calculator generates a comprehensive warning level and basis report based on the causal conclusions obtained by the causal analysis module through weighted calculation. S4. The comprehensive early warning level and the report are fed back to the feature warehouse as incremental samples to evaluate the effectiveness of each model component in the current pipeline and drive the model factory to perform incremental optimization of the components, forming a self-evolving closed loop.

2. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 1, characterized in that: The user entity feature engine constructs a multi-dimensional influence assessment matrix by aggregating user registration information and historical behavior data to extract user influence features. The text semantic feature engine injects industry risk keyword library into a pre-built domain knowledge graph and uses an attention mechanism-enhanced topic matching model to identify specific risk topics. The real-time statistical feature engine dynamically calculates the source sentiment fluctuation variance and posting frequency change rate as dynamic features based on a sliding time window.

3. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 1, characterized in that: When constructing the propagation relationship graph in parallel, the user entity is used as the node and the forward and backward propagation relationship is used as the edge. The node attributes inherit the user influence characteristics, and the edge weight is dynamically calculated based on the forwarding delay attenuation coefficient and content similarity. The propagation relationship graph adopts a dynamic graph structure storage engine to update the state changes of nodes and edges in real time.

4. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 3, characterized in that: The graph reasoning engine performs multi-hop neighbor sampling and subgraph aggregation operations on the propagation relationship graph, extracts information propagation depth features through a random walk algorithm, calculates propagation scale features using a maximum connected subgraph detector, and iteratively solves the centrality index of key nodes based on the weighted PageRank algorithm. The aforementioned features are merged with the user influence features, risk themes, and dynamic features into a structured feature set.

5. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 1, characterized in that: The scenario recognition module employs a lightweight cascaded classifier, which specifically includes: firstly, performing a primary event coarse screening based on the specific risk theme features, and then performing a secondary fine classification by integrating the dissemination scale features, and outputting preset public opinion event tags.

6. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 5, characterized in that: The model scheduler has built-in event type and component mapping rule tables, and the pipeline assembler generates a customized analysis pipeline through directed acyclic graph topology sorting, wherein the component connection order is determined by the preset causal dependencies in the domain knowledge base.

7. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 6, characterized in that: When the customized analysis pipeline runs, a feature segmentation parallel processing mechanism is adopted, which specifically includes: splitting the structured feature set into user entity segments, text semantic segments and graph structure segments according to modality, inputting them into the selected components for parallel computation, and the scoring fusion module weighting and fusing the component output results through learnable weight tensors to generate three types of quantitative scores: source risk score, content risk score and dissemination risk score.

8. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 7, characterized in that: The judgment rules loaded by the rule reasoning engine adopt a hierarchical conditional expression storage structure, which specifically includes: the first layer of rules compares the thresholds of three types of quantitative scores, the second layer of rules associates event type labels, and the initial risk judgment result includes two states: low risk and high risk.

9. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory as described in claim 8, characterized in that: When an event is identified as high-risk, the causal analysis module initiates reverse propagation path tracing. This includes starting from the current event source node based on the propagation relationship graph, visiting key nodes in descending order of centrality index, and finally locating the core propagation nodes and path mutation points that led to the outbreak of public opinion by comparing the differences in dynamic characteristics between nodes.

10. The method for in-depth analysis and early warning of public opinion based on a multimodal model factory according to claim 8, characterized in that: The early warning level calculator constructs a three-dimensional decision matrix, specifically including: The matrix input dimensions include rule hit results, number of key nodes, and incremental risk of propagation. It generates a comprehensive early warning level through a preset fuzzy comprehensive evaluation algorithm, automatically associates path mutation point features with event type labels, and generates a report based on the findings.

Images