Risk transmission prediction method and system based on joint deduction of timing diagram and large model

Abstract

The application provides a risk transmission prediction method and system based on a time sequence diagram and a large model joint deduction, relates to the technical field of artificial intelligence and knowledge engineering, and comprises the following steps: extracting forward-looking time sequence facts based on hierarchical prompt words and a structured constraint mechanism; extracting conflict events and generating time sequence relationships; updating a historical time sequence knowledge graph based on a time sequence relationship and a confidence weighted update strategy; determining a central entity, initiating a structured query to the updated historical time sequence knowledge graph from the central entity, and generating a context knowledge subgraph; converting the context knowledge subgraph into a natural language description with logical relationships, injecting a risk hypothetical event, constructing a prompt word as a new input of a large model, and outputting a structured JSON object containing a complete reasoning chain; and based on a risk path refining and visualizing algorithm with influence weight attenuation, converting the structured JSON object to obtain risk warning information and a visualized transmission path diagram.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and knowledge engineering, and particularly relates to a risk transmission prediction method and system based on joint deduction of time sequence diagram and large model. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] In today's rapid development of globalization and digitization, the complexity and vulnerability of various systems (especially global supply chain networks) are increasingly prominent. A small, local disturbance, such as a production accident of a key supplier, can quickly transmit, amplify, and eventually trigger systemic and catastrophic disruption risks through complex and opaque associated networks. Therefore, early, accurate, and interpretable prediction of such risks has become the core demand for enterprises to maintain competitiveness and security resilience. In traditional risk early warning methods, early statistical model and expert system methods mainly rely on statistical analysis of structured information such as historical financial data, inventory data (such as time series prediction), and combined with industry expert rule base. The shortcomings of such methods are: severe dependence on structured data, unable to utilize the explosive growth of unstructured information such as news, announcements, research reports on the Internet, leading to serious lag in risk perception; at the same time, expert rules are difficult to maintain and extend, and cannot cope with rapidly changing and highly complex modern supply chain networks.

[0004] With the development of technology, there are schemes for risk prediction using machine learning models (such as LSTM) and graph computing technology (such as graph neural network GNN). For example, by constructing a supply chain into a graph, GNN is used to learn how disruption events in history propagate in the graph. Such schemes have improved prediction accuracy compared to traditional methods, but existing machine learning methods still have the following defects: (1) lack of explainability: the prediction process of deep learning models such as GNN is a "black box", which can tell you that the chain break risk probability of company A is 90%, but cannot tell you why it is 90%, where the risk comes from, which links it goes through, and who it finally affects. This makes it difficult for decision makers to trust the results and even more difficult to intervene accurately based on them. (2) Inability to deal with "black swan" events: the upper limit of the ability of such data-driven models is limited by the "historical experience" of the training data. For those "black swan" risk events that have never happened in history or are rare (for example, the emergence of a new type of trade barrier), the model often cannot make any effective prediction due to the lack of learning samples. (3) Limitations of information sources: although some schemes have begun to try to extract events from text using NLP technology, most of them are limited to extracting "strong signal" events with clear trigger words (such as "bankruptcy", "fire"), and their ability to capture "weak signals" hidden in analysts' opinions or industry dynamics is very limited.

[0005] Recently, large language models have shown strong natural language understanding and preliminary reasoning capabilities, and some practitioners have begun to directly ask LLM questions and let it predict the impact of certain events. However, this "naive" method of direct application still has the following problems:

[0006] (1) "Fact illusion" problem: general large language models lack accurate structured knowledge of specific domains (such as the supply chain of a specific enterprise) that is updated in real time. When reasoning, its factual basis is ambiguous and unreliable, and it is prone to "fact illusion", leading to conclusions that "sound reasonable but are not consistent with the facts", which is unacceptable in serious business decisions.

[0007] (2) Shallow and unstable reasoning: without the guidance of structured knowledge, the reasoning process of LLM is often divergent and shallow, making it difficult to perform multi-step and rigorous logical deductions, and the results are unstable and sensitive to changes in prompt words.

[0008] (3) Unable to adapt to changing complex scenarios, lack of industrial-level solutions that deeply integrate time-varying and structured domain knowledge with the powerful reasoning and natural language generation capabilities of large language models. SUMMARY

[0009] The present application proposes a risk transmission prediction method and system based on joint deduction of time sequence diagram and large model to solve the above problems. Through a hierarchical prompt word mechanism, forward-looking time sequence events are extracted from unstructured data, a dynamic time sequence knowledge graph is constructed, and a large model joint deduction framework is combined. The context knowledge subgraph of the time sequence knowledge graph is used as a reliable fact basis for reasoning. By injecting a hypothetical event, the large language model is guided to perform chainable and interpretable causal reasoning, thereby simulating and deducing the complete transmission path of a risk event in a complex network.

[0010] According to some embodiments, the present application adopts the following technical solutions:

[0011] The risk transmission prediction method based on joint deduction of time sequence diagram and large model comprises:

[0012] Obtain a plurality of heterogeneous unstructured data;

[0013] Perform chain reasoning on the heterogeneous unstructured data, identify and extract forward-looking impact events based on a hierarchical prompt word and a structured constraint mechanism, and format them into time sequence facts;

[0014] Retrieve existing facts that conflict with the time sequence facts in the historical time sequence knowledge graph, and filter and build a conflict fact set. For each conflict fact, use time sequence relation algebra to perform conflict judgment to obtain time sequence relations. Update the historical time sequence knowledge graph based on the time sequence relations and a confidence weighted update strategy;

[0015] Initiate a risk transmission prediction request and determine a central entity. Starting from the central entity, initiate a structured query to the updated historical time sequence knowledge graph to generate a context knowledge subgraph;

[0016] Based on a knowledge subgraph-natural language narrative serialization method, convert the context knowledge subgraph into a natural language description with logical relationships. Inject a risk hypothetical event into the natural language description, build a highly structured and formatted output constraint prompt word as a new input for the large model, and output a structured JSON object containing a complete reasoning chain;

[0017] Based on a risk path extraction and visualization algorithm based on impact weight decay, convert the structured JSON object to obtain risk warning information and a visual transmission path graph.

[0018] According to some embodiments, the present application adopts the following technical solutions:

[0019] The risk transmission prediction system based on joint deduction of time sequence diagram and large model comprises:

[0020] The time sequence event extraction module is used for acquiring various heterogeneous unstructured data, performing chain reasoning on the heterogeneous unstructured data, identifying and extracting forward-looking influence events based on hierarchical prompt words and a structured constraint mechanism, and formatting the forward-looking influence events into time sequence facts;

[0021] The dynamic time sequence knowledge graph management module is used for searching existing facts conflicting with the time sequence facts in the historical time sequence knowledge graph, screening and constructing a conflict fact set, performing conflict judgment on each conflict fact by using a time sequence relation algebra to obtain a time sequence relation, and updating the historical time sequence knowledge graph based on the time sequence relation and a confidence weighted updating strategy;

[0022] The large model joint deduction and prediction module is used for initiating a risk transmission prediction request and determining a central entity, initiating a structured query to the updated historical time sequence knowledge graph from the central entity as a starting point to generate a context knowledge subgraph, converting the context knowledge subgraph into a natural language description with logical relations based on a knowledge subgraph-natural language narrative serialization method, injecting risk hypothetical events into the natural language description, constructing a highly structured and formatted output constraint prompt word as a new input of the large model, and outputting a structured JSON object containing a complete reasoning chain.

[0023] The risk transmission path generation and early warning module is used for converting the structured JSON object based on an influence weight attenuation risk path refining and visualization algorithm to obtain risk early warning information and a visual transmission path graph.

[0024] Compared with the prior art, the risk transmission prediction method based on time sequence graph and large model joint deduction has the following beneficial effects:

[0025] The risk transmission prediction method based on time sequence graph and large model joint deduction of the application extracts "forward-looking" time sequence events from unstructured data by using a hierarchical prompt word mechanism and a large language model, updates and constructs a dynamic time sequence knowledge graph, takes a context knowledge subgraph of the time sequence knowledge graph as a credible fact basis for reasoning of the large language model based on a large model joint deduction framework, injects risk hypothetical events to guide the large language model to perform chain-like and explainable causal reasoning, and thus simulates and deduces a complete transmission path of a risk event in a complex network (such as a supply chain network). The method completely realizes how to start from a fuzzy and unstructured "weak signal" influence event, dynamically model the time sequence knowledge graph, and finally generate an explainable, traceable and visual precise risk early warning by using the "hypothesis reasoning" capability of the large language model.

[0026] The risk conduction prediction method based on the joint deduction of the time sequence diagram and the large model of the application, through time sequence event extraction, uses the deep reasoning ability of the large language model to identify and extract weak signals and forward-looking risk events from massive unstructured information that cannot be found by traditional methods. This makes the starting point of risk warning advance from "event has occurred" to "risk is brewing", which saves valuable time window for enterprises to take countermeasures, realizes the foresight of risk perception, and improves the timeliness of warning.

[0027] The risk conduction prediction method based on the joint deduction of the time sequence diagram and the large model of the application constructs a set of time sequence atlas and large model joint deduction framework. The framework does not depend on the existence of similar risk patterns in historical data, but uses the common sense and world knowledge of the large language model to make reasonable and logically consistent hypothetical deduction on new risk events that have never occurred in history. This can solve the problem that existing data-driven models are difficult to deal with "black swan" risks, and basically realize zero-sample reasoning of "black swan" events.

[0028] The risk conduction prediction method based on the joint deduction of the time sequence diagram and the large model of the application is completely different from the "black box" characteristics of traditional prediction models. The output of the application is not only a risk probability, but also a human-readable and logically clear reasoning chain generated by a chain causal reasoning process. Further converting the chain into a visual risk conduction path diagram makes each prediction conclusion traceable, and decision makers can clearly understand the source, conduction path and final impact of the risk, thereby greatly enhancing the trust of the warning result, realizing the "white box" explainability of the prediction process, and enhancing the credibility of the decision.

[0029] The risk conduction prediction method based on the joint deduction of the time sequence diagram and the large model of the application adopts a set of rigorous and credibility-weighted updating algorithm to maintain and update the time sequence knowledge graph. The algorithm can handle the conflict, coverage and fusion of new and old information, ensure that the context knowledge subgraph as the basis of large model reasoning is always dynamically updated, logically consistent and high quality. It provides a solid data guarantee for the accuracy of subsequent reasoning and prediction, ensures the dynamic and consistency of the knowledge base, and improves the accuracy of reasoning.

[0030] The risk conduction prediction method based on the joint deduction of the timing diagram and the large model of the present application is not a single link of the existing risk prediction technology for local optimization, but through the deep coupling of a series of originality processes such as "forward-looking event extraction", "dynamic timing knowledge modeling", "large model joint deduction" and "explainable path generation", a complete, end-to-end, from early signal perception to deep causal deduction of a new generation of risk prediction solution is provided. The core pain points of the traditional method and the existing AI model, such as risk perception lag, prediction process "black box", and inability to cope with new "black swan" risks, are solved. While significantly improving the early warning timeliness and accuracy, the transparency and explainability of the prediction process are guaranteed, which can promote the risk management field to a higher order of intelligent and cognitive decision support, and has high industrial value. BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated herein by reference. The embodiments illustrated in the drawings are provided to explain the present application and are not intended to limit the present application.

[0032] Figure 1 The overall architecture schematic diagram of the risk conduction prediction method based on the joint deduction of the timing diagram and the large model of the present application is shown in the figure.

[0033] Figure 2 The overall flowchart of the risk conduction prediction method based on the joint deduction of the timing diagram and the large model of the present application is shown in the figure.

[0034] Figure 3 The forward-looking impact event extraction flowchart of the present application is shown in the figure.

[0035] Figure 4 The core flowchart of the joint deduction framework of the present application is shown in the figure. DETAILED DESCRIPTION

[0036] The present application will be further described below in conjunction with the drawings and embodiments.

[0037] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0038] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0039] Embodiment 1

[0040] In an embodiment of the present application, a risk transmission prediction method based on joint deduction of timing diagram and large model is provided, and the method steps include:

[0041] Step one: obtain multiple heterogeneous unstructured data, chain reasoning of the heterogeneous unstructured data, identify and extract forward-looking influence events based on hierarchical prompt words and structured constraint mechanism, and format as timing facts;

[0042] Step two: retrieve existing facts conflicting with timing facts in the historical timing knowledge graph, and screen and build a conflict fact set. For each conflict fact, use timing relationship algebra to judge the conflict and obtain the timing relationship. Update the historical timing knowledge graph based on the timing relationship and the confidence weighted update strategy;

[0043] Step three: initiate a risk transmission prediction request and determine a central entity. Starting from the central entity, initiate a structured query to the updated historical timing knowledge graph, generate a context knowledge subgraph, and convert the context knowledge subgraph into a natural language description with logical relationships based on a knowledge subgraph-natural language narrative serialization method. Inject risk hypothetical events into the natural language description, build highly structured and formatted output constraints prompt words as new inputs for the large model, and output structured JSON objects containing complete reasoning chains;

[0044] Step four: convert the structured JSON object based on the risk path refinement and visualization algorithm with influence weight attenuation to obtain risk warning information and visual transmission path graph.

[0045] As an embodiment, the risk transmission prediction method based on joint deduction of timing diagram and large model uses a large language model to extract forward-looking influence events from unstructured data to update a dynamic timing knowledge graph. More importantly, a "large model joint deduction" framework is proposed: it takes the "context knowledge subgraph" of the timing knowledge graph as a reliable fact basis for the large language model to reason, and through the injection of risk hypothetical events, it guides the large language model to perform chain-like and interpretable causal reasoning, thereby simulating and deducing the complete transmission path of risk events in a complex network. The specific implementation process is as follows:

[0046] Step 1: Obtain a variety of heterogeneous unstructured data; chain reasoning on heterogeneous unstructured data, identify and extract forward-looking impact events based on hierarchical prompts and structured constraint mechanisms, and format into time sequence facts;

[0047] The core task of this step 1 is to identify and extract various events that affect the target system (such as the supply chain) from a large amount of heterogeneous and unstructured data sources in real time. The innovation of this process lies in its ability to not only identify "strong signal" events that have occurred, but also to discover and structure "weak signal" and "forward-looking" impact events that are difficult to capture by traditional methods through a method.

[0048] In one specific embodiment, the prompt sets a clear expert role for the LLM model and injects domain knowledge. The prompt breaks down the extraction task of heterogeneous unstructured data into a series of simple, continuous thinking steps, guides the model to chain thinking, and finally provides a JSON Schema constraint as an output format template. The extracted weak signal impact events with predictive properties are finally extracted, and each extracted impact event is formatted as a structured time sequence fact data unit containing event subject, object, relationship, timestamp, source, and confidence metadata.

[0049] Among them, through continuous monitoring of various data sources, including but not limited to industry news portals, relevant department announcements, listed company financial reports, securities company analysis reports, social media, and enterprise internal emails and meeting minutes. Its workflow is different from the event extraction method in the prior art that relies on explicit keywords, but adopts an event extraction method based on hierarchical prompts and structured constraint templates. Instead of asking a general question to a large language model (LLM) at once, this method uses a multi-level, step-by-step prompt chain (Prompt Chaining) to ensure the accuracy and consistency of the extraction, thereby converting the uncertainty of the LLM into a stable and reliable structured data generator. This step is implemented as follows:

[0050] First layer: role playing and domain knowledge injection.

[0051] The prompt sets a clear expert role for the LLM large language model and injects the core domain knowledge required for the task, which enables the subsequent analysis of the LLM to focus on a specific field and use the professional perspective of that field to make judgments. For example, in the supply chain risk early warning scenario, the initial part of the prompt will define: "You are a supply chain risk analysis expert specializing in the semiconductor industry. The key risk categories you need to focus on include: [geopolitical risks, production capacity bottlenecks, technology alternatives, key raw material price fluctuations]." The prompt sets a clear expert role for the LLM large language model and injects the core domain knowledge required for the task, which enables the subsequent analysis of the LLM to focus on a specific field and use the professional perspective of that field to make judgments. For example, in the supply chain risk early warning scenario, the initial part of the prompt will define: "You are a supply chain risk analysis expert specializing in the semiconductor industry. The key risk categories you need to focus on include: [geopolitical risks, production capacity bottlenecks, technology alternatives, key raw material price fluctuations]."

[0052] Second layer: Task decomposition and Chain-of-Thought guidance.

[0053] Prompt words break down complex extraction tasks into a series of simple, sequential thinking steps, guiding the model through a Chain-of-Thought (CoT). This guidance forces the model to demonstrate its logical reasoning process before giving a final conclusion, thereby improving accuracy on complex tasks. For example, the prompt continues: "Please analyze the text by following these steps: (1) Read the entire text and identify all mentioned company entities and geographic locations. (2) For each entity, find the sentence that describes its current or future state. (3) Determine if these state descriptions are related to the key risk categories we are interested in. (4) If related, determine whether this is an 'already occurred fact' or a 'future potential risk.'"

[0054] Third layer: Structured output constraint templates.

[0055] This layer is a crucial step in ensuring output quality and machine readability. The final part of the prompt provides a strict JSON Schema as an output format template, optionally supplemented by one or more "few-shot" examples to further specify the output. This strong constraint ensures that regardless of the input text, the LLM's output is uniform, structured data that can be directly parsed and consumed by downstream modules. A typical output constraint template example is as follows:

[0056] / / ... (previous two layers of prompt content)...

[0057] "Please output your analysis results strictly in the following JSON format, without adding any additional explanations:\n"

[0058] "{\n"

[0059] "\"events\": [\n"

[0060] "{\n"

[0061] "\"event_description\": \"<event text description>\",\n"

[0062] "\"entities\": [{\"name\": \"<company name>\", \"role\": \"subject / object\"}],\n"

[0063] "\"event_type\": \"<already occurred fact | future potential risk>\",\n"

[0064] "\"risk_category\": \"<geopolitical risk |...>\",\n"

[0065] "\"time_window\": {\"start\": \"YYYY-MM-DD\", \"end\": \"YYYY-MM-DD\"},\n"

[0066] "\"confidence_score\":<0.0 to 1.0 float number>,\n"

[0067] "\"source_quote\": \"<original quote>\"\n"

[0068] "}\n"

[0069] "]\n"

[0070] "}".

[0071] By the above method of hierarchical, step-by-step guidance and finally strong constraint output format, the LLM's deep semantic understanding and reasoning ability can be used to extract "weak signal" events such as "B company relies too much on its upstream supplier C, and its production stability has a high risk in the next 6 months" from analyst reports. Each extracted event is formatted as a structured "time sequence fact" data unit containing event subject, object, relationship, timestamp, source, and confidence, providing high-quality input for the next module to build a "foresight" knowledge graph.

[0072] Step 2: Retrieve existing facts that conflict with the time sequence fact in the historical time sequence knowledge graph, and filter and build a conflict fact set. For each conflict fact, use time sequence relation algebra to judge the conflict and get the time sequence relation. Update the historical time sequence knowledge graph based on the time sequence relation and the confidence weighted update strategy.

[0073] Step 2 is responsible for building, updating, storing and dynamically maintaining a temporal knowledge graph (TKG) that accurately reflects the evolution of the real world over time. The core data structure of this temporal knowledge graph is different from the traditional static triple, but adopts an enhanced, confidence-containing five-tuple form: F = (h, r, t, T, C), where h is the head entity, r is the relationship, t is the tail entity, T represents the valid time interval of the fact [T_start, T_end], and C represents the confidence score of the fact [0, 1]. The historical temporal knowledge graph records every confirmed event relationship that is valid in the historical time. For example, a fact (A company, purchase CPU, B company, [2023-01-01, 2023-12-31], 1.0) accurately records a confirmed purchase relationship that is valid in 2023.

[0074] Specifically, when a new temporal fact is extracted, first, all existing facts with the same subject-predicate-object triple (h, r, t) as the temporal fact are retrieved in the historical temporal knowledge graph to form a candidate fact set. From the candidate fact set, all conflicting facts with overlapping time intervals T_existing and T_new are selected to form a conflict fact set. For each existing fact in the conflict fact set, the module uses formal temporal relationship reasoning theory to determine one of the 13 possible temporal relationships between T_new and T_existing. Based on the determined temporal relationship and the confidence of the fact, different historical temporal knowledge graph update strategies defined by formulas are executed.

[0075] As an embodiment, the core function of Step 2 is to implement event-driven dynamic updating of the temporal knowledge graph based on a confidence-based temporal fact updating algorithm, thereby ensuring the consistency, accuracy and timeliness of the graph knowledge. When a new temporal fact F_new = (h, r, t, T_new, C_new) is received from the temporal event extraction module, the module does not simply add it, but performs the following rigorous conflict resolution and fusion process:

[0076] Step (1): Retrieve the conflict fact set;

[0077] First, all existing facts with the same subject-predicate-object triple (h, r, t) as the temporal fact F_new obtained in Step 1 are retrieved in the graph to form a candidate fact set. Then, from the candidate set, all facts with overlapping time intervals T_existing and T_new (i.e., T_existing ∩ T_new ≠ ∅) are selected to form the final conflict fact set S_conflict.

[0078] Step (2): Conflict resolution using temporal relation algebra;

[0079] For each existing fact F_existing in the conflict fact set S_conflict, the module uses Allen's Interval Algebra (a formalized theory of temporal relation reasoning) to accurately determine one of the 13 possible temporal relations between T_new and T_existing (e.g. Equals, Starts, Finishes, Overlaps, During, etc.).

[0080] Step (3): Credibility-weighted historical temporal knowledge graph update strategy and formalized decision;

[0081] Based on the determined temporal relation and the confidence of the facts, different, formula-defined update strategies are executed, rather than vague business rules.

[0082] A: Override & Invalidate strategy:

[0083] This strategy is applied when the confidence of the new temporal fact is significantly higher than the old one. The invention defines a confidence advantage function Adv(C_new, C_existing):

[0084] Adv(C_new, C_existing) = C_new / (C_existing + ε) (where ε is a small constant to prevent division by zero);

[0085] If Adv(C_new, C_existing) > θ_override (e.g. override threshold θ_override = 2.0), then the temporal fact F_new is considered to have a decisive advantage. At this time, for all F_existing overlapping with T_new, the validity of its time interval will be "cut off" by T_new. For example, if T_existing Overlaps T_new, then the end time T_existing_end of F_existing will be updated to T_new_start. The temporal fact F_new is then added to the historical temporal knowledge graph in its entirety. This ensures that information from a more reliable source (e.g. official announcement C=1.0) can override previous market rumors (e.g. C=0.5).

[0086] B: Merge & Enhance strategy:

[0087] This strategy is applied when two facts describe the same continuous state and have similar confidence (i.e. the Adv value is between 1 / θ_override and θ_override). For example, if T_new Meets T_existing (i.e. T_new_start equals T_existing_end), the two facts are merged into one new temporal fact F_merged with a longer time span. Its new confidence C_merged can be computed by a confidence fusion function, e.g. weighted average:

[0088] C_merged = (C_new * len(T_new) + C_existing * len(T_existing)) / (len(T_new) + len(T_existing)) (where len(T) denotes the length of the time interval);

[0089] This ensures that continuous descriptions of the same event from different reports can be integrated.

[0090] C: Tag&Coexist strategy:

[0091] When the new and old facts have content contradictions in the overlapping time interval, but the confidence does not reach a decisive advantage, it is considered that there is a conflict in the information. At this time, both temporal facts F_new and F_existing will be retained in the historical temporal knowledge graph, but they will be attached with a special "conflict" label, and a "conflict pointer" pointing to the ID of the other fact will be created. This enables the high-level reasoning module (such as the third module) to realize that there is uncertainty in the knowledge at this place when performing reasoning, and can make more cautious judgments accordingly.

[0092] Step 3: Initiate risk transmission prediction request and determine central entity, initiate structured query to the updated historical temporal knowledge graph from the central entity as the starting point, generate context knowledge subgraph; based on the knowledge subgraph-natural language narrative serialization method, convert the context knowledge subgraph into natural language description with logical relationship, inject risk hypothetical event in the natural language description, construct highly structured and formatted output constraint prompt words as new input of large model, output structured JSON object containing complete reasoning chain;

[0093] Step 3 proposes a joint inference framework based on contextual knowledge subgraphs and risk hypothetical events using large language models (LLM). The innovation lies in that it does not require fine-tuning or specialized training of the LLM, but rather utilizes and constrains the zero-shot reasoning capabilities of general-purpose LLMs through an engineering approach, thereby achieving "white-box" prediction of risk transmission paths.

[0094] The process initiates a risk transmission prediction request based on temporal facts (e.g., analyzing the impact of "B Company's production halt" on "A Company"), which performs a rigorous, multi-step inference process. The specific implementation process is as follows:

[0095] Step (1): Determine the central entity, and initiate a structured query to the updated historical temporal knowledge graph from the central entity. Extract all entities within the preset hop distance from the central entity and their mutual temporal facts valid up to the current time point. The query returns a local, highly relevant context knowledge subgraph for the current prediction task. This step aims to provide an accurate, complete, and well-defined fact base for subsequent LLM reasoning. First, perform "Contextual Knowledge Subgraph Extraction" operation. Instead of using the entire vast knowledge graph, the following is used:

[0096] A. Determine the central entity: According to the risk transmission prediction request, determine one or more central entities (e.g., "A Company").

[0097] B. Perform N-degree association query: Start from the central entity and initiate a structured query to the historical temporal knowledge graph. This query extracts all entities within the preset hop distance (N degrees, e.g., N=3) from the central entity and their mutual temporal facts valid up to the current time point.

[0098] C. Generate subgraph: The query returns a local, highly relevant context knowledge subgraph for the current task. This subgraph is the complete, trusted background fact set that the LLM needs to refer to for reasoning.

[0099] Step (2): Subgraph-to-Narrative Serialization, which converts the context knowledge subgraph into a natural language description with logical relationships. This step uses an innovative serialization method to convert the machine-readable context knowledge subgraph into a coherent natural language description with logical relationships. The process is as follows:

[0100] A. Determine the narrative center: The center entity of the context knowledge subgraph (e.g., "A Company") is the starting point of the narrative.

[0101] B. Relationship template conversion: Starting from the center entity, perform a breadth-first traversal. Each time a time-series knowledge graph (h, r, t, T, C) is traversed, a time-series fact-to-natural language template converter is called to convert it into a complete sentence. For example, (A Company, Purchases CPU, B Company, [2023-01-01,...], 1.0) will be converted to: "Fact: According to the exact information (confidence 1.0), since January 1, 2023, A Company's main CPU supplier is B Company."

[0102] C. Intelligent insertion of logical conjunctions: When multiple relationships point to the same entity, use logical conjunctions such as "at the same time," "in addition," "its upstream supplier is..." to organize independent sentences into a logically clear and easy-to-understand paragraph for LLM.

[0103] Finally, the input to the LLM is no longer a simple data structure, but a background fact statement similar to an expert analysis report, greatly reducing the cognitive burden of general large models to understand complex domain knowledge.

[0104] Step (3): Injecting the risk hypothetical event into the narrative description, which is a risk event that triggers the risk propagation prediction task. The risk hypothetical event is injected into the narrative description to form a highly structured prompt that includes role-playing, context injection, chain-of-thought guidance, and formatted output constraints. This prompt is used as new input for the large language prediction model, and the output is a structured JSON object that includes a reasoning chain and a final alert.

[0105] After generating the narrative "background fact statement", the following operations are performed:

[0106] A. The risk event that triggers this prediction deduction (e.g., "Risk Hypothetical Event: B Company stops production due to fire, with an estimated recovery time of 6 months") is injected into the background fact statement as new input.

[0107] B. All the above information is submitted to the large language model through a highly structured prompt that includes role-playing, context injection, chain-of-thought guidance, and formatted output constraints.

[0108] The design of this prompt is the core of how the invention "uses" rather than "trains" general LLMs. A typical prompt structure is as follows:

[0109] ### Role ###

[0110] You are a professional supply chain risk analysis expert, and all your reasoning must be strictly based on the "background fact statement" provided below;

[0111] ### Background Facts (Contextual Knowledge Subgraph) ###

[0112] <Insert the story-based natural language description generated by step two here>

[0113] ### Hypothetical Event ###

[0114] <Insert the description of the hypothetical event here>

[0115] ### Task&Chain-of-Thought Instruction ###

[0116] Based on the "background fact statements" and the "assumed event" that occurred, please **step-by-step, chain-like inference** the impact of this event on A Company's future two quarters; your reasoning must form a causal chain, clearly indicating which fact led to which new change;

[0117] ### Output Format Constraint ###

[0118] Please return your complete reasoning process and final conclusion in the following JSON format strictly, do not add any additional explanation;

[0119] <Provide detailed JSON Schema here, including reasoning_chain and final_alert fields>.

[0120] Through this rigorous framework, the invention transforms complex reasoning tasks into a "reading comprehension and continuation" task that LLMs are good at, with clear context and logical guidance. Under the guidance of this framework, the output of the large language model is no longer divergent and unreliable natural language text, but a structured JSON object that contains a complete reasoning process. This object details how each step affects a fact in the graph and triggers the next chain reaction, starting from the initial event, and ultimately assesses the impact on the target entity A Company's core business. This process simulates the deep thinking and logical reasoning process of human experts when faced with complex problems, and does not rely on any modification or training of the model itself.

[0121] Step 4: Based on the risk path extraction and visualization algorithm with influence weight attenuation, convert the structured JSON object to get risk warning information and visualized conduction path diagram.

[0122] The core task of this Step 4 is to convert the complex reasoning results output by the large model joint deduction and prediction module, which are machine-readable, into risk warning information and visualized paths that decision makers can intuitively understand, quickly focus on, and act upon. It is not a simple data format conversion, but through a set of algorithms to refine and enhance the reasoning information.

[0123] In one specific embodiment, according to the structured JSON object containing reasoning_chain (reasoning chain) and final_alert (final alert), a risk path extraction and visualization algorithm based on influence weight attenuation is started. The specific implementation process of this algorithm is as follows:

[0124] Step (1): Construction of the initial conduction path diagram;

[0125] First, the reasoning_chain array is traversed, and each JSON object in the array (representing a reasoning step) is created as a node in the graph. According to their order in the array, directed edges are established between nodes, thus constructing an initial risk transmission path graph. Each node contains original information such as event description and impact analysis for that step.

[0126] Step (2): Quantifying the distribution of impact weights;

[0127] To sort and filter the importance of the path, the algorithm uses the impact_score generated by the large language model for each step of reasoning evaluation (a floating point number in the interval [0, 1]). This score represents the strength of the causal chain or the severity of its impact for the current step. The algorithm uses this impact_score as the initial weight S of its corresponding directed edge.

[0128] Step (3): Path pruning operation based on weight decay to filter key transmission paths with high impact strength;

[0129] To avoid presenting a large number of secondary, interfering decision transmission paths (information overload) in the final visualization result, the algorithm performs a path pruning operation. This operation is based on the calculation of a cumulative impact weight W, which includes:

[0130] A. Weight initialization: Define the cumulative impact weight W(N_0) of the initial risk event (i.e. the first node N_0 in reasoning_chain) as 1.0.

[0131] B. Weight decay propagation: The algorithm propagates along the transmission path and calculates the cumulative impact weight of each subsequent node N_i. The calculation formula can be expressed as:

[0132] W(N_i) = W(N_{i-1}) * S_i

[0133] Where W(N_{i-1}) is the cumulative impact weight of the previous node, and S_i is the weight of the edge from N_{i-1} to N_i (i.e. impact_score). This formula simulates the natural decay of risk energy during transmission.

[0134] C. Perform pruning: A "minimum impact weight threshold θ_prune" (e.g. 0.1, which can be configured by the user) is preset. During the calculation process, once the cumulative impact weight W(N_i) of a node N_i is found to be less than θ_prune, the node N_i and all subsequent transmission paths will be pruned and removed from the graph.

[0135] Through this step, only those key transmission paths that maintain a high influence strength from beginning to end will be retained.

[0136] Step (4): Visualization mapping and highlighting;

[0137] After obtaining the pruned and focused risk transmission path graph, it is mapped into the final visualization chart. To further enhance readability, based on the key transmission path, the calculated weight value is mapped and highlighted to the visual attributes of the chart, obtaining the visualization risk transmission path graph, including:

[0138] A. Node highlighting: the color or size of the node is proportional to the cumulative influence weight W. The final most severely impacted node (W value is still high) will be the most eye-catching in the chart.

[0139] B. Edge highlighting: the thickness or style (such as solid vs. dashed) of the edge is proportional to its initial weight S. This allows the decision maker to see at a glance which step of transmission is the key link of risk amplification.

[0140] Step (5): Generation of structured early warning information;

[0141] Finally, based on the final alert part of the output (including risk level, summary, etc.), it is bound with the visualization risk transmission path graph generated in the previous step. It will generate a final and complete early warning package, which includes:

[0142] A. High-level early warning summary: for example, "Severe level risk: The shutdown of Tier 1 supplier B company will cause a severe impact on your core product line within 2 quarters."

[0143] B. Interactive visualization transmission path chart: users can click on the nodes in the chart to view detailed reasoning descriptions and original evidence sources for that step.

[0144] C. Recommended measures (optional): can be generated by LLM according to risk type and transmission path, or matched by the system according to the preset rule base.

[0145] The present application converts complex AI reasoning results into a clear, focused, clear, and interactive decision support "battle map" through this complete algorithm from original reasoning to final decision support information, improving the readability, credibility, and practical value of early warning information.

[0146] Embodiment 2

[0147] In an embodiment of the present application, a risk transmission prediction system based on joint deduction of time sequence diagram and large model is provided, comprising:

[0148] a time sequence event extraction module, configured to acquire a plurality of heterogeneous unstructured data, perform chain reasoning on the heterogeneous unstructured data, identify and extract forward-looking influence events based on hierarchical prompt words and a structured constraint mechanism, and format the forward-looking influence events into time sequence facts;

[0149] a dynamic time sequence knowledge graph management module, configured to search for existing facts conflicting with the time sequence facts in a historical time sequence knowledge graph, filter and construct a conflict fact set, perform conflict judgment on each conflict fact by using a time sequence relation algebra to obtain a time sequence relation, and update the historical time sequence knowledge graph based on the time sequence relation and a confidence weighted update strategy;

[0150] a large model joint deduction and prediction module, configured to initiate a risk transmission prediction request and determine a central entity, initiate a structured query to the updated historical time sequence knowledge graph starting from the central entity, and generate a context knowledge subgraph; based on a knowledge subgraph-natural language narrative serialization method, convert the context knowledge subgraph into a natural language description with logical relations, inject risk hypothetical events into the natural language description, construct a highly structured and formatted output constraint prompt word as a new input of a large model, and output a structured JSON object containing a complete reasoning chain;

[0151] a risk transmission path generation and early warning module, configured to convert the structured JSON object based on an influence weight attenuation risk path refining and visualization algorithm to obtain risk early warning information and a visual transmission path graph.

[0152] Embodiment 3

[0153] In an embodiment of the present application, a typical electronic product supply chain is taken as a background to illustrate in detail how the risk transmission prediction system based on joint deduction of a time sequence graph and a large model provided by the present application realizes the whole process of risk transmission prediction based on joint deduction of a time sequence graph and a large model from early signal discovery to deep deduction and finally to early warning.

[0154] I. Scene setting:

[0155] (1) Central entity (core enterprise): A company, a smart phone manufacturer.

[0156] (2) Key components: high-performance CPU chips.

[0157] (3) Initial supply chain state (already exists in the historical time sequence knowledge graph):

[0158] Fact 1: (A company, purchase CPU, B company, [2023-01-01,...], 1.0);

[0159] Fact 2: (B Company, purchase wafers, C Company, [2023-01-01,...], 1.0);

[0160] Fact 3: (A Company, has backup CPU supplier, none, [2023-01-01,...], 1.0);

[0161] As an embodiment, based on the above scenario setting, the specific implementation process of the risk transmission prediction method based on the joint deduction of timing diagram and large model of the present application is as follows:

[0162] Step one: weak signal based impact hypothetical event identification and timing fact generation;

[0163] In the process of continuously monitoring the whole network information by the timing event extraction module of the system, an industry analysis report about the region where C company is located is obtained, which is published by a securities company.

[0164] (1) First, the hierarchical prompt word and structured constraint template mechanism in the timing event extraction module is activated. The large language model (LLM) under the role of "semiconductor industry analyst" and CoT guidance does not find strong signal words such as "shutdown" and "interruption", but identifies a key "weak signal" description: "... affected by the increasingly tense labor relations in the region, the salary negotiations between the trade union and the major wafer manufacturer C company have reached a deadlock, analysts generally believe that if an agreement cannot be reached within the next three months, the possibility of a large-scale strike by C company is significantly increasing..."

[0165] (2) The LLM generates a "forward-looking" timing fact with confidence according to the structured output template and transmits it to the next dynamic timing knowledge graph management module. The generated timing fact is as follows:

[0166] {

[0167] "event_description": "C company faces potential risk of large-scale strike",

[0168] "entities": [{"name": "C company", "role": "subject"}],

[0169] "event_type": "future potential risk",

[0170] "risk_category": "labor relations risk",

[0171] "time_window": {"start": "2024-06-01", "end": "2024-08-31"},

[0172] "confidence_score": 0.70,

[0173] "source_quote": "..."

[0174] }。

[0175] Step two: Dynamic updating of the time-series knowledge graph;

[0176] The dynamic time-series knowledge graph management module receives the time-series fact sent by the time-series event extraction module:

[0177] (1) Since this is a completely new "potential risk" event, and there is no conflicting fact in the historical time-series knowledge graph, this new five-tuple is directly added to the historical time-series knowledge graph: F4 = (C company, faces the risk of strike, trade union negotiation deadlock, [2024-06-01, 2024-08-31], 0.70).

[0178] (2) At this time, the updated historical time-series knowledge graph not only records the current stable supply relationship, but also explicitly labels a future, potential risk hypothetical event.

[0179] Step three: Joint inference and prediction based on risk hypothetical events;

[0180] A supply chain risk manager wants to assess "what impact will the strike risk of C company have on the production of A company's mobile phones if it really happens?". Then through system interface interaction, a risk transmission prediction task is started.

[0181] (1) Trigger inference: After initiating the risk transmission prediction request, the task is sent to the large model joint inference and prediction module.

[0182] (2) Extract context knowledge subgraph: The large model joint inference and prediction module takes "A company" as the center and extracts facts 1, 2, 3, and 4 containing A, B, and C companies and their mutual relationships, forming the "background fact statement" for this inference.

[0183] (3) Inject risk hypothetical event: The inference event set by the risk manager is: "Assumption: C company will completely stop wafer production due to large-scale strikes starting from July 1, 2024, and will last for at least one quarter."

[0184] (4) Perform joint inference: The large model joint inference and prediction module submits the narrative "background fact statement" and "hypothetical event" to the general large language model through the designed "chain causal reasoning guidance" prompt words.

[0185] (5) Generate inference chain: Under the guidance of the framework, LLM outputs a structured JSON inference chain:

[0186] {

[0187] "reasoning_chain": [

[0188] { "step": 1, "event": "Company C's wafer production is interrupted", "impact": "Company C is unable to supply wafers to Company B.", "impact_score": 1.0},

[0189] { "step": 2, "event": "Company B loses wafer supply", "impact": "Company B's CPU production line will cease production within 2-4 weeks due to raw material depletion.", "impact_score": 0.95},

[0190] { "step": 3, "event": "Company B's CPU production has ceased", "impact": "Company B will be unable to fulfill its CPU supply contract with Company A.", "impact_score": 1.0},

[0191] { "step": 4, "event": "Company A is unable to obtain CPUs from Company B", "impact": "Because the knowledge base shows that Company A has no alternative suppliers, its smartphone production line will face a supply disruption of core components.", "impact_score": 1.0},

[0192] { "step": 5, "event": "Company A's production line is disrupted", "impact": "This is expected to delay the release of Company A's next-generation flagship phone by at least one quarter and result in significant revenue losses.", "impact_score": 0.85}

[0193] ],

[0194] "final_alert": { "level": "Severe", "summary": "The potential strike risk from Tier 2 supplier C will be directly transmitted to our company within 1-2 months through supplier B, causing a catastrophic impact on our core product lines."}

[0195] }。

[0196] Step four: visualization generation and early warning of risk transmission path;

[0197] The risk transmission path generation and early warning module receives the above structured JSON object:

[0198] (1) Path extraction: The risk transmission path generation and early warning module starts the risk path extraction and visualization algorithm based on influence weight decay. It calculates the cumulative influence weight sequence from step 1 to step 5 of the structured JSON reasoning chain as 1.0 -> 0.95 -> 0.95 -> 0.95 -> 0.8075. Since the weights of all nodes are higher than the set pruning threshold (such as 0.1), the entire critical path is completely retained.

[0199] (2) Visualization generation: According to these weights, a visualization transmission path graph is generated. In the graph, the "C company strike" node is the starting point, and the arrows point to "B company production stop", "A company supply stop", and finally to a "flagship mobile phone release delay" node marked in deep red. The thickness of the edge reflects the impact_score of each step.

[0200] (3) Early warning push: Finally, a complete early warning report containing the above visualization path graph, the summary text in final_alert, and the detailed reasoning steps, is pushed to the decision dashboard of the supply chain risk manager.

[0201] Through this embodiment, the present application completely demonstrates how it starts from a vague, unstructured "weak signal", models through dynamic updating of the time sequence knowledge graph, and finally generates a precise risk early warning that is explainable, traceable, and visualized, by using the "hypothesis reasoning" ability of the large language model. It proves the progress and industrial practical value of the present application compared to the prior art.

[0202] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks.

[0203] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer implemented process, so that the instructions executed on the computer or other programmable data processing devices provide a process for implementing the functions specified in the flowchart Figure 1 one flow or multiple flows and / or the functions specified in the block Figure 1 one flow or multiple flows and / or the functions specified in the block

[0204] The above describes the specific embodiments of the present application in conjunction with the drawings, but is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A risk transmission prediction method based on joint extrapolation of time series diagrams and large-scale models, characterized in that, include: Acquire diverse heterogeneous unstructured data; Chain reasoning is performed on heterogeneous unstructured data. Based on hierarchical prompts and structured constraint mechanisms, forward-looking impact events are identified and extracted, and formatted into time-series facts. By using the LLM model and using prompts to assign clear expert roles to the LLM model and inject domain knowledge, the prompts decompose the task of extracting heterogeneous unstructured data into a series of simple and continuous thinking steps, guiding the model to think in a chain. Finally, JSON Schema constraints are provided as the output format template, and the model is finally extracted to extract forward-looking, predictive weak signal influencing events. Each extracted influencing event is formatted as a structured time-series fact data unit containing metadata of event subject, object, relationship, timestamp, source and confidence level. Retrieve existing facts that conflict with chronological facts from the historical chronological knowledge graph, filter and construct a set of conflicting facts, and use chronological relation algebra to make a conflict judgment for each conflicting fact to obtain the chronological relation. The historical time-series knowledge graph is updated based on a time-series relationship and a confidence-weighted update strategy. The temporal knowledge graph is in the form of a quintuple containing confidence: F = (h, r, t, T, C), where h is the head entity, r is the relation, t is the tail entity, T represents the valid time interval of the fact [T_start, T_end], and C represents the confidence score of the fact [0, 1]. The historical temporal knowledge graph records every confirmed event relation that is valid within a historical time period. When a new temporal fact is extracted, all existing facts with the same subject-verb-object triple (h, r, t) as the temporal fact are first retrieved from the historical temporal knowledge graph to form a candidate fact set. From this candidate fact set, all conflicting facts where the time intervals T_existing and T_new overlap are selected to form a conflicting fact set. For each existing fact in the conflicting fact set, the module uses formal temporal relation reasoning theory to determine one of the 13 possible temporal relations between T_new and T_existing. Based on the determined temporal relation and the confidence of the fact, different historical temporal knowledge graph update strategies defined by formulas are executed. Initiate a risk transmission prediction request and identify the central entity. Starting from the central entity, initiate a structured query to the updated historical time-series knowledge graph to generate a contextual knowledge subgraph. The serialization method based on knowledge subgraph-natural language narrative converts the context knowledge subgraph into a natural language description with logical relationships, injects risk hypothetical events into the natural language description, constructs highly structured and formatted output constraint prompts as new inputs to the large model, and outputs a structured JSON object containing a complete inference chain. Based on the risk path extraction and visualization algorithm that affects weight decay, the structured JSON object is transformed to obtain risk warning information and a visualized transmission path diagram; The structured JSON object is processed using a risk path extraction and visualization algorithm that utilizes the decay of influence weights. This includes: first, traversing the `reasoning_chain` array of inference chains, creating a graph node for each JSON object in the array, and establishing directed edges between nodes based on their order in the array to construct an initial risk transmission path graph. The importance of the paths is then sorted and filtered, with the quantified impact score `impact_score` used as the initial weight of the directed edges. Path pruning is then employed to filter key transmission paths with high impact intensity. Based on these key transmission paths, the calculated weight values ​​are mapped to the visual attributes of the chart to obtain a visualized risk transmission path graph. The color or size of the nodes is set proportionally to their cumulative influence weight, making the most severely impacted nodes the most prominent in the path graph. The thickness or style of the edges is also set proportionally to their initial weights, thus visualizing which step of the transmission is the key link in risk amplification.

2. The risk transmission prediction method based on joint extrapolation of time series diagrams and large-scale models as described in claim 1, characterized in that, Based on this time-series fact, a risk transmission prediction request is initiated, a central entity is identified, and a structured query is initiated to the updated historical time-series knowledge graph starting from the central entity. The query extracts all entities within a preset number of hops related to the central entity and their time-series facts that are still valid up to the current time point. The query results are local context knowledge subgraphs that are highly relevant to the current prediction task.

3. The risk transmission prediction method based on joint extrapolation of time series diagrams and large-scale models as described in claim 1, characterized in that, The serialization method based on knowledge subgraph-natural language narrative converts the context knowledge subgraph into a natural language description with logical relationships. This includes: taking the central entity of the context knowledge subgraph as the starting point of the narrative, performing a breadth-first traversal starting from the central entity, and calling the time-series fact to natural language template converter to convert it into a complete sentence when each time a time-series fact is encountered. When multiple relations point to the same entity, logical connectors are used to organize the independent sentences into a logically clear natural language description that is easy for large language model LLM inference to predict.

4. The risk transmission prediction method based on joint extrapolation of time series diagrams and large-scale models as described in claim 1, characterized in that, Hypothetical risk events are injected into the natural language description. These hypothetical risk events are risk events that trigger the risk propagation prediction task. The injection of these hypothetical risk events into the natural language description forms a highly structured prompt word that includes role-playing, context injection, thought chain guidance, and formatted output constraints. This prompt word serves as a new input to the large language prediction model. The output is a structured JSON object that includes the reasoning_chain and the final_alert.

5. The risk transmission prediction method based on joint extrapolation of time series diagrams and large-scale models as described in claim 1, characterized in that, The final alert (final_alert) is bound to a visual risk transmission path diagram to generate a final, complete alert package, which includes a high-level alert summary, an interactive visual risk transmission path diagram, and recommended measures.

6. A risk transmission prediction system based on joint extrapolation of time series diagrams and large-scale models, characterized in that, Specifically, the risk transmission prediction method based on joint extrapolation of time series diagrams and large models as described in any one of claims 1-5 includes: The time-series event extraction module is used to acquire various heterogeneous unstructured data, perform chain reasoning on the heterogeneous unstructured data, identify and extract forward-looking impact events based on hierarchical prompt words and structured constraint mechanisms, and format them into time-series facts; The dynamic temporal knowledge graph management module is used to retrieve existing facts that conflict with temporal facts in the historical temporal knowledge graph, filter and construct a set of conflicting facts, and for each conflicting fact, use temporal relation algebra to make a conflict judgment to obtain the temporal relation; and update the historical temporal knowledge graph based on the temporal relation and a confidence-weighted update strategy. The large-scale model joint inference and prediction module is used to initiate risk transmission prediction requests and identify the central entity. Starting from the central entity, it initiates a structured query to the updated historical time-series knowledge graph to generate a contextual knowledge subgraph. Based on the knowledge subgraph-natural language narrative serialization method, the contextual knowledge subgraph is converted into a natural language description with logical relationships. Risk hypothetical events are injected into the natural language description, and highly structured and formatted output constraint prompts are constructed as new inputs to the large-scale model. The output is a structured JSON object containing a complete inference chain. The risk transmission path generation and early warning module is used to convert JSON objects based on the risk path extraction and visualization algorithm with diminishing impact weight to obtain risk warning information and a visualized transmission path diagram.

Images