A knowledge graph dynamic optimization method and system based on risk feedback driving

Abstract

The application provides a risk feedback driving-based knowledge graph dynamic optimization method and system, the method comprising: obtaining risk feedback data; taking insurance practitioners as core subjects, constructing a risk knowledge graph of multiple types of entities; when receiving new risk feedback data, dynamically adjusting the knowledge graph structure and outputting an updated risk knowledge graph; introducing a GraphSAGE graph neural network, fusing the node's own features and neighborhood aggregation features to form new node risk features; based on the new node risk features, establishing a joint self-optimization mechanism of relationship edge weight and node weight; based on the risk scoring model feedback result, updating the parameters of the risk scoring model; introducing retrieval enhancement generation technology and a large language model to perform semantic correction on the knowledge graph; based on the semantic corrected risk knowledge graph, outputting a risk identification result of a risk event, and improving the reliability of the risk knowledge graph optimization.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and data-driven intelligent risk control, and in particular to a knowledge graph dynamic optimization method and system based on risk feedback. Background Technology

[0002] With the rapid development of the digital economy and artificial intelligence technologies, the digital transformation of the insurance industry is deepening, the number of insurance practitioners is continuously increasing, and their professional practices are becoming increasingly complex. Risk governance is gradually shifting from traditional manual management to intelligent and digital approaches. Insurance practitioners are involved in multiple aspects of their business operations, including customer service, product sales, fund transactions, complaint handling, and professional management. Any risky behaviors such as illegal sales, false advertising, abnormal fund transactions, or frequent complaints can not only affect the normal operation of insurance institutions but also potentially harm consumers' legitimate rights and interests, and even trigger industry-wide risks. Therefore, establishing an efficient, intelligent, and dynamic risk governance system for insurance practitioners has become an important aspect of the digital governance of the insurance industry.

[0003] Currently, risk identification by insurance practitioners mainly relies on rule matching, statistical analysis, or traditional machine learning models, which integrate data such as basic information, complaint records, and professional records to conduct risk scoring. While these methods can identify some explicit risks, they generally have the following shortcomings: (1) Limited utilization of risk data. Existing technologies typically focus on structured business data as the main object of analysis, lacking unified integration of multi-source heterogeneous data such as public data, industry data, enterprise data, and financial behavior data. The relationships between data are not fully utilized, making it difficult to form a full life-cycle risk profile of insurance practitioners.

[0004] (2) Insufficient knowledge graph update capability. Most existing knowledge graphs are maintained manually or updated periodically. The node attributes, relationship edges and risk tags are not updated in a timely manner, and the graph structure cannot be dynamically corrected according to the latest risk feedback. This results in a deviation between the graph and the actual business, affecting the accuracy of risk analysis.

[0005] (3) Weak risk propagation analysis capability. Traditional risk identification methods usually focus on the analysis of a single entity, lack modeling of the relationship and risk propagation pattern, and have difficulty identifying potential risk propagation paths formed by business relationships, financial relationships, professional relationships, etc., and have insufficient ability to identify complex related risks.

[0006] (4) Risk models lack continuous learning capabilities. Most existing risk scoring models adopt offline training methods, and model parameter updates rely on manual retraining. It is difficult to absorb the latest risk feedback data in a timely manner, and the model is prone to performance degradation due to business changes and the evolution of risk characteristics, resulting in insufficient ability to identify new risk behaviors.

[0007] (5) The level of intelligence in risk analysis is not high. Existing technologies mainly rely on preset rules or fixed algorithms to determine risks, lacking the ability to combine industry knowledge, historical cases and business rules for comprehensive analysis. The generation of risk cause explanation, risk trend prediction and disposal suggestions still depends on human experience, and the ability of intelligent auxiliary decision-making is limited.

[0008] (6) Lack of closed-loop optimization mechanism. Most risk governance systems only complete risk identification or risk warning, and the risk disposal results are not effectively fed back to the knowledge graph, risk model and rule system. They cannot form a continuous learning mechanism of data feedback, knowledge update, model optimization and rule evolution, and it is difficult for the system to maintain a high level of risk identification accuracy during long-term operation.

[0009] Therefore, how to fully utilize risk feedback information to construct a risk governance method for insurance practitioners that can integrate multi-source data, dynamically update knowledge graphs, mine risk propagation relationships, continuously optimize risk models, and combine large language models to achieve intelligent semantic analysis and closed-loop learning has become an urgent technical problem to be solved in this field. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention innovatively proposes a risk feedback-driven dynamic optimization method and system for knowledge graphs, effectively solving the problem of low reliability in dynamic optimization of risk knowledge graphs caused by existing technologies, and effectively improving the reliability of dynamic optimization of risk knowledge graphs.

[0011] The first aspect of this invention provides a risk feedback-driven dynamic optimization method for knowledge graphs, comprising: Risk feedback data is acquired and preprocessed to obtain risk feedback data objects; the risk feedback data objects include insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. With insurance practitioners as the core, a risk knowledge graph of multiple types of entities is constructed; when new risk feedback data is received, the structure of the knowledge graph is dynamically adjusted according to the risk feedback results, and an updated risk knowledge graph is output. Based on the dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse the node's own features with the neighborhood aggregation features to form new node risk features. Based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous update of knowledge graph relations, nodes and risk propagation capabilities. Based on risk feedback data, dynamic graph update results, and node risk characteristics enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to update the parameters of the risk scoring model based on the risk scoring model feedback results. Based on risk feedback data, dynamic graphs, and risk feedback model scoring results, retrieval enhancement generation technology and large language models are introduced to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph. Based on the semantically corrected risk knowledge graph, the risk identification results of risk events are output.

[0012] A second aspect of this invention provides a risk feedback-driven dynamic optimization system for knowledge graphs, comprising: The data feedback acquisition module acquires risk feedback data, preprocesses the risk feedback data to obtain a risk feedback data object; the risk feedback data object includes insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. The knowledge graph update module uses insurance practitioners as the core to construct a risk knowledge graph of multiple types of entities. When new risk feedback data is received, the knowledge graph structure is dynamically adjusted according to the risk feedback results, and the updated risk knowledge graph is output. The risk propagation learning and updating module is based on the dynamically updated risk knowledge graph. It introduces the GraphSAGE graph neural network to fuse the node's own features with the neighborhood aggregation features to form new node risk features. Based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous updating of knowledge graph relations, nodes and risk propagation capabilities. The model training module establishes a risk scoring model feedback and retraining mechanism based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, and updates the parameters of the risk scoring model based on the risk scoring model feedback results. The semantic correction module, based on risk feedback data, dynamic graphs, and risk feedback model scoring results, introduces retrieval enhancement generation technology and large language models to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph. The output module, based on the semantically corrected risk knowledge graph, outputs the risk identification results of risk events.

[0013] The technical solution adopted in this invention has the following technical effects: 1. In this invention, risk feedback data is acquired and preprocessed to obtain risk feedback data objects. A risk knowledge graph of multiple entity types is constructed, with insurance practitioners as the core subject. Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the risk feedback results, outputting an updated risk knowledge graph. Based on the dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse node features with neighborhood aggregation features, forming new node risk features. Based on these new node risk features, a joint self-optimization mechanism for relation edge weights and node weights is established to realize the propagation of knowledge graph relations, nodes, and risks. Synchronous updating of force; based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to update the parameters of the risk scoring model based on the risk scoring model feedback results; based on risk feedback data, dynamic graph, and risk feedback model scoring results, retrieval enhancement generation technology and a large language model are introduced to establish a graph semantic correction mechanism to semantically correct the knowledge graph; based on the semantically corrected risk knowledge graph, the risk identification results of risk events are output, effectively solving the problem of low reliability of risk knowledge graph optimization caused by existing technologies, and effectively improving the reliability of risk knowledge graph optimization.

[0014] 2. The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for knowledge graphs. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk tags, and graph semantics, realizing the transformation of knowledge graphs from static maintenance to dynamic evolution. It can reflect changes in the risk status of insurance practitioners in real time and improve the data integrity, accuracy, and timeliness of risk knowledge graphs.

[0015] 3. The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for the knowledge graph. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk tags, and graph semantics, realizing the transformation of the knowledge graph from static maintenance to dynamic evolution. It can reflect the changes in the risk status of insurance practitioners in real time and improve the data integrity, accuracy, and timeliness of the risk knowledge graph.

[0016] 4. The technical solution of this invention proposes a joint self-optimization method for edge weights and node weights, which synchronously applies the risk feedback results to the relation edge weights and node weights, realizes the coordinated updating of relation strength, node importance and risk propagation capability, constructs a continuous learning mechanism for the knowledge graph, and enables risk relations to be dynamically adjusted according to business feedback, thereby improving the long-term operation effect of the graph.

[0017] 5. In the technical solution of this invention, the actual risk disposal results are used as supervision feedback samples. An incremental learning mechanism is introduced to continuously retrain the risk scoring model, realize the dynamic optimization of model parameters, risk characteristics and scoring rules, so that the risk scoring model can continuously learn new risk patterns, improve the model's generalization ability and long-term prediction performance, and avoid the problem of performance degradation of traditional models after long-term operation.

[0018] 6. The technical solution of this invention combines RAG retrieval enhancement technology and a large language model to perform semantic reasoning and knowledge supplementation on the risk knowledge graph, realizing risk cause explanation, risk label update, risk rule generation, and risk knowledge expansion, thereby improving the interpretability and intelligence level of risk analysis. At the same time, it effectively reduces the risk of large model illusion and improves the quality of risk assessment. Moreover, it constructs a closed-loop optimization mechanism of "risk feedback - knowledge graph update - GraphSAGE risk propagation - joint optimization of edge weights and node weights - risk model retraining - RAG semantic correction - intelligent decision-making", realizing the synergistic optimization of five levels: data, graph, model, knowledge, and decision-making. This enables the risk governance system to have the ability to continuously learn, adapt, evolve, and dynamically optimize, forming a complete intelligent risk governance closed loop.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the method of Embodiment 1 in the present invention; Figure 2 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0022] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0023] Example 1 This invention relates to the field of artificial intelligence and data-driven intelligent risk control technology, specifically to a data modeling and optimization method that integrates knowledge graphs, machine learning, and feedback learning mechanisms, and particularly to a method and system for dynamic updating of knowledge graphs driven by risk feedback signals, adaptive adjustment of relation weights, and continuous optimization of risk assessment models.

[0024] This invention is applicable to scenarios such as financial risk control, insurance anti-fraud, credit risk management, and complex correlation behavior analysis. It can be used to continuously evolve and update knowledge graphs constructed from multi-source heterogeneous data, and to dynamically correct the graph structure and edge weight parameters through feedback data, so as to improve the accuracy and timeliness of risk identification.

[0025] This invention aims to address three core problems existing in the application of knowledge graphs in risk management: (1) The graph relationships are static and cannot be dynamically updated with changes in business; (2) Edge weights and node weights are manually set and lack adaptive capability; (3) The risk model is disconnected from the actual handling results, and a closed-loop learning cannot be formed.

[0026] The overall method adopts a "risk feedback closed-loop driven graph evolution architecture", which consists of five core levels: Risk Feedback Layer; Graph Evolution Layer; Weight Self-Optimization Layer; Risk Scoring Layer; Closed-loop learning layer. The main process is as follows: Risk event generation → Risk model identification → Manual / system handling → Result feedback → Graph structure correction → Edge weight update → Model retraining → New round of risk prediction, forming a continuous evolutionary mechanism of "identification-handling-feedback-correction-optimization".

[0027] Specifically, such as Figure 1 As shown, this invention provides a risk feedback-driven dynamic optimization method for knowledge graphs, including: S1. Obtain risk feedback data, preprocess the risk feedback data to obtain a risk feedback data object; the risk feedback data object includes insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. S2, with insurance practitioners as the core, constructs a risk knowledge graph of multiple types of entities; when new risk feedback data is received, the knowledge graph structure is dynamically adjusted according to the risk feedback results, and an updated risk knowledge graph is output. S3, based on the dynamically updated risk knowledge graph, introduces the GraphSAGE graph neural network to fuse the node's own features with the neighborhood aggregation features to form new node risk features; based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous update of knowledge graph relations, nodes and risk propagation capabilities. S4. Based on risk feedback data, dynamic graph update results and node risk characteristics enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established, and the parameters of the risk scoring model are updated based on the risk scoring model feedback results. S5, based on risk feedback data, dynamic graphs and risk feedback model scoring results, introduces retrieval enhancement generation technology and large language model to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph; S6, based on the semantically corrected risk knowledge graph, outputs the risk identification results of risk events.

[0028] In step S1, risk feedback data is acquired, and the risk feedback data is preprocessed to obtain a risk feedback data object, which specifically includes: S11, standardize the format of risk feedback data from multiple business systems to form a risk feedback dataset with multiple risk feedback records; Specifically, the risk feedback data construction and standardization process in this step aims to transform the real business handling results generated during the risk governance process of insurance practitioners into unified, standardized, and computable risk feedback data, providing standardized input for subsequent dynamic correction of the knowledge graph, self-optimization of relationship weights, and continuous learning of the risk model.

[0029] ① Risk feedback data collection To address the risk governance process throughout the entire lifecycle of insurance practitioners, we continuously collect multi-source risk feedback information and construct a unified risk feedback dataset. The feedback data mainly includes the following categories: (1) Risk warning data The outputs from the risk identification model include risk score, risk level, trigger rules, warning time, risk label, and abnormal behavior characteristics.

[0030] (2) Manual review of data This includes business review information such as manual review comments, audit conclusions, risk confirmation results, rectification suggestions, and penalty levels.

[0031] (3) Business processing data This includes genuine business feedback such as complaint handling results, violation identification results, insurance cancellation status, sales misrepresentation identification, handling results of abnormal practice, and verification results of abnormal funds.

[0032] (4) Continuously track data This includes continuous monitoring information such as the completion status of rectification, changes in subsequent professional conduct, changes in complaints, risk resolution status, and the re-triggering of risk events.

[0033] All feedback data is uniformly connected to the risk feedback database and managed by numbering according to unified data standards, forming a complete risk event lifecycle record.

[0034] ② Standardization of risk feedback data Since the feedback data comes from multiple business systems, and their data formats, encoding methods, and business semantics differ significantly, the feedback data is first standardized and uniformly processed.

[0035] Specifically, it includes: (1) Standardize field names; (2) Unified data type; (3) Standardize the time format; (4) Unified organization coding; (5) Unique identification for all personnel; (6) Standardize risk category coding; (7) Unify risk level coding.

[0036] Meanwhile, missing values, outliers, and duplicate data are automatically cleaned, and data integrity and consistency are verified in conjunction with business rules to improve the quality of feedback data.

[0037] After standardization, a unified risk feedback dataset is formed:

[0038] in: Standardized risk feedback dataset; : The nth risk feedback record.

[0039] S13, map the business risk feedback results in the risk feedback record to risk feedback tags; the feedback tags include unverified risk, low risk confirmed, medium risk confirmed, and high risk confirmed. Specifically, to facilitate model calculations, business feedback results are mapped to a unified risk feedback label.

[0040] Define feedback labels as:

[0041] in: =0: Unverified risk; =1=1: Low risk confirmed; =2: Medium risk confirmed; =3: High risk confirmed.

[0042] S15. Extract risk features based on risk feedback records and risk feedback tags to form a risk feedback feature vector. The risk feedback feature vector includes risk level, risk feedback tag, risk occurrence frequency, risk duration, number of complaints, number of abnormal financial behaviors, number of abnormal professional behaviors, historical penalty records, credit rating results, and changes in business behavior. Specifically, risk features are automatically extracted from the standardized feedback data to form a unified risk feedback feature vector. The characteristics include: risk level, risk label, frequency of risk occurrence, duration of risk, number of complaints, number of abnormal financial activities, number of abnormal professional conduct, historical penalty records, credit rating results, and changes in business behavior. The final result is a standardized risk feedback sample that can be directly input into the knowledge graph update module and the risk model training module. S17. Based on the risk feedback record, risk feedback tag, and risk feedback feature vector, output a risk feedback data object; the risk feedback data object includes the insurance practitioner's unique identifier, risk feedback feature vector, risk feedback tag, risk feedback weight, and feedback time.

[0043] Specifically, after standardization, the system outputs a unified risk feedback data object:

[0044] in: Unique identifier for insurance practitioners; Risk feedback feature vector; Risk feedback tag; Risk feedback weight; Feedback time. The aforementioned risk feedback data will serve as a unified input for dynamic correction of the knowledge graph structure, update of relation edge weights, propagation of node risks, and continuous learning of the risk model, thereby achieving a data closed loop throughout the entire risk governance process and providing a reliable data foundation for the system's self-learning and self-optimization.

[0045] Furthermore, before step S15, which extracts risk features based on risk feedback records and risk feedback tags to form a risk feedback feature vector, the following steps are also included: Step S14: Determine the risk feedback weight based on the importance of the risk events in the risk feedback record; specifically, the determination of the risk feedback weight is as follows:

[0046] in, As risk feedback weight; Risk feedback level; The extent of the impact of the risk event (e.g., the amount involved, the number of customers, etc.); This is a time decay factor used to reflect the importance of recent risk feedback; , , These are the weighting coefficients for risk feedback level, the degree of impact of risk events, and time decay factor, respectively.

[0047] In step S2, the aim is to dynamically modify the knowledge graph structure based on risk feedback results, transforming the risk knowledge graph from static association to dynamic evolution. This allows the graph to continuously reflect changes in the risk relationships of insurance practitioners, improving the accuracy of risk identification and the timeliness of the graph. Step S2 specifically includes: S21, with insurance practitioners as the core subject, construct a risk knowledge graph covering multiple types of entities including personnel, institutions, customers, complaints, professional conduct and risk events; the risk knowledge graph includes a set of entity nodes, a set of entity relationship edges used to represent the business relationships between nodes, a set of node attributes, and time evolution information used to record the changes in nodes and entity relationships; Specifically, with insurance practitioners as the core subject, a risk knowledge graph is constructed by integrating public data, industry data, corporate data, and financial behavior data, covering multiple types of entities such as personnel, institutions, customers, complaints, professional conduct, and risk events.

[0048] Define risk knowledge graph as:

[0049] in: G: Risk Knowledge Graph; V: A collection of entity nodes, including insurance practitioners, institutions, customers, risk events, and other nodes; E: Entity relationship edge set, used to represent the business association relationships between nodes; W: Set of relation edge weights, used to represent the importance of different relations; A: A set of node attributes, including risk tags, risk scores, credit ratings, and other attribute information; T: Time evolution information, used to record the changes in nodes and relationships. Compared to traditional knowledge graphs, this invention introduces a time dimension and dynamic attributes to enable continuous updating and evolution of the knowledge graph.

[0050] S22, Upon receiving new risk feedback data, locate the corresponding node in the knowledge graph based on the risk feedback object, and identify all related relationship edges; Specifically, when the system receives new risk feedback data, it first locates the corresponding node in the knowledge graph based on the feedback object and identifies all the relationship edges associated with it.

[0051] Let the feedback object node be:

[0052] Its set of relationships is as follows:

[0053] S23. Based on the risk feedback results and the relation edge weight threshold, the knowledge graph structure is dynamically adjusted. The dynamic adjustment includes adding risk relations, deleting invalid relations, updating relation attributes, updating node risk labels, updating node risk levels, and updating time attributes. Specifically, the system automatically analyzes newly added relationships, invalid relationships, and changes in relationship strength based on risk feedback results, and dynamically adjusts the knowledge graph structure, including: adding risky relationships, deleting invalid relationships, updating relationship attributes, updating node risk labels, updating node risk levels, and updating time attributes. Through the above methods, the knowledge graph can reflect changes in the risk status of insurance practitioners in real time.

[0054] S24, dynamically update the node risk attributes and the weights of relation edges based on the risk feedback results; Specifically, to accurately describe the strength of risk association, a risk feedback-driven relational edge weight update mechanism is introduced.

[0055] Suppose there is a relation edge (composed of nodes) Pointing to node The initial weights at time t are: ; Based on the risk feedback results, the updated (time t+1) relation weights are defined as follows:

[0056] in: To update the preceding rights; To update the subsequent rights; The learning rate for the risk knowledge graph; Risk feedback level; This is the relationship influence factor, used to indicate the importance of the relationship to risk propagation. The higher the risk feedback level, the more significant the increase in relational edge weights; when the feedback level decreases, the edge weights decay synchronously, keeping the graph up-to-date.

[0057] In addition to relational edge weights, the system synchronously updates node risk attributes.

[0058] Define the node risk vector as follows:

[0059] in: Assess risk level; Risk level; A set of risk labels; This is a record of historical risks. Based on the latest risk feedback, the above attributes are updated in real time and automatically synchronized to the risk profile, knowledge graph and risk model to achieve data consistency across multiple modules.

[0060] Preferably, to ensure the long-term effective operation of the knowledge graph, a self-evolution mechanism for the knowledge graph can also be established.

[0061] Let the threshold values ​​for the relation edges be: ; Then, when a certain relation edge (by node) Pointing to node () is less than the low relation edge weight threshold, i.e. When this occurs, it indicates that the relationship has been invalid for a long time, and the system automatically deletes the corresponding relationship edge; When a relation edge (by a node) Pointing to node () is greater than the threshold of the high relation edge weight, i.e. When this occurs, it indicates that the risk association is continuously strengthening, and the system automatically reinforces this relationship and marks it as a key risk propagation path.

[0062] Meanwhile, for new entities and relationships that emerge from new risk feedback, the system automatically completes node creation, relationship establishment, and attribute initialization, enabling the knowledge graph to continuously expand.

[0063] Ultimately, a dynamic graph update mechanism of "feedback-driven - relationship update - attribute update - structural evolution" is formed, enabling the risk knowledge graph to continuously learn, correct and evolve, providing an accurate and real-time knowledge foundation for subsequent risk propagation analysis, risk prediction and risk model optimization.

[0064] S25, after completing the dynamic correction of the knowledge graph, the system outputs the updated risk knowledge graph.

[0065] Specifically, after completing the dynamic correction of the knowledge graph, the system outputs the updated risk knowledge graph:

[0066] The updated knowledge graph will serve as input for the next step of node risk propagation and GraphSAGE reinforcement learning, enabling risk propagation analysis, risk evolution prediction, and continuous model optimization.

[0067] In step S3, based on the dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to aggregate neighborhood features and analyze risk propagation for insurance practitioner nodes. This enables adaptive updating of node risk features, improving the accuracy and generalization ability of risk identification. Step S3 specifically includes: S31, based on the dynamically updated risk knowledge graph, obtain the node risk propagation neighborhood of a certain node; Specifically, after completing the dynamic correction of the knowledge graph structure, the system obtains a dynamic knowledge graph:

[0068] Each insurance professional corresponds to a node.

[0069] Node risk originates not only from its own behavior but is also influenced by the risk status of associated nodes. Therefore, the node risk propagation neighborhood is defined as:

[0070] in: :node The first-order neighborhood; : with nodes Neighboring nodes that have a relationship. The system prioritizes high-impact neighbors for risk propagation analysis based on relationship weights.

[0071] S32 adopts the GraphSAGE fixed neighborhood sampling strategy, selects K neighboring nodes from the node risk propagation neighborhood, and aggregates the sampling neighborhood of the K neighboring nodes through the aggregation function; Specifically, due to the large size of actual knowledge graphs, the computational complexity would increase rapidly if all neighbors participated in the computation. Therefore, this invention adopts the GraphSAGE fixed neighborhood sampling strategy.

[0072] Let the number of samples be K, then the number of nodes The corresponding sampling neighborhood is:

[0073] satisfy:

[0074] Sampling criteria include: relational edge weights, risk levels, risk feedback frequency, and time decay factors. High-risk, highly correlated, and recently active nodes are prioritized for retention to improve the efficiency of risk propagation analysis.

[0075] GraphSAGE neighborhood aggregation: Set nodes The feature vector of the l-th layer is represented as:

[0076] First, aggregate the sampling neighborhood:

[0077] in, For nodes The neighborhood aggregation result at layer l; AGG is the aggregation function; For nodes The feature vector at the l-th layer; For nodes , neighboring nodes; A sampling area consisting of K neighboring nodes The aggregation function of this invention preferably uses a Mean Aggregator:

[0078] It can also be replaced with Max Pooling, LSTM Aggregator, or AttentionAggregator, depending on business needs. S33, GraphSAGE graph neural network fuses the node’s own features with the features of its neighborhood aggregation to form new node risk features; Specifically, the fusion of node-specific features with neighborhood aggregation features (through multiple GraphSAGE iterations, a high-dimensional risk representation containing both the node's own risk information and neighborhood-propagated risk information, i.e., new node risk features) is as follows:

[0079] Where L is the number of network layers in the GraphSAGE graph neural network; The final risk embedding vector; updated node risk features for:

[0080] in, The training weights are for the l-th layer; Concat represents feature concatenation; For activation functions; For nodes The feature vector at the l-th layer; S34, based on the generated new node risk characteristics, recalculate the node risk propagation index and the enhanced comprehensive risk score.

[0081] Specifically, based on the node embedding vectors generated by the GraphSAGE graph neural network, the system recalculates the node risk propagation index.

[0082] Define the propagation risk value:

[0083] in, For nodes Propagation risk probability in GraphSAGE graph neural network; Risk classification weights; For bias terms; The enhanced comprehensive risk score is calculated as follows:

[0084] in, For nodes The enhanced comprehensive risk score; The results are from the calculation of the risk scoring model (four-dimensional dynamic risk index model). For weight fusion.

[0085] By integrating individual risk with neighborhood-borne risk, the system achieves more accurate dynamic risk assessment. The system automatically updates the knowledge graph node attributes based on the latest risk scores: risk score, risk level, risk tag, risk propagation intensity, and update time. Simultaneously, the updated risk information is fed back to the knowledge graph, providing input for the next round of dynamic correction, thus enabling: A closed-loop learning mechanism consisting of risk feedback → graph correction → GraphSAGE propagation → risk enhancement → graph update.

[0086] This step outputs the final enhanced node risk representation (node ​​risk characteristics):

[0087] And the enhanced risk score set:

[0088] The above results will serve as input for the next step, the large-scale model intelligent explanation and risk management decision-making module, to generate risk cause explanation, risk trend prediction, and intelligent management suggestions.

[0089] Furthermore, in step S3, based on the new node risk characteristics, a joint self-optimization mechanism of relation edge weights and node weights is established to achieve synchronous updates of knowledge graph relations, nodes, and risk propagation capabilities. (This step, based on risk feedback results and the node representation enhanced by GraphSAGE, establishes a joint self-optimization mechanism of relation edge weights and node weights to achieve synchronous updates of knowledge graph relation strength, node importance, and risk propagation capabilities, enabling the risk knowledge graph to continuously learn and adaptively optimize based on business feedback, thereby improving the accuracy of risk identification and the ability to analyze risk propagation.) Specifically: S35, the weight of the relation edge is updated as follows:

[0090] in, For nodes With neighboring nodes The updated weights of the relation edges between them; For nodes With neighboring nodes The weights of the relation edges before the update; For nodes New risk feedback level; For nodes With neighboring nodes New factors influencing the propagation of relationships; The learning rate for the edge weights of the GraphSAGE graph neural network; Specifically, after the GraphSAGE graph neural network completes the risk propagation between nodes, it dynamically adjusts the edge weights of the relationships between nodes based on the risk feedback results. The stronger the risk feedback, the faster the edge weights increase; if there is no risk feedback for a long period, the edge weights gradually decay. The edge weight decay model is as follows:

[0091] in: The time decay coefficient ensures that the knowledge graph can automatically eliminate invalid relationships.

[0092] S36, the node weight update is as follows:

[0093] in, For nodes Updated weights; For nodes Weights before update; Output a comprehensive risk score for the GraphSAGE graph neural network; The learning rate of the nodes in the GraphSAGE graph neural network; Specifically, the GraphSAGE graph neural network computes the enhanced node representations. Subsequently, the node weights are updated synchronously based on the node risk scores. This update method allows the node weights to gradually approach the actual risk level. Each insurance practitioner node in the knowledge graph corresponds to a node weight, which represents the importance of that node in the entire risk propagation network.

[0094] Let the set of nodes be The node weight is then defined as: ; in: For the node at time t Weight; Node weights comprehensively reflect a node's risk level, historical risk frequency, risk propagation capability, and business impact. During system initialization, initial node weights are calculated based on a four-dimensional dynamic risk index model.

[0095] S37, the objective function for the joint optimization of edge weights and node weights is as follows:

[0096] Where L is the joint optimization objective function; Optimize the loss for node weights; Optimize the loss for relational edge weights; , These are the balance coefficients for the node weight optimization loss and the relation edge weight optimization loss, respectively.

[0097] Specifically, considering that changes in node risk affect the relationship propagation capability, and that relationship propagation in turn affects node risk, this invention establishes a joint optimization objective function. The node optimization loss is defined as:

[0098] The edge weight optimization loss is defined as:

[0099] in, These parameters are consistent with the node weights and maximum values; for example, they can all be values ​​between 0 and 1. The target node weights are adjusted based on risk feedback. The target edge weights are adjusted based on risk feedback. For nodes With neighboring nodes The weight of the edges relating to each other. The system uses the gradient descent algorithm to continuously optimize the node weights and edge weights to achieve joint convergence.

[0100] After the joint optimization is completed, the system synchronously updates the knowledge graph, including: node weight, node risk score, risk label, relationship edge weight, risk propagation strength, and update timestamp. If a node maintains a low-risk state multiple times in a row, its weight will be automatically reduced; if a node triggers high-risk events repeatedly, its weight will be increased and the relevant relationships will be strengthened.

[0101] The entire knowledge graph is continuously completed: node update → edge weight update → risk propagation → joint optimization → graph evolution, forming a dynamic graph.

[0102] After completing the joint optimization, the optimized knowledge graph is output:

[0103] in: : The optimized set of relation edge weights; : The optimized set of node weights. The above results will serve as input for the next step, the large-scale model intelligent interpretation and risk management decision-making module, to achieve risk interpretation, trend prediction, and intelligent decision-making.

[0104] In step S4, based on risk feedback data, dynamic graph update results, and the node risk representation enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to continuously optimize risk model parameters, update risk features, and improve model performance, forming a closed-loop learning system covering risk identification, risk feedback, model update, and re-prediction. Step S4 specifically involves: S41. After completing the risk handling, the actual business feedback results are fed back and used to construct a risk feedback sample set together with historical risk samples. The risk feedback sample set includes newly added training datasets and historical training datasets. The risk feedback samples include risk feature vectors of insurance practitioners, real risk feedback labels, and the number of feedback samples.

[0105] Specifically, after the system completes risk handling, the actual business feedback results are automatically fed back to the model training module to jointly construct an incremental training dataset with historical risk samples.

[0106] Let the risk feedback sample set be defined as:

[0107] in: The risk characteristic vector for insurance practitioners includes basic credit, business behavior, relationship network, financial behavior, and GraphSAGE enhanced features; Labels for accurate risk feedback; : Number of feedback samples. Risk labels are automatically generated based on the actual handling results of the business, including: normal, low risk, medium risk, and high risk. By continuously accumulating risk feedback data, we can expand the scale of training samples and improve the model's generalization ability. S42 employs an incremental learning approach, fusing newly added feedback samples from the newly added training dataset with historical training samples from the historical training dataset; and sets different training weights based on the time interval between the risk feedback data and the current moment. Specifically, to avoid the computational overhead of retraining the model every time, this invention adopts an incremental learning approach, which fuses newly added feedback samples with historical samples.

[0108] Let the historical training samples be:

[0109] The newly added feedback samples are:

[0110] The new training dataset is then defined as:

[0111] At the same time, a time weight is introduced based on the freshness of the feedback data:

[0112] in: The time interval between the feedback time and the current time. Time decay coefficient. Recent feedback samples have higher training weights, enabling the model to quickly adapt to the latest changes in risk.

[0113] S43, using the newly added training dataset as input, optimize the parameters of the risk scoring model; the specific objective function of the risk scoring model is:

[0114] in, For risk scoring model parameters The objective function value at that time; For the first The weights of each feedback sample; N is the total number of feedback samples; For the first Feedback Sample In parameters Risk scoring model at that time; The loss function; For the first Feedback Sample The corresponding actual risk score.

[0115] Specifically, the gradient descent algorithm can be used to update the model parameters:

[0116] in: The learning rate for the risk scoring model; : Gradient of the objective function. Through multiple rounds of iterative training, the model gradually learns new risk characteristics, improving prediction accuracy. After training is complete, the system automatically evaluates the performance of the updated model.

[0117] Let the model predict the value as follows: The authenticity label is Y. Accuracy, precision, recall, and F1 score can be used as evaluation metrics, or other types of metrics can be used; this embodiment does not impose any restrictions. If the model performance reaches a preset threshold, the online model is updated; otherwise, training continues until the requirements are met.

[0118] Once the model is validated, the system automatically replaces the current risk scoring model and updates it synchronously. Parameters of the four-dimensional dynamic risk index model (risk scoring model); GraphSAGE node representation; Risk labeling system; Risk scoring rules. The new risk scoring model immediately participates in the next round of risk scoring, enabling continuous online optimization of the risk model. Let the final risk scoring model be expressed as:

[0119] in: Optimized model parameters; The latest training dataset. Preferably, the present invention establishes a closed-loop continuous learning mechanism covering the entire risk governance process: risk identification → risk handling → feedback collection → graph update → GraphSAGE enhancement → model retraining → risk re-prediction; by continuously introducing new risk feedback samples, the model can continuously learn, knowledge can continuously accumulate, rules can continuously optimize, and risk scoring can continuously iterate, enabling the risk governance system to adapt to changes in the risk characteristics of insurance practitioners and improve long-term operational stability and risk prediction capabilities.

[0120] After the model is retrained, a new generation of risk scoring model will be output:

[0121] in: Optimized model parameters; Updated risk knowledge graph; GraphSAGE enhances node representation; Latest risk score results. The updated model is put back into operation, forming a complete closed loop of "risk feedback - graph correction - GraphSAGE enhancement - joint optimization - model retraining - prediction again", enabling the risk scoring model to continuously learn, adaptively optimize and intelligently evolve.

[0122] In step S5, based on risk feedback data, dynamic graphs, and GraphSAGE-enhanced node representations, this step introduces Retrieval-Augmented Generation (RAG) technology and Large Language Model (LLM) to establish a graph semantic correction mechanism. This enables automatic understanding of risk knowledge, intelligent reasoning of risk relationships, automatic correction of risk labels, and continuous optimization of risk rules, thereby improving the semantic expression capability of the knowledge graph and the level of intelligent risk governance. Step S5 specifically includes: S51, Establish a risk knowledge base as the data source for RAG retrieval; the knowledge base includes: insurance industry regulatory system, insurance company business management system, typical risk cases in the insurance industry, historical risk disposal records, risk label system, risk rule base, and dynamic risk knowledge graph. Specifically, to improve the accuracy of large-scale model inference, the system establishes a unified risk knowledge base as the data source for RAG retrieval. Let the unified knowledge base be represented as:

[0123] in: Unified knowledge base; : No. Knowledge unit. All knowledge is represented using a unified vector database to support subsequent semantic retrieval.

[0124] S52, After receiving a risk event, construct a semantic query based on the risk characteristics and risk score in the risk feedback information; Specifically, after receiving a risk event, the system first constructs a semantic query based on the risk feedback information:

[0125] in, A semantic query language for construction; Risk characteristics; To output a risk score. S53 uses vector similarity retrieval to obtain the k most relevant knowledge sets from the risk knowledge base; the retrieval content includes: similar risk cases, similar complaint events, similar abnormal financial behavior, industry regulatory basis, risk disposal experience, and risk transmission patterns; Specifically, vector similarity retrieval is then used to obtain the most relevant set of knowledge from the knowledge base. :

[0126] in: Semantic similarity function; Return the top k most relevant pieces of knowledge.

[0127] The RAG mechanism provides authentic and reliable external knowledge support for large models, effectively reducing model illusion and improving the accuracy of risk analysis.

[0128] S54: Dynamic graph information, new node risk features, and RAG retrieval results are input into the large language model to perform comprehensive semantic reasoning and generate risk explanation results. Specifically, the system inputs dynamic graph information, GraphSAGE node representations, and RAG retrieval results into the large language model to conduct comprehensive semantic reasoning.

[0129] The input is represented as:

[0130] in: Dynamic graph; GraphSAGE node embedding; RAG retrieval knowledge set. Comprehensive analysis using a large-scale model: Causes of risk formation, risk propagation paths, influencing factors, risk trends, and the basis for risk governance. The final risk interpretation results are generated.

[0131] in, : Results of risk semantic interpretation.

[0132] S55. Based on the output of the large model, perform semantic correction on the knowledge graph; the semantic correction includes: adding risk tags, correcting node attributes, supplementing relation semantics, correcting relation types, updating risk propagation rules, and adding risk knowledge. Specifically, semantic corrections are made to the knowledge graph based on the output of the large model.

[0133] Let the node attribute update function be:

[0134] The relational semantics are updated as follows:

[0135] in, For the updated node attributes; The node attributes before the update; This is the attribute correction amount; For nodes With neighboring nodes Updated relational semantics; For nodes With neighboring nodes Relational semantics before update; This is a semantic correction value for relations. After semantic correction, the knowledge graph not only updates its structure, but also improves its semantic expression capabilities.

[0136] Preferably, the system can also continuously learn the risk rule base based on the analysis results of the large model. Let the risk rule set be:

[0137] Based on the latest risk feedback and semantic reasoning results, the big model automatically identifies: new risk patterns, new risk labels, new risk propagation rules, and new business rules, enabling the continuous evolution of risk rules. The newly added rule is represented as follows:

[0138] in, For the updated risk rules; The risk rules were in effect before the update. This is a newly added rule.

[0139] S56. After completing the semantic correction, output a new risk knowledge graph.

[0140] After semantic correction, the system outputs a new knowledge graph:

[0141] Among them, S is newly added: semantic knowledge set. Simultaneously, the following updates are implemented: risk tags, risk rules, risk explanation templates, GraphSAGE node representations, and risk scoring models, achieving collaborative optimization of the knowledge graph, GraphSAGE, and risk model.

[0142] This step ultimately outputs an enhanced intelligent risk knowledge system:

[0143] in: : A semantically corrected risk knowledge graph; Optimized risk rule base; A continuously updated risk knowledge base; : Optimized risk scoring model parameters. The above results will be fed back to the risk identification module, forming the following process: Risk Feedback → Graph Correction → GraphSAGE Enhancement → Model Retraining → RAG Semantic Correction → Risk Identification.

[0144] Furthermore, this step is driven by risk feedback data and takes risk knowledge graphs, GraphSAGE graph neural networks, risk scoring models and RAG enhanced large models as the core to build a closed-loop optimization mechanism covering "data feedback - graph update - risk propagation - model optimization - semantic correction - intelligent decision-making". This enables the risk governance system to continuously learn, adapt, evolve and optimize intelligently, and continuously improve the risk identification, risk prediction and risk disposal capabilities.

[0145] After each risk event is handled, the system automatically feeds back the risk handling results to the risk governance platform and initiates a closed-loop optimization process.

[0146] Let the risk feedback results be represented as a risk feedback set. The risk feedback set includes multiple risk feedback messages. Each risk feedback message includes the risk level, handling result, handling effect, and time information. The system completes the optimization in the following steps: Risk feedback → Risk knowledge graph update → GraphSAGE enhanced node risk feature update → Risk scoring model update → RAG semantic knowledge base update → Intelligent risk decision results, forming a closed loop covering the entire chain of data, graph, model, knowledge and decision.

[0147] The system automatically updates the following based on feedback results: node attributes, node weights, risk labels, edge weight relationships, risk propagation paths, and graph semantics. The updated knowledge graph is represented as follows:

[0148] Where S is the semantic knowledge set. The graph has undergone both structural and semantic updates, providing a more accurate data foundation for the next round of risk analysis.

[0149] After the knowledge graph is updated, the GraphSAGE graph neural network resamples. ) Neighboring nodes, and compute new node representations:

[0150] The system updates synchronously based on the latest node representations: risk propagation characteristics, node embedding vectors, risk propagation strength, and risk impact scope, continuously enhancing the model's ability to identify complex associated risks.

[0151] By combining the newly added feedback samples, the risk scoring model is incrementally trained, the model parameters are updated, and the four-dimensional dynamic risk index (risk scoring model) is recalculated to achieve dynamic optimization of risk scoring.

[0152] RAG knowledge continues to evolve: the system automatically writes the latest risk cases into the knowledge base. At the same time, the large model is continuously optimized based on the new cases: risk tags, risk rules, risk interpretation templates, and risk handling experience, achieving continuous knowledge accumulation and semantic evolution.

[0153] In step S6, when the system receives a new risk event, it no longer relies on the initial model, but instead calls the latest optimized model: risk knowledge graph, GraphSAGE node representation (node ​​risk features), risk scoring model, RAG knowledge base, and large language model; and automatically generates: risk cause analysis, risk level determination, risk propagation path, risk trend prediction, and risk disposal suggestions, forming a new risk governance result.

[0154] In summary, this invention constructs a closed-loop learning framework covering the entire risk governance process: Risk feedback → Knowledge graph update → GraphSAGE propagation learning → Risk model retraining → RAG semantic correction → AI intelligent decision-making → Risk handling → Feedback collection. Each time the system completes a risk handling process, it performs a knowledge update, model optimization, and rule evolution, enabling the entire risk governance system to form a continuous learning, adaptive optimization, and iterative intelligent operating mechanism. Ultimately, it outputs a continuously optimized intelligent risk governance system.

[0155] in: Dynamically evolving risk knowledge graph; GraphSAGE enhances node representation; Continuously optimize the risk scoring model; RAG Enhanced Knowledge Base; The results of intelligent risk decision-making are then re-entered into the risk identification stage, forming a continuously iterative intelligent governance system.

[0156] To illustrate this plan more clearly, we will use the specific example of monitoring the behavior of insurance practitioners as an example: 1. Implementation Scenarios An insurance institution has accessed the insurance practitioner risk governance platform built by this invention. The platform relies on a trusted data space to aggregate public data, industry data, enterprise data, and financial behavior data to conduct continuous risk monitoring on insurance agent A.

[0157] The data collected by the platform in real time includes: (1) Basic identity information of the agent; (2) Professional conduct data, including the employing institution, business volume, commission income and resignation records; (3) Customer complaint data, including the number of complaints, the reasons for the complaints, and the results of the complaint handling; (4) Public information on administrative penalties and judicial proceedings; (5) Fund behavior data, including abnormal fund flows, transaction frequency and abnormal fund transfers; (6) Historical risk management records and risk feedback information.

[0158] The system first standardizes the multi-source data and constructs a unified risk feature vector, which serves as the input to the risk identification model.

[0159] 2. Risk Feedback Data Construction As shown in Table 1 below, an agent exhibited the following abnormal behaviors over a consecutive 30-day period: Table 1: Record of Abnormal Behavior of Insurance Agents

[0160] The system automatically generates risk feedback data:

[0161] Where: ID=A000001; X represents the risk feature vector; Y represents the risk feedback label (high risk); ω=0.92; t represents the time of risk occurrence. The above feedback data is then fed into the feedback learning module.

[0162] 3. Risk knowledge graph is dynamically updated. The system locates Agent A's node in the knowledge graph. Based on newly added risk feedback: new risk tags, high-frequency complaints, abnormal financial behavior, and cross-institutional practice; and simultaneously updates: node risk level, node risk attributes, and node historical risk records; and also updates the nodes associated with Agent A: affiliated institution, client relationship, and historical business relationship. Recalculate the edge weights of the association relationships:

[0163] The dynamic graph update is complete.

[0164] 4. GraphSAGE Risk Propagation Analysis The system uses agent A as the central node and performs GraphSAGE aggregation calculations on its neighboring nodes. The neighborhood includes: agents from the same institution, historical collaborators, individuals with similar complaints, and individuals with similar financial activities. Calculate the new node representation:

[0165] The analysis results show that: Agent A has three other agents with highly similar risk characteristics.

[0166] The GraphSAGE risk propagation index increased from 0.46 to 0.81, indicating that the risk has a high probability of propagation.

[0167] 5. Joint optimization of edge weights and node weights The system updates synchronously based on risk feedback results: Node weight:

[0168] The edge weights of the relationships are shown in Table 2: Table 2: Correspondence Table Before and After Relationship Edge Weight Update

[0169] After the update, the knowledge graph regenerates key risk transmission paths.

[0170] 6. Risk model feedback retraining The system adds this risk event to the training set. The new training sample is represented as follows:

[0171] After retraining the risk scoring model, the risk identification accuracy was 95.1%, an improvement of 3.7 percentage points compared to the previous model (91.4%). Simultaneously, the F1 score improved by approximately 4%.

[0172] 7. Semantic Correction of Large Models Enhanced by RAG The system automatically accesses the RAG knowledge base, retrieving content including: historical complaint cases, industry regulatory provisions, similar abnormal fund cases, and risk disposal cases; then, it uses a large language model for reasoning; the output includes: Causes of risk: Frequent complaints and abnormal financial behavior have led to a rapid increase in risk.

[0173] Risk trend forecast: The risk level will continue to rise over the next 30 days.

[0174] Risk management recommendations: Suspend new business, conduct special risk audits, strengthen monitoring of fund behavior, and include them in the list of key personnel. Meanwhile, the large model adds: 2 risk labels; 1 risk rule; and updates the semantic relationships of the knowledge graph.

[0175] 8. Closed-loop continuous optimization After the risk is resolved, the platform automatically collects the following feedback information: whether the high risk was confirmed, the actual resolution result, changes in subsequent complaints, the status of risk elimination, and the evaluation results of management personnel. The system restarts the closed-loop optimization process: risk feedback → knowledge graph update → GraphSAGE propagation → edge weight optimization → model retraining → RAG semantic correction → AI intelligent decision-making, achieving synchronous optimization of knowledge graph, risk model, large language model and risk rules.

[0176] After continuous operation, the platform can continuously adjust the risk propagation path, risk scoring results, and risk handling strategies based on new risk feedback, realizing the transformation of risk governance for insurance practitioners from static identification to dynamic perception, from rule-driven to data intelligence-driven, and from post-event handling to pre-event early warning.

[0177] 9. Implementation Results After adopting the method of this invention, the monitoring of insurance practitioners' behavior has achieved intelligent management throughout the entire process, and the results are shown in Table 3: Table 3: Correspondence Table of Effects Before and After Implementation

[0178] This embodiment verifies that the present invention can realize continuous monitoring, dynamic analysis, intelligent early warning and closed-loop governance of the risk behavior of insurance practitioners. It has good real-time performance, self-learning ability and scalability, and can be widely applied to risk management scenarios of insurance agents, insurance brokers and other insurance practitioners. It can also be extended to the field of personnel risk governance in the banking, securities and other financial industries.

[0179] In this invention, risk feedback data is acquired and preprocessed to obtain risk feedback data objects. A risk knowledge graph of multiple entity types is constructed, with insurance practitioners as the core subject. Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the feedback results, outputting an updated risk knowledge graph. Based on this dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse node features with neighborhood aggregation features, forming new node risk features. Based on these new node risk features, a joint self-optimization mechanism for relation edge weights and node weights is established to realize the propagation capabilities of knowledge graph relationships, nodes, and risks. The system synchronously updates risk knowledge graphs; based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to update the parameters of the risk scoring model based on the feedback results; based on risk feedback data, dynamic graphs, and risk feedback model scoring results, retrieval enhancement generation technology and a large language model are introduced to establish a graph semantic correction mechanism to semantically correct the knowledge graph; based on the semantically corrected risk knowledge graph, the risk identification results of risk events are output, effectively solving the problem of low reliability in risk knowledge graph optimization caused by existing technologies, and effectively improving the reliability of risk knowledge graph optimization.

[0180] The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for knowledge graphs. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk labels, and graph semantics, realizing the transformation of knowledge graphs from static maintenance to dynamic evolution. It can reflect changes in the risk status of insurance practitioners in real time and improve the data integrity, accuracy, and timeliness of risk knowledge graphs.

[0181] The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for knowledge graphs. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk labels, and graph semantics, realizing the transformation of knowledge graphs from static maintenance to dynamic evolution. This mechanism can reflect changes in the risk status of insurance practitioners in real time, improving the data integrity, accuracy, and timeliness of the risk knowledge graph.

[0182] The technical solution of this invention proposes a joint self-optimization method for edge weights and node weights, which synchronously applies the risk feedback results to the relation edge weights and node weights, thereby achieving coordinated updates of relation strength, node importance, and risk propagation capability. This constructs a continuous learning mechanism for the knowledge graph, enabling risk relations to be dynamically adjusted based on business feedback and improving the long-term operating effect of the graph.

[0183] The technical solution of this invention uses real risk disposal results as supervision feedback samples, introduces an incremental learning mechanism, and continuously retrains the risk scoring model to achieve dynamic optimization of model parameters, risk characteristics and scoring rules. This enables the risk scoring model to continuously learn new risk patterns, improve the model's generalization ability and long-term prediction performance, and avoid the problem of performance degradation of traditional models after long-term operation.

[0184] This invention combines RAG retrieval enhancement technology and a large language model to perform semantic reasoning and knowledge supplementation on the risk knowledge graph. This enables the explanation of risk causes, updating of risk labels, generation of risk rules, and expansion of risk knowledge, improving the interpretability and intelligence of risk analysis. Simultaneously, it effectively reduces the risk of large model illusions and enhances the quality of risk assessment. Furthermore, it constructs a closed-loop optimization mechanism of "risk feedback—knowledge graph update—GraphSAGE risk propagation—joint optimization of edge weights and node weights—risk model retraining—RAG semantic correction—intelligent decision-making," achieving synergistic optimization across five levels: data, graph, model, knowledge, and decision-making. This enables the risk governance system to possess continuous learning, adaptive evolution, and dynamic optimization capabilities, forming a complete intelligent risk governance closed loop.

[0185] Example 2 like Figure 2 As shown, the present invention also provides a risk feedback-driven dynamic optimization system for knowledge graphs, comprising: The data feedback acquisition module acquires risk feedback data, preprocesses the risk feedback data to obtain a risk feedback data object; the risk feedback data object includes insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. The knowledge graph update module uses insurance practitioners as the core to construct a risk knowledge graph of multiple types of entities. When new risk feedback data is received, the knowledge graph structure is dynamically adjusted according to the risk feedback results, and the updated risk knowledge graph is output. The risk propagation learning and updating module is based on the dynamically updated risk knowledge graph. It introduces the GraphSAGE graph neural network to fuse the node's own features with the neighborhood aggregation features to form new node risk features. Based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous updating of knowledge graph relations, nodes and risk propagation capabilities. The model training module establishes a risk scoring model feedback and retraining mechanism based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, and updates the parameters of the risk scoring model based on the risk scoring model feedback results. The semantic correction module, based on risk feedback data, dynamic graphs, and risk feedback model scoring results, introduces retrieval enhancement generation technology and large language models to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph. The output module, based on the semantically corrected risk knowledge graph, outputs the risk identification results of risk events.

[0186] The implementation process of the data feedback acquisition module, graph update module, risk propagation learning update module, model training module, semantic correction module, and output module in this embodiment two corresponds to the method steps in embodiment one, and will not be repeated here.

[0187] In this invention, risk feedback data is acquired and preprocessed to obtain risk feedback data objects. A risk knowledge graph of multiple entity types is constructed, with insurance practitioners as the core subject. Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the feedback results, outputting an updated risk knowledge graph. Based on this dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse node features with neighborhood aggregation features, forming new node risk features. Based on these new node risk features, a joint self-optimization mechanism for relation edge weights and node weights is established to realize the propagation capabilities of knowledge graph relationships, nodes, and risks. The system synchronously updates risk knowledge graphs; based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to update the parameters of the risk scoring model based on the feedback results; based on risk feedback data, dynamic graphs, and risk feedback model scoring results, retrieval enhancement generation technology and a large language model are introduced to establish a graph semantic correction mechanism to semantically correct the knowledge graph; based on the semantically corrected risk knowledge graph, the risk identification results of risk events are output, effectively solving the problem of low reliability in risk knowledge graph optimization caused by existing technologies, and effectively improving the reliability of risk knowledge graph optimization.

[0188] The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for knowledge graphs. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk labels, and graph semantics, realizing the transformation of knowledge graphs from static maintenance to dynamic evolution. It can reflect changes in the risk status of insurance practitioners in real time and improve the data integrity, accuracy, and timeliness of risk knowledge graphs.

[0189] The technical solution of this invention constructs a risk feedback-driven dynamic correction mechanism for knowledge graphs. Based on the risk feedback results, it automatically updates node attributes, relationship edges, risk labels, and graph semantics, realizing the transformation of knowledge graphs from static maintenance to dynamic evolution. This mechanism can reflect changes in the risk status of insurance practitioners in real time, improving the data integrity, accuracy, and timeliness of the risk knowledge graph.

[0190] The technical solution of this invention proposes a joint self-optimization method for edge weights and node weights, which synchronously applies the risk feedback results to the relation edge weights and node weights, thereby achieving coordinated updates of relation strength, node importance, and risk propagation capability. This constructs a continuous learning mechanism for the knowledge graph, enabling risk relations to be dynamically adjusted based on business feedback and improving the long-term operating effect of the graph.

[0191] The technical solution of this invention uses real risk disposal results as supervision feedback samples, introduces an incremental learning mechanism, and continuously retrains the risk scoring model to achieve dynamic optimization of model parameters, risk characteristics and scoring rules. This enables the risk scoring model to continuously learn new risk patterns, improve the model's generalization ability and long-term prediction performance, and avoid the problem of performance degradation of traditional models after long-term operation.

[0192] This invention combines RAG retrieval enhancement technology and a large language model to perform semantic reasoning and knowledge supplementation on the risk knowledge graph. This enables the explanation of risk causes, updating of risk labels, generation of risk rules, and expansion of risk knowledge, improving the interpretability and intelligence of risk analysis. Simultaneously, it effectively reduces the risk of large model illusions and enhances the quality of risk assessment. Furthermore, it constructs a closed-loop optimization mechanism of "risk feedback—knowledge graph update—GraphSAGE risk propagation—joint optimization of edge weights and node weights—risk model retraining—RAG semantic correction—intelligent decision-making," achieving synergistic optimization across five levels: data, graph, model, knowledge, and decision-making. This enables the risk governance system to possess continuous learning, adaptive evolution, and dynamic optimization capabilities, forming a complete intelligent risk governance closed loop.

[0193] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A risk feedback-driven dynamic optimization method for knowledge graphs, characterized in that, include: Obtain risk feedback data, preprocess the risk feedback data, and obtain a risk feedback data object; The risk feedback data objects include insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. With insurance practitioners as the core, construct a risk knowledge graph for multiple types of entities; Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the risk feedback results, and an updated risk knowledge graph is output. Based on the dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse the node's own features with the neighborhood aggregation features to form new node risk features. Based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous update of knowledge graph relations, nodes and risk propagation capabilities. Based on risk feedback data, dynamic graph update results, and node risk characteristics enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established to update the parameters of the risk scoring model based on the risk scoring model feedback results. Based on risk feedback data, dynamic graphs, and risk feedback model scoring results, retrieval enhancement generation technology and large language models are introduced to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph. Based on the semantically corrected risk knowledge graph, the risk identification results of risk events are output.

2. The knowledge graph dynamic optimization method based on risk feedback as described in claim 1, characterized in that, Obtain risk feedback data, preprocess the risk feedback data to obtain risk feedback data objects, specifically including: Risk feedback data from multiple business systems is standardized in format to form a risk feedback dataset with multiple risk feedback records. The business risk feedback results in the risk feedback record are mapped to risk feedback tags; the feedback tags include unverified risk, low risk confirmed, medium risk confirmed, and high risk confirmed. Risk features are extracted based on risk feedback records and risk feedback tags to form a risk feedback feature vector. The risk feedback feature vector includes risk level, risk feedback tag, risk occurrence frequency, risk duration, number of complaints, number of abnormal financial behaviors, number of abnormal professional behaviors, historical penalty records, credit rating results, and changes in business behavior. Based on risk feedback records, risk feedback tags, and risk feedback feature vectors, a risk feedback data object is output; the risk feedback data object includes a unique identifier for insurance practitioners, a risk feedback feature vector, a risk feedback tag, a risk feedback weight, and a feedback time.

3. The knowledge graph dynamic optimization method based on risk feedback as described in claim 2, characterized in that, Before extracting risk features from risk feedback records and risk feedback tags to form a risk feedback feature vector, the following steps are also included: The risk feedback weights are determined based on the importance of the risk events in the risk feedback records; specifically, the determination of the risk feedback weights is as follows: in, As risk feedback weight; Risk feedback level; The extent of the impact of the risk event; This is the time decay factor; , , These are the weighting coefficients for risk feedback level, the degree of impact of risk events, and time decay factor, respectively.

4. The knowledge graph dynamic optimization method based on risk feedback as described in claim 1, characterized in that, Using insurance practitioners as the core, a risk knowledge graph of multiple entity types is constructed. Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the risk feedback results, and the updated risk knowledge graph is output as follows: With insurance practitioners as the core subject, a risk knowledge graph is constructed that covers multiple types of entities, including personnel, institutions, customers, complaints, professional conduct, and risk events. The risk knowledge graph includes a set of entity nodes, a set of entity relationship edges that represent the business relationships between nodes, a set of node attributes, and time evolution information that records the changes in nodes and entity relationships. Upon receiving new risk feedback data, the corresponding node in the knowledge graph is located based on the risk feedback object, and all related relationship edges are identified. Based on the risk feedback results and the relation edge weight threshold, the knowledge graph structure is dynamically adjusted. The dynamic adjustment includes adding risk relations, deleting invalid relations, updating relation attributes, updating node risk labels, updating node risk levels, and updating time attributes. The node risk attributes and the weights of the relation edges are dynamically updated based on the risk feedback results. After completing the dynamic correction of the knowledge graph, the system outputs the updated risk knowledge graph.

5. The knowledge graph dynamic optimization method based on risk feedback as described in claim 1, characterized in that, Based on the dynamically updated risk knowledge graph, a GraphSAGE graph neural network is introduced to fuse node features with neighborhood aggregation features, forming new node risk features, specifically including: Based on the dynamically updated risk knowledge graph, obtain the node risk propagation neighborhood of a certain node; The GraphSAGE fixed neighborhood sampling strategy is adopted to select K neighboring nodes from the node risk propagation neighborhood, and then the K neighboring nodes are aggregated by the sampling neighborhood through the aggregation function; GraphSAGE graph neural network fuses the node's own features with the features of its neighborhood aggregation to form new node risk features; Based on the newly generated node risk characteristics, the node risk propagation index and the enhanced comprehensive risk score are recalculated.

6. The knowledge graph dynamic optimization method based on risk feedback as described in claim 5, characterized in that, The sampling neighborhood aggregation is specifically as follows: in, For nodes The neighborhood aggregation result at layer l; AGG is the aggregation function; For nodes Feature vectors in the l-th layer; For nodes , neighboring nodes; The sampling area consists of K neighboring nodes; The fusion of node-specific features and neighborhood aggregation features is specifically as follows: Where L is the number of network layers in the GraphSAGE graph neural network; The final risk embedding vector; updated node risk features for: in, The training weights are for the l-th layer; Concat represents feature concatenation; For activation functions; For nodes Feature vectors in the l-th layer; The calculation of the node risk propagation value is as follows: in, For nodes Propagation risk probability in GraphSAGE graph neural network; Risk classification weights; For bias terms; The enhanced comprehensive risk score is calculated as follows: in, For nodes The enhanced comprehensive risk score; The results are from the risk scoring model. For weight fusion.

7. The knowledge graph dynamic optimization method based on risk feedback as described in claim 5, characterized in that, Based on the new node risk characteristics, a joint self-optimization mechanism of relation edge weights and node weights is established to achieve synchronous updates of knowledge graph relations, nodes, and risk propagation capabilities. Specifically: The weights of the relation edges are updated as follows: in, For nodes With neighboring nodes The updated weights of the relation edges between them; For nodes With neighboring nodes The weights of the relation edges before the update; For nodes New risk feedback level; For nodes With neighboring nodes New factors influencing the propagation of relationships; The learning rate for the edge weights of the GraphSAGE graph neural network; The node weight update is as follows: in, For nodes Updated weights; For nodes Weights before update; Output a comprehensive risk score for the GraphSAGE graph neural network; The learning rate of the nodes in the GraphSAGE graph neural network; The objective function for the joint optimization of edge weights and node weights is as follows: Where L is the joint optimization objective function; Optimize the loss for node weights; Optimize the loss for relational edge weights; , These are the balance coefficients for the node weight optimization loss and the relation edge weight optimization loss, respectively.

8. The knowledge graph dynamic optimization method based on risk feedback as described in claim 1, characterized in that, Based on risk feedback data, dynamic graph update results, and node risk characteristics enhanced by GraphSAGE, a risk scoring model feedback and retraining mechanism is established. Specifically, the parameters of the risk scoring model are updated based on the feedback results: After risk management is completed, the actual business feedback results are fed back and used to construct a risk feedback sample set together with historical risk samples. The risk feedback sample set includes newly added training datasets and historical training datasets. The risk feedback samples include risk feature vectors of insurance practitioners, real risk feedback labels, and the number of feedback samples. An incremental learning approach is adopted to fuse newly added feedback samples in the newly added training dataset with historical training samples in the historical training dataset; and different training weights are set according to the time interval between the risk feedback data and the current time. Using the newly added training dataset as input, the parameters of the risk scoring model are optimized; the specific objective function of the risk scoring model is as follows: in, For risk scoring model parameters The objective function value at that time; For the first The weights of each feedback sample; N is the total number of feedback samples; For the first Feedback sample In parameters Risk scoring model at that time; The loss function; For the first Feedback sample The corresponding actual risk score.

9. The knowledge graph dynamic optimization method based on risk feedback as described in claim 1, characterized in that, Based on risk feedback data, dynamic knowledge graphs, and risk feedback model scoring results, this paper introduces retrieval enhancement generation technology and a large language model to establish a knowledge graph semantic correction mechanism. Specifically, the semantic correction of the knowledge graph involves: A risk knowledge base is established as the data source for RAG retrieval; the knowledge base includes: insurance industry regulatory system, insurance company business management system, typical risk cases in the insurance industry, historical risk disposal records, risk label system, risk rule base, and dynamic risk knowledge graph; Upon receiving a risk event, a semantic query is constructed based on the risk characteristics and risk score in the risk feedback information. Vector similarity retrieval is used to obtain the k most relevant knowledge sets from the risk knowledge base; the retrieval content includes: similar risk cases, similar complaint events, similar abnormal financial behaviors, industry regulatory basis, risk disposal experience, and risk transmission patterns; By inputting dynamic graph information, new node risk features, and RAG retrieval results into a large language model, comprehensive semantic reasoning is performed to generate risk explanation results. Based on the output of the large model, the knowledge graph is semantically corrected; the semantic correction includes: adding risk tags, correcting node attributes, supplementing relation semantics, correcting relation types, updating risk propagation rules, and adding risk knowledge. After semantic correction, a new risk knowledge graph is output.

10. A risk feedback-driven dynamic optimization system for knowledge graphs, characterized in that, include: The data feedback acquisition module acquires risk feedback data, preprocesses the risk feedback data, and obtains a risk feedback data object. The risk feedback data objects include insurance practitioners, risk feedback feature vectors, risk feedback tags, risk feedback weights, and feedback time. The graph update module uses insurance practitioners as the core subject to construct a risk knowledge graph for multiple types of entities; Upon receiving new risk feedback data, the knowledge graph structure is dynamically adjusted based on the risk feedback results, and an updated risk knowledge graph is output. The risk propagation learning and updating module is based on the dynamically updated risk knowledge graph. It introduces the GraphSAGE graph neural network to fuse the node's own features with the neighborhood aggregation features to form new node risk features. Based on the new node risk features, a joint self-optimization mechanism of relation edge weights and node weights is established to realize the synchronous updating of knowledge graph relations, nodes and risk propagation capabilities. The model training module establishes a risk scoring model feedback and retraining mechanism based on risk feedback data, dynamic graph update results, and node risk features enhanced by GraphSAGE, and updates the parameters of the risk scoring model based on the risk scoring model feedback results. The semantic correction module, based on risk feedback data, dynamic graphs, and risk feedback model scoring results, introduces retrieval enhancement generation technology and large language models to establish a graph semantic correction mechanism to perform semantic correction on the knowledge graph. The output module, based on the semantically corrected risk knowledge graph, outputs the risk identification results of risk events.

Images