Credit risk knowledge base construction method fusing internal examination risk control auditing multi-source data

Abstract

The invention relates to the related technical field of data processing and credit risk management and control, in particular to a credit risk knowledge base construction method fusing internal review risk control auditing multi-source data, and provides a double-path intelligent incremental updating mechanism to execute total calculation for a new risk subject and to execute total calculation for a historical risk subject. By calculating the distance between the new data feature vector and the historical feature confidence interval, whether the risk change belongs to quantitative change or qualitative change is intelligently judged, if the risk change is quantitative change, lightweight vector fine adjustment is performed, and if the risk change is qualitative change, full-amount recalculation is triggered, so that the normal update overhead is greatly reduced while the knowledge accuracy is guaranteed, and the risk change rate is improved. Risk knowledge is extracted based on the real-time updated atlas and is stored in a layered manner, decision feedback data is further combined, and risk rule parameters are adaptively adjusted and optimized through a multi-objective optimization algorithm. According to the method, the updating efficiency, decision timeliness and adaptive capacity of the credit risk knowledge base are remarkably improved.

Description

Technical Field

[0001] This invention relates to the technical fields of data processing and credit risk management, and in particular to a method for constructing a credit risk knowledge base that integrates multi-source data from internal audit, risk control, and auditing. Background Technology

[0002] As enterprises become increasingly digitalized, credit risk manifests itself in more complex and dynamic ways, with risk signals often hidden within interconnected networks of data from multiple sources, including internal audits, business risk control, and external audits. However, existing credit risk knowledge base construction technologies face a significant common technical bottleneck when dealing with the dynamic evolution of risk situations: the difficulty in balancing real-time knowledge updates with computational efficiency, resulting in a prominent "static knowledge" problem. Specifically, existing technical solutions mainly employ two update modes, both with inherent flaws: First, the full batch update mode, while ensuring the integrity of the knowledge base, requires recalculating all data for each update, leading to huge computational resource costs and long update cycles (typically several days or even weeks). This makes it impossible to respond in real-time to dynamically generated risk data, resulting in poor timeliness of the knowledge base. Second, some solutions attempt simple incremental update strategies, which reduce computational burden but generally lack the ability to intelligently judge the degree of data change, failing to effectively distinguish between the "quantitative" and "qualitative" changes in risk. For historical risk subjects, whether it is a minor fluctuation in characteristics or a major risk event, a uniform update method is used. This can easily lead to the knowledge base either missing major risk changes due to insufficient updates or introducing noise and wasting computing power due to excessive updates.

[0003] Therefore, there is an urgent need in this field for a core technology that can intelligently balance update efficiency and knowledge accuracy, namely an incremental update mechanism that can adaptively select the optimal update path based on the characteristics of data changes, so as to fundamentally solve the problem of static knowledge and realize the truly real-time, efficient and accurate evolution of the credit risk knowledge base.

[0004] To address the aforementioned issues, a search revealed a patent with publication number CN110288172A that discloses a credit management system based on risk control. The patent proposes "using an economic development analysis module to classify national economic development trends, and then using a credit limit setting module to determine the corresponding total credit limit. This avoids situations where large amounts of credit are released and difficult to recover when the national economic development trend is unfavorable, reduces the likelihood of corporate cash flow disruptions, facilitates effective risk control of corporate credit sales, and promotes stable corporate development." However, this application primarily relies on cyclical economic indicators such as national GDP for credit limit regulation. This method not only suffers from slow data updates and an inability to reflect the real-time risk dynamics of micro-entities, but also exhibits rigid decision-making logic, making it difficult to address complex and hidden risk transmission.

[0005] In light of this, in-depth research into the aforementioned issues led to the creation of this case. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a credit risk knowledge base that integrates multiple sources of data from internal audit, risk control, and auditing, in order to solve the problem in the prior art that it is difficult to construct a credit risk knowledge base that can deeply integrate multiple sources of data from internal audit, risk control, and auditing, and has the ability to perform in-depth analysis, dynamic evolution, and adaptive optimization.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control, comprising the following steps:

[0008] Step S1, Data Acquisition and Fusion: Collect heterogeneous data from internal audit, risk control, and audit systems, and perform semantic alignment and standardization on the heterogeneous data to generate structured risk data;

[0009] Step S2, Dynamic Knowledge Graph Construction and Update: Based on the credit risk domain ontology, the structured risk data is constructed into a multi-source data association graph, and the graph is updated in real time using an incremental update mechanism;

[0010] Step S3, Risk Knowledge Generation: Extract risk entities, risk relationships, and risk transmission paths from the updated multi-source data association graph to generate structured risk knowledge elements. Step S4, Layered Knowledge Base Construction: Store the risk knowledge elements in a basic layer, core layer, and application layer according to data abstraction levels, forming a three-layer knowledge base architecture.

[0011] The incremental update mechanism in step S2 is a dual-path update strategy, which specifically includes: when the new input data corresponds to a new risk subject in the knowledge base, a full calculation path is triggered to perform full feature extraction and graph embedding on the new risk subject; when the new input data corresponds to a historical risk subject in the knowledge base, a lightweight update path is triggered to determine the degree of risk change based on the existing feature information of the historical risk subject, and a differentiated update is performed accordingly.

[0012] Preferably, the lightweight update path specifically includes:

[0013] Based on the historical data of the aforementioned historical risk subjects, determine the confidence interval of their risk characteristic vector;

[0014] Calculate the distance between the feature vector of the new input data and the center point of the historical risk subject features;

[0015] If the distance is less than a preset threshold, the risk change is determined to be within the expected range, and a lightweight linear adjustment is made to the graph embedding vector of the historical risk subject.

[0016] If the distance is greater than or equal to a preset threshold, it is determined that the risk characteristics have changed drastically, and the complete calculation path is triggered for the historical risk subject.

[0017] Preferably, the method for determining the confidence interval includes: calculating the mean and standard deviation of the feature vector of the set of adjacent nodes of the historical risk subject to define the confidence interval.

[0018] Preferably, the semantic alignment and standardization processing of heterogeneous data in step S1 specifically includes:

[0019] Entity recognition and relation extraction from unstructured text using a Transformer-based pre-trained language model;

[0020] By employing ontology mapping technology, entities and relationships extracted from different data sources are mapped to the concepts and attributes defined in the unified ontology of the credit risk domain.

[0021] Preferably, the method further includes step S5, risk decision application and feedback optimization:

[0022] Utilizing the knowledge base for risk identification and decision support;

[0023] Collect application feedback data for risk decision-making, including at least risk control effectiveness indicators and business impact indicators;

[0024] Based on the feedback data, the risk rule parameters in the core layer and / or application layer are adaptively optimized to achieve a dynamic balance between risk control and business development.

[0025] Preferably, the adaptive optimization of risk rule parameters is based on a multi-objective optimization algorithm, which dynamically adjusts the rule threshold to optimize the comprehensive objective function consisting of risk control benefits and business impact costs.

[0026] Preferably, in step S1, a federated learning-based architecture is used for cross-institutional data collaboration, specifically including:

[0027] Edge computing nodes are deployed locally at the data sources of each participating institution for local data feature extraction.

[0028] A secure parameter aggregation algorithm is used to aggregate model parameter updates from each node on a central server without moving the original data.

[0029] Preferably, the method further includes a blockchain-based audit and traceability step: generating cryptographic hash values ​​for key operations such as data collection, knowledge updating, and risk decision-making, and storing them in the blockchain to achieve traceability and immutability of the operations.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. Through a dual-path update strategy, a differentiated update logic is innovatively designed for "new risk subjects" and "historical risk subjects." Especially for massive historical subjects, the system can intelligently identify the "quantitative change" and "qualitative change" of risk characteristics by using lightweight update path judgment based on confidence intervals. For most common "quantitative change" data, the system only needs to make rapid fine adjustments with extremely low computational cost. Only when the data indicates that the risk status has undergone a "qualitative change" is a comprehensive recalculation triggered. This mechanism reduces the update latency of the knowledge base from hours to minutes or even seconds, while reducing the normal computational resource consumption by more than an order of magnitude. Thus, without significantly increasing hardware costs, it ensures that the knowledge base can reflect the latest risk status in real time.

[0032] 2. The system can discover newly emerging risk associations (breadth) that traditional static knowledge bases cannot capture, remain sensitive to subtle changes in the risk characteristics of historical subjects, and provide early warnings of the accumulation and evolution trends of risks (depth). The application layer makes decisions based on this high-quality dynamic knowledge, which improves the accuracy of risk warnings and reduces misjudgments and omissions caused by information lag.

[0033] 3. The system's "intelligence" focus shifts from complex and difficult-to-control decision-making models to continuous optimization of the knowledge base content itself. By collecting feedback on the actual effects of decisions (such as early warning accuracy and business impact), the system can periodically and automatically adjust the rule parameters (such as risk thresholds) in the core knowledge base. This makes the risk control strategy no longer a rigid set of static rules, but a "living" system that can continuously learn from business practices and self-calibrate. It can dynamically balance the relationship between risk control and business development. For example, it can automatically relax restrictions to promote business when the false alarm rate is high, and automatically tighten the strategy to ensure safety when risks rise, ultimately maximizing risk and return. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall architecture of the present invention;

[0035] Figure 2 This is a schematic diagram of the data acquisition and fusion steps of the present invention;

[0036] Figure 3 This is a schematic diagram of the steps involved in constructing and updating the dynamic knowledge graph of this invention.

[0037] Figure 4 This is a schematic diagram of the risk knowledge generation and construction steps of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1-4 This invention provides a technical solution: a method for constructing a credit risk knowledge base that integrates multi-source data from internal audit and risk control, comprising the following steps:

[0040] Step S1: Data Acquisition and Fusion

[0041] S11. Data Acquisition: Raw heterogeneous data is acquired from the enterprise's internal audit system (collecting audit reports and internal control deficiency records), risk control system (collecting real-time transaction logs and risk event alarms), and external audit system (collecting audit opinions and compliance inspection results) through various methods such as system API interfaces, database log capture, and file transfer. To address cross-institutional data privacy issues, a federated learning-based architecture is preferred: edge computing nodes are deployed within each participating institution. These nodes perform feature extraction on sensitive raw data locally (e.g., using a multilayer perceptron (MLP) to extract time-series features of transaction logs), and only the encrypted model feature parameters are uploaded to the central server for secure aggregation (e.g., using FedAvg or FedProx algorithms), without moving the raw data throughout the process.

[0042] S12. Semantic Alignment and Standardization:

[0043] S121. Entity and Relation Extraction: For unstructured text (such as audit reports), a Transformer-based pre-trained language model (such as BERT-base) is fine-tuned to identify and extract "risk entities" (such as "Customer A" and "overdue events") and "risk relationships" (such as "trigger" and "association"). To enhance extraction performance, Graph Convolutional Networks (GCNs) can be introduced to model syntactic dependencies in the text, improving the accuracy of relation extraction.

[0044] S122. Ontology Mapping: A unified credit risk domain ontology is pre-constructed using tools such as Protégé, defining core concepts such as "customer," "risk event," "product," and "compliance measures," along with their attributes and relationships. The entities and relationships extracted from various data sources in the previous step are aligned to the standard concepts defined in this ontology through semantic similarity calculation (such as cosine similarity) and rule mapping. For example, the "transaction anomaly" event in the risk control system and the "internal control failure" event in the internal audit system are uniformly mapped to the ontology's concept of "operational risk event."

[0045] Step S2: Dynamic Knowledge Graph Construction and Update

[0046] Step S21, Initial Graph Construction: Import the standardized structured risk data generated in Step S1 into a graph database (such as Neo4j). Construct an initial multi-source data association graph using entities as nodes, relationships as edges, and the attributes of entities and relationships as attributes.

[0047] Step S22, Dual-path incremental update mechanism:

[0048] Step S221, New Risk Entity Identification and Complete Calculation Path: When new input data flows in, the system first searches the graph based on the entity's unique identifier (such as the enterprise's unified social credit code). If the entity does not exist, it is identified as a new risk entity, triggering the complete calculation path. Full feature extraction (including basic attributes, behavioral features, etc.) is performed on this entity, and a graph embedding algorithm (such as TransE, RotatE) is used to embed it into the graph's vector space, establishing its association with existing entities in the graph.

[0049] Step S222, Lightweight update path for historical risk subjects: If the new data corresponds to a historical risk subject, the lightweight update path is triggered.

[0050] Confidence interval construction: Based on the feature vectors of the directly related nodes of the historical subject in the graph (such as its past risk events, industry affiliation, etc.), the mean of these vectors is calculated as the feature center point O, and the standard deviation σ is calculated. A dynamic confidence interval is defined in this way, for example, a hyperspherical region centered at O ​​with a radius of 2σ.

[0051] Distance calculation and update decision: Calculate the Euclidean distance D between the new data feature vector V and the center point O.

[0052] If D < the preset threshold T (e.g., T = 0.3, which can be verified through historical data), then the risk change represented by the new data is considered to be within the expected range. In this case, linear interpolation is used to fine-tune the map embedding vector of the historical subject: V_new = α * V_old + (1 - α * V (where α is a decay factor close to 1, such as 0.9), to achieve fast and low-cost updates.

[0053] If D ≥ T, then the risk characteristics of the subject are considered to have changed drastically (such as the emergence of a completely new high-risk type), and a complete calculation path is immediately triggered for it to perform full feature extraction and graph embedding to ensure the accuracy of knowledge.

[0054] Step S3: Risk Knowledge Generation

[0055] Step S31, Knowledge Element Extraction: Using a graph query language (such as Cypher), extract "entity-relationship-entity" triples from the graph, along with attributes, to form standardized risk knowledge elements. For example, query and extract "(Customer A)-[occurred]->(Overdue Event B){Amount: 1 million, Time: 2023-10-01}".

[0056] Step S32, Transmission Path Mining: Using graph algorithms (such as Breadth-First Search (BFS) or Shortest Path Algorithm), the transmission chains of risk in the entity network are mined. For example, the transmission path "Customer A experiences financial difficulties -> Credit rating of its guarantor B declines -> Financing costs of enterprise C in the supply chain increase" is mined, and the confidence level of the path is calculated.

[0057] Step S4: Building a Layered Knowledge Base

[0058] Step S41, Base Layer: Use a relational database (such as MySQL) to store the raw knowledge elements extracted from the graph, as well as the raw data after standardization.

[0059] Step S42, Core Layer: Utilize high-performance caching (such as Redis) and model file storage systems. In this layer, the knowledge from the foundational layer is abstracted and refined to form general risk rules (such as "customers with three consecutive overdue payments are automatically marked as high-risk") and risk models (such as customer credit scoring models).

[0060] Step S43, Application Layer: Utilize a combination of graph databases and relational databases to store personalized risk strategies and real-time decision results tailored to specific business scenarios (such as credit approval and post-loan monitoring).

[0061] Step S5: Risk Decision Application and Feedback Optimization

[0062] Step S51, Risk Decision Application: When a business system initiates a request (such as a credit application), the application layer calls the corresponding scenario-based rules and combines them with the core layer's model for comprehensive judgment, outputting the risk level and decision recommendations (such as "Approve," "Reject," or "Manual Review"). The decision-making process requires low latency (e.g., response within 500ms).

[0063] Step S52, Feedback Optimization: The system continuously collects feedback on the actual effects of decisions, including risk control effectiveness indicators (such as early warning accuracy rate and risk loss recovery amount) and business impact indicators (such as business success rate and customer satisfaction). Based on this feedback data, the system periodically (e.g., weekly) dynamically adjusts the threshold parameters of risk rules in the core layer using multi-objective optimization algorithms (such as genetic algorithms and particle swarm optimization). For example, if the false alarm rate of a rule is too high, its triggering conditions are automatically relaxed, thereby achieving a dynamic balance between risk control and business development.

[0064] Working Principle: First, through data collection and fusion steps, semantic alignment and standardization are performed on heterogeneous data from multiple sources such as internal audit, risk control, and auditing using pre-trained language models and ontology mapping technology to generate unified structured risk data, laying the foundation for knowledge construction. Next, in the core step of dynamic knowledge graph construction and updating, the structured data is constructed into a multi-source association graph based on the credit risk domain ontology. A dual-path incremental update mechanism is adopted to achieve intelligent real-time maintenance of the graph—when new data corresponds to a new risk subject, a full calculation path is triggered to ensure complete information entry; when it corresponds to a historical risk subject, the distance between the confidence intervals of the new data features and historical features is calculated to intelligently determine whether the risk change is a "quantitative change" or a "qualitative change," thereby selecting to perform a lightweight partial update or trigger a full update. This approach fundamentally solves the problem of static knowledge, ensuring the timeliness and accuracy of the knowledge base with low computational overhead. Then, it extracts risk entities, relationships, and transmission paths from the real-time updated knowledge graph, generating structured risk knowledge elements. These elements are then stored and managed hierarchically according to the foundational, core, and application layers, achieving an orderly organization of knowledge from abstract to concrete. Finally, in the risk decision-making application and feedback optimization phase, the application layer utilizes dynamic knowledge for accurate and rapid risk identification and decision-making. Simultaneously, the system continuously collects decision feedback data and adaptively adjusts the rule parameters in the core knowledge base based on a multi-objective optimization algorithm, forming an autonomous closed loop of "decision-feedback-optimization." Ultimately, this enables risk management strategies to dynamically balance business development and risk control, achieving an evolution from a static knowledge base to an intelligent adaptive system.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control, characterized in that... Includes the following steps: Step S1, Data Acquisition and Fusion: Collect heterogeneous data from internal audit, risk control, and audit systems, and perform semantic alignment and standardization on the heterogeneous data to generate structured risk data; Step S2, Dynamic Knowledge Graph Construction and Update: Based on the credit risk domain ontology, the structured risk data is constructed into a multi-source data association graph, and the graph is updated in real time using an incremental update mechanism; Step S3, Risk Knowledge Generation: Extract risk entities, risk relationships, and risk transmission paths from the updated multi-source data association graph to generate structured risk knowledge elements. Step S4, Layered Knowledge Base Construction: Store the risk knowledge elements in a basic layer, core layer, and application layer according to data abstraction levels, forming a three-layer knowledge base architecture. The incremental update mechanism in step S2 is a dual-path update strategy, which specifically includes: when the new input data corresponds to a new risk subject in the knowledge base, a full calculation path is triggered to perform full feature extraction and graph embedding on the new risk subject; when the new input data corresponds to a historical risk subject in the knowledge base, a lightweight update path is triggered to determine the degree of risk change based on the existing feature information of the historical risk subject, and a differentiated update is performed accordingly.

2. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 1, characterized in that, The lightweight update path specifically includes: Based on the historical data of the aforementioned historical risk subjects, determine the confidence interval of their risk characteristic vector; Calculate the distance between the feature vector of the new input data and the center point of the historical risk subject features; If the distance is less than a preset threshold, the risk change is determined to be within the expected range, and a lightweight linear adjustment is made to the graph embedding vector of the historical risk subject. If the distance is greater than or equal to a preset threshold, it is determined that the risk characteristics have changed drastically, and the complete calculation path is triggered for the historical risk subject.

3. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 2, characterized in that, The method for determining the confidence interval includes: calculating the mean and standard deviation of the feature vector of the set of adjacent nodes of the historical risk subject to define the confidence interval.

4. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 1, characterized in that, Step S1, which involves semantic alignment and standardization of heterogeneous data, specifically includes: Entity recognition and relation extraction from unstructured text using a Transformer-based pre-trained language model; By employing ontology mapping technology, entities and relationships extracted from different data sources are mapped to the concepts and attributes defined in the unified ontology of the credit risk domain.

5. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 1, characterized in that, The method also includes step S5, risk decision application and feedback optimization: Utilizing the knowledge base for risk identification and decision support; Collect application feedback data for risk decision-making, including at least risk control effectiveness indicators and business impact indicators; Based on the feedback data, the risk rule parameters in the core layer and / or application layer are adaptively optimized to achieve a dynamic balance between risk control and business development.

6. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 5, characterized in that, The adaptive optimization of risk rule parameters is based on a multi-objective optimization algorithm, which dynamically adjusts the rule threshold to optimize the comprehensive objective function consisting of risk control benefits and business impact costs.

7. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 1, characterized in that, In step S1, a federated learning-based architecture is used for cross-institutional data collaboration, specifically including: Edge computing nodes are deployed locally at the data sources of each participating institution for local data feature extraction. A secure parameter aggregation algorithm is used to aggregate model parameter updates from each node on a central server without moving the original data.

8. The method for constructing a credit risk knowledge base integrating multi-source data from internal audit and risk control as described in claim 1, characterized in that, The method also includes a blockchain-based audit and traceability step: generating cryptographic hash values ​​for key operations such as data collection, knowledge updating, and risk decision-making, and storing them in the blockchain to achieve traceability and immutability of the operations.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images