Knowledge graph-based financial risk correlation mining model construction method and system

Abstract

The application discloses a kind of financial risk association mining model construction method and system based on knowledge graph, it is related to financial risk data mining and knowledge graph construction technical field, including acquisition enterprise finance and associated multi-source data, through standardization API interface docking;Clean up and preprocess structured and unstructured data, fill in missing values, delete duplicate data and convert unstructured data;Build the financial risk knowledge graph containing entity, attribute, relationship, store with graph database;Extract three kinds of financial risk characteristics;With graph neural network as foundation to build model, cross-validation divides dataset training;Correlation analysis risk result, trace source and transmission path and generate report;Periodically iterated update model and graph, set cycle and trigger condition.The application improves data reliability, multi-source integration and fourfold check filtering problem data;Enhance risk mining accuracy, entity alignment and attention mechanism optimize model, adapt to different industries.

Description

Technical Field

[0001] This invention relates to the field of financial risk data mining and knowledge graph construction technology, and in particular to a method and system for constructing a financial risk association mining model based on knowledge graphs. Background Technology

[0002] Traditional corporate financial risk assessments often rely on manual analysis and single data sources, focusing on indicators such as the debt-to-equity ratio and net profit in the company's own financial statements, while neglecting the impact of related data such as supply chain transactions, market sentiment, and regulatory penalties. For example, a company may have normal short-term financial indicators, but its upstream core suppliers have a large number of overdue payment records. Traditional methods struggle to identify this cross-entity risk transmission, leading to delayed risk warnings. Furthermore, unstructured data processing is inefficient, and manual extraction of key information is time-consuming and error-prone, resulting in the omission of a large amount of valuable risk-related information. This prevents the formation of a complete risk assessment dimension and creates significant blind spots in risk identification.

[0003] Existing knowledge graph-based financial risk analysis solutions have significant shortcomings in entity processing and model construction. During knowledge graph construction, inconsistencies in the representation of the same entity across different data sources (e.g., "XX Group Co., Ltd." versus "XX Group") are not effectively addressed. Entity alignment relies heavily on simple text matching, failing to incorporate comprehensive judgments based on multiple attributes such as unified social credit code and registered address. This leads to duplicate or missed entity identifications, resulting in biased risk association detection. Furthermore, the allocation of model feature weights lacks scientific basis, assigning equal weight to secondary sentiment features and core financial indicators, weakening the model's sensitivity to key risk factors. Some models lack attention mechanisms, paying insufficient attention to entity relationships with high risk transmission probabilities in the knowledge graph, resulting in poor generalization ability and a significant drop in prediction accuracy when facing differences in financial risk characteristics across different industries (e.g., manufacturing and services).

[0004] The existing system also suffers from a lack of model iteration mechanisms and poor module synergy. Most financial risk models have not been updated for extended periods after deployment, making them unable to adapt to new financial risks (such as virtual asset risks in the digital economy) and regulatory policy adjustments. When new risk transmission patterns emerge in the industry, the models struggle to respond effectively. Furthermore, data flow between system modules lacks standardization, and the data collection process does not implement strict compliance and timeliness checks, leading to the entry of false financial data and outdated historical data into subsequent processing. The model training and risk assessment modules are loosely connected, lacking dual verification of prediction results (historical data verification and expert review). Some high-risk enterprises are misjudged as low-risk due to model bias, increasing business decision-making risks. In addition, the system lacks visual interaction and operational traceability functions, making it difficult for financial analysts to intuitively view risk correlation paths, and regulatory personnel cannot trace data access and model adjustment records, failing to meet data security management requirements. Summary of the Invention

[0005] The present invention proposes a method and system for constructing a financial risk association mining model based on knowledge graphs to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a financial risk association mining model based on a knowledge graph, comprising the following steps: Collect corporate financial and related data, including annual financial statements, quarterly operating reports, bank statements, supply chain transaction data, public market sentiment data, and industry regulatory penalty information, encompassing both structured and unstructured data; connect to data sources through standardized API interfaces; The collected data is cleaned and preprocessed. Missing values ​​in structured data are filled with the mean and mode, respectively, and duplicate data is removed based on key fields. Key information in unstructured data is extracted using natural language processing technology and converted into structured data tables through word segmentation, part-of-speech tagging, and entity recognition. Construct a knowledge graph of corporate financial risk, with core entity types including corporate entities, financial indicators, counterparties, regulatory agencies, and risk events. It is stored in a graph database and supports adding, deleting, modifying, querying, and batch importing and exporting of entities and relationships. Three categories of financial risk characteristics are extracted: basic characteristics, correlation characteristics, and public opinion characteristics. All characteristics are then standardized. Based on knowledge graphs, a financial risk association mining model is constructed using graph neural network algorithms. The model is input with standardized features and entity relationship matrices, and outputs three levels of financial risk: low, medium, and high, after passing through an input layer, graph convolutional layer, attention layer, and fully connected layer. Perform correlation analysis on the model output results, combine knowledge graphs to trace the source of risks, explore the risk transmission path, and generate a risk correlation analysis report; The model and knowledge graph are updated regularly. Updates are triggered when more than 1,000 new data entries are added, regulatory policies are adjusted, or new risk events occur. During the update, data collection, preprocessing, and feature extraction steps are performed, model parameters are fine-tuned, and the knowledge graph is updated.

[0007] Furthermore, it also includes a step of weighting financial risk characteristics. This weighting is calculated using a multi-dimensional assessment formula, which is as follows: in For the first The weights of each financial risk characteristic; For the first The risk significance of each feature is calculated by the frequency and contribution of the feature in historical risk events; For the first The correlation between each characteristic and financial risk is calculated using the Pearson correlation coefficient to determine the degree of linear correlation between the characteristic value and the risk level. This represents the total number of financial risk characteristics.

[0008] Furthermore, it also includes performing entity alignment operations during the knowledge graph construction process to eliminate the differences in the representation of the same entity in different data sources. The entity alignment step first extracts the attribute set of entities from each data source. The attribute set of enterprise entities includes the unified social credit code, enterprise name, registered address, and legal representative. The attribute set of financial indicator entities includes the indicator name, calculation method, and statistical period. Then, the attribute similarity between entities is calculated. The similarity of the text attribute enterprise name is calculated using the edit distance algorithm, and the similarity of the numerical attribute registered capital is calculated using the Euclidean distance algorithm. Finally, a similarity threshold is set.

[0009] Furthermore, when collecting corporate financial and related data, a new data quality verification step has been added: Completeness verification compares the data with preset fields set according to the data source type. Corporate financial statements contain core fields. If the missing rate exceeds 5%, the data will be reconnected. If the data is still missing, the record will be marked. Accuracy verification compares the data with relevant data through cross-validation. If the deviation exceeds 10%, two or more financial specialists will be asked to conduct independent reviews. Timeliness verification sets a 30-day or 7-day interval threshold according to the data type. Compliance verification is based on relevant regulations, and non-compliant data is removed.

[0010] Furthermore, when constructing the financial risk association mining model, an attention mechanism is introduced to optimize the graph neural network layer. By calculating the importance of the association between entities, attention weights are assigned, and three parameters are extracted: the proportion of transaction amount between adjacent entities, the duration of cooperation, and historical risk association records. The risk transmission probability is calculated as a weight to highlight the relationship between entities with high risk transmission probability. In the later stage of model training, a dropout layer is introduced to randomly discard some neuron connections.

[0011] Furthermore, when performing correlation analysis on the risk results output by the model, a dual verification step for risk levels is added: The first step is historical data verification, which selects data from no fewer than 1,000 companies with complete risk records from the past 5 years as the verification set, compares the predicted results with actual risk events, and calculates the accuracy, recall, and F1 score. If any indicator fails to meet the standard, the model parameters are adjusted. The second step is expert verification, in which no fewer than 3 financial risk assessment experts with more than 5 years of experience independently review the list of high-risk companies, risk transmission paths, and correlation analysis reports, and calculate the deviation rate. If the deviation exceeds 10%, the feature weights are reassigned and the model is fine-tuned. Only the results that pass the verification can be used for business decisions.

[0012] Furthermore, when periodically iterating and updating the model and knowledge graph, a new update effect evaluation step is added. The evaluation indicators include the entity coverage and relation integrity of the knowledge graph, and the prediction accuracy and recall of the model. The entity coverage and relation integrity are calculated as the proportion of the number of newly added entities and relations after the update to the total number before the update, respectively. The model indicators are calculated using the test set containing newly added enterprise data after the update. If the entity coverage is lower than 15% or the relation integrity is lower than 20%, the data collection scope is optimized. If the model prediction accuracy drops by more than 5% or the recall drops by more than 8%, the model is rolled back and a new iteration plan is formulated.

[0013] Furthermore, it includes the following modules: The data acquisition module is configured with a standardized API interface library containing multiple data sources, and authentication is required for calls; it has a built-in task scheduler that supports preset periods and manual triggering modes, and the collected data is temporarily stored in a Redis temporary cache, which is automatically cleaned up after 24 hours of inactivity; The data preprocessing module connects to the acquisition module and includes built-in data cleaning and unstructured data conversion tools. The cleaning tools include units for handling missing values, duplicate data, and outlier data, while the conversion tools integrate natural language processing algorithms to convert unstructured data into structured data tables. The processed data undergoes format validation and is then transmitted to the knowledge graph construction and feature extraction module. The knowledge graph construction module connects to the preprocessing module and includes entity definition, relation definition, entity alignment, and graph database storage units; it pre-defines five types of core entities and three types of basic relations; the entity alignment unit completes matching through multiple sub-units and uses distributed graph database storage; The feature extraction module, connected to the preprocessing module, includes feature classification and extraction, standardization, and weight allocation units. It extracts features according to three categories, maps feature values ​​using a standardization method, calculates feature weights according to a specified formula and supports manual fine-tuning, and transmits the results to the model training module. The model training module connects the feature extraction and knowledge graph modules and contains multi-functional units. It adopts a hybrid graph neural network architecture, divides the dataset through hierarchical cross-validation, optimizes parameters with the Adam optimizer, and integrates attention mechanisms and overfitting control methods. The risk assessment module connects the model training and knowledge graph modules and contains multiple core units; it supports batch and single-enterprise risk assessment, combines knowledge graphs to trace the source of risks, explore transmission paths, automatically generate risk analysis reports in multiple formats, and perform dual verification. The model iteration module connects multiple modules and includes units for iteration triggering, update execution, effect evaluation, optimization, and rollback. It supports periodic and condition-triggered updates, executes according to a fixed process, evaluates key indicators of the knowledge graph and model, and optimizes the data source or rolls back the model and formulates new iteration plans based on the evaluation results. The data storage module adopts a distributed architecture and is divided into five functional storage areas, each storing corresponding data; all storage areas are encrypted.

[0014] Furthermore, it also includes a risk association strength calculation module, which is connected to the knowledge graph construction module and the risk assessment module; the module includes a parameter acquisition unit, an intensity calculation unit, and a result application unit; the parameter acquisition unit extracts transaction frequency, cooperation association duration, and historical risk transmission probability between entities from the knowledge graph construction module. For entities and Total number of transactions in the past 12 months For entities and The number of years since the partnership was established. For entities in the past 3 years Entities due to risk events The number of times financial irregularities occurred accounted for the majority of entities The proportion of total risk events; the intensity calculation unit uses the formula Calculate the association strength, where For entities With entity The strength of the financial risk association; This is the transaction frequency weighting coefficient; This refers to the correlation duration weighting coefficient; This is the weighting coefficient for the risk transmission coefficient; , , All passed Standardized mapping.

[0015] Furthermore, it also includes a user interaction module, connected to the data storage module and risk assessment module; this module comprises three main units: interface display, access control, and operation log; the interface display unit is divided into a knowledge graph visualization area, a risk report area, and a task management area: the knowledge graph visualization area uses a guide map to display entities and relationships; the risk report area supports multi-condition retrieval; the task management area supports initiating data collection, model training, and risk assessment tasks, and displays task progress; the access control unit divides users into three roles, assigning permissions according to the model, with financial analysts, regulators, and system administrators each having corresponding operation permissions.

[0016] Compared with existing technologies, the beneficial effects of this invention are: At the data processing level, this invention achieves comprehensive integration and rigorous quality control of multi-source financial and related data, effectively solving the problems of single data dimensions and inconsistent quality in traditional methods. Through standardized API interfaces, it connects to various data sources such as enterprise financial systems, supply chain platforms, and public opinion monitoring, covering both structured and unstructured data. Simultaneously, it performs four checks—completeness, accuracy, timeliness, and compliance—filtering out false data and illegal information, significantly improving data reliability. This provides a comprehensive and high-quality data foundation for risk mining, avoiding misjudgments due to missing or biased data.

[0017] At the knowledge graph construction and model optimization levels, the accuracy of risk association mining and the model's generalization ability are significantly improved. The entity alignment process combines multiple attributes and multiple algorithms to calculate similarity, effectively avoiding mining bias caused by entity duplication or omission. Feature weight allocation quantifies the impact of key indicators through multi-dimensional evaluation, and the attention mechanism focuses on entity relationships with high risk transmission probability, weakening interference from low correlation, enabling the model to accurately capture cross-entity risk transmission (such as the transmission of defaults between upstream and downstream of the supply chain). At the same time, through dropout layers and incremental training, the model's adaptability to different industries and new risks is enhanced, solving the pain points of poor generalization and weak identification of associated risks in traditional models.

[0018] At the risk analysis and system adaptation level, a full-process upgrade is achieved, encompassing risk tracing, iterative optimization, and secure interaction. The risk assessment module traces the source and transmission path of risks, generating detailed reports with response recommendations to provide clear guidance for business decisions. Model iteration incorporates effectiveness evaluation, automatically optimizing or rolling back when standards are not met, ensuring the system's long-term adaptability to risk changes. The user interaction module displays risk correlations through visual graphs, combined with RBAC access control and operation logs, satisfying the operational needs of different roles while ensuring traceable data access and compliance with data security regulations. Overall, this significantly improves the comprehensiveness, timeliness, and reliability of financial risk assessment. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the method for constructing a financial risk association mining model based on knowledge graphs proposed in this invention. Figure 2 This is a schematic block diagram of the system for constructing a financial risk association mining model based on knowledge graphs proposed in this invention. Figure 3 A bar chart comparing the weights of core financial risk characteristics; Figure 4 A line chart showing the monthly trend of entity association strength in a knowledge graph; Figure 5 Grouped bar chart comparing the accuracy of financial risk identification using different methods; Figure 6A bar chart comparing the risk warning lead time before and after model iteration. Detailed Implementation

[0020] 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.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 6 A method for constructing a financial risk association mining model based on knowledge graphs, comprising the following steps: We collect enterprise financial data and related data, with data sources including annual financial statements, quarterly operating reports, bank transaction records, supply chain transaction data, public market sentiment data, and industry regulatory penalty information. The data formats are divided into structured and unstructured data. Structured data specifically includes Excel spreadsheets and database tables, while unstructured data specifically includes PDF reports and text-based sentiment data. The collection process connects to various data sources through standardized API interfaces, which support HTTPS protocol and JSON data format. During data transmission, a data verification mechanism is used, performing an integrity check every 100 data entries. The collected data undergoes cleaning and preprocessing. For structured data, numerical missing data is imputed using the mean, while categorical missing data is imputed using the mode. Duplicate data is identified by uniqueness checks of key fields, including the company's unique identifier and transaction serial number. Data with identical key fields is considered duplicate and is deleted. For unstructured data, natural language processing technology is used to extract key information. Word segmentation, part-of-speech tagging, and entity recognition are performed sequentially to obtain financial terms and risk-related expressions, transforming the unstructured data into a structured data table containing entities, attributes, and relationships. A knowledge graph of enterprise financial risk is constructed, defining the core entity types as enterprise entities, financial indicators, counterparties, regulatory agencies, and risk events. Enterprise entity attributes include registered capital, industry, and credit rating; financial indicator attributes include numerical value, calculation period, and industry average; and counterparty attributes include enterprise size and credit rating. The relationships between entities are defined as follows: "transactional" relationship between enterprises and counterparties, "indicator attribution" relationship between enterprises and financial indicators, and "event association" relationship between enterprises and risk events. "Transactional" relationship attributes include transaction amount, transaction frequency, and settlement method; "event association" relationship attributes include risk event level and scope of impact. A graph database is used to store the knowledge graph, supporting CRUD operations and batch import / export of entities and relationships. Financial risk characteristics were extracted from the preprocessed data, including three categories: basic characteristics such as debt-to-equity ratio, current ratio, net profit growth rate, and net cash flow; correlation characteristics such as overdue payment rates of upstream and downstream enterprises in the supply chain, credit default frequency of counterparties, and cooperation stability scores; and public opinion characteristics such as frequency of negative public opinion mentions, correlation with regulatory penalties, and media attention index. All characteristics were standardized and processed using... The standardization method maps eigenvalues ​​to the [0,1] interval. The standardization calculation formula is as follows: ,in These are the original eigenvalues. This is the minimum value of the feature. This is the maximum value of the feature; A financial risk correlation mining model is constructed. Based on a knowledge graph, a graph neural network algorithm is used to build the model architecture. The model input is the standardized financial risk features and the entity relationship matrix of the knowledge graph. The model architecture includes an input layer, a graph convolutional layer, an attention layer, a fully connected layer, and an output layer. The graph convolutional layer is used to learn the risk correlation features between entities, the attention layer is used to strengthen the weights of key entity relationships, and the fully connected layer is used to integrate features and output the enterprise's financial risk level. The risk level is divided into low risk, medium risk, and high risk. The dataset is divided into training set and test set using cross-validation, with the training set accounting for 70% and the test set accounting for 30%. The learning rate and iteration number parameters of the model are adjusted with the risk level prediction accuracy as the optimization objective. The model outputs risk results and performs correlation analysis. By combining entity relationships in the knowledge graph, the source of risk is traced. Risk sources include upstream supplier default, abnormal financial indicators, and changes in regulatory policies. The risk transmission path is explored. A typical transmission path is supplier default → raw material shortage → production stagnation → cash flow shortage → escalation of financial risk. A risk correlation analysis report is generated, which includes the enterprise risk level, a list of related entities, a risk transmission path diagram, a risk impact assessment, and response recommendations. The model and knowledge graph are updated regularly, with a regular update cycle of once a month. The update trigger conditions are set as more than 1,000 new financial data entries, adjustments to industry regulatory policies, and the emergence of new financial risk events. The update process sequentially executes data collection, data preprocessing, and feature extraction steps. The model parameters are fine-tuned based on the new data, and the entity attributes and relationships in the knowledge graph are updated simultaneously. After the update, the model performance and knowledge graph coverage before and after the update are compared to ensure that the system performance meets business requirements after the update.

[0024] This invention also includes a step of weighting financial risk characteristics. The feature input is optimized by quantifying the impact of different characteristics on financial risk. The weighting is calculated using a multi-dimensional evaluation formula, which is as follows: in For the first The weights of each financial risk characteristic range from [0,1]; For the first The risk significance of a feature is calculated by the frequency and contribution of the feature in historical risk events; the higher the frequency and the greater the contribution, the greater the risk significance. The larger the value, the range of values ​​is [0, 10]. For the first The correlation between each characteristic and financial risk is calculated using the Pearson correlation coefficient, which determines the degree of linear correlation between the characteristic value and the risk level. A higher correlation indicates a greater risk level. The larger the value, the range of values ​​is [0,1]. This represents the total number of financial risk features. Through this weighting, features with a greater impact on risk, such as the debt-to-equity ratio and net cash flow, are given higher weights in model training. The weight of the debt-to-equity ratio is typically allocated to 0.15-0.2, and the weight of net cash flow is typically allocated to 0.12-0.18. Secondary features, such as the media attention index, are allocated to 0.03-0.05. This improves the model's sensitivity to key risk factors and reduces the interference of secondary features on the model's prediction results.

[0025] Claim 3, the method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, is characterized in that it further includes performing entity alignment operations during the knowledge graph construction process to eliminate the differences in the representation of the same entity in different data sources; the entity alignment step first extracts the attribute sets of entities in each data source, the attribute set of enterprise entities includes unified social credit code, enterprise name, registered address, and legal representative, and the attribute set of financial indicator entities includes indicator name, calculation caliber, and statistical period; then, it calculates the attribute similarity between entities, the similarity of the text attribute enterprise name is calculated using the edit distance algorithm, the smaller the edit distance, the higher the similarity, and the similarity of the numerical attribute registered capital is calculated using the Euclidean distance algorithm, the smaller the Euclidean distance, the higher the similarity; finally, it sets a similarity threshold, the similarity threshold for text attributes is set to 0.85, and the similarity threshold for numerical attributes is set to 0.9, entities with similarity exceeding the threshold are determined to be the same entity, their attributes and associations are merged, duplicate entity entries are deleted, ensuring the uniqueness and consistency of entities in the knowledge graph, and reducing the risk association mining bias caused by entity duplication or representation differences.

[0026] In this invention, when collecting enterprise financial data and related data, a data quality verification step is also included. The verification dimensions cover data completeness, accuracy, timeliness, and compliance. Data completeness verification checks for missing fields by comparing with a preset field list. The preset field list is set according to the data source type. The preset fields for enterprise financial statements include total assets, total liabilities, operating revenue, and net profit. If the proportion of missing fields exceeds 5%, the data source is reconnected. Missing fields that are still not supplemented after reconnection are marked as "to be supplemented" and recorded in the data quality report. Data accuracy verification uses cross-validation, comparing the revenue data in the financial statements with the payment data in the bank transaction records, and comparing the supply chain transaction data. If the discrepancy between the purchase amount and accounts payable data exceeds 10%, it is marked as abnormal data and a manual review is initiated. The manual review must be completed independently by at least two finance specialists. Data timeliness verification checks the interval between data collection time and data generation time. The interval threshold for financial statements and operating reports is set at 30 days, and the interval threshold for bank statements and transaction data is set at 7 days. Historical data with intervals exceeding the threshold must be supplemented with the latest data. Data compliance verification is based on the "Enterprise Accounting Standards", "Data Security Law" and "Personal Information Protection Law". Data containing un-anonymized personal information, false financial indicators and illegal transaction records is filtered out. Data that fails the compliance verification is directly removed and does not proceed to the subsequent preprocessing steps.

[0027] In this invention, when constructing the financial risk association mining model, an attention mechanism is introduced to optimize the graph neural network layer. The attention mechanism assigns different attention weights based on the importance of the association between entities. Specifically, for each entity in the knowledge graph, three parameters are extracted: the proportion of transaction amount between the entity and its neighboring entities, the duration of cooperation, and historical risk association records. The proportion of transaction amount is the ratio of the entity's annual transaction amount with its neighboring entities to its total annual transaction amount. The duration of cooperation is the number of years the entities have established a cooperative relationship. Historical risk association records are the number of times in the past three years that one entity's risk has affected the other. Based on these three parameters, the risk transmission probability between entities is calculated: Risk Transmission Probability = 0.4 × Proportion of Transaction Amount + 0.3 × Cooperation The model employs a duration coefficient plus 0.3 × historical risk correlation coefficient. The duration coefficient is set to 0.3 for 1-3 years, 0.6 for 3-5 years, and 0.9 for over 5 years. The historical risk correlation coefficient is set to 0.1 for 0 records, 0.5 for 1-2 records, and 0.9 for over 3 records. Risk transmission probability is used as an attention weight, ensuring the model focuses on high-risk transmission probability entities during learning, while mitigating the impact of low-correlation entities. Furthermore, a dropout layer is introduced in the later stages of model training, with a dropout probability of 0.2, randomly discarding 20% ​​of neural connections to prevent overfitting of training data and improve the model's predictive adaptability to financial risks across different industries and enterprise sizes.

[0028] In this invention, when performing correlation analysis on the risk results output by the model, a risk level verification step is also included. The verification method adopts a dual verification mechanism. The first step is historical data verification, selecting enterprise data containing complete risk event records from the past 5 years as the verification set. The sample size of the verification set is no less than 1000 enterprises. The model's risk prediction results for the enterprises in the verification set are compared with the actual risk events, and three indicators are calculated: prediction accuracy, recall, and F1 score. Prediction accuracy is the proportion of samples whose predicted risk level matches the actual risk level out of the total sample size. Recall is the proportion of actual high-risk samples that are correctly predicted as high-risk by the model. The F1 score is the harmonic mean of accuracy and recall. If the accuracy is lower than 85%, the F1 score is calculated. If either the recall rate is below 80% or the F1 score is below 82%, the system will return to the model training step to adjust the learning rate and attention weight parameters. The second step is expert verification, in which a verification team of at least three financial risk assessment experts with no less than five years of experience will be invited to independently review the high-risk enterprise list, risk transmission path, and correlation analysis report output by the model. The expert judgment deviation rate will be calculated, which is the proportion of samples whose expert judgment is inconsistent with the model result to the total number of samples. If the deviation rate exceeds 10%, feature weight reallocation and model fine-tuning will be initiated. During the fine-tuning process, priority will be given to adjusting the model parameters corresponding to the features with large deviations pointed out by the experts. Only the risk results that have passed the verification can be used for business decision-making.

[0029] In this invention, the periodic iterative updates of the model and knowledge graph also include an update effect evaluation step. Evaluation metrics include entity coverage, relational integrity, and the model's prediction accuracy and recall. Entity coverage is the ratio of newly added entities after the update to the total number of entities in the knowledge graph before the update; newly added entities include newly connected enterprises, newly emerging risk events, and newly added regulatory agencies. Relational integrity is the ratio of newly added relationships after the update to the total number of relationships in the knowledge graph before the update; newly added relationships include newly established transaction relationships and newly associated risk event relationships. Model prediction accuracy and recall are calculated using the updated test set, which includes newly added enterprise data during the update cycle. If entity coverage is below 15% or relational integrity is below 20%, the data collection scope is optimized by adding new industry enterprise data sources and new risk event data sources. If the model prediction accuracy decreases by more than 5% or the recall decreases by more than 8% compared to before the update, the model is rolled back to the previous version. Parameter adjustments and data quality issues during the update process are analyzed, and a new iterative plan including supplementary data collection and parameter gradient adjustments is developed to ensure that the performance of the updated model and knowledge graph is not lower than the level before the update.

[0030] This invention includes the following modules: The data acquisition module is configured with a standardized API interface library, which includes interfaces for enterprise financial systems, bank transaction systems, supply chain management platforms, public opinion monitoring platforms, and regulatory information release platforms. All interfaces support HTTPS protocol and JSON data format. Identity authentication is performed during interface calls using the OAuth2.0 protocol. The module has a built-in data acquisition task scheduler that supports two modes: preset period and manual trigger. The preset period can be set to 2:00 AM daily or 3:00 AM every Sunday. Manual triggering requires user permission verification before it can be initiated. The collected data is temporarily stored in a temporary cache using a Redis database. The cached data retention period is set to 24 hours, and it is automatically cleared if it is not accessed within this time limit. The data preprocessing module connects to the data acquisition module and includes built-in data cleaning and unstructured data conversion tools. The data cleaning tools include missing value handling, duplicate data handling, and outlier data handling units. The missing value handling unit automatically identifies numerical and categorical data and calls the corresponding filling methods. The duplicate data handling unit allows users to define key fields and perform uniqueness checks. The outlier data handling unit allows users to set deviation thresholds and mark outlier data. The unstructured data conversion tool integrates natural language processing algorithms: Jieba word segmentation, CRF part-of-speech tagging, and the BERT-BiLSTM-CRF model for entity recognition. The converted structured data table includes entity ID, attribute name, attribute value, and data source field. The processed data is then transmitted to the knowledge graph construction module and the feature extraction module, respectively. Data format verification is performed during transmission to ensure the data meets the input requirements of the receiving module. The knowledge graph construction module, connected to the data preprocessing module, includes entity definition units, relationship definition units, entity alignment units, and graph database storage units. The entity definition unit pre-defines five core entity types: enterprise entities, financial indicators, counterparties, regulatory agencies, and risk events. New entity types and corresponding attributes can be added, with attribute settings supporting numeric, text, and date types. The relationship definition unit pre-defines three basic relationship types: "transactions," "indicator affiliation," and "event associations," with customizable relationship names, attributes, and constraints. The entity alignment unit includes attribute extraction subunits, similarity calculation subunits, and threshold determination subunits. The attribute extraction subunit automatically extracts the core attributes of entities, the similarity calculation subunit uses edit distance and Euclidean distance algorithms, and the threshold determination subunit supports custom similarity thresholds for text and numeric attributes. The graph database storage unit uses the Neo4j distributed graph database, supporting batch import of entities and relationships at a rate of up to 1000 records per second. It also supports the Cypher query language, enabling complex entity relationship traversal queries. The feature extraction module, connected to the data preprocessing module, includes a feature classification extraction unit, a feature standardization unit, and a feature weight allocation unit. The feature classification extraction unit extracts features in three categories: basic features, related features, and public opinion features. Basic feature extraction calls a financial indicator calculation function; related feature extraction connects to a knowledge graph query interface to obtain related entity data; and public opinion feature extraction calls a public opinion analysis interface to obtain negative mention data. The feature standardization unit uses a minimum weighting algorithm. Standardization methods, through formulas Map the eigenvalues ​​to the interval [0,1], where These are the original eigenvalues. This is the minimum value of the feature in the entire dataset. The maximum value of the feature in the full dataset is standardized to generate a feature standardization report, which records the original extreme values ​​and standardization parameters of each feature. The feature weight allocation unit calls the multi-dimensional evaluation formula in claim 2 to automatically calculate the weight of each feature. It supports manual adjustment of the weight coefficients, with an adjustment range of ±0.05. After adjustment, a weight allocation table is generated and transmitted to the model training module along with the standardized features. The model training module, connected to the feature extraction and knowledge graph construction modules, includes a model architecture building unit, a dataset partitioning unit, a parameter optimization unit, an attention mechanism unit, and an overfitting control unit. The model architecture building unit employs a graph neural network architecture, specifically a hybrid architecture combining graph convolutional networks and graph attention networks. The input layer receives standardized features and an entity relationship matrix. Three graph convolutional layers are set, each outputting 128-dimensional features, and two graph attention layers are set, each with four attention heads. The dataset partitioning unit uses hierarchical cross-validation, partitioning by industry to ensure consistent proportions of samples from each industry in the training and test sets. The training set comprises 70% of the data, and the test set comprises... The parameter optimization unit aims to improve the accuracy of risk level prediction. It uses the Adam optimizer with an initial learning rate of 0.001. The learning rate decreases by 10% every 5 iterations, and the number of iterations is set to 100. Training stops when the accuracy on the test set does not improve for 5 consecutive iterations. The attention mechanism unit calculates entity association weights according to the method in claim 5 and automatically allocates attention resources. The overfitting control unit introduces a dropout layer with a fixed dropout probability of 0.2 and uses L2 regularization with a regularization coefficient of 0.001. The trained model is stored in the model library, which is managed by version and retains model files from the last 5 versions, supporting version rollback. The risk assessment module, connected to the model training module and knowledge graph construction module, includes a model invocation unit, a risk tracing unit, a path mining unit, a report generation unit, and a result verification unit. The model invocation unit supports batch and single-enterprise risk assessments; batch assessments can process data from 1000 enterprises at a time, and single-enterprise assessments have a response time of ≤3 seconds. The risk tracing unit, combined with the knowledge graph, uses Cypher queries to locate risk-related entities, trace the source of risk, and output the risk contribution of the source entity. The path mining unit uses a depth-first search algorithm to mine risk transmission paths of length ≤3, outputting the risk transmission probability of each entity in the path. The report generation unit automatically generates a risk association analysis report, which includes basic enterprise information, risk level and confidence level, a list of related entities, a transmission path diagram, risk impact assessment, and response suggestions. The report format supports PDF, Word, and Excel. The result verification unit performs the dual verification as described in claim 6, automatically compares the verification results with historical data, generates a verification report, supports importing expert verification opinions, and triggers a model fine-tuning instruction when verification fails, pushing it to the model training module. The model iteration module, connected to the data acquisition module, knowledge graph construction module, and model training module, includes an iteration trigger unit, an update execution unit, an effect evaluation unit, and an optimization and rollback unit. The iteration trigger unit supports both periodic and conditional triggers. Periodic triggers are executed at midnight on the 1st of each month, while conditional triggers include situations such as adding more than 1000 new data entries, regulatory policy adjustments, and the emergence of new risk events. The update execution unit automatically executes the process of "data acquisition → preprocessing → feature extraction → knowledge graph update → model fine-tuning." Fine-tuning uses incremental training, with 20% of the training set being new data and the remainder being the original training set data. The effect evaluation unit calculates the entity coverage and relation integrity of the knowledge graph, as well as the model's prediction accuracy and recall, and generates an evaluation report. The optimization and rollback unit automatically expands the data source type when entity coverage is <15% or relation integrity is <20%, and automatically rolls back to the previous model version when model accuracy drops >5% or recall drops >8%, while simultaneously analyzing the reasons for update failures and generating optimization solutions. The data storage module connects to all modules of the system and adopts a distributed storage architecture, divided into five functional storage areas: temporary cache area, preprocessed data area, knowledge graph data area, model data area, and report data area. All storage areas use the AES-256 encryption algorithm to encrypt data, and the encryption key is uniformly managed by the key management system and automatically rotated every 7 days. It supports data retrieval by multiple dimensions, and the search conditions include enterprise ID, data date, risk level, and report type. The search response time is ≤1 second, and all data access operations are recorded to generate access logs.

[0031] This invention also includes a risk association strength calculation module, connected to the knowledge graph construction module and the risk assessment module, used to quantify the degree of financial risk association between entities in the knowledge graph, providing a quantitative basis for risk tracing and path mining; the module includes a parameter acquisition unit, an intensity calculation unit, and a result application unit; the parameter acquisition unit extracts the transaction frequency between entities from the knowledge graph construction module (… ), duration of cooperation ( ), historical risk transmission probability ( ), For entities and Total number of transactions in the past 12 months For entities and Number of years since the partnership was established (accurate to 0.1 years). For entities in the past 3 years Entities due to risk events The number of times financial irregularities occurred accounted for the majority of entities The proportion of total risk events; the intensity calculation unit uses the formula Calculate the association strength, where For entities With entity The financial risk correlation strength, with a value range of [0,1]; This is the transaction frequency weighting coefficient, with a value of 0.4, because transaction frequency directly reflects the degree of business dependence between entities, and has the highest weighting. The association duration weighting coefficient is 0.3. The longer the association duration, the higher the possibility of risk transmission between entities. This is the risk transmission coefficient weighting coefficient, with a value of 0.3. Historical transmission records are an important reference for future risk transmission. , , All passed Standardize and map to the [0,1] interval; the result application unit will... Entity pairs with a correlation strength ≥ 0.7 are marked as "high correlation strength," those between 0.4 and 0.7 as "medium correlation strength," and those < 0.4 as "low correlation strength." The marking results are then transmitted to the risk assessment module. This module prioritizes tracing the paths of entities with high correlation strength to reduce the interference of low-correlation entities in the analysis, thereby improving the efficiency and accuracy of risk transmission path discovery. For example, when a company is identified as high-risk, the analysis of its related entities is prioritized. For counterparties with a trade ratio of ≥0.7, the source of risk spread can be quickly identified.

[0032] This invention also includes a user interaction module, connected to the data storage module and the risk assessment module, providing a visual operation interface and permission management functions to meet the operational needs of users with different roles. The module includes an interface display unit, a permission control unit, and an operation log unit. The interface display unit is divided into three core areas: a knowledge graph visualization area, a risk report area, and a task management area. The knowledge graph visualization area uses a force-directed graph to display entities and relationships, with node size corresponding to the risk correlation of the entity (the higher the correlation, the larger the node), and edge thickness corresponding to... The system supports dragging nodes, zooming in and out of views, and clicking on entities to view attributes and relationships. The risk report area allows searching reports by company name, risk level, and generation date, previewing key report information, and supports online viewing and downloading. The task management area supports initiating data collection tasks, model training tasks, and risk assessment tasks, displaying task progress. The access control unit categorizes users into three roles: financial analysts, regulators, and system administrators. Financial analysts can only view risk reports for their industry and initiate single-company assessment tasks; regulators can view industry-wide data and batch assessment tasks; and system administrators have full module operation permissions. Permission allocation uses a role-based access control (RBAC) model, supporting the addition, modification, and deletion of permissions. The operation log unit records all user operations, including user ID, user role, operation time, operation module, operation content, data access scope, and operation result (success / failure). The log is retained for 3 years, supports searching by user, time, and operation type, and cannot be modified or deleted, ensuring traceability and complying with the data access audit requirements of the Data Security Law.

[0033] The following examples further illustrate the specific implementation of this system: Example: Application of Financial Risk Correlation Mining in Large Manufacturing Groups This embodiment is applied to a large manufacturing group with 12 subsidiaries, covering the entire industrial chain from raw material procurement, component production, and finished product assembly. The supply chain includes more than 200 upstream and downstream enterprises. This method is used to build a financial risk correlation mining model to identify potential risks of the group and its supply chain. The system deployment and operation details are as follows.

[0034] I. Detailed Implementation of Methods and Steps Data Collection: Data is collected via standardized API interfaces connecting to the group's ERP system, bank transaction system, supply chain management platform, industry opinion platform, and local regulatory platform. The group's ERP system provides annual financial statements and quarterly operating reports; the bank transaction system provides bank statements for the past 12 months; the supply chain management platform provides purchase orders, payment records, and supplier credit data; the industry opinion platform provides negative public opinion regarding the group and its suppliers; and the local regulatory platform provides regulatory penalty information. The interfaces use HTTPS protocol and JSON format, automatically triggering data collection at 2 AM daily, with integrity checks performed every 100 data entries to ensure no data loss.

[0035] Data preprocessing: In structured data processing, the accounts receivable turnover rate of subsidiary A was missing, specifically a numerical one. This was filled using the average of 5.2 times / year from other subsidiaries in the same industry within the group. Supplier B's transaction records contained three duplicate entries with identical transaction serial numbers; these were deleted. In unstructured data processing, for a PDF-formatted regulatory penalty document involving supplier C's environmental violations, key information such as "environmental penalty," "fine of 500,000 yuan," and "rectification period of 3 months" was extracted using Jieba word segmentation. The BERT-BiLSTM-CRF model identified the entity supplier C and the environmental penalty event, converting it into a structured data table containing entity ID, attribute name, attribute value, and data source.

[0036] Knowledge Graph Construction: Core entity types are defined as group entity, subsidiary, supplier, financial indicator, and risk event. Subsidiary attributes include registered capital, product line, and credit rating; supplier attributes include company size, years of cooperation, and historical default count. Entity relationships include transaction relationships between subsidiaries and suppliers, and indicator attribution relationships between subsidiaries and financial indicators. Transaction relationship attributes include annual transaction amount of 50 million yuan and transaction frequency of 12 times / year; indicator attribution attributes include current ratio of 1.5 and calculation period of quarters. During entity alignment, supplier XX Machinery Manufacturing Co., Ltd. and XX Machinery Company have the same Unified Social Credit Code. The text attribute similarity calculated using the edit distance algorithm is 0.92, and the numerical attribute similarity calculated using the Euclidean distance algorithm is 0.98. Both similarities exceed the thresholds (text attribute threshold 0.85, numerical attribute threshold 0.9), thus they are determined to be the same entity and merged.

[0037] Feature Extraction: Three categories of financial risk features were extracted. Basic features included a group debt-to-asset ratio of 65%, a current ratio of 1.2, and a net profit growth rate of -3%. Related features included an overdue payment rate of 8% for upstream supplier D and a cooperation stability score of 0.6. Public opinion features included a frequency of 5 mentions of negative public opinion about the group per month and a correlation with regulatory penalties of 0.7. All features were processed using the min-max standardization method. The standardization calculation formula is as follows: Taking the debt-to-equity ratio as an example, the original value is 65%. In the full dataset, the minimum value for this indicator is 40%, and the maximum value is 80%. Substituting these values ​​into the formula yields... =(65-40) / (80-40)=0.625.

[0038] Feature weight allocation: The weights of each feature are calculated using a multi-dimensional evaluation formula, the formula is as follows: Taking the three characteristics of debt-to-asset ratio, net profit growth rate, and negative public opinion as examples, the debt-to-asset ratio is the first characteristic, and its risk significance is significant. =8. Correlation with financial risk =0.9; Net profit growth rate is the second characteristic, indicating significant risk. =5. Relevance =0.6; Negative public opinion is the third characteristic, indicating significant risk. =4. Relevance =0.5. First calculate the denominator. Then calculate the weights of each feature and the weight of the debt-to-equity ratio separately. =(8×0.9) / 12.2≈0.589, weight of net profit growth rate =(5×0.6) / 12.2≈0.246, negative public opinion weight =(4×0.5) / 12.2≈0.164, ensuring that key financial indicators have a higher weight.

[0039] Model Construction and Training: A hybrid graph neural network architecture combining GCN and GAT is adopted. The model input consists of standardized financial risk features and a knowledge graph entity relationship matrix, with the entity relationship matrix having a dimension of 100×100. The model architecture includes 3 graph convolutional layers and 2 graph attention layers. Each graph convolutional layer outputs 128-dimensional features, and each graph attention layer has 4 attention heads. The dataset is divided using hierarchical cross-validation, with the training set containing 70% historical data from 500 companies and the test set containing 30%. The Adam optimizer is used, with an initial learning rate of 0.001, decaying by 10% every 5 iterations, for a total of 100 iterations. When the attention mechanism is introduced, the transaction amount between subsidiary A and supplier D accounts for 40%, the cooperation duration of 5 years corresponds to a cooperation duration coefficient of 0.9, and the historical risk association once corresponds to a historical risk association coefficient of 0.5. The risk transmission probability = 0.4×0.4+0.3×0.9+0.3×0.5=0.16+0.27+0.15=0.58. This probability is used as the attention weight to strengthen the relationship between the two. At the same time, a dropout layer is set with a dropout probability of 0.2 to prevent the model from overfitting.

[0040] Risk Analysis and Iteration: The model outputs that subsidiary A has a medium risk level. Based on the knowledge graph, the source of the risk is traced back to supplier D's overdue payments, which led to a cash flow shortage for subsidiary A, subsequently causing overdue accounts payable. The risk transmission path is supplier D → subsidiary A → downstream distributor E. Model and knowledge graph iterations are performed on the 1st of each month. Adding 1200 new supply chain transaction data entries reached the update trigger threshold, increasing the number of overdue payments for supplier D in the knowledge graph from 1 to 2. Based on the new data, the model's learning rate was fine-tuned to 0.0008. After the update, the model's prediction accuracy improved by 3% compared to before the update.

[0041] II. Data Representation and Interpretation Table 1 Comparison of Traditional Risk Assessment Methods for Manufacturing Groups with Key Indicators of This Invention Table 1 Comparison of Traditional Risk Assessment Methods for Manufacturing Groups with Key Indicators of This Invention Table 1 shows that traditional risk assessment methods rely on single financial data and lack correlation mining capabilities, resulting in a supply chain risk identification rate of only 45%. Risk tracing requires manual review of transaction records and financial statements, taking 4 hours. Furthermore, the model's generalization ability is only 60% when facing companies with different product lines. Simultaneously, the filtering of false data is incomplete, leading to 8 false risk assessments per month. This invention, through multi-source data integration and knowledge graph correlation mining, improves the supply chain risk identification rate to 92%. Relying on the graph to automatically locate the source and transmission path of risks reduces tracing time to 20 minutes. Attention mechanisms and dropout layer optimizations enable the model to achieve 88% generalization ability across product lines. Four-fold data quality verification filters out 98% of false data, reducing monthly false risk assessments to once. This fully adapts to the risk mining needs of complex supply chains in manufacturing groups, significantly improving the accuracy and efficiency of risk identification.

[0042] Reference Figure 3 This chart clearly presents the weight allocation of core financial risk features, with the debt-to-equity ratio having the highest weight, followed by net cash flow, and the remaining features having tiered weights. It intuitively verifies the rationality of the weight allocation algorithm and provides direction for feature optimization.

[0043] Reference Figure 4 This chart shows the monthly changes in the financial risk correlation between the company and its core suppliers, with a continuous increase from January to May, consistent with the business context. This verifies the effectiveness of the entity correlation strength calculation and provides support for risk warning and strategy adjustment.

[0044] Reference Figure 5 This chart compares the accuracy of risk identification across multiple industries using traditional methods with that of this invention. The accuracy of this invention is nearly 90%, demonstrating outstanding cross-industry generalization ability. It provides data support for fine-tuning industry feature weights and adapts to industry differences.

[0045] Reference Figure 6 This chart shows the risk warning lead time before and after model iteration. After iteration, it increased to 28-32 days, with differences observed among different types of enterprises. This demonstrates the value of the iteration mechanism, allowing sufficient time for risk management.

[0046] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a financial risk association mining model based on knowledge graphs, characterized in that, Includes the following steps: Collect corporate financial and related data, including annual financial statements, quarterly operating reports, bank statements, supply chain transaction data, public market sentiment data, and industry regulatory penalty information, encompassing both structured and unstructured data; connect to data sources through standardized API interfaces; The collected data is cleaned and preprocessed. Missing values ​​in structured data are filled with the mean and mode, respectively, and duplicate data is removed based on key fields. Key information in unstructured data is extracted using natural language processing technology and converted into structured data tables through word segmentation, part-of-speech tagging, and entity recognition. Construct a knowledge graph of corporate financial risk, with core entity types including corporate entities, financial indicators, counterparties, regulatory agencies, and risk events. It is stored in a graph database and supports adding, deleting, modifying, querying, and batch importing and exporting of entities and relationships. Three categories of financial risk characteristics are extracted: basic characteristics, correlation characteristics, and public opinion characteristics. All characteristics are then standardized. Based on knowledge graphs, a financial risk association mining model is constructed using graph neural network algorithms. The model is input with standardized features and entity relationship matrices, and outputs three levels of financial risk: low, medium, and high, after passing through an input layer, graph convolutional layer, attention layer, and fully connected layer. Perform correlation analysis on the model output results, combine knowledge graphs to trace the source of risks, explore the risk transmission path, and generate a risk correlation analysis report; The model and knowledge graph are updated regularly. Updates are triggered when more than 1,000 new data entries are added, regulatory policies are adjusted, or new risk events occur. During the update, data collection, preprocessing, and feature extraction steps are performed, model parameters are fine-tuned, and the knowledge graph is updated.

2. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, It also includes a step of weighting financial risk characteristics. The weighting is calculated using a multi-dimensional assessment formula, which is as follows: in For the first The weights of each financial risk characteristic; For the first The risk significance of each feature is calculated by the frequency and contribution of the feature in historical risk events; For the first The correlation between each characteristic and financial risk is calculated using the Pearson correlation coefficient to determine the degree of linear correlation between the characteristic value and the risk level. This represents the total number of financial risk characteristics.

3. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, It also includes performing entity alignment operations during the knowledge graph construction process to eliminate the differences in the representation of the same entity in different data sources. The entity alignment step first extracts the attribute set of entities from each data source. The attribute set of enterprise entities includes the unified social credit code, enterprise name, registered address, and legal representative. The attribute set of financial indicator entities includes the indicator name, calculation method, and statistical period. Then, the attribute similarity between entities is calculated. The similarity of the text attribute enterprise name is calculated using the edit distance algorithm, and the similarity of the numerical attribute registered capital is calculated using the Euclidean distance algorithm. Finally, a similarity threshold is set.

4. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, When collecting corporate financial and related data, a new data quality verification step has been added: Completeness verification compares the data with preset fields set according to the data source type. Corporate financial statements contain core fields. If the missing rate exceeds 5%, the data needs to be reconnected. If the data is still missing, the record is marked. Accuracy verification compares the data with related data through cross-validation. If the deviation exceeds 10%, at least two financial specialists will be asked to conduct an independent review. Timeliness verification sets a 30-day or 7-day interval threshold according to the data type. Compliance verification is based on relevant regulations, and non-compliant data is removed.

5. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, When constructing the financial risk association mining model, an attention mechanism is introduced to optimize the graph neural network layer. By calculating the importance of the association between entities, attention weights are assigned. Three parameters are extracted from adjacent entities: the proportion of transaction amount, the duration of cooperation, and the historical risk association records. The risk transmission probability is calculated as a weight to highlight the relationship between entities with high risk transmission probability. In the later stages of model training, a dropout layer is introduced to randomly discard some neuron connections.

6. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, When performing correlation analysis on the risk results output by the model, a new dual verification step for risk levels is added: The first step is historical data verification, which selects data from no less than 1,000 enterprises with complete risk records in the past 5 years as the verification set, compares the predicted results with the actual risk events, and calculates the accuracy, recall, and F1 score. If any indicator fails to meet the standard, the model parameters are adjusted. The second step is expert verification, in which no less than 3 financial risk assessment experts with more than 5 years of experience independently review the list of high-risk enterprises, risk transmission paths, and correlation analysis reports, and calculate the deviation rate. If the deviation exceeds 10%, the feature weights are reassigned and the model is fine-tuned. Only the results that pass the verification can be used for business decisions.

7. The method for constructing a financial risk association mining model based on knowledge graphs according to claim 1, characterized in that, When periodically iterating and updating the model and knowledge graph, a new update effect evaluation step is added. The evaluation indicators include the entity coverage and relation completeness of the knowledge graph, and the prediction accuracy and recall of the model. The entity coverage and relation completeness are calculated as the proportion of the number of newly added entities and relations after the update to the total number before the update, respectively. The model indicators are calculated using the test set containing newly added enterprise data after the update. If the entity coverage is lower than 15% or the relation completeness is lower than 20%, the data collection scope is optimized. If the model prediction accuracy drops by more than 5% or the recall drops by more than 8%, the model is rolled back and a new iteration plan is formulated.

8. A system for constructing a financial risk association mining model based on knowledge graphs as described in any one of claims 1-7, characterized in that, Includes the following modules: The data acquisition module is configured with a standardized API interface library containing multiple data sources, and authentication is required for calls; it has a built-in task scheduler that supports preset periods and manual triggering modes, and the collected data is temporarily stored in a Redis temporary cache, which is automatically cleaned up after 24 hours of inactivity; The data preprocessing module connects to the acquisition module and includes built-in data cleaning and unstructured data conversion tools. The cleaning tools include units for handling missing values, duplicate data, and outlier data, while the conversion tools integrate natural language processing algorithms to convert unstructured data into structured data tables. The processed data undergoes format validation and is then transmitted to the knowledge graph construction and feature extraction module. The knowledge graph construction module connects to the preprocessing module and includes entity definition, relation definition, entity alignment, and graph database storage units. Five types of core entities and three types of basic relationships are preset; entity alignment units are matched through multiple sub-units and stored using a distributed graph database. The feature extraction module, connected to the preprocessing module, includes feature classification and extraction, standardization, and weight allocation units. It extracts features according to three categories, maps feature values ​​using a standardization method, calculates feature weights according to a specified formula and supports manual fine-tuning, and transmits the results to the model training module. The model training module connects the feature extraction and knowledge graph modules and contains multi-functional units. It adopts a hybrid graph neural network architecture, divides the dataset through hierarchical cross-validation, optimizes parameters with the Adam optimizer, and integrates attention mechanisms and overfitting control methods. The risk assessment module connects the model training and knowledge graph modules and contains multiple core units; it supports batch and single-enterprise risk assessment, combines knowledge graphs to trace the source of risks, explore transmission paths, automatically generate risk analysis reports in multiple formats, and perform dual verification. The model iteration module connects multiple modules and includes units for iteration triggering, update execution, effect evaluation, optimization, and rollback. It supports periodic and condition-triggered updates, executes according to a fixed process, evaluates key indicators of the knowledge graph and model, and optimizes the data source or rolls back the model and formulates new iteration plans based on the evaluation results. The data storage module adopts a distributed architecture and is divided into five functional storage areas, each storing the corresponding data. All storage areas are encrypted.

9. The knowledge graph-based financial risk association mining model construction system according to claim 8, characterized in that, It also includes a risk association strength calculation module, which is connected to the knowledge graph construction module and the risk assessment module; the module contains a parameter acquisition unit, an intensity calculation unit, and a result application unit; the parameter acquisition unit extracts transaction frequency, cooperation association duration, and historical risk transmission probability between entities from the knowledge graph construction module. For entities and Total number of transactions in the past 12 months For entities and The number of years since the partnership was established. For entities in the past 3 years Entities due to risk events The number of times financial irregularities occurred accounted for the majority of entities The proportion of total risk events; the intensity calculation unit uses the formula Calculate the association strength, where For entities With entity The strength of the financial risk association; This is the transaction frequency weighting coefficient; This refers to the correlation duration weighting coefficient; This is the weighting coefficient for the risk transmission coefficient; , , All passed Standardized mapping.

10. The system for constructing a financial risk association mining model based on knowledge graphs according to claim 8, characterized in that, It also includes a user interaction module, which connects to the data storage module and the risk assessment module. This module consists of three main units: interface display, access control, and operation log. The interface display unit is divided into a knowledge graph visualization area, a risk report area, and a task management area: the knowledge graph visualization area uses a guide map to display entities and relationships; the risk report area supports multi-condition retrieval. The task management area supports initiating data collection, model training, and risk assessment tasks, and displays task progress; the permission control unit divides users into three roles and assigns permissions according to the model, with financial analysts, regulators, and system administrators having corresponding operation permissions.

Images