Supply chain transaction credit data risk management method and device

Abstract

The invention provides a supply chain transaction credit data risk management method and device, and belongs to the technical field of supply chain management. The method provided by the invention comprises the following steps: acquiring credible data; verifying the authenticity of the credible data, performing cleaning, standardization and association processing on the verified data, and constructing a multi-dimensional feature variable; based on the multi-dimensional characteristic variables, through relation graph analysis, machine learning model operation and rule engine judgment, fraud identification, credit scoring, credit line calculation and differentiated interest rate pricing are carried out, and a comprehensive risk assessment result is output; taking the comprehensive risk assessment result as a reference, monitoring key links of supply chain transactions in real time, comparing latest transaction data with the reference, triggering corresponding early warning and automatic disposal actions according to an abnormal level, and recording a disposal process and a result to form feedback data; and comparing the evaluation feedback data with a historical prediction result, and performing iterative optimization on the machine learning model and the multi-dimensional feature variable construction process by using the feedback data.

Description

Technical Field

[0001] This application relates to the field of supply chain management technology, and in particular to a method and apparatus for risk management of supply chain transaction credit data. Background Technology

[0002] Against the backdrop of accelerated industrial digitalization and the booming development of fintech, the digital transformation of supply chain transactions has generated massive amounts of usable data, providing a foundation for supply chain finance innovation. Conducting risk management of supply chain transaction credit data has become crucial to ensuring the healthy development of the supply chain finance market.

[0003] Currently, the mainstream supply chain finance risk control solutions in the industry mainly fall into four categories: First, offline credit granting schemes based on the core enterprise's "1+N" model, which rely on the core enterprise's credit to provide financing for its upstream and downstream SMEs and depend on manual review of paper documents; second, traditional credit granting schemes based on real estate mortgages or warehouse receipt pledges, using static assets as financing collateral; third, early online credit systems, which rely only on limited static data such as financial statements and central bank credit reports, and use simple models for credit scoring; and fourth, isolated "data silo" systems, which only access a portion of single data sources. However, these existing solutions have the following shortcomings: the "1+N" offline model depends on the cooperation of the core enterprise, has high information verification costs, and cannot cover multiple levels of suppliers, lacking process management in risk control; the real estate or warehouse receipt pledge model has a narrow scope of application, with high risks and poor asset liquidity in movable asset pledges; early online systems have narrow data dimensions and poor timeliness, cannot verify the authenticity of transactions, have weak model predictive capabilities, and still use the "entity credit" logic, which is not friendly to SMEs; and "data silo" systems suffer from incomplete risk views and low efficiency in cross-institutional collaboration. These shortcomings result in high operating costs, difficulty in risk management, and narrow service coverage for supply chain finance businesses, making it impossible to meet the financing needs of SMEs and the risk management requirements of financial institutions.

[0004] Therefore, there is an urgent need for a method to build a highly automated and intelligent risk management system based on real transaction data, so as to reduce operating and risk costs, improve the efficiency of credit approval and risk management, identify and control risks in real time, and expand the scope of inclusive financial services. Summary of the Invention

[0005] In view of this, this application provides a supply chain transaction credit data risk management method and apparatus to build a highly automated and intelligent risk management system based on real transaction data, thereby reducing operating and risk costs, improving the efficiency of credit approval and risk management, identifying and controlling risks in real time, and expanding the scope of inclusive financial services.

[0006] Specifically, this application is implemented through the following technical solution: The first aspect of this application provides a method for managing supply chain transaction credit data risks, the method comprising: Collect multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs; The authenticity of the trusted data is verified based on the blockchain transaction ID. The verified data is then cleaned, standardized, and correlated to construct multi-dimensional feature variables. Based on the constructed multi-dimensional feature variables, through relationship graph analysis, machine learning model operation and rule engine judgment, fraud identification, credit scoring, credit limit calculation and differentiated interest rate pricing are performed, and a comprehensive risk assessment result is output. Based on the comprehensive risk assessment results, key links in the entire supply chain transaction process are monitored in real time. The latest transaction data monitored is compared with the benchmark. According to the level of abnormality, corresponding early warning and automated handling actions are triggered, and the handling process and results are recorded to form feedback data. The feedback data is compared and evaluated with historical prediction results, and the feedback data is used to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables.

[0007] A second aspect of this application provides a supply chain transaction credit data risk management device, the device comprising a data acquisition module, a data construction module, a processing module, and an optimization module; The acquisition module is used to collect multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs. The construction module is used to verify the authenticity of the trusted data based on the blockchain transaction ID, clean, standardize and correlate the verified data, and construct multi-dimensional feature variables. The processing module is used to perform fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing based on the constructed multi-dimensional feature variables through relationship graph analysis, machine learning model operation, and rule engine judgment, and output a comprehensive risk assessment result. The processing module is also used to monitor key links in the entire supply chain transaction process in real time based on the comprehensive risk assessment results, compare the latest transaction data monitored with the benchmark, trigger corresponding early warnings and automated handling actions according to the anomaly level, and record the handling process and results to form feedback data. The optimization module is used to compare and evaluate the feedback data with historical prediction results, and to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables using the feedback data.

[0008] The supply chain transaction credit data risk management method and apparatus provided in this application collect multi-source heterogeneous data related to supply chain transactions and hash key transaction data onto the blockchain. Utilizing the immutable and traceable characteristics of blockchain, it generates trusted data with transaction IDs, fundamentally solving the problems of difficulty in verifying data authenticity, frequent fraudulent trade, and repeated pledging in traditional models. This provides a trusted data foundation for "transaction credit" to replace "entity credit," breaking the predicament of SMEs struggling to obtain financing due to a lack of collateral and high-quality financial statements. After verifying data authenticity based on blockchain transaction IDs, multi-dimensional feature variables are constructed through cleaning, standardization, and correlation processing. This integrates static enterprise attributes, dynamic transaction behaviors, and supply chain network relationships, achieving a shift from "single-point risk control" to "chain / network risk control." It overcomes the shortcomings of incomplete risk views and low collaborative efficiency caused by "data silos," making risk assessment more comprehensive and forward-looking. Through relationship graph analysis, machine learning model calculations, and rule engine judgment, it conducts fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing, replacing a large amount of traditional manual offline review work. This system shortens credit decision-making time from "weeks / days" to "hours / minutes," significantly improving approval and risk management efficiency. Simultaneously, through dynamic credit limits and differentiated pricing models, it achieves "one policy per account, one price per transaction," reducing operating costs for financial institutions while meeting the high-frequency, rapid turnover needs of the supply chain. Based on comprehensive risk assessment results, it monitors key aspects of the entire transaction process in real time, comparing the latest data with the benchmark and triggering early warnings and automated handling according to anomaly levels. This represents a shift from "static assessment" to "dynamic risk control," solving the problem of the lag in traditional risk control's post-event risk exposure. It upgrades risk identification from "post-event remediation" to "in-event intervention" and even "pre-event prediction," effectively reducing bad debt losses. By comparing feedback data with historical prediction results, it iteratively optimizes the machine learning model and feature construction process, forming a self-improving intelligent closed loop. This allows the system to continuously adapt to new risk patterns, constantly improving the accuracy of risk identification and the scientific nature of credit decisions, ensuring the stability and effectiveness of long-term services. Ultimately, it achieves the core objectives of significantly reducing operating and risk costs, expanding the service scope for SMEs, and strengthening supply chain resilience. Attached Figure Description

[0009] Figure 1 A flowchart of the supply chain transaction credit data risk management method provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of the supply chain transaction credit data risk management device provided in Embodiment 2 of this application. Detailed Implementation

[0010] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0011] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0012] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0013] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0014] Figure 1 This is a flowchart illustrating the supply chain transaction credit data risk management method provided in Embodiment 1 of this application. Please refer to... Figure 1 The method provided in this embodiment may include: S101. Collect multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs.

[0015] Specifically, multi-source heterogeneous data refers to a comprehensive collection of data obtained from multiple different sources throughout the entire supply chain transaction process, with varying data formats and types. Multi-source heterogeneous data specifically includes: core enterprise ERP system data (including purchase orders, sales orders, inbound slips, and acceptance slips), upstream and downstream enterprise ERP / B2B platform data (such as received orders and delivery notes), logistics platform (TMS) and IoT device data (including waybill numbers, real-time GPS location, transportation trajectory, warehouse temperature and humidity, etc.), electronic contract platform data (such as signed electronic contracts and their corresponding hash values), invoice and tax platform data (including the authenticity, authentication status, and cancellation status of VAT invoices), bank / payment system data (such as historical fund flows and accounts receivable collection records), and publicly available data sources (including business registration information, judicial judgments, industry sentiment, and macroeconomic indices). Multi-source heterogeneous data includes structured data (such as order amounts and transaction times stored in database tables), semi-structured data (such as logistics tracking information in JSON / XML format), and unstructured data (such as text data such as scanned copies of contracts and invoices).

[0016] Furthermore, key transaction data is core information extracted from multi-source heterogeneous data that has core value for verifying the authenticity of transactions and assessing credit risk. Specifically, key transaction data includes transaction identifiers (order ID, contract number, invoice number, waybill number, etc., unique identifiers throughout the entire transaction process), core transaction information (transaction amount, information of both parties (supplier / purchaser name, identity), transaction subject matter (goods type, quantity, specifications)), performance key information (delivery time, logistics tracking nodes, acceptance results, payment time, payment amount, overdue days, etc.), and rights-related information (accounts receivable amount, payment period, ownership of debt, and other rights information directly related to credit). The blockchain transaction ID (TxID) is a unique identifier generated by the blockchain system when the hash value of key transaction data is written to the blockchain; it serves as a "trust anchor" for data authenticity. Trustworthy data is transaction-related data that has been verified and guaranteed for authenticity and integrity through blockchain storage; it is a combination of "original data + corresponding blockchain transaction ID."

[0017] In practical implementation, the data sources for the entire supply chain transaction process are determined, including the core enterprise ERP system, upstream and downstream enterprise ERP / B2B platforms, logistics platforms (TMS) and IoT devices, electronic signature platforms, invoice and tax platforms, banking / payment systems, and public data sources. Structured, semi-structured, and unstructured raw data are collected from various data sources through API interfaces, push notifications, or crawling. A parser matching the data source type is called to read and parse the raw data, extracting raw fields and corresponding values. From the parsed raw data, core value data such as order ID, transaction amount, contract number, invoice number, waybill number, transaction party information, key performance information, and accounts receivable information are selected to form a key transaction dataset. Pre-set encryption algorithms such as SHA-256 and SM3 are used to calculate a unique hash value (digital fingerprint) for the extracted key transaction data. A blockchain interface (such as AntChain or WeBank Chain) is called to write the generated hash value into the blockchain network. The blockchain system generates a unique blockchain transaction ID (TxID) for this evidence storage operation, forming an immutable proof of existence. The original key transaction data is bound to the corresponding blockchain transaction ID (TxID) to form trusted data with blockchain transaction ID.

[0018] S102. Based on the blockchain transaction ID, verify the authenticity of the trusted data, clean, standardize and correlate the verified data, and construct multi-dimensional feature variables.

[0019] Specifically, multi-dimensional feature variables are a comprehensive set of features constructed through feature engineering, covering a company's static attributes, dynamic transaction behavior, and network relationships. Static features refer to the company's relatively stable and low-frequency basic attribute information, reflecting its inherent qualifications, including registered capital, main business type, years of establishment, industry, and historical average profit margin, serving as the foundation for assessing the company's basic operational strength. Dynamic transaction features are timeliness features calculated based on historical and real-time transaction behavior data, aggregated according to preset time windows (e.g., 7 days, 30 days, 90 days, 12 months), reflecting the company's operational vitality and fulfillment capabilities, including summed transaction amounts, average order amount, transaction frequency, on-time delivery rate, number and days of overdue payments, and month-on-month / year-on-year transaction volume growth rates, dynamically capturing changes in the company's operational status. Relationship network features are association features calculated based on supply chain transaction chain data and supplier network graphs, reflecting the company's position and dependencies within the supply chain network, including transaction concentration (dependence on a single customer / supplier), supply chain network centrality indicators, dependence on core enterprises, and the proportion of transactions with related parties.

[0020] In practice, the validated data is cleaned, standardized, and correlated to construct multi-dimensional feature variables, including: (1) Construct a unified data model and a central mapping rule base. Based on the central mapping rule base, map fields of data from different sources to standard fields and simultaneously perform operations such as format unification, value domain conversion, missing value completion, and outlier handling. The central mapping rule base predefines the correspondence between fields of each data source and standard fields of the system.

[0021] Specifically, the unified data model is a standardized data organization framework defined within the system. It is used to standardize the storage structure, field meanings, and relationship logic of multi-source heterogeneous data, and serves as the basic framework for achieving data fusion. The unified data model clearly defines the data classification methods (such as transaction, enterprise basic, and logistics), field naming conventions, data types (such as string, number, and date), field constraints (such as not null and unique), and the relationships between fields (such as linking order data and invoice data through order number).

[0022] Furthermore, the central mapping rule base is a collection that stores the correspondence between various data source fields and system standard fields, as well as data processing rules. The central mapping rule base includes field mapping rules, format unification rules, value domain conversion rules, and data completion and anomaly handling rules. Among them, field mapping rules clearly define the correspondence between the original fields of each data source and the standard fields, such as mapping the source system's "PO_AMT" to the standard field "order_amount" and "supplier name" to "supplier_name". Format unification rules specify standardized formats for dates (unified as "YYYY-MM-DD HH:MM:SS"), monetary units, quantity units, etc. Value domain conversion rules unify status descriptions with the same meaning but different expressions into preset enumerated values, such as unifying "qualified", "PASS", and "good" into "QUALIFIED". Data completion and anomaly handling rules specify the methods for filling missing values ​​(such as mean filling and logical deduction filling) and the rules for identifying, correcting, or removing outliers (such as threshold detection).

[0023] In practice, the parser matching the data source type is invoked to read and parse the raw data, extracting the raw fields and their corresponding values. Based on the pre-defined field correspondence in the central mapping rule base, each raw field is mapped to a system standard field. Date formats from different sources are unified to pre-defined standard dates, monetary units and quantity units are unified to pre-defined standard units, and status descriptions and result identifiers with the same meaning but different expressions from different data sources are uniformly mapped to pre-defined enumerated values. For fields with missing values ​​after mapping, they are filled based on the statistical mean of similar data or deduced based on business logic, or their status is marked. Outlier identification is performed through statistical distribution analysis, and the identified outliers are corrected, removed, or marked separately.

[0024] Specifically, the process involves determining whether the collected raw data is structured, semi-structured (JSON / XML files), or unstructured, and clarifying the format characteristics of each data source. Based on the data source type, the corresponding parser (e.g., structured data parser, JSON parser, text parser) is invoked to read and parse the raw data, extracting the original field names and their corresponding values ​​to form a set of raw data fields. The central mapping rule base is queried, and based on the preset correspondence between the raw fields of each data source and the system standard fields, the extracted raw fields are mapped one by one to the corresponding system standard fields. Date representations from different sources (e.g., "2024-09-25", "25 / 09 / 2024") are uniformly converted to a preset standard date format (e.g., "2024-09-25"); monetary units (e.g., yuan, ten thousand yuan) and quantity units (e.g., pieces, boxes) are uniformly converted to the system's preset standard units; and status descriptions and result identifiers with the same meaning but different expressions from different data sources (e.g., "qualified", "PASS", "good") are uniformly mapped to preset enumerated values ​​(e.g., "QUALIFIED"). The system iterates through the mapped standard fields, identifies fields with missing values, and processes them according to preset rules: numeric fields are filled using the statistical mean of similar data; logically related fields are filled based on business logic deduction; and fields that cannot be filled are marked as "pending verification." Outliers exceeding reasonable ranges are identified through statistical distribution analysis (such as calculating standard deviation and interquartile range); correctable outliers (such as obvious data entry errors) are corrected; severely deviated outliers without a basis for correction are removed; and outliers requiring continued observation are marked separately.

[0025] (2) Based on risk control business knowledge, candidate fields related to risk assessment and credit decision are initially screened from standard fields. The importance of each candidate field is quantitatively evaluated through feature selection algorithm. Redundant and target fields are eliminated to determine the key field set. The target fields are candidate fields whose importance is lower than the preset value.

[0026] Specifically, candidate fields are a set of fields directly related to risk assessment and credit granting decisions, initially selected from standardized fields based on risk control business knowledge. Candidate fields typically include historical overdue payment counts, order amounts, transaction frequency, supplier establishment years, logistics fulfillment rate, payment timeliness, contract fulfillment rate, company registered capital, and transaction volume growth rate.

[0027] Furthermore, the key field set is the set of core fields that are determined after quantitative evaluation of candidate fields using feature selection algorithms, eliminating redundant fields and target fields whose importance is lower than the preset value, and which have high predictive value for risk assessment and credit granting decisions and are free of redundancy.

[0028] In practice, risk control experts, considering the management needs of credit risk, fraud risk, and operational risk, initially screen directly relevant candidate fields from standard fields. These candidate fields include at least one of the following: historical overdue number of transactions, order amount, transaction frequency, supplier establishment years, logistics fulfillment rate, payment timeliness, and contract fulfillment rate. The variance of each candidate field is calculated, and fields with stable variance values ​​approaching zero are eliminated, retaining fields with data discriminative power. A risk assessment target variable is constructed, and correlation coefficients between each candidate field and the target variable are calculated through correlation analysis. Simultaneously, the degree of collinearity among candidate fields is analyzed, and redundant fields are eliminated based on the correlation coefficients and collinearity. A tree model is used to train the remaining candidate fields, and the tree model outputs a ranking of the feature importance of each candidate field. Finally, the set of key fields is determined by combining the business judgments of risk control experts with the quantitative screening results.

[0029] Specifically, risk control experts, considering the management needs of credit risk, fraud risk, and operational risk, initially screen directly relevant candidate fields from standard fields, including at least one of the following: historical overdue payment count, order amount, transaction frequency, supplier establishment years, logistics fulfillment rate, payment timeliness, and contract fulfillment rate. They calculate the variance of each candidate field, eliminating fields with stable variance values ​​approaching zero, retaining only candidate fields with data discriminatory power. They construct risk assessment target variables such as "whether there is a default" and "whether there is fraudulent behavior," and calculate the correlation coefficient between each candidate field and the target variables through correlation analysis. The process involves analyzing the degree of collinearity among candidate fields; eliminating fields with extremely low correlation to the target variable based on the calculated correlation coefficients; eliminating redundant fields that are highly collinear with other fields based on the collinearity analysis results; selecting a random forest model or a gradient boosting decision tree model as the tree model, inputting the remaining candidate fields into the tree model for training, and ranking the feature importance of each candidate field through the model output, retaining the top-ranked candidate fields; and combining the business judgment of risk control experts with the quantitative screening results of the above variance analysis, correlation analysis, collinearity analysis, and tree model feature importance ranking to determine the final set of key fields.

[0030] (3) Using the purchase order number or contract number of the core enterprise as the core identifier, the full-process data of the direct trading parties are associated to construct a first-level transaction chain. Through the supplier network map, the transaction data of upstream second-level and above suppliers are penetrated and associated to form a complete transaction chain.

[0031] Specifically, the primary transaction chain refers to the basic transaction link formed by using the core enterprise's purchase order number or contract number as the core identifier and linking the core enterprise with the direct trading party (primary supplier) throughout the entire transaction process.

[0032] The main link in the primary transaction chain is "core enterprise - primary supplier", and the associated data includes standardized full-process data such as order details, delivery records, logistics trajectory, acceptance certificates, invoice information, and payment records.

[0033] Furthermore, the complete transaction chain builds upon the primary transaction chain by using a supplier network graph to achieve a penetrating connection of transaction data from upstream second-tier and higher suppliers, ultimately forming a full-link transaction system extending from the core enterprise to the most upstream multi-level suppliers. Starting with the primary transaction chain, the complete transaction chain uses a supplier network graph to progressively connect the transaction data of upstream second-tier suppliers of the primary suppliers, and upstream third-tier suppliers of the second-tier suppliers, integrating them to form a complete link of "core enterprise—primary supplier—secondary supplier—…—Nth-level supplier".

[0034] In specific implementation, the purchase order number or contract number of the core enterprise is selected as the core identifier for constructing the transaction chain, and the corresponding matching fields of the core identifier in each data source are determined. Based on the core identifier, the full-process transaction data between the core enterprise and the direct trading parties is associated. The full-process transaction data includes order details, delivery records, logistics tracks, acceptance certificates, invoice information, and payment records. The upstream suppliers, main raw material source information, enterprise equity relationship data, and public bidding information in the supplier master data are collected to construct a supplier network graph containing multi-level supplier relationships. The upstream second-level suppliers of the first-level suppliers are located through the supplier network graph, the identity identifiers of the second-level suppliers are extracted, and the transaction data between the second-level suppliers and the first-level suppliers is associated. The transaction data between upstream third-level and above suppliers is associated. The transaction data after the first-level transaction chain and the multi-level penetration association are integrated to form a complete transaction chain.

[0035] Specifically, the purchase order number or contract number of the core enterprise is selected as the core identifier for constructing the transaction chain. Matching fields corresponding to this core identifier in various data sources are identified, and the field mapping relationships are clarified. Based on the determined core identifier and matching fields, the cleaned and standardized transaction data of the core enterprise and its direct trading partners (Tier 1 suppliers) are linked, including order details, shipping records, logistics tracks, acceptance certificates, invoice information, and payment records, forming a Tier 1 transaction chain of "core enterprise - Tier 1 supplier". Information on "upper-level suppliers", "main raw material sources", enterprise equity relationships, and public bidding information are collected from the supplier master data. This information is then organized and linked to construct a supplier network graph containing multi-level supplier relationships. The supplier network graph is used to locate the upstream Tier 2 suppliers of Tier 1 suppliers, extract their identity identifiers, and link the cleaned and standardized transaction data between Tier 2 suppliers and Tier 1 suppliers based on these identifiers. Following the Tier 2 supplier association logic, the transaction data between upstream Tier 3 and above suppliers is sequentially located and linked, achieving a layer-by-layer upward penetration and extension of the transaction chain. By integrating primary transaction chain data and all transaction data after multi-level penetration and association, a complete transaction chain is formed, extending from the core enterprise to upstream multi-level suppliers, and outputting a transaction chain data set containing enterprise information at each node and transaction information throughout the entire process.

[0036] (4) Integrate standard field data, complete transaction chain data and external data to form a panoramic dataset. Based on the panoramic dataset, calculate static features, dynamic transaction features and relationship network features in sequence through statistical analysis, time window aggregation and indicator derivation, and construct multi-dimensional feature variables.

[0037] Specifically, based on the frequency of change, relatively stable basic data in the panoramic dataset are selected, and static features are extracted through statistical analysis. These static features include the company's registered capital, main business type, historical average profit margin, years of establishment, and industry. Multiple time window dimensions are preset, and based on the dynamic data in the panoramic dataset, aggregation calculations are performed according to the time windows to form dynamic transaction features. Based on the transaction chain data and supplier network graph data in the panoramic dataset, relationship network features are calculated through indicator derivation. The static features, dynamic transaction features, and relationship network features are standardized and integrated to form a multi-dimensional feature variable set.

[0038] In practice, the system selects relatively stable, low-frequency basic data from the panoramic dataset and extracts information such as registered capital, main business type, historical average profit margin, years of establishment, and industry to form a static feature set through statistical analysis. Multiple time windows (7 days, 30 days, 90 days, 12 months, etc.) are preset to extract dynamic data such as transaction flow and performance records from the panoramic dataset. Aggregation calculations are performed for each time window, including transaction summation, average order amount statistics, transaction frequency counting, on-time delivery rate calculation, payment overdue days statistics, and transaction volume month-on-month / year-on-year growth rate calculation, forming a dynamic transaction feature set. Based on the complete transaction chain data and supplier network graph data in the panoramic dataset, information such as transaction concentration, supply chain network centrality indicators, core enterprise dependence, and transaction proportion with related parties are calculated through indicator derivation, forming a relationship network feature set. The static, dynamic, and relationship network features are standardized to eliminate dimensional differences between different features. Invalid and duplicate features are removed, and the three types of features are integrated to form a multi-dimensional feature variable set that meets the input requirements of the risk assessment model.

[0039] S103. Based on the constructed multi-dimensional feature variables, through relationship graph analysis, machine learning model operation and rule engine judgment, fraud identification, credit scoring, credit limit calculation and differentiated interest rate pricing are performed, and a comprehensive risk assessment result is output.

[0040] Specifically, the comprehensive risk assessment result is a holistic evaluation conclusion formed by integrating core decision-making information such as fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing, based on multi-dimensional feature variables, after relational graph analysis, machine learning model calculation, and rule engine judgment. The comprehensive risk assessment result includes fraud identification results, credit score and risk level, recommended credit limit, and differentiated interest rate quote.

[0041] In specific implementation, based on the relational network features and transaction chain data in the multi-dimensional feature variables, entities are extracted, relationships are defined and extracted, attributes are assigned to entities and relationships, and then stored in a graph database. Abnormal structures are detected through graph computing technology to identify fraud risks. The multi-dimensional feature variables are input into a preset machine learning model, and the model outputs the enterprise's comprehensive credit score and default probability through calculation. Based on the core indicators in the multi-dimensional feature variables, a suggested credit limit is calculated. Taking into account risk premium, market adjustment factors, and customer value discounts, differentiated interest rate pricing is determined. A preset rule set is executed to integrate fraud identification results, credit scores, suggested credit limits, differentiated interest rate quotes, and rule engine judgment results to form a comprehensive risk assessment result. The preset rule set includes direct rejection if the applicant is a dishonest person subject to enforcement, composite scoring rules based on risk features, and graph-based association rules.

[0042] Specifically, based on the relational network features and transaction chain data in multi-dimensional feature variables, entities such as enterprises, individuals, orders, and funds are extracted; relationships such as "belonging to," "supply," "guarantee," and "shareholding" are defined and extracted; entities are assigned attributes such as registered capital and industry, and relationships are assigned attributes such as annual transaction volume and cooperation duration; entities, relationships, and corresponding attributes are stored in a graph database; abnormal structures such as circular trade and densely connected fraud groups are detected through graph computing technology, fraud risks are identified, and fraud identification results are output. Furthermore, multi-dimensional feature variables are input into preset machine learning models such as XGBoost, LightGBM, and LSTM; the model calculations output the enterprise's comprehensive credit score (e.g., 0-100 points) and the corresponding default probability. The process of calculating suggested credit limits and differentiated interest rate pricing will be described in the following embodiments and will not be repeated here. The system executes a set of preset rules, including a simple hard rule that "if the applicant is a dishonest person subject to enforcement, the application will be rejected directly," a composite scoring rule based on risk characteristics, and a graph-based association rule. It integrates fraud identification results, comprehensive credit scores, default probabilities, suggested credit limits, differentiated interest rate quotes, and rule engine judgment results to form a comprehensive risk assessment result that includes risk levels and core supporting evidence.

[0043] For example, in one embodiment, the rules in the preset rule set are a collection of "condition-action" pairs used to implement automated logical judgments. The basic structure is as follows: ``` WHEN <when ​​the condition is met> IF THEN <The Action Performed> ``` The preset rule set includes simple rules, composite scoring rules, and correlation rules. Among them, Simple rules (hard rules) can be represented as: ``` WHEN New transaction request enters The applicant for IF has been included in the 'List of Dishonest Judgment Debtors'. THEN automatically rejected the application and triggered a high-risk warning. ``` Composite rules (scoring rules) can be expressed as: ``` WHEN calculates transaction risk score IF transaction amount > 300% of the average transaction amount over the past three months THEN Risk Score +20 The IP address used for the IF transaction does not match the city where the transaction is registered. THEN Risk Score +15 IF trading hours are between 2:00 AM and 5:00 AM. THEN Risk Score +10 ``` Association rules can be represented as: ``` WHEN analyzes corporate relationship graphs If the company has more than three first-tier related parties that are marked as 'high-risk'. Then the risk score is increased by 30, and it is marked as 'suspected associated risk'. ``` It's important to note that rules are typically configured and modified through a visual interface, without requiring code changes. Rules can be easily added, disabled, or adjusted. Preset rule sets support setting rule priority and execution order.

[0044] Optionally, based on the core indicators among the multi-dimensional feature variables, a suggested credit limit is calculated, including: extracting the core indicators from the multi-dimensional feature variables and calculating a benchmark credit limit by combining them with industry coefficients; the core indicators include the company's average annual transaction volume, industry type, and debt-to-asset ratio; mapping the comprehensive credit score to a risk adjustment coefficient, wherein the risk adjustment coefficient is positively correlated with the credit score and ranges from 0 to 2; introducing an adjustment factor, and determining the suggested credit limit by multiplying the benchmark credit limit, the risk adjustment coefficient, and the adjustment factor; the adjustment factor includes policy guidance, the strategic importance of the customer, and collateral / guarantee status.

[0045] Specifically, the system extracts the company's average annual transaction volume, industry type, and debt-to-asset ratio from multi-dimensional feature variables as core indicators for calculating credit limits. Based on these core indicators and pre-set industry coefficients (different coefficients for different industries), the initial benchmark credit limit is calculated using the formula "Benchmark Credit Limit = Average Annual Transaction Volume × Industry Coefficient × (1 - Debt-to-Asset Ratio Adjustment Ratio)". The system obtains the company's comprehensive credit score output by the machine learning model and converts it into a risk adjustment coefficient according to pre-set mapping rules. This coefficient ranges from 0 to 2, with higher comprehensive credit scores corresponding to larger risk adjustment coefficients. Based on current policy requirements, policy-oriented adjustment factors are identified (e.g., a coefficient of 1.1-1.3 under inclusive finance policies), and adjustments are made according to the customer's cooperation level, etc. The business contribution determines the customer's strategic importance adjustment factor (e.g., a coefficient of 1.2-1.5 for core customers), and the collateral status adjustment factor is determined based on the type and value of the collateral (e.g., a coefficient of 1.0-1.4 for sufficient collateral, and a coefficient of 0.8-1.0 for no collateral). The base credit limit, risk adjustment coefficient, and each adjustment factor are multiplied to obtain the suggested credit limit, using the formula: "Suggested Credit Limit = Base Credit Limit × Risk Adjustment Coefficient × Policy Guidance Adjustment Factor × Customer Strategic Importance Adjustment Factor × Collateral Status Adjustment Factor". Furthermore, it is ensured that the calculated suggested credit limit falls within the system's preset minimum and maximum credit limits. If it exceeds the range, the final suggested credit limit is output after adjusting according to boundary values.

[0046] Optionally, a differentiated interest rate pricing method is determined by comprehensively considering risk premium, market adjustment factor, and customer value discount, including: determining a base interest rate based on funding costs, operating costs, and target profit margin; determining a risk premium based on the risk level corresponding to the comprehensive credit score; setting a market adjustment factor in conjunction with the intensity of market competition; setting a customer value discount based on customer value; calculating the sum of the base interest rate, risk premium, and market adjustment factor; and determining the difference between the sum and the customer value discount as the differentiated interest rate pricing method.

[0047] Specifically, the process involves calculating the funding costs of financial institutions and the operating costs of supply chain finance, combined with a pre-set target profit margin, to derive the base interest rate using the formula: "Base Interest Rate = Funding Cost + Operating Cost + Target Profit Margin". The process also involves obtaining the comprehensive credit score of enterprises from machine learning models and determining their risk level (e.g., A / B / C) according to pre-defined rules; matching the corresponding risk premium based on the risk level, with A at 0.5%, B at 1.0%, and C at 2.0%. The competitive landscape of the current supply chain finance market is analyzed. If market competition is intense, a positive market adjustment factor of 0-0.3% is set (to increase interest rates to balance risk); if market competition is moderate, a negative market adjustment factor of -0.2%-0% is set (to lower interest rates to attract customers). Customer value is assessed based on indicators such as the length of cooperation between the customer and the financial institution, historical transaction volume, and business contribution. A customer value discount of 0.3%-0.5% is set for high-value core customers, and a customer value discount of 0-0.2% is set for ordinary customers. No customer value discount is set for new customers with no prior cooperation history. Furthermore, first calculate the sum of the base rate, risk premium, and market adjustment factor, then subtract the customer value discount from this sum. The formula is "differentiated interest rate pricing = base rate + risk premium ± market adjustment factor - customer value discount", and finally output the differentiated interest rate pricing result.

[0048] S104. Based on the comprehensive risk assessment results, real-time monitoring is conducted on key links in the entire supply chain transaction process. The latest monitored transaction data is compared with the benchmark, and corresponding early warnings and automated handling actions are triggered according to the anomaly level. The handling process and results are recorded to form feedback data.

[0049] Specifically, the monitoring scope of the entire supply chain transaction process is clearly defined, covering key stages such as order generation, shipment, in-transit transportation, receipt, acceptance, invoice issuance and authentication, and payment. Core data indicators to be monitored at each stage are identified. The latest transaction data for each key stage is collected in real time through API interfaces and data pushes, including logistics status updates, acceptance results, invoice status, and payment progress, ensuring data synchronization with business processes. The latest transaction data collected in real time is compared with benchmark data in the comprehensive risk assessment results (such as preset logistics timeliness thresholds, payment deadlines, and acceptance standards) to identify anomalies that deviate from the benchmark (such as logistics delays exceeding 24 hours, unacceptable acceptance, voided invoices, and overdue payments). Anomalies are classified into three levels—low, medium, and high—based on their severity and impact on risk (e.g., minor logistics delays are low-level, and voided invoices are high-level). Low-level anomalies automatically send alerts to the corresponding account manager's workbench and simultaneously record the anomaly information; medium-level anomalies automatically suspend the unreleased credit line for the borrower and send alert SMS and emails to relevant personnel; high-level anomalies automatically trigger the post-loan investigation process, freeze the remaining credit line, and simultaneously notify financial institutions, core enterprises, and other relevant participants. The system records the trigger time, anomaly content, handling actions, executing entity, and handling results (such as anomaly resolution, risk confirmation, and loss amount) in real time, forming structured feedback data.

[0050] S105. The feedback data is compared and evaluated with the historical prediction results, and the feedback data is used to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables.

[0051] Specifically, the generated feedback data (including details of abnormal events, handling actions, final risk outcomes, and loss amounts) is collected and aligned with the corresponding historical prediction results output by the above embodiments (including fraud identification conclusions, credit scores, default probabilities, and risk levels) according to key fields such as transaction ID and enterprise identifier, forming a "prediction-actual" comparison dataset. Based on the comparison dataset, model prediction indicators are calculated, including accuracy, recall, precision, F1 score, and default probability prediction error, to evaluate the prediction accuracy and effectiveness of machine learning models (credit scoring models, fraud identification models, etc.); at the same time, the explanatory power and correlation of multi-dimensional feature variables to actual risk outcomes are verified. According to the evaluation results, invalid, redundant, or low-correlation feature variables are screened and removed; new effective features (such as adding derivative features related to abnormal handling) are discovered and added to the feature variable set; parameters in the feature engineering process (such as time window length and indicator calculation logic) are adjusted to optimize the feature generation process. Using the labeled actual risk results from the comparison dataset as new training labels, and combining them with optimized multi-dimensional feature variables, the original machine learning model is incrementally trained or fully retrained, updating model parameters and feature weights to improve the model's adaptability to new risk patterns. The optimized model and feature variable set are tested using a validation dataset to verify whether their predictive performance is improved compared to before optimization. If the preset standards are not met, the feature or model training parameters are readjusted until the requirements are met. The validated optimized model and feature variable construction process is deployed to the production environment, replacing the original version, for a new round of risk assessment, forming a closed-loop optimization mechanism.

[0052] The method provided in this embodiment collects multi-source heterogeneous data related to supply chain transactions and hashes key transaction data onto the blockchain. Leveraging the immutable and traceable characteristics of blockchain, it generates trusted data with transaction IDs, fundamentally solving the problems of difficulty in verifying data authenticity, frequent fraudulent trade, and repeated pledging in traditional models. This provides a trusted data foundation for "transaction credit" to replace "entity credit," breaking the predicament of SMEs struggling to obtain financing due to a lack of collateral and high-quality financial statements. After verifying data authenticity based on blockchain transaction IDs, multi-dimensional feature variables are constructed through cleaning, standardization, and correlation processing. This integrates enterprise static attributes, dynamic transaction behaviors, and supply chain network relationships, achieving a shift from "single-point risk control" to "chain / network risk control." This overcomes the shortcomings of incomplete risk views and low collaborative efficiency caused by "data silos," making risk assessment more comprehensive and forward-looking. Utilizing relationship graph analysis, machine learning model calculations, and rule engine judgment, it conducts fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing, replacing a large amount of traditional manual offline review work and streamlining credit decision-making. The timeframe has been shortened from "weeks / days" to "hours / minutes," significantly improving the efficiency of approval and risk management. Simultaneously, through dynamic credit limits and differentiated pricing models, a "one-customer-one-policy, one-order-one-price" approach is achieved, reducing operating costs for financial institutions while meeting the high-frequency, rapid turnover needs of the supply chain. Real-time monitoring of key stages throughout the transaction process, based on comprehensive risk assessment results, compares the latest data with the benchmark and triggers alerts and automated handling according to anomaly levels. This represents a shift from "static assessment" to "dynamic risk control," resolving the lag in traditional risk control's post-event risk exposure. Risk identification has been upgraded from "post-event remediation" to "in-event intervention" and even "pre-event prediction," effectively reducing bad debt losses. By comparing feedback data with historical prediction results, the machine learning model and feature construction process are iteratively optimized, forming a self-improving intelligent closed loop. This allows the system to continuously adapt to new risk patterns, constantly improving the accuracy of risk identification and the scientific nature of credit decisions, ensuring the stability and effectiveness of long-term services. Ultimately, this achieves the core objectives of significantly reducing operating and risk costs, expanding the service scope for SMEs, and strengthening supply chain resilience.

[0053] Corresponding to the aforementioned embodiment of a supply chain transaction credit data risk management method, this application also provides an embodiment of a supply chain transaction credit data risk management device.

[0054] Figure 2 This is a schematic diagram of the supply chain transaction credit data risk management device provided in Embodiment 2 of this application. Please refer to... Figure 2 The device provided in this embodiment includes a data acquisition module 210, a data construction module 220, a processing module 230, and an optimization module 240. The acquisition module 210 is used to acquire multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs. The construction module 220 is used to verify the authenticity of the trusted data based on the blockchain transaction ID, clean, standardize and correlate the verified data, and construct multi-dimensional feature variables. The processing module 230 is used to perform fraud identification, credit scoring, credit limit calculation and differentiated interest rate pricing based on the constructed multi-dimensional feature variables through relationship graph analysis, machine learning model operation and rule engine judgment, and output a comprehensive risk assessment result. The processing module 230 is also used to monitor key links in the entire supply chain transaction process in real time based on the comprehensive risk assessment results, compare the latest transaction data monitored with the benchmark, trigger corresponding early warning and automated handling actions according to the anomaly level, and record the handling process and results to form feedback data. The optimization module 240 is used to compare and evaluate the feedback data with historical prediction results, and to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables using the feedback data.

[0055] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0056] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0057] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0058] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for managing supply chain transaction credit data risk, characterized in that, The method includes: Collect multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs; The authenticity of the trusted data is verified based on the blockchain transaction ID. The verified data is then cleaned, standardized, and correlated to construct multi-dimensional feature variables. Based on the constructed multi-dimensional feature variables, through relationship graph analysis, machine learning model operation and rule engine judgment, fraud identification, credit scoring, credit limit calculation and differentiated interest rate pricing are performed, and a comprehensive risk assessment result is output. Based on the comprehensive risk assessment results, key links in the entire supply chain transaction process are monitored in real time. The latest transaction data monitored is compared with the benchmark. According to the level of abnormality, corresponding early warning and automated handling actions are triggered, and the handling process and results are recorded to form feedback data. The feedback data is compared and evaluated with historical prediction results, and the feedback data is used to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables.

2. The method according to claim 1, characterized in that, The validated data is cleaned, standardized, and correlated to construct multi-dimensional feature variables, including: A unified data model and a central mapping rule base are constructed. Based on the central mapping rule base, fields of data from different sources are mapped to standard fields, and operations such as format unification, value range conversion, missing value completion, and outlier handling are performed synchronously. The central mapping rule base predefines the correspondence between fields of each data source and standard fields of the system. Based on risk control business knowledge, candidate fields related to risk assessment and credit granting decisions are initially screened from standard fields. The importance of each candidate field is quantitatively evaluated through feature selection algorithms. Redundant and target fields are eliminated to determine the key field set. The target fields are candidate fields whose importance is lower than a preset value. Using the core enterprise's purchase order number or contract number as the core identifier, the entire process data of the direct trading parties is linked to build a primary transaction chain. Through the supplier network graph, the transaction data of upstream secondary and above suppliers are penetrated and linked to form a complete transaction chain. Standard field data, complete transaction chain data, and external data are integrated to form a panoramic dataset. Based on the panoramic dataset, static features, dynamic transaction features, and relationship network features are calculated sequentially through statistical analysis, time window aggregation, and indicator derivation to construct multi-dimensional feature variables.

3. The method according to claim 2, characterized in that, Based on the central mapping rule base, fields from different data sources are mapped to standard fields, and operations such as format unification, value range conversion, missing value completion, and outlier handling are performed synchronously, including: Call the parser that matches the data source type to read and parse the raw data, and extract the raw fields and their corresponding values; Based on the pre-defined field correspondence in the central mapping rule base, each original field is mapped to a system standard field; Unify the date format from different sources to a preset standard date, unify the monetary unit and quantity unit to a preset standard unit, and unify the status descriptions and result identifiers with the same meaning but different expressions from different data sources to preset enumeration values. For fields with missing values ​​after mapping, fill them based on the average of similar data, or based on business logic deduction, or mark their status. Outlier identification is performed through statistical distribution analysis, and the identified outliers are corrected, removed, or marked separately.

4. The method according to claim 2, characterized in that, Based on risk control business knowledge, candidate fields related to risk assessment and credit granting decisions are initially screened from standard fields. A feature selection algorithm is then used to quantitatively evaluate the importance of each candidate field, eliminating redundant and target fields to determine the key field set, including: Risk control experts, taking into account the management needs of credit risk, fraud risk, and operational risk, initially screened directly relevant candidate fields from the standard fields; the candidate fields include at least one of the following: historical overdue number, order amount, transaction frequency, supplier establishment years, logistics fulfillment rate, payment timeliness, and contract fulfillment rate. Calculate the variance of each candidate field, remove fields with stable variance values ​​that tend to be 0, and retain fields with data discrimination. Construct a target variable for risk assessment, calculate the correlation coefficient between each candidate field and the target variable through correlation analysis, and analyze the degree of collinearity among the candidate fields. Based on the correlation coefficient and degree of collinearity, eliminate redundant fields. A tree model is used to train the remaining candidate fields, and the tree model outputs the feature importance ranking of each candidate field. Based on the business judgments and quantitative screening results of comprehensive risk control experts, the final set of key fields is determined.

5. The method according to claim 2, characterized in that, Using the core enterprise's purchase order number or contract number as the core identifier, and linking the entire process data of direct trading parties, a primary transaction chain is constructed. Through a supplier network graph, transaction data from upstream second-tier and higher suppliers is further linked to form a complete transaction chain, including: Select the purchase order number or contract number of the core enterprise as the core identifier for constructing the transaction chain, and determine the corresponding matching field of the core identifier in each data source. Based on the core identifier, the entire process of transaction data between the core enterprise and the direct trading parties is linked; the entire process of transaction data includes order details, shipping records, logistics trajectory, acceptance certificates, invoice information, and payment records; Collect information on upstream suppliers, main raw material sources, corporate equity relationships, and public bidding information from the supplier master data to construct a supplier network graph that includes multi-level supplier relationships; The supplier network map is used to locate the upstream second-tier suppliers of the first-tier suppliers, extract the identity identifiers of the second-tier suppliers, and associate the transaction data between the second-tier suppliers and the first-tier suppliers; it is also used to associate the transaction data between upstream third-tier and above suppliers. By integrating the primary transaction chain and the transaction data after multi-level penetration and correlation, a complete transaction chain is formed.

6. The method according to claim 2, characterized in that, Based on the aforementioned panoramic dataset, through statistical analysis, time window aggregation, and indicator derivation, static features, dynamic transaction features, and relationship network features are calculated sequentially to construct multi-dimensional feature variables, including: Based on the frequency of change, relatively stable basic data in the panoramic dataset are selected, and static features are extracted through statistical analysis. The static features include the company's registered capital, main business type, historical average profit margin, years of establishment, and industry. Multiple time window dimensions are preset, and based on dynamic data in the panoramic dataset, aggregation calculations are performed according to the time windows to form dynamic transaction features; Based on the transaction chain data and supplier network graph data in the panoramic dataset, the relationship network features are calculated through indicator derivation. The static features, dynamic transaction features, and relationship network features are standardized and integrated to form a multi-dimensional feature variable set.

7. The method according to claim 1, characterized in that, Based on the constructed multi-dimensional feature variables, through relationship graph analysis, machine learning model calculation, and rule engine judgment, fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing are performed, outputting a comprehensive risk assessment result, including: Based on the relational network features and transaction chain data in multi-dimensional feature variables, entities are extracted, relationships are defined and extracted, attributes are assigned to entities and relationships and stored in a graph database, and abnormal structures are detected and fraud risks are identified through graph computing technology. Input multi-dimensional feature variables into a preset machine learning model, and output the enterprise's comprehensive credit score and default probability through model calculation; Based on the core indicators among the multi-dimensional feature variables, the recommended credit limit is calculated. The differentiated interest rate pricing is determined by comprehensively considering risk premium, market adjustment factors, and customer value discounts. The system executes a preset rule set, integrating fraud identification results, credit scores, suggested credit limits, differentiated interest rate quotes, and rule engine judgment results to form a comprehensive risk assessment result. The preset rule set includes direct rejection if the applicant is a dishonest person subject to enforcement, composite scoring rules based on risk characteristics, and association rules based on graphs.

8. The method according to claim 7, characterized in that, Based on the core indicators among the aforementioned multi-dimensional feature variables, a suggested credit limit is calculated, including: Extract core indicators from multi-dimensional feature variables and combine them with industry coefficients to calculate the benchmark amount; the core indicators include the company's average annual transaction volume, industry type, and asset-liability ratio. The comprehensive credit score is mapped to a risk adjustment factor, which is positively correlated with the credit score and has a value range of 0-2. An adjustment factor is introduced, and the product of the benchmark amount, the risk adjustment coefficient, and the adjustment factor is determined as the recommended credit line; the adjustment factor includes policy guidance, the strategic importance of the customer, and collateral and guarantees.

9. The method according to claim 7, characterized in that, Taking into account risk premium, market adjustment factors, and customer value discounts, differentiated interest rate pricing is determined, including: The base interest rate is determined based on the cost of capital, operating costs, and target profit margin. The risk premium is determined based on the risk level corresponding to the comprehensive credit score. Set market adjustment factors based on the intensity of market competition, and set customer value discounts based on customer value; The sum of the base interest rate, risk premium, and market adjustment factor is calculated, and the difference between the sum and the customer value discount is determined as the differentiated interest rate pricing.

10. A supply chain transaction credit data risk management device, characterized in that, The device includes an acquisition module, a construction module, a processing module, and an optimization module; The acquisition module is used to collect multi-source heterogeneous data related to supply chain transactions, extract key transaction data and generate hash values, write the hash values ​​into the blockchain to form an immutable proof of existence, and obtain trusted data with blockchain transaction IDs. The construction module is used to verify the authenticity of the trusted data based on the blockchain transaction ID, clean, standardize and correlate the verified data, and construct multi-dimensional feature variables. The processing module is used to perform fraud identification, credit scoring, credit limit calculation, and differentiated interest rate pricing based on the constructed multi-dimensional feature variables through relationship graph analysis, machine learning model operation, and rule engine judgment, and output a comprehensive risk assessment result. The processing module is also used to monitor key links in the entire supply chain transaction process in real time based on the comprehensive risk assessment results, compare the latest transaction data monitored with the benchmark, trigger corresponding early warnings and automated handling actions according to the anomaly level, and record the handling process and results to form feedback data. The optimization module is used to compare and evaluate the feedback data with historical prediction results, and to iteratively optimize the machine learning model and the construction process of multi-dimensional feature variables using the feedback data.

Images