Data processing and feature scoring method and device based on supply chain heterogeneous network

Abstract

The application discloses a data processing and feature scoring method and device based on a supply chain heterogeneous network, and relates to the technical field of supply chain data management.The application is executed by a physical device with a data acquisition port, a data cache medium, a central computing unit and a data output port, a dynamic updating supply chain heterogeneous transaction network including multiple types of nodes and edges is constructed, the network is divided into a credit turnover frequency sub-network representing transaction frequency between enterprises, a credit turnover amount sub-network representing transaction scale, a financing frequency sub-network representing financing activity and a financing amount sub-network representing financing scale according to business dimensions, multi-dimensional features are extracted, preprocessed and output to a standardized score through two-layer mapping, and the model weight is iteratively optimized through supervised learning.The application overcomes the technical defects of homogeneous networks losing heterogeneous information, and realizes global graph data processing and quantitative output of multi-source heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of supply chain data management technology. More specifically, this invention relates to a data processing and feature scoring method and apparatus based on heterogeneous supply chain networks. Background Technology

[0002] The supply chain is a core industrial organization form that connects production, distribution, and consumption. With the development of digital technology, the entire supply chain process generates massive amounts of business flow data. How to effectively process and quantify this multi-source heterogeneous data is a core technological requirement in the digital transformation of industries.

[0003] Currently, quantitative analysis methods for supply chain participants mainly rely on static structured data such as corporate financial statements and bank statements. However, for non-core small and medium-sized business entities in the supply chain, due to their smaller scale of operation and imperfect digital management systems, there are often technical problems such as limited data dimensions, delayed information acquisition, and difficulty in verifying data authenticity. This leads to existing data processing solutions relying excessively on the credit endorsement of core entities when conducting quantitative analysis, making it difficult to achieve accurate quantitative evaluation of small and medium-sized entities throughout the entire chain, thus limiting the depth of industrial digitalization. At the same time, although some data analysis models based on complex networks have emerged in existing technologies, significant technical bottlenecks still exist.

[0004] First, existing models are mostly limited to the modeling paradigm of homogeneous networks. In homogeneous networks, all nodes and edges are treated as a single type. However, in real-world supply chain digitization scenarios, the system includes multiple entities (nodes) such as core purchasing enterprises, multi-level suppliers, and financial service institutions, as well as various complex interaction relationships (edges) such as credit transfer, financing and lending, geographical proximity, and equity affiliation. Using homogeneous network modeling will lose a lot of key heterogeneous information and will fail to distinguish the differences in the impact of different types of business activities on data correlation and transmission.

[0005] Secondly, there is a lack of in-depth exploration of the topological characteristics of heterogeneous networks. Since the supply chain is essentially a typical heterogeneous information network, there are complex nonlinear interactions between nodes. Existing technologies mostly focus on the attribute analysis of a single entity, making it difficult to capture the intricate node-edge-node path characteristics in heterogeneous networks. As a result, the transmission effect of data association on different entities and different business chains cannot be quantitatively calculated.

[0006] Finally, the multi-dimensional heterogeneous behavioral characteristics are not fully utilized. When digital credit certificates in the supply chain circulate between different types of nodes, their dynamic behavioral data, such as cycle, frequency, approval records, and clearing status, have completely different related connotations in a heterogeneous environment. At present, there is a lack of systematic technical means to transform these high-frequency behavioral data into effective standardized computable features within a unified heterogeneous network framework.

[0007] In summary, how to construct a data analysis method that can deeply integrate heterogeneous transaction networks in the supply chain, multi-dimensional profiling indicators, and dynamic circulation behavior characteristics to provide more accurate, real-time, and interpretable quantitative analysis tools is a technical problem that urgently needs to be solved in the field of supply chain data management. Summary of the Invention

[0009] This invention provides a data processing and feature scoring method based on heterogeneous supply chain networks. By constructing a heterogeneous supply chain transaction network with multiple types of nodes and edges and dividing it into sub-networks for feature extraction and two-layer mapping processing, it overcomes the technical defect of homogeneous networks losing heterogeneous information and realizes global graph data processing and quantitative output of multi-source heterogeneous data.

[0010] This invention provides a data processing and feature scoring device based on a heterogeneous supply chain network. Through a physical hardware architecture consisting of a data acquisition port, a data buffer medium, a central computing unit, and a data output port, it realizes the physical execution of the entire process of electrical reading, partitioned storage, graph computation processing, and scoring signal output of multi-source heterogeneous data streams in the supply chain.

[0011] To achieve these objectives and other advantages according to the present invention, a data processing and feature scoring method based on a heterogeneous supply chain network is provided. The method is executed via a physical device equipped with a data acquisition port, a data buffer medium, a central processing unit, and a data output port. The method includes: Through the data acquisition port, the system reads the digital credit transfer certificate data stream, financing business flow data stream, and enterprise business registration and credit data stream within a preset period from the external supply chain finance business system in real time, and writes the data stream into the data cache medium for persistent storage. The central processing unit constructs a dynamically updated supply chain heterogeneous transaction network based on the data stored in the data cache medium. G = ( V , E ), where the node set V Including core procurement enterprise nodes, Tier 1 supplier nodes, Tier 2 supplier nodes, Z Tier-1 supplier nodes and financial institution nodes, each node is associated with a multi-dimensional attribute vector including industry, tier, and credit limit, edge setE It includes directed credit certificate transfer relationship edges, directed financing business relationship edges, and undirected edges based on geographical location. Each edge is associated with a behavioral attribute vector including transaction amount, frequency, and cycle. Through this multi-type node and multi-type edge network architecture, the heterogeneous interaction characteristics between different entities in the supply chain are preserved, and the node attributes and edge weights are incrementally updated according to the natural month cycle. Based on the relationships within the heterogeneous transaction network of the supply chain, the central computing unit divides it into four sub-networks according to business dimensions: a credit transfer frequency sub-network representing the frequency of inter-enterprise transactions, a credit transfer amount sub-network representing the transaction scale, a financing frequency sub-network representing financing activity, and a financing amount sub-network representing the scale of capital acquisition. Based on the relationships between various types of nodes in the four sub-networks, multi-dimensional credit characteristic indicators are automatically extracted, including at least enterprise attributes and core enterprise affinity indicators, supply chain flow behavior indicators, and network topology characteristic indicators. The central processing unit performs distribution skewness calculation, long-tail smoothing transformation, and dimensionless standardization on all extracted indicators. Based on domain experience, it determines the internal weights of each dimension indicator. Then, through the first-level dimension mapping, the indicators are aggregated into major category dimension scores. Through the second-level comprehensive mapping, the scores are linearly weighted according to a fixed weight ratio of 60%, 20%, and 20%, and finally, the enterprise credit risk score that falls within the range of [0, 100] is output. The central processing unit sends the credit risk score, risk level, and early warning marker to an external risk control display terminal, business approval terminal, or early warning device through the data output port for display and triggering. The central processing unit loads negative samples of defaulting, litigated, and blacklisted companies, and iteratively corrects the weights of each level of indicators through a supervised learning algorithm, outputs the optimal weights, and updates the credit scoring model.

[0012] Preferably, when constructing the heterogeneous transaction network of the supply chain, the central computing unit writes node data into a continuous physical storage block of the data cache medium in a fixed-length structure format, and each node is accompanied by attribute fields such as scale, industry, enterprise nature, province, city / county, registered capital, total credit line, available credit line, whether it is listed, and whether it has issued bonds. The central processing unit also establishes three types of edge relationship records in the data cache medium: Credit transfer relationship edge: pointing from upstream enterprise to downstream enterprise, with attributes including transfer direction, single transaction amount, total amount, cumulative average amount, time span, annualized number of transfers, and total number of transfers. The central processing unit appends a directed edge record to the data cache medium for each transfer record. Financing relationship edge: pointing from financial institutions to financing enterprises, with attributes including financing amount, interest rate, term, and total number of times. The central processing unit writes these information one by one according to the financing records. Geographic association edge: An undirected edge is established based on whether the province, city or district where the enterprise is located is the same. The central processing unit directly reads from the enterprise attribute field and generates association tags, which are then stored in the data cache medium.

[0013] Preferably, the central processing unit calculates the edge weights of each of the four sub-networks based on the edge data in the data cache medium, and all calculation results are still written to different physical partitions of the data cache medium for partitioned storage. Credit transfer subnetwork: Based on the credit transfer relationship edges stored in the data cache medium, the total number of times is counted by grouping by cube root and receiver, and used as the edge weight; Credit transfer amount sub-network: Based on the credit transfer relationship edges stored in the data cache medium, the transfer amounts between the same enterprise pairs are summed and used as edge weights; Financing frequency subnetwork: Based on the financing relationship edges stored in the data cache medium, the total number of financing requests is calculated by grouping financing demanders and financial institutions, and used as the edge weight; Financing Amount Subnetwork: Based on the financing relationship edges stored in the data cache medium, the financing amounts between the same enterprise and financial institutions are summed and used as edge weights; The central processing unit reads new business data monthly, updates the nodes and edges in the data cache medium, and synchronously updates the weights of the four sub-networks and edges.

[0014] Preferably, the central processing unit, based on the four sub-networks and edge weights already stored in the data cache medium, further reads the enterprise basic data and flow relationship data from the medium and extracts the following credit indicators: Enterprise attribute indicators: province, industry, whether it is a core enterprise, whether it is a first-tier supplier, whether it is a second-tier supplier, whether it is a third-tier supplier, enterprise size, whether there are joint buyers, number of joint buyers, whether it has issued bonds, whether it is listed, available credit line, total credit line, and total number of credit institutions; Core enterprise intimacy indicators: average ratio of credit transfer amount to credit opening amount at each level, average hierarchical distance between suppliers and core enterprises, average credit transfer amount of suppliers, and average credit opening amount; Among them, the average ratio of credit transfer amount to opening amount at all levels = (1 / M 1)×Σ(Credit Transfer Amount) l / Amount opened × 100), l = 1 to M 1, M1 represents the total number of credit transfer records; the average distance between suppliers and core enterprises = (1 / M 2)×Σ(flow distance) j ), j = 1 to M 2, M 2 represents the total number of records for hierarchical distance statistics; the average supplier credit turnover amount = (1 / M 3) × Σ (Credit Transfer Amount) p ), p = 1 to M 3, M 3 represents the total number of supplier transaction amounts; the average credit opening amount = (1 / K )×Σ(Credit Opening Amount) q ), q = 1 to K , K Total number of credit records opened; The central processing unit will extract all completed metrics and store them in the metric zone of the data cache medium.

[0015] Preferably, the central processing unit, based on the enterprise attribute indicators and core enterprise affinity indicators already stored in the data cache medium, and in conjunction with the credit transfer records, opening records, and clearing records stored in the medium, further calculates transfer behavior indicators, approval behavior indicators, and clearing behavior indicators: Circulation behavior indicators: total number of times suppliers transfer in, total amount of ... Approval behavior indicators: success rate of the circulation process, number of rejections in the circulation process, percentage of rejections in the circulation process, success rate of the opening process, number of rejections in the opening process, percentage of rejections in the opening process; Clearing behavior indicators: number of successful clearing, clearing success rate, whether there are any overdue clearing records, average overdue clearing days, and average cycle from opening date to clearing; The central processing unit calls the internal hardware arithmetic unit to complete all calculations, and adds the calculated flow, approval, and clearing indicators to the indicator area of ​​the data cache medium.

[0016] Preferably, the central processing unit, based on the circulation, approval, and clearing indicators already stored in the data cache medium, further reads financing data and asset freeze data from the medium and extracts the following financing and constraint indicators: Financing behavior indicators: total amount of supplier financing, average amount of supplier financing per transaction, average interest rate of supplier financing per transaction, average number of days of supplier financing per transaction, number of times supplier financing is conducted, success rate of supplier financing process, number of supplier financing rejections, and percentage of supplier financing rejections. Financing constraint indicators: frozen amount of credit outflow, number of times credit outflow is frozen, and number of times partial credit outflow is frozen; The central processing unit performs filtering and cumulative calculations on the frozen records, and appends the results to the indicator area of ​​the data cache medium.

[0017] Preferably, the central processing unit, based on the circulation, approval, clearing indicators, financing and constraint indicators already stored in the data cache medium, further combines the four sub-networks and edge weights already stored in the data cache medium to perform graph calculations and extract network topology feature indicators: Circulation amount network metrics: in-degree, out-degree, proximity centrality, eigenvector centrality, PageRank index; Financing frequency network metrics: Degree centrality, Betweenness centrality, Closeness centrality, Eigenvector centrality, PageRank index; The central processing unit temporarily stores intermediate calculation results such as adjacency matrix, eigenvector, reciprocal distance, and weight ratio in the high-speed cache area of ​​the data cache medium. After completing the graph calculation, it writes the topological feature indicators into the indicator area of ​​the data cache medium.

[0018] Preferably, the central processing unit performs data preprocessing and two-layer mapping scoring calculation based on enterprise attribute indicators, core enterprise affinity indicators, circulation behavior indicators, approval behavior indicators, clearing behavior indicators, financing behavior indicators, financing constraint indicators, and network topology characteristic indicators already stored in the data cache medium, following these steps: The distribution skewness of each extracted indicator is calculated, with a skewness threshold of 2. For indicators with a skewness greater than 2, the following steps are performed: ln ( x +1) Logarithmic transformation, followed by mean-variance standardization to dimensionlessly process all processed indicators. The calculation formula is as follows: z = ( x - μ ) / σ ,in μ The mean of the indicators. σ The standard deviation of the indicator is used, and the internal weights of the indicator after processing are confirmed based on domain experience. Perform the first-level dimensional mapping: Divide the indicators into three dimensions: credit circulation links, enterprise attributes and financing channels, and supply chain network characteristics. Then, perform a weighted summation and normalization of the indicators within each dimension. The calculation formula is as follows: The company in the dimension k Credit score Z k : , in k = 1,2,3, where 1 is the credit circulation link dimension, 2 is the enterprise attributes and financing channels dimension, and 3 is the supply chain network characteristics dimension; J k To be classified as a dimension k The collection of all indicators; Perform the second-level comprehensive mapping: linearly weight the scores of the three dimensions according to a fixed weight ratio to obtain the final credit risk score. The calculation formula is as follows: Corporate Credit Score Y = 60% × Credit Circulation Link Dimension Score + 20% × Enterprise Attributes and Financing Channel Dimension Score + 20% × Supply Chain Network Characteristics Dimension Score; The central processing unit uses its internal hardware floating-point unit to perform all operations, including multiplication, division, logarithms, and summations, and stores the final credit risk score in the data cache medium.

[0019] Preferably, the central processing unit performs a supervised learning model optimization step based on the final credit risk score stored in the data cache medium: Data on known defaulters, litigants, and blacklisted negative companies is collected and matched with the full sample of companies in the supply chain to identify negative companies in the supply chain network. At the same time, an equal number of normal companies are extracted to form a monitoring sample set with balanced positive and negative labels, which is then stored in the data cache medium. The supervised sample set is trained using a supervised learning algorithm, and the feature importance ranking of each indicator is automatically calculated. Based on the feature importance, the weight of indicators with high feature importance is automatically increased, and the weight of indicators with low feature importance is decreased. Repeat the iterative optimization process until the model converges or reaches the preset maximum number of iterations, and output the optimized set of index weights. The optimized set of weights is written over the model parameter area of ​​the data cache medium, and the credit risk score of all enterprises is recalculated based on the optimized weights. The central computing unit writes the iterative optimization process, model convergence status, and importance ranking of each indicator feature into the running log area of ​​the data cache medium, and pushes it to the external monitoring terminal via the data output port.

[0020] A data processing and feature scoring device based on heterogeneous supply chain networks includes: The data acquisition port is used to electrically read supply chain data streams and write them to storage. Data caching medium for persistent storage of nodes, edges, subnetworks, metrics, weights, and scores; The central processing unit, electrically connected to the data buffer medium, is used to execute the method; The data output port is used to electrically output the score and warning signal to the risk control terminal, approval terminal or warning device.

[0021] The present invention has at least the following beneficial effects: First, this invention constructs a heterogeneous supply chain transaction network that includes multiple types of nodes and multiple types of edges, and divides it into four sub-networks that represent transaction frequency, transaction scale, financing activity, and capital acquisition scale. It automatically extracts multi-dimensional feature indicators and outputs scores through two-layer mapping, thus fully preserving the heterogeneous interaction characteristics between different entities in the supply chain. This overcomes the technical defect of homogeneous networks losing heterogeneous information and realizes global, automated processing and accurate quantitative evaluation of multi-source heterogeneous data in the supply chain.

[0022] Secondly, this invention achieves the physical execution of the entire process of electrical reading, persistent storage, computational processing, and scoring output of supply chain data streams through a physical hardware architecture consisting of a data acquisition port, a data cache medium, a central processing unit, and a data output port. The data acquisition port reads multi-source data streams from external business systems in real time through a standard communication protocol. The data cache medium adopts a multi-partition architecture to achieve data classification and storage. The central processing unit integrates a hardware floating-point arithmetic unit and a graph computing acceleration unit to perform core operations at hardware-level speed. The data output port electrically outputs the scoring results and warning signals to external terminals, thus forming a complete process from data input to signal output.

[0023] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating one technical solution of the present invention; Figure 2 This is a schematic diagram illustrating the construction framework of a heterogeneous supply chain transaction network according to a technical solution of the present invention. Figure 3 This is a schematic diagram of a credit transfer sub-network construction framework according to a technical solution of the present invention; Figure 4 This is a schematic diagram of the credit transfer amount sub-network construction framework of one technical solution of the present invention; Figure 5 This is a schematic diagram of the financing sub-network construction framework of one technical solution of the present invention; Figure 6This is a schematic diagram of the financing amount sub-network construction framework of one technical solution of the present invention; Figure 7 A schematic diagram of a financing amount sub-network constructed based on sub-samples and the full sample, as an example of the present invention; Figure 8 This is a dynamic demonstration diagram of the financing amount sub-network generation process in an example of the present invention. Figure 9 This is a schematic diagram of the output scoring range of an example of the present invention; Figure 10 This is a schematic diagram illustrating the distribution of the importance of various indicator features during the supervised learning iterative optimization process, which is an example of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0026] It should be understood that terms such as "having," "comprising," and "including" as used herein do not exclude the presence or addition of one or more other elements or combinations thereof. It should be noted that the experimental methods described in the following embodiments, unless otherwise specified, are conventional methods and therefore should not be construed as limiting the invention.

[0027] like Figure 1 As shown, this invention provides a data processing and feature scoring method based on a heterogeneous supply chain network. The method is executed through a physical device equipped with a data acquisition port, a data buffer medium, a central processing unit, and a data output port. The method includes: The central processing unit reads the digital credit transfer certificate data stream, financing business flow data stream, and enterprise business registration and credit data stream within a preset period from the external supply chain finance business system in real time through the data acquisition port. The data acquisition port is electrically connected to the external business system, receives structured business data packets through a standard data interface protocol, and writes the data stream into the data cache medium for persistent storage. The central processing unit constructs a dynamically updated supply chain heterogeneous transaction network based on the data stored in the data cache medium. G = ( V , E A supply chain heterogeneous transaction network is a complex network structure that includes various types of nodes and edges. Unlike homogeneous networks where all nodes and edges are of a single type, supply chain heterogeneous transaction networks can preserve the heterogeneous interaction characteristics between different entities in the supply chain. The node set... V Including core procurement enterprise nodes, Tier 1 supplier nodes, Tier 2 supplier nodes, ZTier 1 supplier nodes and financial institution nodes, core procurement enterprise nodes represent the initial demanders and credit issuers in the supply chain, tier 1 supplier nodes represent suppliers that directly connect with core procurement enterprises, and tier 2 supplier nodes represent enterprises that provide raw materials or product support to tier 1 suppliers. ，Z Tier-3 supplier nodes represent suppliers at or above the third tier in the supply chain, while financial institution nodes represent banks or non-bank financial entities providing financing support to supply chain enterprises. Each node is associated with a multi-dimensional attribute vector, specifically, each node is associated with a multi-dimensional attribute vector including industry, tier, and credit limit. This multi-dimensional attribute vector is used to fully describe the node's characteristic information, providing static background information for the node in subsequent calculations. E This includes directed credit transfer relationship edges (direction from upstream enterprises to downstream enterprises, representing the transfer direction of accounts receivable credit in the supply chain), directed financing business relationship edges (direction from financial institutions to financing enterprises, representing the flow of funds), and undirected edges based on geographical location (connecting enterprises registered in the same province, city, or district). Each edge is associated with a behavioral attribute vector including transaction amount, frequency, and cycle, used to record the quantitative characteristics of interaction behavior between nodes. Through this multi-type node and multi-type edge network architecture, the heterogeneous interaction characteristics between different entities in the supply chain are preserved, effectively overcoming the defect of traditional isomorphic network modeling that treats nodes and edges as a single type and loses key heterogeneity information. Incremental updates of node attributes and edge weights are performed according to the natural month cycle. New business data added in the current month is read, and incremental updates are performed on the node set and edge set, including adding nodes, adding edges, updating the attribute vector field values ​​of existing nodes, updating the behavioral attribute vector field values ​​of existing edges, and recalculating edge weights. This captures the dynamic evolution process of the supply chain network, updating only the newly added or changed data at a fixed cycle, rather than rebuilding the entire network every time. Based on the relationships within the heterogeneous transaction network of the supply chain, the central computing unit divides it into four sub-networks according to business dimensions: a credit transfer frequency sub-network representing the frequency of inter-enterprise transactions, a credit transfer amount sub-network representing the transaction scale, a financing frequency sub-network representing financing activity, and a financing amount sub-network representing the scale of capital acquisition. In the credit transfer frequency sub-network, nodes represent enterprises participating in credit transfer activities, and directed edges point from the credit initiator to the credit recipient. The weight of each edge is the total number of credit transfers between the two enterprises within the statistical period. This sub-network is used to represent the frequency of inter-enterprise transactions. In the credit transfer amount subnetwork, the definitions of nodes and directed edges are the same as before, but the weight of the edge is the total amount of credit transfer between the two enterprises within the statistical period. This subnetwork is used to represent the transaction scale. In the financing frequency subnetwork, the nodes include financing demanders and financial institutions. The directed edges point from financing demanders to financial institutions. The weight of the edge is the total number of financing transactions between the two within the statistical period. This subnetwork is used to represent the financing activity. In the financing amount subnetwork, the definitions of nodes and directed edges are the same as before, but the weight of the edge is the sum of the financing amounts between the two within the statistical period. This subnetwork is used to represent the scale of fund acquisition. Based on the relationships between various types of nodes in the four sub-networks, the central computing unit automatically extracts multi-dimensional credit characteristic indicators. The extracted indicators include at least three major categories: The first category is enterprise attributes and the closeness to the core enterprise, including the enterprise's industry, hierarchical position, credit limit, and average hierarchical distance from the core enterprise; the second category is supply chain flow behavior indicators, including the number and amount of voucher transfers in and out, flow days, success rate of flow links, and clearing success rate; the third category is network topology characteristic indicators, including the node's in-degree, out-degree, proximity centrality, eigenvector centrality, and PageRank index in each sub-network, used to characterize the node's position in the heterogeneous network topology. The central processing unit performs data preprocessing on all extracted indicators, including distribution skewness calculation, long-tail smoothing transformation, and dimensionless standardization. Distribution skewness calculation measures the symmetry of the indicator data distribution and can be performed using the quantile method. For indicators with skewness greater than a preset threshold, long-tail smoothing transformation is performed. The mean-variance standardization method is used to perform dimensionless standardization on each indicator to eliminate the magnitude difference between different indicators. The internal weights of each dimension indicator are determined based on domain experience, which may include manual judgment by supply chain finance experts, statistical analysis of indicators and default events in historical data, etc. Then, the indicators are aggregated into major dimension scores through the first-level dimension mapping. The indicators are weighted and summed according to their respective dimensions and normalized to generate three major dimension scores: credit circulation link dimension score, enterprise attribute and financing channel dimension score, and supply chain network characteristic dimension score. Through the second-level comprehensive mapping, the scores are linearly weighted according to a fixed weight ratio of 60%, 20%, and 20%, with 60% for the credit circulation link dimension score, 20% for the enterprise attribute and financing channel dimension score, and 20% for the supply chain network characteristic dimension score. The final output falls within [0, The credit risk score of enterprises within the interval

[100] can also be modified by applying a nonlinear correction function to the linear weighted result, such as applying a Sigmoid-type squeezing correction to the scores near the two ends of the interval, so that the score distribution is more in line with the actual risk distribution law. The higher the score, the stronger the enterprise's performance capability and the lower the credit risk in the heterogeneous transaction network of the supply chain. The central processing unit sends the credit risk score, risk level, and early warning marker to an external risk control display terminal, business approval terminal, or early warning device through the data output port for display and triggering. The risk control display terminal presents the score and risk level to risk control personnel in a visual interface. The business approval terminal uses the score as a reference for credit approval. The early warning device triggers an audible and visual early warning signal when the score is lower than a preset warning line. The early warning triggering conditions can be configured as the score being lower than a threshold, the score continuously decreasing by more than a preset value, or key indicators deteriorating significantly. The central processing unit loads negative samples of defaulting, litigated, and blacklisted companies to construct a supervised sample set. The weights of each level of indicators are iteratively corrected through a supervised learning algorithm. The negative company labels and normal company labels are trained. The supervised learning algorithm can be an ensemble learning method such as XGBoost or random forest, or a linear model method such as logistic regression or support vector machine, or an optimization method based on neural networks. The iterative convergence condition is that the change in discrimination between two adjacent iterations is less than a preset threshold or the maximum number of iterations is reached. After the iteration is completed, the optimal weights are output and the credit scoring model is updated.

[0028] In the above technical solution, by constructing a heterogeneous transaction network of supply chain with multiple types of nodes and edges, the heterogeneous interaction characteristics between different entities in the supply chain are fully preserved. This effectively overcomes the technical defect of traditional homogeneous network modeling, which treats all nodes and edges as a single type and loses heterogeneous information. By dividing the network into four sub-networks according to business dimensions, decoupled analysis and independent quantification of different business dimensions are achieved. On this basis, enterprise attributes and core enterprise affinity indicators, supply chain flow behavior indicators, and network topology characteristic indicators are automatically extracted. Through a standardized two-layer mapping scoring model and supervised learning iterative optimization, global, automated processing and accurate quantitative evaluation of multi-source heterogeneous data in the supply chain are realized, which significantly improves the accuracy, dynamic adaptability and interpretability of the evaluation results.

[0029] In another technical solution, when constructing the heterogeneous supply chain transaction network, the central computing unit writes node data into a continuous physical storage block of the data cache medium in a fixed-length structure format. A schematic diagram of the construction framework of the heterogeneous supply chain transaction network is shown below. Figure 2 As shown, specifically: First, define multi-type entity node sets. Based on the business logic of the supply chain finance platform, the parties involved in transactions and financing are abstracted as heterogeneous nodes. As mentioned above, the types of heterogeneous nodes include: 1. Core purchasing enterprise nodes, which act as the initial demand party and credit initiator in the supply chain; 2. Multi-level supplier nodes, including first-tier suppliers directly connected to the core purchasing enterprise, second-tier suppliers providing raw materials or product support to higher-level suppliers, and... Z Tier 1 suppliers; 3. Financial institution nodes, banks or other non-bank financial entities that provide financing support to enterprises in the supply chain. At the same time, multi-dimensional attribute vectors are configured for each node, including enterprise size, industry, enterprise nature, province, city / county, registered capital, total credit line, available credit line, whether listed, and whether bonds have been issued. Each node is accompanied by attribute fields of size, industry, enterprise nature, province, city / county, registered capital, total credit line, available credit line, whether listed, and whether bonds have been issued. Second, establish multi-dimensional relation edge mapping. The central processing unit also defines the interaction relationships between different nodes and constructs a set of directed or undirected edges in the data cache medium, establishing three types of edge relationship records: Credit transfer relationship edge: pointing from upstream enterprise to downstream enterprise, representing the supplier's transfer of part or all of its credit for accounts receivable to upstream supplier or partner. For example, a core purchasing enterprise issues a purchase order to a first-tier supplier, and the first-tier supplier delivers goods or services to the core purchasing enterprise according to the order requirements, forming a transfer relationship edge from the first-tier supplier to the core purchasing enterprise. At the same time, when the first-tier supplier purchases raw materials from the second-tier supplier, the transfer relationship edge points from the second-tier supplier to the first-tier supplier. The attributes of the credit transfer relationship edge include the direction of credit transfer (one-way), the amount of accounts receivable transferred, the time span of accounts receivable transfer, the total amount of transfer, the cumulative average transfer amount (average number of transactions per transaction), the annualized number of transfers, and the total number of transfers. The central processing unit appends a directed edge record to the data cache medium for each transfer record. Financing Relationship Edge: This refers to the financial institution pointing to the supply chain enterprise with funding needs, representing the supplier to apply for financing from banks or other financial institutions using its accounts receivable. The attributes of the financing relationship edge include financing direction (one-way), financing amount, financing days, financing interest rate, total number of financings (determined according to data distribution), and annualized number of financings. The central processing unit writes the information one by one according to the financing records. Geographic association edges: Undirected edges are established based on whether the province, city or district where the enterprise is located are the same, representing whether suppliers and suppliers, and suppliers and core purchasing enterprises are located in the same province, city or county. Based on the geographical location information of the enterprise, the central processing unit directly reads from the enterprise attribute field and generates association tags and stores them in the data cache medium.

[0030] The data cache medium fully preserves the credit transfer behavior, financing behavior, and geographical proximity relationships of enterprises in the supply chain.

[0031] In the above technical solution, node data is written into continuous physical storage blocks of the data cache medium in a fixed-length structure format. Each node is configured with complete attribute fields including scale, industry, enterprise nature, province, city / county, registered capital, total credit line, available credit line, whether it is listed, and whether it has issued bonds. This achieves structured and unified storage of the static characteristics of each participant in the supply chain, which greatly improves the read and write efficiency and storage stability of node data. By establishing standardized recording rules for credit flow relationship edges, financing relationship edges, and undirected edges based on geographical location in the data cache medium, the credit flow relationship edges include flow direction, single transaction amount, total amount, cumulative average amount, time span, annualized flow times, and total flow times. The financing relationship edges include financing amount, interest rate, term, and total number of times. This completely preserves the interaction records between various types of enterprises in the supply chain, ensuring that the heterogeneous network can truly reproduce the business interaction relationships of the supply chain.

[0032] In another technical solution, the central processing unit calculates the edge weights of each of the four sub-networks based on the edge data in the data cache medium, and all calculation results are still written to different physical partitions of the data cache medium for partitioned storage. Credit Transaction Subnetwork: In the credit transaction subnetwork, nodes represent enterprises participating in credit transaction activities, which can be either initiators or recipients. The starting point of an edge represents the initiator of the credit transaction, and the ending point of the edge represents the recipient. Based on the credit transaction relationship edges stored in the data cache medium, the total number of credit transactions occurring within a period is statistically analyzed by grouping initiators and recipients, and this sum is used as the edge weight. The credit transaction subnetwork is constructed as follows: Figure 3 As shown, the specific steps are as follows: (1) Collect all credit transfer records within the statistical period, including information on the issuer and the recipient; (2) For each credit transfer record, extract the issuer information. A and the recipient B And add a path from the network. A point to B The directed edges, (3) summarizing A and B The number of credit transfers between edges is used as the edge weight, calculated using the formula: Edge weight (number of credit transfers) = ; Credit Transfer Amount Subnetwork: In the credit transfer amount subnetwork, nodes represent enterprises participating in credit transfer activities, which can be either initiators or recipients. The starting point of an edge represents the initiator of the credit transfer, and the ending point of the edge represents the recipient. Based on the credit transfer relationship edges stored in the data cache medium, the credit transfer amounts occurring between pairs of the same enterprise within a statistical period are summed and used as edge weights. The credit transfer amount subnetwork is constructed as follows: Figure 4 As shown, the specific steps are as follows: (1) Collect all credit transfer records within the statistical period, including information on the issuer and the recipient; (2) For each credit transfer record, extract the issuer information. A and the recipient B And add a path from the network. A point to B The directed edges, (3) summarizing A and B All credit transfer amounts between them are used as the weight of the edge, calculated using the formula: Edge Weight (Credit Transfer Amount) = ; Financing Frequency Subnetwork: In the financing frequency subnetwork, nodes represent the entities participating in financing activities, namely financing demanders and financial institutions. The starting point of an edge represents the financing demander, and the ending point represents the financial institution. Based on the financing relationship edges stored in the data cache medium, the total number of financing transactions occurring within a period is grouped by financing demander and financial institution, and used as the edge weight. The financing frequency subnetwork is constructed as follows: Figure 5 As shown, the specific steps are as follows: (1) Collect all financing records within the statistical period, including information on financing issuers and financial institutions; (2) For each financing record, extract the financing demander. A and financial institutions B And add a path from the network. A point to B The directed edges, (3) summarizing A and B All the number of financing rounds between them are used as the weight of the edge, and the formula is: Edge weight (number of financing rounds) = ; Financing Amount Subnetwork: In the financing amount subnetwork, nodes represent the entities participating in financing activities, namely, financing demanders and financial institutions. The starting point of an edge represents the financing demander, and the ending point represents the financial institution. Based on the financing relationship edges stored in the data cache medium, the financing amounts occurring between the same enterprise and financial institution within a statistical period are summed and used as edge weights. The financing amount subnetwork is constructed as follows: Figure 6 As shown, the specific steps are as follows: (1) Collect all financing records within the statistical period, including information on financing issuers and financial institutions; (2) For each financing record, extract the financing demander. A and financial institutions B And add a path from the network. A point to B The directed edges, (3) summarizing A and B All financing amounts between them are used as the weight of the edge, calculated using the formula: Edge Weight (Financing Amount) = ; The central computing unit reads new business data monthly and sets up a monthly update mechanism to regularly update the heterogeneous transaction network of the supply chain. At the end of each month, it integrates the credit flow, financing and other data accumulated in that month and updates the node attributes (such as first-tier suppliers and second-tier suppliers) and edge attributes (such as financing amount adjustments) in the data cache medium, or adds new nodes and edges (such as new suppliers joining or new financing business being carried out), captures the dynamic changes of the network, and updates the four sub-networks and edge weights in sync. Through this monthly incremental update mechanism, the heterogeneous transaction network of the supply chain and its four derived sub-networks can keep synchronized with the actual business data and accurately reflect the dynamic changes of the supply chain structure.

[0033] In the above technical solution, the heterogeneous transaction network is divided into four subnetworks according to business dimensions: credit transfer frequency subnetwork, credit transfer amount subnetwork, financing frequency subnetwork, and financing amount subnetwork. Each subnetwork is assigned a business meaning representing transaction frequency, transaction size, financing activity, and capital acquisition scale, respectively. This achieves fine-grained decomposition of heterogeneous interaction relationships within the heterogeneous network, completely decoupling the core business dimensions of the original heterogeneous network. This avoids mutual interference between different dimensional features, enabling more accurate extraction of effective information from each business dimension. Each subnetwork calculates edge weights according to its corresponding business logic. The credit transfer frequency subnetwork is grouped by the issuer and the recipient to calculate the total number of transfers. The credit transfer amount subnetwork sums the transfer amounts between the same enterprise and the recipient. The financing frequency subnetwork is grouped by the financing demander and the financial institution to calculate the total number of transfers. The financing amount subnetwork sums the financing amounts between the same enterprise and the financial institution. The edge weight calculation results of the four subnetworks are stored in different physical partitions of the data cache medium, supporting parallel access without interference. A monthly incremental update mechanism ensures that the subnetwork model is synchronized with the actual business status in real time, avoiding feature extraction deviations caused by model lag.

[0034] In another technical solution, the central processing unit, based on the four sub-networks and edge weights already stored in the data cache medium, further reads the enterprise's basic data and flow relationship data from the medium and extracts the following credit indicators: Enterprise attribute indicators aim to construct a scoring benchmark by combining static background information and external credit performance of enterprises. These indicators include: province of origin, industry, whether it is a core enterprise, whether it is a Tier 1 supplier, Tier 2 supplier, Tier 3 supplier, enterprise size, presence of joint buyers, number of joint buyers, whether it has issued bonds, whether it is listed on the stock exchange, available credit line, total credit line, and total number of credit institutions. Specifically: The province is directly read from the province field of the company's registration information; The industry category is directly read from the enterprise's industry classification field, and the enterprise's industry category is determined according to the "National Industrial Classification of Economic Activities" (GB / T4754—2017) standard. Whether a company is a core enterprise is determined by whether it is the initiator of the credit transfer process, i.e., whether it is the initiator of each credit transfer at its inception. If so, it is marked as yes. Whether a company is a Tier 1 supplier indicates its hierarchical position in the supply chain. This is determined by whether the company has a record of receiving credit transfers from core enterprises; if so, it is marked as yes. Whether a company is a Tier 2 supplier indicates its position in the supply chain. This is determined by whether the company has a record of receiving credit transfers from Tier 1 suppliers. If so, it is marked as Tier 2 supplier. Whether a company is a Tier 3 supplier indicates its position in the supply chain. This is determined by whether the company has a record of receiving credit transfers from Tier 2 suppliers. If so, it is marked as yes. Company size is calculated based on the company's total assets. Whether there are joint buyers, whether the core enterprise shares buyers with other enterprises. A joint buyer refers to a group consisting of a joint buyer entity (platform-certified enterprise) and an unspecified number of joint buyer subsidiaries (platform-certified enterprises), which facilitates the joint use of the joint buyer entity's credit limit by various enterprises in credit business. Credit can be split under the joint buyer model. Whether there are joint buyers can be determined by judging whether the core enterprise is associated with a joint buyer entity. Whether a company has issued bonds or not, by matching the full name of the supplier or core enterprise in the company's bond issuance list, it can be determined whether the company has ever issued bonds. Whether a company is listed or not is determined by matching the full name of the supplier or core enterprise in the list of listed companies. Available credit line is the unused credit line of an enterprise. It is the available credit line granted to the supplier or core enterprise by the supply chain finance platform and financial institutions. This limit is affected by the amount of credit turnover. Total credit line refers to the total credit limit that an enterprise possesses, which is the total credit line granted to the supplier or core enterprise by the supply chain finance platform and financial institutions. Total number of credit granting institutions: The total number of financial institutions that have granted credit lines to the enterprise; This refers to the total number of financial institutions that have granted credit lines to the enterprise. The core enterprise intimacy index aims to measure the degree of closeness between the target enterprise and the core enterprise. The core enterprise intimacy index includes the average ratio of credit transfer amount to credit opening amount at each level, the average hierarchical distance between suppliers and the core enterprise, the average credit transfer amount of suppliers, and the average credit opening amount. The average ratio of credit transfer amount to opening amount at each level is used to measure a company's payment ability; the calculation formula is as follows: C 1 = (1 / M 1)×Σ(Credit Transfer Amount) l / Amount opened × 100), l = 1 to M 1, M 1 represents the total number of credit transfer records and the total amount of credit transfers. l For the first lThe credit transfer amount of each recipient, and the opening amount is the corresponding opening amount; The average distance between suppliers and core enterprises reflects the average order of difference between them and the degree of their connection. The calculation formula is as follows: C 2 = (1 / M 2)×Σ(flow distance) j ), j = 1 to M 2, M 2 represents the total number of records for hierarchical distance statistics, and the flow distance. j For the first j The distance of a credit transfer record, i.e., the hierarchical distance; The average amount of supplier credit turnover reflects the average amount a supplier receives through the supply chain over a certain period. The calculation formula is as follows: C 3 = (1 / M 3) × Σ (Credit Transfer Amount) p ), p = 1 to M 3, M 3 represents the total number of supplier turnover records and the credit turnover amount. p For the first p The amount of credit transfer in the supplier's transaction records; Average credit opening amount, used to reflect the average amount of credit opened by a core enterprise to its suppliers within a certain period, is calculated using the following formula: C 4 = (1 / K )×Σ(Credit Opening Amount) q ), q = 1 to K , K Total number of credit records opened, and amount of credit opened. q For the first q The amount of credit opened in each credit opening record; The central processing unit will extract all completed metrics and store them in the metric zone of the data cache medium.

[0035] The aforementioned technical solution extracts enterprise attribute indicators, including province, industry, whether it is a core enterprise, whether it is a first- to third-tier supplier, enterprise size, presence of joint buyers, number of joint buyers, whether it has issued bonds, whether it is listed, available credit line, total credit line, and total number of credit institutions. It also extracts core enterprise closeness indicators, including the average ratio of credit transfer amount to opening amount at each level, the average hierarchical distance between suppliers and core enterprises, the average credit transfer amount of suppliers, and the average credit opening amount. This allows for a comprehensive portrayal of the enterprise's static background information, external financing performance, and the depth of its connection with core credit without relying on traditional financial statements. It achieves standardized extraction of basic information and core business characteristics of supply chain participants, ensuring horizontal comparability of characteristic data between different entities. Through categorical statistics of core enterprise closeness indicators, it accurately quantifies the depth of connection between the entity and core nodes in the supply chain, supplementing the missing industrial chain connection dimension in traditional evaluation schemes. This effectively solves the information asymmetry problem caused by missing or low-quality financial data for small and medium-sized business entities.

[0036] In another technical solution, the central processing unit, based on the enterprise attribute indicators and core enterprise affinity indicators already stored in the data cache medium, and in conjunction with the credit transfer records, opening records, and clearing records stored in the medium, further calculates transfer behavior indicators, approval behavior indicators, and clearing behavior indicators: The circulation behavior indicators are used to quantify and statistically analyze the entire process of enterprise credit circulation vouchers being transferred in and out. These indicators include: total number of supplier circulation transfers in, total total amount of supplier circulation transfers in, total number of supplier circulation transfers out, total total amount of supplier circulation transfers out, average number of days for circulation to the next level supplier, average number of days for circulation to the previous level supplier, percentage of circulation transfers out to circulation transfers in (radius), percentage of circulation transfers in to circulation transfers out, average number of days for circulation to financing, and total number of circulations to financing. Specifically: Total number of transfers to suppliers: This involves counting the total number of supply chain vouchers transferred to the supplier within a certain period and summing the number of transfers to each supplier within a specified statistical period. Total amount transferred to suppliers: The total amount transferred to the supplier's supply chain within a certain period; The sum of all credit transfer amounts to the target supplier within a specified statistical period. The total number of times a supplier transfers out is calculated by summing the total number of supply chain vouchers issued by the supplier within a certain period and for each supplier within a specified statistical period. Total amount of credit transferred out by the supplier; total amount of supply chain amount transferred out by the supplier within a certain period; sum of all credit transfer amounts transferred out from the target supplier within a specified statistical period. The average number of days it takes for a product to be transferred to the next level of supplier is used to measure the efficiency of the supplier in fulfilling the order. This is calculated by summing all the transfer days and dividing by the number of transfer records. ; The average number of days it takes for the resource to be transferred from the previous level to this supplier is used to measure the efficiency of the supplier in obtaining the required resources. This is calculated by summing all transfer days and dividing by the number of transfer records. ; The percentage of transfers out to transfers in is used to measure a supplier's reach and market influence in the supply chain. It is calculated by dividing the total number of transfers out by the total number of transfers in. ; The percentage of credit received out of total credit transactions is used to measure the relative relationship between the number of credit transactions received and the number of credit transactions transferred by the supplier. It is calculated by dividing the total number of credit transactions received by the supplier by the total number of credit transactions transferred out. ; The average number of days from transaction to financing is used to measure how quickly a supplier converts a transaction in the supply chain into financing. First, the time interval from transaction to financing for each business link is calculated, and then the average of these time intervals is calculated over a specified statistical period. The total number of times the supply chain is transferred to financing is used to measure how frequently suppliers use the supply chain for financing. For each supplier, within the specified statistical period, the number of times the supply chain is transferred to financing is counted in all its flow records. Approval behavior indicators are used to statistically analyze the review results of the two business processes: circulation and opening. These indicators include the success rate of the circulation process, the number of rejections in the circulation process, the percentage of rejections in the circulation process, the success rate of the opening process, the number of rejections in the opening process, and the percentage of rejections in the opening process. Specifically: Success rate in the circulation process; percentage of successful transactions in the supply chain. ; Number of rejections in the circulation process: The number of records rejected during the circulation process in the supply chain. Within a specified statistical period, the number of records with the business process name "circulation" and the credit circulation status of the recipient is "rejected" or "rejected upon review". The percentage of items rejected during the circulation process; the proportion of items rejected during the circulation of the supply chain. ; Success rate of the account opening process, percentage of successfully opened supply chains. ; Number of rejections at the opening stage, number of rejections at the supply chain opening stage, and the number of records where the credit transfer status of the recipient is "rejected" or "rejected upon review" within a specified statistical period. The percentage of applications rejected at the opening stage; the percentage of applications rejected at the supply chain stage. ; The clearing behavior indicators are used to quantify the clearing performance of core enterprises. These indicators include the number of successful clearing transactions, clearing success rate, whether there are any overdue clearing records, average overdue days, and average period from invoice opening date to clearing. Specifically: Successful clearing count: The number of times clearing was successfully completed within a certain period. The clearing status of each enterprise within a specified statistical period is determined. If the clearing status is "Successful Clearing" (clearing status equals 2), the count is incremented by 1. The total number of records with a clearing status of "Successful Clearing" within a certain period is counted. The clearing status is successful (clearing status = 2). i Indicate each supplier, n This refers to the total number of suppliers within the statistical period. Liquidation success rate is used to measure a company's financial management capabilities. , i Indicate each supplier, n This refers to the total number of suppliers within the statistical period. Are there any overdue settlement records? Are there any overdue settlement behaviors? In each established enterprise, if there are records where the settlement completion time is later than the planned settlement date, return "1" to mark it as existing; otherwise, return "0" to mark it as not existing. Average overdue settlement days are used to measure the degree of operational delays for companies in supply chain finance. , i Indicate each supplier, n It refers to the total number of all suppliers within the statistical period. m The number of recipients for a specified statistical period; The average time from opening date to clearing is used to measure the speed of a company's cash flow in the supply chain. , i Indicate each supplier, n This refers to the total number of suppliers within the statistical period. The central processing unit calls the internal hardware arithmetic unit to complete all calculations, and adds the calculated flow, approval, and clearing indicators to the indicator area of ​​the data cache medium.

[0037] In the aforementioned technical solution, by tracking the entire lifecycle behavior of credit transfer vouchers, a standardized calculation system for dynamic behavioral indicators across the entire lifecycle of supply chain transfer, approval, and clearing is established. This transforms high-frequency dynamic behaviors in the credit transfer process into quantifiable standardized characteristics. Transfer behavior indicators include the total number and amount of transfers into the supplier's portfolio, the total number and amount of transfers out of the portfolio, the average number of days to transfer to the next and previous levels, the extent of influence, and the average number of days and total number of transfers to financing. This allows for a comprehensive quantification of a company's transfer activity, market influence, and financing conversion efficiency. Approval behavior indicators include... The indicators include the success rate, number of rejections, and rejection rate in the circulation and issuance stages, reflecting the company's approval status and degree of obstruction in the supply chain from the perspectives of document receipt and document issuance. The clearing behavior indicators include the number of successful clearing, clearing success rate, whether there are overdue clearing records, average overdue clearing days, and average cycle from the issuance date to clearing. These indicators characterize the core company's financial management capabilities from the two dimensions of performance completion and timeliness. The above three types of indicators complement each other and can capture dynamic credit signals that are difficult to reflect by traditional static indicators, effectively improving the timeliness and accuracy of the evaluation results.

[0038] In another technical solution, the central processing unit, based on the circulation, approval, and clearing indicators already stored in the data cache medium, further reads financing data and asset freeze data from the medium and extracts the following financing and constraint indicators: Financing behavior indicators are used to quantify the scale, cost, and efficiency of suppliers using supply chain finance to raise funds. These indicators include total supplier financing amount, average financing amount per supplier financing transaction, average financing interest rate per supplier financing transaction, average financing days per supplier financing transaction, number of supplier financing transactions, success rate of supplier financing process, number of supplier financing rejections, and percentage of supplier financing rejections. Total supplier financing amount: The total amount of funds raised by suppliers using supply chain finance. , i Indicate each supplier, N This refers to the number of times a supplier receives financing during the statistical period; Average financing amount per supplier transaction; average financing amount per supplier company over a certain period. , i Indicate each supplier, N This refers to the number of times a supplier receives financing during the statistical period; Average financing rate per supplier transaction; average financing rate per supplier transaction over a certain period. , i Indicate each supplier, N This refers to the number of times a supplier receives financing during the statistical period; Average financing days per supplier financing transaction: The average number of days a supplier company takes to finance a single transaction within a certain period. The financing days for each transaction are calculated by grouping by month, quarter, year, and target supplier, and the average value is calculated. The historical average represents the cumulative average number of financing days per supplier per transaction up to each half-year. Supplier financing frequency refers to the number of times a supplier engages in financing activities within a certain period. It is calculated by grouping suppliers by month, quarter, year, and target supplier. The historical average represents the total number of financing activities for each supplier accumulated to each half-year. Supplier financing success rate: the percentage of suppliers who successfully obtain financing within a certain period. ; The number of times a supplier's financing application was rejected is the number of times a supplier was rejected in the financing process within a certain period. It is calculated by grouping suppliers by month, quarter, year, and target supplier. The number of "platform rejection", "financial institution rejection", "enterprise rejection", "loan disbursement rejection" and "overdue return" for each supplier's financing application is calculated. The historical average represents the total number of times a supplier was rejected in all financing events in each half-year. The percentage of suppliers rejected in the financing process; the proportion of suppliers whose financing applications were rejected within a certain period. ; Financing constraint indicators are used to identify situations where a company's asset liquidity is restricted due to legal disputes or financial crises. Financing constraint indicators include the amount of credit outflow frozen, the number of times credit outflows are frozen, and the number of times a portion of credit outflows are frozen. The amount frozen during credit outflows; the total amount of funds frozen during a certain period due to credit transfers. , i Indicate each supplier, N This refers to the number of times a supplier receives financing during the statistical period; The number of times credit outflows are frozen is the number of times credit transfers are frozen within a certain period. The frozen credits are filtered from the company's credit transfer records, and the number of frozen transactions is counted. Monthly data represents the cumulative number of frozen transactions in the current month. Credit outflow partial freezing count refers to the number of times a portion of the credit transfer amount is frozen within a certain period. It is calculated by filtering credits in a partially frozen state from the company's credit transfer records and counting the number of frozen transactions. Monthly data represents the cumulative number of partially frozen transactions within the current month.

[0039] The central processing unit performs filtering and cumulative calculations on the frozen records, and appends the results to the indicator area of ​​the data cache medium.

[0040] In the aforementioned technical solution, by extracting financing behavior indicators such as total supplier financing amount, average financing amount per transaction, average financing interest rate per transaction, average financing days per transaction, number of financing transactions, success rate of financing stages, number of financing rejections, and rejection rate, the financing capacity, financing cost, cash flow pressure, and financial market acceptance of supply chain participants are standardized and quantified. This comprehensively reflects the company's capital operation status from four dimensions: financing scale, financing cost, financing cycle, and financing success rate. By extracting financing constraint indicators such as frozen amount, number of times frozen, and number of times partially frozen in credit outflows, the solution can identify situations where the company's asset liquidity is restricted due to legal disputes or financial crises. This allows for early detection of corporate capital pressure, legal dispute risks, and financial crisis signals, providing quantifiable abnormal state characteristics for risk warning. The financing behavior indicators and financing constraint indicators work together; the former reflects the company's normal capital operation status, while the latter captures abnormal signals. Together, they provide multi-dimensional data support for the scoring model to identify potential financial risks, further improving the completeness of the feature system.

[0041] In another technical solution, the central processing unit, based on the circulation, approval, clearing indicators, financing and constraint indicators already stored in the data cache medium, further combines the four sub-networks and edge weights already stored in the data cache medium to perform graph calculations and extract network topology feature indicators: The transfer amount network is a credit transfer amount network built based on the transfer amount between companies. It can be accumulated to different time granularities such as monthly, semi-annual, quarterly and annual. The nodes of this network are all enterprises with transfer behavior. The directed edges point from the open side to the receiver, and the weight of the edge is the transfer amount. The transfer amount network indicators include in-degree, out-degree, proximity centrality, eigenvector centrality and PageRank index. In-degree measures the number of other nodes that are connected to a given node, reflecting the node's activity level as a credit recipient. ,in u Indicates the recipient, v Indicates cube root, A uv For the receiver in the adjacency matrix u Pointing cube root v Element; Out-degree measures the number of shortest paths through a given node. ,in v Indicates cube root, w Indicates the recipient, A vw For cube root v Point to the receiver w Element; Proximity centrality measures the reciprocal of the distances between a node and all other nodes in a network. It reflects how close a node is to the center of the network; a higher proximity centrality indicates that information or credit can reach other nodes more quickly from that node. ,in S This represents the total number of nodes in the network. u Indicates the recipient, v Indicates cube root, d ( v , u ) is a node v To the node u The shortest path distance, V It is the set of all nodes in the network; Eigenvector centrality calculates the importance of nodes using the eigenvectors of the network's adjacency matrix. Its core idea is that nodes connected to important nodes are themselves important, reflecting a node's overall influence in the credit transfer network. ,in u Indicates the recipient, v Indicates cube root, λ The largest eigenvalue of the adjacency matrix; The PageRank index measures the relative importance of a node within the entire network. A node's importance depends not only on the number of links directly pointing to it, but also on the importance of the nodes that these links reside on themselves, reflecting the node's global pivotal position in the credit transfer network. ,in D The damping coefficient is 0.85, and the neighboring ( v ) refers to all pointers to nodes v The set of nodes, S This represents the total number of nodes in the network. u Indicates the recipient, v Indicates cube root; The Financing Frequency Network is a supply chain financing network based on all companies that have engaged in financing activities. The nodes of this network are financing demanders and financial institutions. Directed edges point from financing demanders to financial institutions, and the weight of each edge is the number of financing transactions. The Financing Frequency Network metrics include degree centrality, betweenness centrality, proximity centrality, eigenvector centrality, and PageRank index. Degree centrality measures the total number of connections a node makes in a financing network, reflecting the degree of a node's participation in financing activities. Betweenness centrality measures the frequency with which a company node is on the shortest path between any two other nodes, reflecting the node's bridging role and information transmission capability in the financing chain; Proximity centrality measures the reciprocal of the distance between a node and all other nodes in the network. It is calculated in the same way as the proximity centrality of the circulating amount network. ,in S This represents the total number of nodes in the network. u Indicates the recipient, v Indicates cube root, d ( v , u ) is a node v To the node u The shortest path distance, V It is the set of all nodes in the network; Eigenvector centrality calculates the importance of nodes using the eigenvectors of the adjacency matrix of the financing frequency network. Nodes connected to important nodes are also important themselves. The calculation method is the same as that for the eigenvector centrality of the turnover amount network. ,in u Indicates the recipient, v Indicates cube root, λ The largest eigenvalue of the adjacency matrix; The PageRank index measures the global importance of a node in a financing network. A node's importance depends not only on the number of links directly pointing to it, but also on the importance of the nodes to which those links reside. The calculation method is the same as the PageRank index for the Circulating Funds Network. ,in D The damping coefficient is 0.85, and neighbor(v) refers to all nodes pointing to it. v The set of nodes, S This represents the total number of nodes in the network. u Indicates the recipient, v Indicates cube root; The central processing unit temporarily stores intermediate calculation results such as adjacency matrix, eigenvector, reciprocal distance, and weight ratio in the high-speed cache area of ​​the data cache medium. After completing the graph calculation, it writes the topological feature indicators into the indicator area of ​​the data cache medium.

[0042] In the above technical solution, standardized topological feature extraction rules are established by performing graph computation on the circulation amount network and the financing frequency network. Topological feature indicators such as in-degree, out-degree, proximity centrality, eigenvector centrality, and PageRank index of nodes are extracted. This enables in-depth mining and precise quantification of the pivotal position, transmission influence, and global importance of the subject in the global supply chain network. In-degree and out-degree reflect the connection scale of the node as a credit recipient and credit transmitter, respectively. Proximity centrality measures the ease with which a node can reach all other nodes in the network. Eigenvector centrality comprehensively considers the importance of neighboring nodes. The PageRank index integrates the number of direct links and the quality of link sources. It can identify implicit associations and risk transmission effects that traditional behavioral indicators cannot capture from the global perspective of the industrial chain. This solves the defects of insufficient global association information mining and only focusing on the attributes of a single subject in traditional solutions, and further improves the comprehensiveness of the feature system and the accuracy of the evaluation results.

[0043] In another technical solution, the central processing unit, based on enterprise attribute indicators, core enterprise affinity indicators, circulation behavior indicators, approval behavior indicators, clearing behavior indicators, financing behavior indicators, financing constraint indicators, and network topology characteristic indicators already stored in the data cache medium, performs data preprocessing and two-layer mapping scoring calculation according to the following steps: Because supply chain transaction data exhibits a highly long-tail distribution, skewness calculations are performed on each extracted indicator. Skewness measures the symmetry of the indicator data distribution. The central processing unit calculates the skewness value for each indicator, setting a skewness threshold of 2. For indicators with a skewness greater than 2, it indicates a severe right-skewed long-tail characteristic in the data distribution, meaning that extremely high values ​​from a few companies can significantly affect the stability of the score. ln ( x +1) Logarithmic transformation smooths the impact of extreme values ​​on score stability, compresses differences in high-value intervals, and narrows the gap in low-value intervals. Then, mean-variance standardization is applied to dimensionlessly process all processed indicators. The calculation formula is as follows: z = ( x - μ ) / σ ,in μ The mean of the indicators. σAs the standard deviation of the indicators, after standardization, all indicators are converted into a standard distribution with a mean of zero and a standard deviation of one, eliminating the incomparability caused by differences in magnitude between different indicators. Based on domain experience, the internal weights of the processed indicators are confirmed. For key indicators that have been proven to be highly correlated with credit risk in business practice (such as clearing overdue records, PageRank index, etc.), a higher weight contribution coefficient is set, and for indicators with weak correlation or high information redundancy, a lower weight contribution coefficient is set, thereby ensuring that the dimension score generated by the first layer of mapping can accurately reflect the corporate credit side under that dimension. Perform the first-level dimensional mapping: aggregate the scores from indicators to major categories. Indicators are divided into three dimensions: credit circulation links, enterprise attributes and financing channels, and supply chain network characteristics. The credit circulation links dimension includes circulation behavior indicators, approval behavior indicators, clearing behavior indicators, financing behavior indicators, and financing constraint indicators. The enterprise attributes and financing channels dimension includes enterprise attribute indicators and core enterprise closeness indicators. The supply chain network characteristics dimension includes network topology characteristic indicators. The indicators within each dimension are weighted, summed, and normalized to calculate the credit score for that dimension. The calculation formula is as follows: The company in the dimension k Credit score Z k : , in k = 1,2,3, where 1 is the credit circulation link dimension, 2 is the enterprise attributes and financing channels dimension, and 3 is the supply chain network characteristics dimension; J k To be classified as a dimension k The collection of all indicators; The second layer of comprehensive mapping is performed: linear weighting and correction of dimensional scores to comprehensive credit scores, establishing a comprehensive credit scoring mapping function, and using preset dimensional weight ratios for final synthesis. The scores of the three dimensions are linearly weighted according to fixed weight ratios, with the credit circulation link dimension score assigned a weight of 60%, the enterprise attributes and financing channels dimension score assigned a weight of 20%, and the supply chain network characteristics dimension score assigned a weight of 20%, to obtain the final credit risk score. The calculation formula is as follows: Corporate Credit Score Y = 60% × Credit Circulation Link Dimension Score + 20% × Enterprise Attributes and Financing Channel Dimension Score + 20% × Supply Chain Network Characteristics Dimension Score; After performing linear weighted calculation, the final credit score is mapped to the real number range [0, 100]. A higher score indicates a stronger ability to fulfill obligations and lower credit risk for the enterprise in the heterogeneous transaction network of the supply chain, thereby providing financial institutions with quantitative risk control decision support. The central computing unit uses its internal hardware floating-point computing unit to complete all operations such as multiplication, division, logarithm, and summation, and stores the final credit risk score in the data cache medium.

[0044] In the above technical solution, by clarifying the standardized execution process of feature preprocessing and the two-layer mapping scoring model, logarithmic transformation is performed on indicators with skewness greater than the threshold to smooth the long-tail distribution characteristics of supply chain data, eliminating the interference of extreme values ​​on the scoring results. The mean-variance standardization method eliminates the incomparability between indicators of different magnitudes, ensuring fair integration of multi-source heterogeneous indicators on the same scale. The first-layer dimensional mapping weights and aggregates indicators according to three dimensions: credit circulation links, enterprise attributes and financing channels, and supply chain network characteristics. The second-layer comprehensive mapping integrates the three dimensions for scoring according to a fixed weight ratio, realizing hierarchical aggregation and systematic deep integration of features. This two-layer mapping architecture has a clear scoring logic and strong interpretability. It retains the fine weight structure within each dimension and achieves interpretable aggregation between dimensions through a fixed ratio. The fixed weight ratio ensures the stability and consistency of the scoring results. The output standardized score has horizontal comparability and business practicality, solving the problems of insufficient interpretability and large fluctuations in scoring results of traditional black box models.

[0045] In another technical solution, the central processing unit performs a supervised learning model optimization step based on the final credit risk score stored in the data cache medium: The central processing unit first collects negative samples and constructs a sample set. It collects data on known defaults, lawsuits, and blacklisted companies from external databases or data interfaces, matches them with the full sample of companies in the supply chain, identifies negative companies in the supply chain network, and at the same time extracts an equal number of normal companies from the full sample of companies in the supply chain that have no negative records and are operating normally, forming a supervisory sample set with balanced positive and negative labels, which is then stored in the data cache medium. The supervised sample set is trained using a supervised learning algorithm. The algorithm evaluates the information gain or Gini impurity contribution of each indicator in distinguishing between negative and normal enterprises. It automatically calculates the feature importance ranking of each indicator. The higher the feature importance value, the greater the contribution of the indicator to identifying credit risk. Based on the feature importance, the weight of indicators with high feature importance is automatically increased, and the weight of indicators with low feature importance is decreased. After completing one round of weight adjustment, the central computing unit recalculates the credit risk score of the sample enterprises and evaluates the model's discrimination index. If the discrimination index does not converge and the preset maximum number of iterations is not reached, the feature importance assessment and weight adjustment steps are repeated, and the optimization process is repeated until the model converges or the preset maximum number of iterations is reached, and the optimized optimal index weight set is output. The optimized set of weights is written over the model parameter area of ​​the data cache medium, and the credit risk score of all enterprises is recalculated based on the optimized weights. The central computing unit writes the iterative optimization process, model convergence status, and importance ranking of each indicator feature into the running log area of ​​the data cache medium, and pushes it to the external monitoring terminal via the data output port.

[0046] In the above technical solution, real negative samples such as defaulters, litigants, and blacklisted companies are introduced as monitoring signals to construct a monitoring sample set with balanced positive and negative labels. The supervised learning algorithm automatically calculates the feature importance ranking of each indicator for risk differentiation and adjusts the weight coefficients of each indicator in a directional manner according to the importance. It automatically increases the weight of indicators with high importance and decreases the weight of indicators with low importance, realizing adaptive iterative optimization of model weights. This optimization method effectively corrects the possible deviation between the initial preset weights and the actual risk patterns, solves the problems of insufficient accuracy of traditional manual experience weights and inability to adapt to dynamic changes in business, and continuously improves the model's ability to capture credit deterioration signals and risk differentiation. By fully recording the entire model optimization process, the traceability and transparency of the optimization process are ensured, guaranteeing the long-term effectiveness and reliability of the scoring model.

[0047] A data processing and feature scoring device based on heterogeneous supply chain networks includes: The data acquisition port is the data input interface module of the device. This port is electrically connected to the data server of the external supply chain finance business system through a standard communication bus or network interface. According to the preset data reading protocol, it periodically requests and receives digital credit transfer certificate data streams, financing business transaction data streams, and enterprise business registration and credit data streams. The data acquisition port includes a data buffer circuit and a signal level conversion circuit. After converting the received external data streams into a signal format compatible with the device's internal bus, it sequentially writes the data streams into the specified base address offset area in the data buffer medium for storage according to the writing timing.

[0048] The data caching medium is the device's persistent data storage module. This medium is implemented using non-volatile storage devices and is internally divided into multiple logically independent storage areas, including: a node data area for persistently storing the attribute vectors of all network nodes; an edge data area for persistently storing all edge relationship records and behavioral attribute vectors; a sub-network partition for partitioning and storing the node sets, edge sets, and edge weights of four sub-networks; an indicator area for storing all extracted enterprise attribute indicators, affinity indicators, circulation behavior indicators, approval behavior indicators, clearing behavior indicators, financing behavior indicators, financing constraint indicators, and topological feature indicators; a model parameter area for storing the weights of indicators at each level and dimension weights; a runtime log area for storing the iterative process and convergence status of model optimization; and a high-speed cache area for temporarily storing intermediate results such as adjacency matrices, eigenvectors, and shortest path distances during graph calculation.

[0049] The central computing unit (CCU) is the core computing and control module of the device. Electrically connected to the data cache medium, it reads data from various storage areas via an internal high-speed bus. The CCU integrates a hardware floating-point unit, a hardware arithmetic unit, and a graph computing acceleration unit. The hardware floating-point unit performs floating-point operations such as logarithmic transformation, standard deviation calculation, and weighted average. The hardware arithmetic unit performs basic operations such as summation, counting, and comparison. The graph computing acceleration unit performs graph algorithms such as in-degree, out-degree, shortest path search, and PageRank iteration in parallel on the adjacency matrix. Following the described methods and steps, the CCU sequentially executes the construction of a heterogeneous supply chain transaction network, subnetwork segmentation, multi-dimensional indicator extraction, data preprocessing, two-layer mapping score calculation, and supervised learning optimization, ultimately generating a corporate credit risk score.

[0050] The data output port is the device's data output interface module. This port is electrically connected to the central processing unit, receives the calculated credit risk score, risk level, and early warning marker data, and electrically outputs the above data to an external risk control display terminal, business approval terminal, or early warning device via a standard communication protocol. The risk control display terminal displays the score in a visual chart format, the business approval terminal embeds the score into the credit approval workflow, and the early warning device triggers an audible and visual alarm upon receiving an early warning marker signal.

[0051] In the above technical solution, the physical hardware architecture of data acquisition port, data cache medium, central computing unit and data output port realizes the physical execution of the entire process of electrical reading, persistent storage, computing and processing and scoring output of supply chain data flow. The data acquisition port reads multi-source data streams from external business systems in real time through standard communication protocols. The data cache medium is a non-volatile storage device and adopts a multi-partition storage architecture to realize the classification management and isolated storage of data. The central computing unit integrates a hardware floating-point computing unit and a graph computing acceleration unit to perform core computing at hardware-level speed. The data output port electrically outputs the scoring results and early warning signals to the risk control display terminal or early warning device, which constitutes a complete process from data input to signal output.

[0052] Example 1: like Figure 1 As shown in the illustration, this invention provides an exemplary method for data processing and feature scoring based on heterogeneous supply chain networks. The specific scenario of this invention is a supply chain data processing and feature scoring model based on heterogeneous networks, comprising the following steps: Step 1: Building the dataset This embodiment first collects full business data from September 14, 2017 to August 7, 2024 based on the business logic of a large supply chain finance platform. The dataset includes 122,933 enterprises, of which 87,369 have credit transfer records, including 2,114 core enterprises, 59,604 first-tier suppliers, 24,515 second-tier suppliers and 1,136 third-tier suppliers.

[0053] In terms of data governance, the following cleaning logic is implemented for raw data: 1. Missing value handling: For records where the original credit holder is empty, manual supplementation or cause analysis is performed according to the business flow; 2. Outlier correction: For outlier data where the total credit amount or available amount is negative, it is uniformly processed to zero; For redundant credit records of the same enterprise, they are split or deleted through business scenario analysis; 3. Standardization processing: The terminology of fields such as the definition of supply chain virtual accounts and funder balances is standardized, and the calculation formulas are clarified.

[0054] Step 2: Construct a heterogeneous transaction network model for the supply chain This embodiment constructs a dynamic heterogeneous transaction network. G = ( V , E ), where the node set V This includes core enterprise nodes, multi-level supplier nodes, and banking and financial institution nodes.

[0055] In the edge set EIn the construction of the system, three types of related edges are defined: 1. Credit transfer edge (directed): determined by the direction of credit transfer, with attributes including transfer amount, transfer days, and annualized transfer frequency; 2. Financing relationship edge (directed): pointed from financial institutions to enterprises in need of funds, with attributes including financing amount, financing interest rate, and financing days; 3. Geographical related edge (undirected): established based on the province, city, or district where the enterprise is registered.

[0056] The aforementioned heterogeneous network is further divided into four sub-networks: credit transfer frequency sub-network, credit transfer amount sub-network, financing frequency sub-network, and financing amount sub-network. By summarizing the weights of each edge within the statistical period, a foundation is provided for subsequent topology feature extraction. Using a monthly step size, newly added credit transfer data is integrated monthly to achieve dynamic evolution of the network topology. An example of the network graph is shown below. Figure 7 As shown, the left image is a network built based on subsamples, and the right image is a network built based on the full sample. A dynamic demonstration of the network generation process is shown below. Figure 8 As shown.

[0057] Step 3: Extraction of Multi-Dimensional Enterprise Profile Indicator Features This embodiment automatically extracts multi-dimensional credit feature indicators based on the heterogeneous network constructed in step 2, including at least 61 indicators in total, such as enterprise attributes and core enterprise closeness indicators, supply chain flow behavior indicators, and network topology feature indicators. The indicator extraction logic is as follows: 3.1: Extracting Enterprise Attributes and Core Enterprise Intimacy Indicators: This step aims to construct a scoring benchmark based on the enterprise's static background and external credit performance. Specifically, it extracts: 1. Basic Enterprise Characteristics: This includes the province where the enterprise is located, its industry, whether it is a core enterprise, and its hierarchical position in the supply chain (first- to third-tier suppliers); it also identifies whether there are joint buyers, the number of joint buyers, and the enterprise's size; 2. External Financing Source Characteristics: This involves determining whether the enterprise has issued bonds, whether it is a listed company, and extracting the enterprise's available credit limit, total credit limit, and total number of credit institutions on the platform; 3. Core Enterprise Intimacy Characteristics: This calculates the average proportion of credit transfer amounts and opening amounts at each level between the target enterprise and the core enterprise, as well as the average hierarchical distance between them in the topology network, to measure the enterprise's payment ability and the depth of its binding with core credit.

[0058] 3.2: Extracting Supply Chain Circulation Behavior Indicators: This step characterizes a company's operational capabilities and integrity by tracking the entire lifecycle of circulation vouchers. Specifically, it extracts: 1. Circulation Behavior Characteristics: Statistically, it tracks the number of inflows / outflows and the total amount of circulation vouchers within the statistical period, calculates the average number of days and total number of transactions from circulation to financing, and uses the radiation index to measure the company's market influence within the supply chain; 2. Approval and Clearing Behavior Characteristics: For the circulation and issuance stages, it calculates the success rate, number of rejections, and rejection rate; simultaneously, at the core enterprise level, it extracts whether clearing was successful, the number of successful clearing transactions / success rate, and focuses on quantifying whether there are overdue clearing records, the average overdue clearing days, and the total cycle days from issuance to clearing; 3. Financing Behavior Characteristics: It quantifies the total amount of supplier financing, the average financing amount per transaction, the average financing interest rate, the average financing days, and the success / rejection ratio of the financing stage to identify the company's cash flow pressure and market acceptance of financing.

[0059] 3.3: Extracting Network Topology Feature Indicators: This step uses global graph algorithms and anomaly tracking to identify systemic risks and asset constraints. Specifically, it includes: 1. Network Topology Indicators: Based on the flow of funds network, graph computing algorithms are used to extract the in-degree, out-degree, proximity centrality, eigenvector centrality, and PageRank index of nodes, reflecting the pivotal position and importance of nodes in credit flow; 2. Financing Network Topology Indicators: Based on the financing frequency network, the degree centrality, betweenness centrality, proximity centrality, eigenvector centrality, and PageRank index of nodes are calculated to identify the transmission influence of nodes in the financing chain; 3. Enterprise Financing Constraint Features: Through anomaly tracking, the total amount of frozen funds and the number of frozen transactions in the outflow supply chain are extracted as important early warning signals for identifying corporate legal disputes, financial crises, and limited asset liquidity.

[0060] Step 4: Calculate credit score based on two-layer mapping function 4.1: Feature Preprocessing and Standardization: Due to the highly long-tailed distribution characteristics of supply chain transaction data, skewness calculation is first performed on each indicator extracted from S2, with a preset threshold of 2. For indicators with a skewness greater than 2, a logarithmic transformation operation is performed. ln ( x +1), to smooth the impact of extreme values ​​on score stability, and then use the mean-variance standardization method to dimensionlessly process all processed indicators. The calculation formula is as follows: z = ( x - μ ) / σ ,in μ The mean of the indicators. σ The standard deviation of the indicator; The above formula is used to eliminate the magnitude difference between indicators of different magnitudes, ensuring the fairness of the subsequent scoring process.

[0061] 4.2: Performing the First-Level Mapping: Aggregating Indicator Scores to Major Dimensions: Based on the aforementioned indicator system, the indicators are divided into three core dimensions: credit circulation links, enterprise attributes and financing channels, and supply chain network characteristics. Using internal weights determined based on domain experience, a weighted summation is performed on the indicators within each dimension to calculate the preliminary scores for each major dimension. During the mapping process, higher weights are assigned to key indicators (such as clearing overdue records and PageRank index) to ensure that the dimension scores generated by the first-level mapping accurately reflect the enterprise's creditworthiness within that dimension. The mapping formula is as follows: , in k = 1,2,3, where 1 is the credit circulation link dimension, 2 is the enterprise attributes and financing channels dimension, and 3 is the supply chain network characteristics dimension; J k To be classified as a dimension k The collection of all indicators; 4.3: Performing the Second Layer Mapping: Linear Weighting and Correction of Dimension Scores to Comprehensive Credit Score: Establish a comprehensive credit score mapping function, using preset dimension weight ratios for final synthesis. The specific mapping logic is as follows: assign 60% weight to the credit circulation link dimension score, 20% weight to the enterprise attribute and financing channel dimension score, and 20% weight to the supply chain network characteristic dimension score. The mapping formula is as follows: Corporate credit score Y = 60% × credit circulation link dimension score + 20% × corporate attributes and financing channel dimension score + 20% × supply chain network characteristics dimension score; After performing linear weighted calculation, the final credit score is mapped to the real number range [0, 100]. A higher score indicates a stronger ability to fulfill obligations and lower credit risk within the heterogeneous transaction network of the supply chain, thus providing financial institutions with quantitative risk control decision support. The output score range in this example is as follows: Figure 9 As shown.

[0062] Step 5: Weight optimization and validation based on negative labels This embodiment introduces a monitoring signal consisting of 91 negative suppliers (including blacklisted companies and companies involved in serious litigation). Through matching, it was found that 9 of these companies exist in the current full sample.

[0063] 5.1 Model Validation: As shown in Table 1, the credit scores of these 9 negative companies were all at the bottom of the entire sample in the year they were penalized, and their scores were significantly lower than those of normal companies. For example, the negative company with company ID ca071ad9-5649-4418-ba70-43e31f969270 had a score of 16.33 in 2021, the year it was penalized, and ranked 8223rd.

[0064] Table 1 5.2: Supervised Learning Iterative Optimization: The XGBoost algorithm is used to train the negative company labels to obtain the importance distribution of each indicator, such as... Figure 10 As shown, the algorithm automatically identifies the characteristics of the financing and circulation stages that contribute the most to risk identification. Based on this, it automatically adjusts the weight coefficients in step 4, enabling the model to more accurately capture credit deterioration signals caused by circulation interruption or overdue settlement.

[0065] In summary, by constructing a dynamic heterogeneous transaction network, this embodiment of the invention fully explores data credit in supply chain finance, realizes dynamic and multi-dimensional data quantification and feature scoring of non-core enterprises, and significantly improves the accuracy of evaluation results.

[0066] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0067] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A data processing and feature scoring method based on heterogeneous supply chain networks, characterized in that, The method is executed through a physical device equipped with a data acquisition port, a data buffer medium, a central processing unit, and a data output port. The method includes: Through the data acquisition port, the system reads the digital credit transfer certificate data stream, financing business flow data stream, and enterprise business registration and credit data stream within a preset period from the external supply chain finance business system in real time, and writes the data stream into the data cache medium for persistent storage. The central processing unit constructs a dynamically updated supply chain heterogeneous transaction network based on the data stored in the data cache medium. G = ( V , E ), where the node set V Including core procurement enterprise nodes, Tier 1 supplier nodes, Tier 2 supplier nodes, Z Tier-1 supplier nodes and financial institution nodes, each node is associated with a multi-dimensional attribute vector including industry, tier, and credit limit, edge set E It includes directed credit certificate transfer relationship edges, directed financing business relationship edges, and undirected edges based on geographical location. Each edge is associated with a behavioral attribute vector including transaction amount, frequency, and cycle. Through this multi-type node and multi-type edge network architecture, the heterogeneous interaction characteristics between different entities in the supply chain are preserved, and the node attributes and edge weights are incrementally updated according to the natural month cycle. Based on the relationships within the heterogeneous transaction network of the supply chain, the central computing unit divides it into four sub-networks according to business dimensions: a credit transfer frequency sub-network representing the frequency of inter-enterprise transactions, a credit transfer amount sub-network representing the transaction scale, a financing frequency sub-network representing financing activity, and a financing amount sub-network representing the scale of capital acquisition. Based on the relationships between various types of nodes in the four sub-networks, multi-dimensional credit characteristic indicators are automatically extracted, including at least enterprise attributes and core enterprise affinity indicators, supply chain flow behavior indicators, and network topology characteristic indicators. The central processing unit performs distribution skewness calculation, long-tail smoothing transformation, and dimensionless standardization on all extracted indicators. Based on domain experience, it determines the internal weights of each dimension indicator. Then, through the first-level dimension mapping, the indicators are aggregated into major category dimension scores. Through the second-level comprehensive mapping, the scores are linearly weighted according to a fixed weight ratio of 60%, 20%, and 20%, and finally, the enterprise credit risk score that falls within the range of [0, 100] is output. The central processing unit sends the credit risk score, risk level, and early warning marker to an external risk control display terminal, business approval terminal, or early warning device through the data output port for display and triggering. The central processing unit loads negative samples of defaulting, litigated, and blacklisted companies, and iteratively corrects the weights of each level of indicators through a supervised learning algorithm, outputs the optimal weights, and updates the credit scoring model.

2. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 1, characterized in that, When constructing the heterogeneous transaction network of the supply chain, the central computing unit writes node data into a continuous physical storage block of the data cache medium in a fixed-length structure format. Each node is accompanied by attribute fields such as scale, industry, enterprise nature, province, city / county, registered capital, total credit line, available credit line, whether it is listed, and whether it has issued bonds. The central processing unit also establishes three types of edge relationship records in the data cache medium: Credit transfer relationship edge: pointing from upstream enterprise to downstream enterprise, with attributes including transfer direction, single transaction amount, total amount, cumulative average amount, time span, annualized number of transfers, and total number of transfers. The central processing unit appends a directed edge record to the data cache medium for each transfer record. Financing relationship edge: pointing from financial institutions to financing enterprises, with attributes including financing amount, interest rate, term, and total number of times. The central processing unit writes these information one by one according to the financing records. Geographic association edge: An undirected edge is established based on whether the province, city or district where the enterprise is located is the same. The central processing unit directly reads from the enterprise attribute field and generates association tags, which are then stored in the data cache medium.

3. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 2, characterized in that, The central processing unit calculates the edge weights of each of the four sub-networks based on the edge data in the data cache medium. All calculation results are then written to different physical partitions of the data cache medium for partitioned storage. Credit transfer subnetwork: Based on the credit transfer relationship edges stored in the data cache medium, the total number of times is counted by grouping by cube root and receiver, and used as the edge weight; Credit transfer amount sub-network: Based on the credit transfer relationship edges stored in the data cache medium, the transfer amounts between the same enterprise pairs are summed and used as edge weights; Financing frequency subnetwork: Based on the financing relationship edges stored in the data cache medium, the total number of financing requests is calculated by grouping financing demanders and financial institutions, and used as the edge weight; Financing Amount Subnetwork: Based on the financing relationship edges stored in the data cache medium, the financing amounts between the same enterprise and financial institutions are summed and used as edge weights; The central processing unit reads new business data monthly, updates the nodes and edges in the data cache medium, and synchronously updates the weights of the four sub-networks and edges.

4. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 3, characterized in that, Based on the four sub-networks and edge weights already stored in the data cache medium, the central processing unit further reads the enterprise's basic data and flow relationship data from the medium and extracts the following credit indicators: Enterprise attribute indicators: province, industry, whether it is a core enterprise, whether it is a first-tier supplier, whether it is a second-tier supplier, whether it is a third-tier supplier, enterprise size, whether there are joint buyers, number of joint buyers, whether it has issued bonds, whether it is listed, available credit line, total credit line, and total number of credit institutions; Core enterprise intimacy indicators: average ratio of credit transfer amount to credit opening amount at each level, average hierarchical distance between suppliers and core enterprises, average credit transfer amount of suppliers, and average credit opening amount; Among them, the average ratio of credit transfer amount to opening amount at all levels = (1 / M 1)×Σ(Credit Transfer Amount) l / Amount opened × 100), l = 1 to M 1, M 1 represents the total number of credit transfer records; the average distance between suppliers and core enterprises = (1 / M 2)×Σ(flow distance) j ), j = 1 to M 2, M 2 represents the total number of records for hierarchical distance statistics; the average supplier credit turnover amount = (1 / M 3) × Σ (Credit Transfer Amount) p ), p = 1 to M 3, M 3 represents the total number of supplier transaction amounts; the average credit opening amount = (1 / K )×Σ(Credit Opening Amount) q ), q = 1 to K , K Total number of credit records opened; The central processing unit will extract all completed metrics and store them in the metric zone of the data cache medium.

5. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 4, characterized in that, The central processing unit, based on the enterprise attribute indicators and core enterprise affinity indicators already stored in the data cache medium, and in conjunction with the credit transfer records, opening records, and clearing records stored in the medium, further calculates transfer behavior indicators, approval behavior indicators, and clearing behavior indicators: Circulation behavior indicators: total number of times suppliers transfer in, total amount of ... Approval behavior indicators: success rate of the circulation process, number of rejections in the circulation process, percentage of rejections in the circulation process, success rate of the opening process, number of rejections in the opening process, percentage of rejections in the opening process; Clearing behavior indicators: number of successful clearing, clearing success rate, whether there are any overdue clearing records, average overdue clearing days, and average cycle from opening date to clearing; The central processing unit calls the internal hardware arithmetic unit to complete all calculations, and adds the calculated flow, approval, and clearing indicators to the indicator area of ​​the data cache medium.

6. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 5, characterized in that, Based on the circulation, approval, and clearing indicators already stored in the data cache medium, the central processing unit further reads financing data and asset freeze data from the medium and extracts the following financing and constraint indicators: Financing behavior indicators: total amount of supplier financing, average amount of supplier financing per transaction, average interest rate of supplier financing per transaction, average number of days of supplier financing per transaction, number of times supplier financing is conducted, success rate of supplier financing process, number of supplier financing rejections, and percentage of supplier financing rejections. Financing constraint indicators: frozen amount of credit outflow, number of times credit outflow is frozen, and number of times partial credit outflow is frozen; The central processing unit performs filtering and cumulative calculations on the frozen records, and appends the results to the indicator area of ​​the data cache medium.

7. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 6, characterized in that, The central processing unit, based on the circulation, approval, clearing indicators, financing and constraint indicators already stored in the data cache medium, and further combining the four sub-networks and edge weights already stored in the data cache medium, performs graph calculations and extracts network topology feature indicators: Circulation amount network metrics: in-degree, out-degree, proximity centrality, eigenvector centrality, PageRank index; Financing frequency network metrics: Degree centrality, Betweenness centrality, Closeness centrality, Eigenvector centrality, PageRank index; The central processing unit temporarily stores intermediate calculation results such as adjacency matrix, eigenvector, reciprocal distance, and weight ratio in the high-speed cache area of ​​the data cache medium. After completing the graph calculation, it writes the topological feature indicators into the indicator area of ​​the data cache medium.

8. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 7, characterized in that, The central processing unit, based on enterprise attribute indicators, core enterprise affinity indicators, circulation behavior indicators, approval behavior indicators, clearing behavior indicators, financing behavior indicators, financing constraint indicators, and network topology characteristic indicators already stored in the data cache medium, performs data preprocessing and two-layer mapping scoring calculation according to the following steps: The distribution skewness of each extracted indicator is calculated, with a skewness threshold of 2. For indicators with a skewness greater than 2, the following steps are performed: ln ( x +1) Logarithmic transformation, followed by mean-variance standardization to dimensionlessly process all processed indicators. The calculation formula is as follows: z = ( x - μ ) / σ ,in μ The mean of the indicators. σ The standard deviation of the indicator is used, and the internal weights of the indicator after processing are confirmed based on domain experience. Perform the first-level dimensional mapping: Divide the indicators into three dimensions: credit circulation links, enterprise attributes and financing channels, and supply chain network characteristics. Then, perform a weighted summation and normalization of the indicators within each dimension. The calculation formula is as follows: The company in the dimension k Credit score Z k : ， in k = 1,2,3, where 1 is the credit circulation link dimension, 2 is the enterprise attributes and financing channels dimension, and 3 is the supply chain network characteristics dimension; J k To be classified as a dimension k The collection of all indicators; Perform the second-level comprehensive mapping: linearly weight the scores of the three dimensions according to a fixed weight ratio to obtain the final credit risk score. The calculation formula is as follows: Corporate Credit Score Y = 60% × Credit Circulation Link Dimension Score + 20% × Enterprise Attributes and Financing Channel Dimension Score + 20% × Supply Chain Network Characteristics Dimension Score; The central processing unit uses its internal hardware floating-point unit to perform all operations, including multiplication, division, logarithms, and summations, and stores the final credit risk score in the data cache medium.

9. The data processing and feature scoring method based on heterogeneous supply chain networks according to claim 8, characterized in that, The central processing unit performs supervised learning model optimization steps based on the final credit risk score stored in the data cache medium: Data on known defaulters, litigants, and blacklisted negative companies is collected and matched with the full sample of companies in the supply chain to identify negative companies in the supply chain network. At the same time, an equal number of normal companies are extracted to form a monitoring sample set with balanced positive and negative labels, which is then stored in the data cache medium. The supervised sample set is trained using a supervised learning algorithm, and the feature importance ranking of each indicator is automatically calculated. Based on the feature importance, the weight of indicators with high feature importance is automatically increased, and the weight of indicators with low feature importance is decreased. Repeat the iterative optimization process until the model converges or reaches the preset maximum number of iterations, and output the optimized set of index weights. The optimized set of weights is written over the model parameter area of ​​the data cache medium, and the credit risk score of all enterprises is recalculated based on the optimized weights. The central computing unit writes the iterative optimization process, model convergence status, and importance ranking of each indicator feature into the running log area of ​​the data cache medium, and pushes it to the external monitoring terminal via the data output port.

10. A data processing and feature scoring device based on a heterogeneous supply chain network, characterized in that, include: The data acquisition port is used to electrically read supply chain data streams and write them to storage. Data caching medium for persistent storage of nodes, edges, subnetworks, metrics, weights, and scores; A central processing unit, electrically connected to the data buffer medium, is used to execute the method according to any one of claims 1-9; The data output port is used to electrically output the score and warning signal to the risk control terminal, approval terminal or warning device.

Images