Enterprise credit report automatic acquisition and credit business co-processing system

Abstract

The invention relates to the technical field of financial science and technology and big data processing, and discloses an enterprise credit report automatic acquisition and credit business co-processing system, which converts a request into a context vector through a business request analysis module; the dynamic strategy route generation module calculates an acquisition priority based on the vector and generates a targeted task configuration file; the multi-source heterogeneous data acquisition module executes data acquisition; the data conflict arbitration module constructs a bipartite graph topological structure, eliminates conflicts among sources by using trust propagation iteration, and outputs cleaned feature vectors; the credit risk quantitative evaluation module calculates a credit score based on nonlinear mapping; and when the score falls into a gray interval, the gray closed-loop feedback module calculates decision ambiguity and feeds back an offset correction vector, and triggers supplementary collection of the uncollected data source. According to the method, on-demand data acquisition and truth value automatic discovery are realized, the boundary sample decision problem is effectively solved through a gray closed-loop mechanism, and the automation level of a credit system is improved.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and big data processing technology, specifically a system for automatically acquiring corporate credit reports and collaboratively processing credit business. Background Technology

[0002] With the rapid development of digital inclusive finance, commercial banks and financial institutions are increasingly relying on multi-source external data to build comprehensive borrower profiles when handling corporate lending. In order to obtain multi-dimensional data such as corporate registration information, tax records, legal proceedings, and operating cash flow, financial institutions need to connect to dozens or even hundreds of third-party credit data sources or government data interfaces.

[0003] In the existing enterprise credit data collection and processing workflow, the system adopts a static, predefined strategy for data acquisition. Regardless of the specific credit business scenario—whether it's a small, quick loan or a large business loan—and regardless of the industry characteristics of the applicant company, the system tends to perform a full scan and request of all external data sources. This extensive collection model ignores the differentiated data dimension requirements of different business scenarios, which not only increases data procurement costs and API call fees, generating a large amount of redundant data that does not contribute substantially to current decision-making, but also severely consumes the system's network bandwidth and server processing resources due to excessive concurrent requests during peak business periods, causing service response delays.

[0004] Furthermore, when the system obtains credit data for the same company from multiple heterogeneous channels, inconsistencies and even contradictions often arise due to differences in data update frequency, collection criteria, and source reliability (e.g., discrepancies in annual revenue disclosed by different platforms). Existing data integration solutions mostly employ simple rule-based weighting or manually specified priority strategies for conflict resolution. They lack a mathematical model capable of automatically assessing the credibility of each data source and extracting objective truth values ​​from conflicting data without manual annotation. This results in the generated corporate credit feature vector being mixed with noise, reducing the predictive accuracy of subsequent risk assessment models.

[0005] On the other hand, traditional credit approval processes follow a unidirectional, linear logic: data collection, feature extraction, model scoring, and threshold determination are performed sequentially. When the credit score output by the model falls into the gray area (i.e., the gray zone) between automatic approval and automatic rejection, existing systems lack automated response mechanisms and can only transfer such cases to manual review or directly reject loans based on rigid rules. This approach not only increases the workload of credit review personnel and reduces business processing efficiency, but also, due to the lack of incremental information supplementation mechanisms for fuzzy samples, is prone to inadvertently rejecting high-quality customers or mistakenly allowing potentially risky customers to pass due to insufficient information, making it difficult to maintain robust decision-making in complex credit business scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automatic enterprise credit report acquisition and credit business collaborative processing system. It solves the problems in existing credit business caused by the high cost of external data acquisition and excessive system resource consumption due to extensive full data collection, numerical conflicts between multi-source heterogeneous data and the difficulty in discovering the true value under unsupervised conditions, and the lack of an automated closed-loop retesting mechanism for credit applications in the gray area of ​​decision-making, which leads to high dependence on manual review and insufficient decision robustness.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions, including a business request parsing module, a dynamic policy routing generation module, a multi-source heterogeneous data acquisition module, a data conflict arbitration and structured parsing module, a credit risk quantitative assessment module, and a gray-scale closed-loop feedback module.

[0008] The business request parsing module is configured to receive credit trigger request messages, parse them, extract business metadata, convert the metadata into a numerical business context vector, and pass the business context vector to the dynamic policy routing generation module. The dynamic policy routing generation module is configured to generate collection priority probabilities for multiple external data sources based on the business context vector through linear mapping and normalization calculations, and generate a task configuration file containing the target data source address based on the collection priority probabilities. The multi-source heterogeneous data acquisition module is configured to initiate data requests to multiple external data sources according to the task configuration file, obtain the original messages, and output a multi-source heterogeneous dataset containing the original messages. The data conflict arbitration and structured parsing module is configured to receive the multi-source heterogeneous dataset, construct an observation matrix, eliminate numerical conflicts between multiple external data sources through iterative calculations based on a trust propagation graph, and output a ground truth cleaned enterprise credit feature vector. The credit risk quantification assessment module is configured to receive the enterprise credit feature vector and the business context vector, calculate the default probability prediction value through a nonlinear mapping network, and convert the default probability prediction value into a quantified credit score. The grayscale closed-loop feedback module is configured to compare the credit score with a preset decision threshold. When the credit score falls into the grayscale decision range, the decision ambiguity is calculated and a bias correction vector is generated. The bias correction vector is then fed back to the dynamic strategy routing generation module to trigger a supplementary data collection process for external data sources that have not been collected.

[0009] Furthermore, in terms of business feature processing, this invention achieves standardized vector mapping of unstructured business data by maintaining a business metadata dictionary. The business request parsing module performs one-hot encoding on discrete feature sets to generate binary sub-vectors; for continuous feature sets, it performs maximum-minimum normalization using pre-stored historical statistical lower and upper bound thresholds to generate normalized scalar feature values. The business request parsing module then concatenates the binary sub-vectors with the scalar feature values ​​to construct a feature vector that represents the current credit business context, providing a numerical basis for subsequent dynamic routing decisions.

[0010] Furthermore, in terms of data acquisition strategy optimization, this invention abandons the traditional full-data acquisition mode and adopts a context-based dynamic routing mechanism. The dynamic strategy routing generation module utilizes the stored routing weight matrix and routing bias vector to perform matrix multiplication and vector addition operations, projecting the business context vector onto the data source activation space to obtain an intermediate feature vector. Subsequently, an exponential function and summation normalization are applied to the intermediate feature vector to calculate a probability distribution vector with a sum of one. Each component in this probability distribution vector represents the acquisition priority of the corresponding external data source in the current business scenario. The system only activates data sources with an acquisition priority probability greater than a preset threshold, thereby reducing acquisition costs and system load while ensuring data validity.

[0011] Furthermore, in terms of multi-source heterogeneous data acquisition, this invention is compatible with both structured and unstructured data sources. The multi-source heterogeneous data acquisition module sends query requests to external data sources providing structured data services through an application programming interface adapter; simultaneously, by loading a headless browser kernel, it simulates client behavior by sending Hypertext Transfer Protocol requests to external data sources providing web document services and captures the document stream responded by the server, thereby achieving broad coverage of various credit reporting data sources.

[0012] Furthermore, in terms of data conflict arbitration, this invention introduces a truth discovery mechanism based on a graph model. The data conflict arbitration and structured parsing module maps data records into a bipartite graph topology containing a set of source nodes and a set of object nodes. In this bipartite graph topology, a weighted connection is established between the corresponding source node and object node only when the external data source provides valid observations of the credit indicator object. The system performs iterative calculations on the bipartite graph topology: in the truth estimation step, the source node reputation determined in the previous iteration is used as a weight to perform a weighted average of the observations to update the estimated truth value of the credit indicator object; in the reputation update step, the difference between the observation and the estimated truth value and the weighted error loss are calculated, and the source node reputation is updated through logarithmic mapping. When the iteration converges, the output estimated truth value is the cleaned enterprise credit feature vector, thus effectively solving the problem of inconsistency between multi-source data.

[0013] Furthermore, in terms of risk quantification assessment, this invention employs a deep learning model for nonlinear risk fitting. The credit risk quantification assessment module first performs feature fusion, cascading the business context vector and the enterprise credit feature vector to construct a comprehensive risk input vector. Subsequently, a multilayer perceptron network is used to perform linear transformation and correct linear unit activation on the comprehensive risk input vector, extracting hidden layer features. The output is then mapped to a probability range of zero to one using the Sigmoid function to obtain the predicted default probability value. In addition, the system uses a scoring conversion logic, based on the logarithm of the default odds and the doubling ratio factor, to convert the probability value into a credit score that aligns with business intuition.

[0014] Furthermore, regarding decision-making closed-loop feedback, this invention establishes a re-decision mechanism for gray-scale intervals. The gray-scale closed-loop feedback module determines whether the credit score falls within the closed interval formed by the automatic pass threshold and the automatic rejection threshold. If it falls within this gray-scale interval, the decision ambiguity is calculated based on the degree of distance of the score from the center score, and a bias correction vector is constructed accordingly. This bias correction vector uses a retest gain coefficient to weight the uncollected data sources, and the result is superimposed on the internal parameters of the dynamic policy routing generation module, forcibly triggering a supplementary data collection task. This mechanism enables the system to automatically acquire more dimensions of information for fuzzy samples, achieving a dynamic closed loop of initial review, retesting, and final review.

[0015] The enterprise credit report automatic acquisition and credit business collaborative processing system provided by this invention realizes on-demand allocation of data collection through dynamic routing driven by business context, reducing data acquisition costs; it realizes truth discovery of multi-source heterogeneous data under unsupervised conditions through conflict arbitration based on trust propagation graph, improving the accuracy of credit data; and it solves the problem of credit application processing at the decision boundary through gray-scale closed-loop feedback mechanism, improving the automation and reliability of the credit approval system.

[0016] This invention provides a system for automatically acquiring enterprise credit reports and collaboratively processing credit business. It has the following beneficial effects: 1. This invention converts business metadata into a context vector through a business request parsing module, and the dynamic policy routing generation module calculates the collection priority probability of external data sources based on this vector. Collection tasks are generated only for data sources with a probability higher than the activation threshold. By collecting data on demand based on business context, the blindness of performing a full scan of all data sources in traditional credit systems is avoided, the number of unnecessary external interface calls and data procurement costs are reduced, and the occupation of system network bandwidth and processing resources by concurrent collection is reduced.

[0017] 2. This invention utilizes a data conflict arbitration and structured parsing module to construct a bipartite graph topology of source nodes and object nodes. It also uses an iterative algorithm based on a trust propagation graph to synchronously update the estimated true value of credit indicator objects and the credibility of the data source. In the absence of manual annotation of true values, it can automatically identify and eliminate numerical conflicts and noise interference between different external data sources, ensuring that the output corporate credit feature vector has high credibility, thereby providing a high-quality data foundation for subsequent credit risk assessment.

[0018] 3. This invention monitors whether the credit score falls within the gray-scale decision range between automatic approval and automatic rejection through a gray-scale closed-loop feedback module. It also calculates the decision ambiguity for gray-scale samples to generate a bias correction vector, which is fed back to the routing module to trigger a supplementary data collection process. This breaks the traditional one-way linear approval process and enables automatic retesting and incremental information acquisition for fuzzy boundary samples. It reduces the proportion of manual review and avoids false rejections or false releases due to insufficient information, thereby improving the system's decision robustness in complex business scenarios. Attached Figure Description

[0019] Figure 1 This is a block diagram of the system logic functional module structure of one embodiment of the present invention; Figure 2 This is a schematic diagram of the business feature vectorization and dynamic policy routing generation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of data conflict arbitration and truth discovery based on trust propagation graph according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the credit risk quantitative assessment and gray-scale closed-loop feedback mechanism according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides an automatic enterprise credit report acquisition and credit business collaborative processing system, which may include: a collaborative processing server, a credit business management terminal, multiple external data source servers, and a data storage device.

[0022] The collaborative processing server, as the core computing device of this system, is deployed within the secure network domain of the financial institution. The collaborative processing server establishes a bidirectional data communication connection with the credit business management terminal through a first communication network, and establishes one-way or bidirectional data request connections with multiple external data source servers through a second communication network. The first communication network is an internal local area network (LAN) based on the TCP / IP protocol, and this network environment is within the protection boundary of a hardware firewall. The second communication network is a wide area network (WAN) or the Internet, used to realize data interaction with external nodes. At the connection boundary between the first and second communication networks, a network security gateway device is deployed, configured to perform access control list (ACL) filtering and network address translation (NAT) operations.

[0023] The credit business management terminal is a human-computer interaction device used by credit personnel to initiate query requests and receive decision results. This terminal is equipped with a front-end application to receive user-inputted credit application metadata. This metadata includes, but is not limited to, the company name, unified social credit code, loan amount requested, credit product type identifier, and urgency level identifier. The credit business management terminal encapsulates this metadata into a standardized trigger request data packet and sends it to the collaborative processing server via Hypertext Transfer Security Protocol (HTTPS).

[0024] Multiple external data source servers are clusters of computing devices independently maintained by different data holding institutions. These external data source servers specifically include: a first type of credit data server, a second type of judicial data server, and a third type of commercial data server. The first type of credit data server is maintained by the government credit reporting center, providing structured data query services based on application programming interfaces (APIs). The second type of judicial data server is maintained by the judicial administration department, providing public information retrieval services based on Hypertext Markup Language (HTML) documents. The third type of commercial data server is maintained by third-party data service providers, providing data streams in specific formats based on token authentication. These different types of servers employ heterogeneous data transmission protocols and message format standards.

[0025] The collaborative processing server, at the hardware level, includes at least one central processing unit (CPU), volatile memory (RAM), non-volatile memory (ROM), and a network interface controller (NIC). The non-volatile memory stores computer-executable instructions. When these instructions are executed by the CPU, the collaborative processing server performs specific logical steps for data acquisition, conflict arbitration, risk calculation, and collaborative feedback. The collaborative processing server is internally configured with a multi-threaded concurrent processing environment to maintain network sessions with multiple external data source servers simultaneously.

[0026] The data storage device is connected to the collaborative processing server via a high-speed peripheral interconnect bus or Fibre Channel. Logically, the data storage device is divided into a relational database storage area and a vector database storage area. The relational database storage area stores structured business request logs, raw data collection messages, and the final generated credit report documents. The vector database storage area is specifically configured to store high-dimensional feature vectors. In this embodiment, for each enterprise entity of the processing object, the system constructs and stores a business context vector. and corporate credit feature vector .

[0027] Among them, business context vector Defined as , express 3D real space, The preset number of business feature dimensions. Enterprise credit feature vector. Defined as , express 3D real space, This represents the number of dimensions for the cleaned credit metrics. The vector database storage area is configured with a vector index structure, supporting fast write and retrieval operations on the above vectors to support subsequent policy routing calculations and risk model operations.

[0028] The communication links between the collaborative processing server and multiple external data source servers are encrypted using Transport Layer Security (TLS). The collaborative processing server is equipped with a digital certificate management unit to store and verify the authentication certificates required for communication with the first-class credit data server. The collaborative processing server is also configured with a dynamic IP proxy pool interface to switch the network egress IP address when accessing the second-class judicial data server, adapting to the access frequency restriction policies of the external servers.

[0029] See attached document Figure 1 The enterprise credit report automatic acquisition and credit business collaborative processing system of the present invention is logically divided into: a business request parsing module, a dynamic strategy routing generation module, a multi-source heterogeneous data acquisition module, a data conflict arbitration and structured parsing module, a credit risk quantitative assessment module, and a business collaborative interaction module. These modules are connected through a high-speed data bus within the system, enabling bidirectional transmission of control commands and business data.

[0030] The business request parsing module is configured as the system's entry processing unit. This module includes a message unpacker and a feature vectorization unit. The message unpacker receives trigger request data packets from the credit business management terminal and decrypts and verifies the integrity of Hypertext Transfer Security Protocol (HTTPS) messages. The feature vectorization unit extracts discrete business metadata from the verified message payload and maps this discrete business metadata into a numerical business context vector. Business context vector The control signal is transmitted to the dynamic policy routing generation module.

[0031] The dynamic policy routing generation module includes a policy calculation unit and a task configuration generator. The policy calculation unit stores predefined routing weight parameters and is used to receive the business context vector. The system calculates collection priority values ​​for different data sources. Based on the calculated priority values, the task configuration generator generates a task configuration file containing a list of target data source addresses, access token parameters, and the number of concurrent threads. The dynamic policy routing generation module is also configured with a feedback receiving interface, which dynamically adjusts the routing weight parameters and regenerates the task configuration file when it receives a gray-scale retest instruction from the credit risk quantification assessment module.

[0032] The multi-source heterogeneous data acquisition module is the execution unit that performs specific data capture operations. Internally, this module integrates an application programming interface (API) adapter submodule and a webpage document parser submodule. The API adapter submodule is configured to send a structured query request conforming to RESTful specifications to the first type of credit data server according to the task configuration file. The webpage document parser submodule is configured to load a headless browser kernel according to the task configuration file, simulate a client sending a Hypertext Transfer Protocol (HTTP) request to the second type of judicial data server, and capture the document stream of the server's response. The output of the multi-source heterogeneous data acquisition module is connected to the data conflict arbitration and structured parsing module, outputting a multi-source heterogeneous dataset containing the original messages.

[0033] The data conflict arbitration and structured parsing module includes a field mapping engine and a trust propagation graph calculation unit. The field mapping engine extracts standardized credit indicator values ​​from multi-source heterogeneous datasets according to preset regular expression matching rules. The trust propagation graph calculation unit is configured to construct a bipartite graph structure containing data source nodes and data item nodes, and execute iterative calculation logic to eliminate numerical conflicts between different data sources. The data conflict arbitration and structured parsing module ultimately outputs a cleaned and denoised enterprise credit feature vector. .

[0034] The credit risk quantification assessment module includes a feature transformation layer, a risk calculation layer, and a gray-scale determination logic unit. The feature transformation layer is used to transform the enterprise's credit feature vector... The mapping is to hidden layer features. The risk calculation layer is used to calculate a single numerical risk probability value based on the hidden layer features. The grayscale decision logic unit stores a first decision threshold and a second decision threshold, which are used to compare the calculated risk probability value with the first and second decision thresholds. When the risk probability value is within the grayscale range defined between the first and second decision thresholds, the grayscale decision logic unit generates a supplementary acquisition instruction containing a missing dimension identifier and sends this supplementary acquisition instruction to the dynamic policy routing generation module through the feedback path.

[0035] The business collaboration interaction module is configured as the system's exit processing unit. This module includes a message encapsulator and a message queue client. The message encapsulator is used to encapsulate the determined risk probability value output by the credit risk quantification assessment module and the enterprise credit feature vector output by the data conflict arbitration and structured parsing module. The final report is merged and encapsulated into JavaScript Object Notation (JSON) format. The message queue client pushes the final report to the credit business management terminal or external credit core system via an asynchronous communication protocol.

[0036] Specifically, the business request parsing module is configured to execute feature vectorization processing logic. Internally, this module stores a pre-defined business dimension mapping table, which defines the correspondence between credit metadata and numerical ranges. When a trigger request data packet is received, the business request parsing module extracts the business type label, industry classification label, and credit limit range value. For discrete labels, the module uses one-hot encoding to convert them into binary sequences; for continuous values, it uses a maximum-minimum normalization method to map them to a closed interval between zero and one. Through these processes, the business request parsing module constructs... Dimensional business context vector . Defined as: ; in, Represents the first in the business context vector Each feature component; The number of business feature dimensions; This represents the transpose operation of a matrix or vector.

[0037] The dynamic policy route generation module is configured to be based on business context vectors. Calculate the access weight for each external data source. This dynamic policy routing generation module stores the routing weight matrix. and routing bias vector . The dimension is , The dimension is ,in This indicates the total number of external data sources that the system can connect to. This refers to the number of business feature dimensions. The dynamic policy routing generation module first performs a linear transformation operation to obtain an intermediate feature vector. : ; Subsequently, the dynamic policy routing generation module processes the intermediate feature vector. The final policy probability vector is calculated by applying the Softmax normalization function. . The first in Each component Defined as the first The probability of prioritizing the collection of each external data source in the current task is calculated using the following formula: ; in, Represented by natural constant An exponential function with base 0; and These represent the intermediate feature vectors respectively. The Middle The and the first The value of an external data source; This represents the sum of intermediate feature index values ​​corresponding to all data sources. This term serves as a normalization factor to ensure that the sum of the probabilities of all collection priorities is 1. This represents the total number of external data sources. The dynamic policy routing generation module will calculate this. Compare with a preset activation threshold, only when When the value is greater than the activation threshold, the corresponding number will be... Add an external data source to the task configuration file.

[0038] The data conflict arbitration and structured parsing module is configured to resolve numerical inconsistencies in multi-source heterogeneous data. This module executes iterative computation logic based on a trust propagation graph. Definitions For the first The external data source is for the first The observations provided by each credit data item. Definition For the first In the nth iteration Global reputation of an external data source. Definition For the first In the nth iteration The estimated true value of each credit data item. In each iteration step, the data conflict arbitration and structured parsing module first calculates the estimated true value based on the current data source reputation. .

[0039] ; in, To provide the first A collection of indexes from all external data sources for each credit data item; Indicates the first The external data source in the previous round (i.e., the first) Global reputation in round (iteration); This represents the sum of observations weighted by reputation level. This represents the sum of the reputation scores of all relevant data sources, used for weighted averaging.

[0040] Subsequently, the data conflict arbitration and structured parsing module updates the reputation of each data source based on the calculated estimated truth value. This data conflict arbitration and structured parsing module first calculates the... The external data source in the first Weighted loss value in the next iteration : ; in, For the first The set of data items covered by an external data source; For the first The variance of all observations for each credit data item; It is a non-zero smoothing constant; representing the smoothing constant for the set of data items. All credit data items Perform a traversal summation operation. Based on this weighted loss value, the data conflict arbitration and structured parsing module updates the global reputation of the data source using a logarithmic function. : ; in, Represents the natural logarithm function; Indicates the first The external data source in the first The weighted loss value in the next iteration; Indicates the first The external data source in the first The weighted loss value in the next iteration; Indicates all The sum of the weighted loss values ​​of each external data source in the current round.

[0041] The data conflict arbitration and structured parsing module repeatedly executes the above truth estimation and reputation update steps until the change in global reputation is less than a preset convergence threshold. The estimated truth value at the end of the iteration. Constitutes the final corporate credit feature vector .

[0042] The credit risk quantitative assessment module is configured to be based on corporate credit feature vectors. Calculate the risk probability value. This credit risk quantification assessment module includes feature extraction layer parameters. , and output layer parameters , The credit risk quantitative assessment module first calculates the hidden layer feature vector. : ; in, This represents the activation function of the linear rectifier unit. Subsequently, the credit risk quantification assessment module calculates and outputs the risk probability value using the Sigmoid function. The calculation formula is: ; in, This represents the Sigmoid activation function, and its expression is: ; Indicates the output layer parameters Transpose of; This represents the output layer parameters.

[0043] The credit risk quantitative assessment module will Output to the grayscale determination logic unit. The grayscale determination logic unit then makes a judgment. Whether it falls within the first determination threshold Second determination threshold Defined closed interval If the value falls within this closed interval, it is determined to be in a grayscale state, and a supplementary data collection process for the missing dimension is triggered; otherwise, it is determined to be in a definite state, and... Send to the business collaboration and interaction module.

[0044] See attached document Figure 2 The business feature vectorization processing logic provided by this invention is first executed by the business request parsing module, which is configured to convert unstructured credit application metadata into a computer-computable mathematical vector representation.

[0045] The business request parsing module first performs field-level parsing on the received credit trigger request message. Internally, it maintains a business metadata dictionary, which defines all business dimension fields supported by the system and their data types. The parsing process decomposes the message payload into discrete and continuous feature sets. Discrete feature sets include credit product type identifiers (e.g., working capital loans, fixed asset loans, bill discounting), enterprise industry classification codes (e.g., GB / T 4754 standard codes), and guarantee method identifiers (e.g., credit, mortgage, pledge, guarantee). Continuous feature sets include the application amount, loan term (in months), and business processing urgency score (integer levels from 1 to 10).

[0046] For discrete feature sets, the business request parsing module performs one-hot encoding. The system predefines a fixed state space for each discrete dimension. Taking credit product types as an example, let the total number of product types supported by the system be... If the product type in the current request corresponds to the first one in the state space... Number of positions (of which) If the business request parsing module generates a request with a length of [length missing], then the business request parsing module will generate a request with a length of [length missing]. The binary subvector, where the first... The element at one position has a value of 1, and the elements at the remaining positions have a value of 0. For industry classification codes, the system uses a hash mapping method to convert them into fixed-dimensional sparse sub-vectors to handle high-dimensional sparse features.

[0047] For continuous feature sets, the business request parsing module performs a min-max normalization operation. The system stores historically calculated lower bound thresholds for each continuous dimension. and upper bound threshold For the original values ​​in the current request The business request parsing module calculates the normalized feature values. : ; like If the request exceeds the preset upper and lower bounds, the business request parsing module will perform truncation processing, and... Forced constraints Within a closed interval. This step eliminates the order-of-magnitude difference between different physical units (such as the amount in yuan versus the period in months), ensuring the numerical stability of subsequent matrix operations.

[0048] After completing the independent processing of each of the above dimensions, the business request parsing module performs a vector concatenation operation. Following a predefined feature arrangement order, it concatenates all generated binary sub-vectors, sparse sub-vectors, and normalized scalar feature values ​​end-to-end to form the final vector. Dimensional business context vector .

[0049] This business context vector It is then passed to the dynamic policy routing generation module, serving as the sole input variable that determines subsequent data collection behavior.

[0050] The dynamic policy routing generation module is configured to generate refined collection instructions for each external data source based on the business context vector. This module has a pre-built set of learnable routing parameters, including a routing weight matrix. and routing bias vector These parameters define the nonlinear mapping relationship between the high-dimensional business feature space and the data source access space.

[0051] The dynamic policy route generation module first receives data from the business request parsing module. Dimensional business context vector Then, matrix multiplication and vector addition operations are performed to project the business context features onto the data source activation space, using formulas. Calculate the intermediate feature vector .

[0052] Subsequently, the dynamic policy routing generation module processes the intermediate feature vector. The Softmax normalization function is applied to eliminate the dimensional differences in activation intensity from different data sources and to generate a probability distribution-based acquisition priority, which is then calculated using the formula... Calculate the first Collection priority probability of external data sources .

[0053] After calculating the sampling priority probability vector Then, the dynamic policy routing generation module executes the task configuration generation logic. This dynamic policy routing generation module stores preset activation thresholds. For any _th An external data source, if and only if At this time, the external data source is marked as active. For external data sources in an active state, the dynamic policy routing generation module further... The numerical value of the collected parameters is dynamically calculated.

[0054] Specifically, the system is configured with a retry count scaling factor. and base timeout The dynamic policy route generation module is the first... Calculate the maximum number of retries from external data sources. and request timeout threshold . Calculated as and The floor value of the product, i.e. ; Calculated as This calculation logic ensures that higher-priority data sources have more retry opportunities and a longer tolerance for waiting time during network fluctuations, thereby guaranteeing the success rate of acquiring critical data. Finally, the dynamic policy routing generation module outputs a task configuration file containing the above parameters to the multi-source heterogeneous data acquisition module.

[0055] See attached document Figure 3 The data conflict arbitration and structured parsing module involved in this invention first performs scenario modeling on the raw data obtained from different channels, mapping discrete data records into an observation matrix.

[0056] The data conflict arbitration and structured parsing module defines all external data sources connected to the system as a source set. ,in This represents the total number of external data sources. The enterprise credit metrics to be determined are defined as a set of objects. ,in This indicates the number of dimensions in the cleaned credit metric. Due to objective differences in the update frequency, collection channels, and information focus of various external data sources, the number of dimensions varies for the same credit metric object. (e.g., a company's paid-in capital), different external data sources Different observation records will be provided. Data conflict arbitration and structured parsing module definition. For data source Regarding credit indicator targets The observed value. It should be noted that this observed value... It is a value that has undergone unit unification and format cleaning, but has not yet been judged as true.

[0057] Based on the above definitions, the data conflict arbitration and structured parsing module constructs the observation matrix. Observation matrix It is a dimension A sparse matrix, where row indices correspond to credit metric objects and column indices correspond to external data sources. Since not all external data sources can cover all credit metrics, when external data sources... No information was provided regarding the credit indicator targets. When information is available, the observed values Marked as null or missing (NaN). The data conflict arbitration and structured parsing module traverses each row of the observation matrix. If a row contains at least two non-null observations, and the absolute value of the difference between these two observations is greater than the preset tolerance error, then the data is considered null. When this happens, the system defines the state as a data conflict scenario.

[0058] The data conflict arbitration and structured parsing module further categorizes data conflict scenarios into numerical conflicts and categorical conflicts. For numerical conflicts (such as total assets and debt ratio), the observed values... Belonging to the continuous real number domain, conflicts manifest as statistical dispersion of values; for categorical conflicts (such as business status, credit rating), the observed values... Belonging to a discrete symbol set, conflicts manifest as inconsistencies in state labels. For the aforementioned conflict scenarios, the data conflict arbitration and structured parsing module does not directly employ simple majority voting or arithmetic averages, as this ignores the inherent differences in credibility between different data sources. Therefore, the observation matrix... As input, it is transmitted to the subsequent trust propagation graph computation unit, where the potential truth value is derived through the graph model.

[0059] The data conflict arbitration and structured parsing module provided by this invention is configured to construct a bipartite graph topology based on the aforementioned observation matrix, and perform iterative calculations on the bipartite graph topology to derive potential truth values, thereby deriving a unique truth solution in multi-source data conflict scenarios.

[0060] The trust propagation graph computation unit first constructs a bipartite graph structure containing two types of nodes. The first type of nodes is defined as the set of source nodes. Each node in this set of source nodes corresponds to a unique external data source. The total number of external data sources is The second type of node is defined as a collection of object nodes. Each node in this set of object nodes corresponds to a unique credit indicator object. The number of dimensions for the cleaned credit metrics is 10. The trust propagation graph calculation unit is based on the observation matrix. An edge connection is established between non-empty elements in the array if and only if the first element is empty. The external data source provided information about the first... Observations of each credit data item At that time, in external data sources With credit indicator objects Create an undirected edge between them. The weight of this undirected edge is initially set to the value of the observation. This topology establishes a constraint that only allows information exchange between source nodes and object nodes, while prohibiting direct communication between nodes of the same type.

[0061] Based on the established bipartite graph, the trust propagation graph computation unit initiates the iterative computation process, first performing initialization operations and setting the number of iterations. And calculate the variance of the observations for each credit indicator object. Since there is a lack of prior reliability information in the initial stage of computation, the trust propagation graph computation unit calculates the initial global reputation of all source nodes. Set to a preset uniform value (e.g.) This indicates that all external data sources are assumed to have equal prior confidence in the initial stage of computation.

[0062] Subsequently, the algorithm enters an iterative phase, alternately executing the truth value estimation of credit indicator objects and the source node reputation update steps. In the object node truth value estimation step, for any credit indicator object... The trust propagation graph computation unit aggregates the current reputation information of all source nodes connected to it and calculates the weighted average truth value. Definition In order to be related to credit indicators A set of indexes for all external data sources connected by edges. (Using the formula...) Calculate the first The estimated true value in the next iteration .

[0063] In the source node reputation update step, for any external data source The trust propagation graph computation unit calculates the sum of deviations between all the observations it provides and the current estimated true value. Definition To connect with external data sources The set of data items for all object nodes connected by edges. The trust propagation graph calculation unit first uses the formula... Calculate the first Weighted loss value of each external data source Based on the calculated weighted loss value, the trust propagation graph calculation unit uses the formula... Update the global reputation of the source node. .

[0064] The trust propagation graph calculation unit repeats the above two steps until the convergence condition is met. The convergence condition is defined as the L2 norm of the global reputation change across all external data sources being less than a preset convergence threshold. ,Right now When this condition is met, the system stops iterating and changes the current set of estimated true values. The final cleaned corporate credit feature vector was determined. The amount.

[0065] See attached document Figure 4 The credit risk quantification assessment module involved in this invention is configured to receive a high-confidence corporate credit feature vector output from the data conflict arbitration and structured parsing module, and transform it into a standardized credit risk score through a nonlinear mapping network.

[0066] The credit risk quantitative assessment module first performs feature fusion and preprocessing operations. The input layer receives two parts of data: one part is ground truth cleaning. 3D corporate credit feature vector The other part is reserved. The original business context vector of dimension The credit risk quantitative assessment module concatenates these two vector components to construct a dimension... Comprehensive risk input vector This fusion mechanism ensures that the model makes a comprehensive judgment not only based on the company's objective credit status (such as assets and litigation), but also on the current business application background (such as loan urgency and product type).

[0067] Subsequently, the comprehensive risk input vector The data is fed into a feature crossover and extraction network. To capture the nonlinear combination relationships between features of different dimensions (e.g., the impact of the ratio of total liabilities to registered capital on risk), the credit risk quantification assessment module employs a multilayer perceptron (MLP) structure as the feature extractor. This network contains at least one hidden layer. Definition For dimension The risk characteristic weight matrix For dimension The bias vector. The hidden feature vectors of the hidden layer. The calculation is as follows: ; in, This represents a nonlinear activation function, which is selected in this embodiment. A function to address the vanishing gradient problem and accelerate convergence; This parameter represents the number of neurons in the hidden layer, and it determines the model's capacity to abstract complex risk patterns.

[0068] After deep transformation by the feature extraction network, the credit risk quantification assessment module performs risk probability prediction. The output layer outputs high-dimensional hidden feature vectors. Mapped to a scalar, and the output value compressed using the Sigmoid function. The interval is used to obtain the predicted value of the default probability. : ; in, Indicates the output layer parameters Transpose of; Indicates the output layer parameters; Represented by natural constant An exponential function with base 0; It quantifies the estimated probability of current applicant companies defaulting on their loans.

[0069] To meet the interpretability needs of business personnel and the universality of risk control strategies, the credit risk quantitative assessment module further includes scoring conversion logic. This scoring conversion logic converts the predicted default probability value... Credit scores mapped to integers The system is preset with a baseline score. and doubling ratio factor The conversion parameters constitute the credit score. The formula for calculating the credit score is as follows: ; in, This represents the final credit score, usually a positive integer. This represents the preset baseline score, corresponding to the standard score value under a specific odds. This represents the doubling ratio factor, used to control the amount of score increase when the default odds double; Indicated by A logarithmic function with base 0; This is the predicted probability of default. This represents the ratio of the probability of non-default to the probability of default, i.e., the default odds. This formula indicates the predicted default probability value. The lower the score, the higher the default penalty, and the higher the calculated credit score. The higher the value, the more directly it reflects the company's creditworthiness.

[0070] This formula ensures a linear relationship between credit score and the logarithm of default odds. The calculated... As the final quantitative risk indicator, it is directly pushed to the credit approval system as a basis for decision-making, and also transmitted as a feedback signal to the subsequent gray-scale closed-loop feedback module for dynamic adjustment of the system's routing strategy.

[0071] The grayscale closed-loop feedback module involved in this invention is configured to classify the confidence level of the aforementioned generated credit score, and trigger dynamic secondary data routing when the state is determined to be uncertain, thereby forming a closed-loop control system.

[0072] The grayscale closed-loop feedback module has a pre-set set of decision thresholds, including automatic pass thresholds. and automatic rejection threshold ,in, The grayscale closed-loop feedback module receives the credit score. Then, execute the three-level decision logic: if The system directly generates the command; if The system directly generates a rejection command; if The system defines the current state as a gray-scale decision interval. Requests falling within this gray-scale decision interval indicate that the currently collected combination of external data sources is insufficient to support a high-confidence risk assessment, resulting in excessive information entropy.

[0073] For business requests falling into the gray-scale decision-making range, the gray-scale closed-loop feedback module activates the closed-loop retesting mechanism. This mechanism aims to uncover potential high-value data sources that have not yet been activated, supplementing feature dimensions. First, the gray-scale closed-loop feedback module calculates the current decision ambiguity. Its quantification formula is as follows: ; in, Indicates the fuzziness of the decision, with a value range of . The larger the value, the closer the score is to the middle of the fuzzy zone; The center score of the grayscale interval is represented by the following formula: .

[0074] Subsequently, the grayscale closed-loop feedback module generates a feedback adjustment signal for the dynamic policy routing generation module. To force the system to collect new data increments, this grayscale closed-loop feedback module constructs a bias correction vector. This is used to temporarily adjust routing bias parameters. The first bias correction vector... Activation bias increment The calculation is as follows: ; in, Indicates that for the first Incremental activation bias of an external data source; This is a preset retest gain coefficient used to control the adjustment intensity; Let the second state indicator be a binary state indicator variable, if the first... If an external data source has already been collected in the previous round of routing, then ,otherwise This formula ensures that the system only adds activation weights to external data sources that have not been collected, and the more ambiguous the current decision, the greater the increase in weight.

[0075] The grayscale closed-loop feedback module will calculate the... This feedback is sent to the dynamic policy route generation module. The dynamic policy route generation module then updates its internal route bias vector accordingly. Based on the updated parameters, the system re-executes the Softmax normalization and data acquisition task generation steps, thereby activating a new set of external data sources for supplementary data acquisition. The newly acquired data will be processed by the data conflict arbitration and structured parsing module before updating the enterprise credit feature vector. This triggers a recalculation of the credit risk feature extraction and scoring model, generating a new credit score. The retest process is subject to the maximum number of iterations. The constraint, if it is reached If the score remains in the gray area after the second cycle, the system will forcibly transfer it to the manual review queue.

Claims

1. An enterprise credit report automatic acquisition and credit business collaborative processing system, characterized in that, include: The business request parsing module is configured to receive a credit trigger request message, parse the credit trigger request message and extract business metadata, convert the business metadata into a business context vector in numerical form, and pass the business context vector to the dynamic policy routing generation module. The dynamic policy routing generation module is configured to generate collection priority probabilities for multiple external data sources based on the business context vector through linear mapping and normalization calculation, and generate a task configuration file containing the target data source address based on the collection priority probabilities. The multi-source heterogeneous data acquisition module is configured to initiate data requests to multiple external data sources according to the task configuration file, obtain raw messages, and output a multi-source heterogeneous dataset containing the raw messages. The data conflict arbitration and structured parsing module is configured to receive the multi-source heterogeneous dataset, construct an observation matrix, eliminate numerical conflicts between multiple external data sources through iterative calculation based on the trust propagation graph, and output a truth-cleaned enterprise credit feature vector. The credit risk quantification assessment module is configured to receive the enterprise credit feature vector and the business context vector, calculate the default probability prediction value through a nonlinear mapping network, and convert the default probability prediction value into a quantified credit score. The grayscale closed-loop feedback module is configured to compare the credit score with a preset decision threshold. When the credit score falls into the grayscale decision range, it calculates the decision ambiguity and generates a bias correction vector. The bias correction vector is then fed back to the dynamic strategy routing generation module to trigger a supplementary data collection process for external data sources that have not been collected.

2. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The business request parsing module internally maintains a business metadata dictionary, which defines discrete feature sets and continuous feature sets; For the discrete feature set, the business request parsing module performs a one-hot encoding operation to generate binary sub-vectors; For the continuous feature set, the business request parsing module performs a maximum and minimum value normalization operation using the pre-stored historical statistical lower and upper bound thresholds to generate normalized scalar feature values; The business request parsing module concatenates the binary subvector and the scalar feature value in a predefined order to form the business context vector.

3. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The dynamic policy route generation module stores a route weight matrix and a route bias vector; The dynamic policy routing generation module first performs matrix multiplication and vector addition operations to project the business context vector onto the data source activation space to obtain an intermediate feature vector. Subsequently, the dynamic strategy routing generation module applies an exponential function and summation normalization to the intermediate feature vector to calculate a probability distribution vector with a sum of one. Each component in the probability distribution vector represents the collection priority probability of the corresponding external data source. The dynamic policy routing generation module compares the collection priority probability with a preset activation threshold, marks the external data source corresponding to the collection priority probability being greater than the activation threshold as active, and writes it into the task configuration file.

4. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module integrates an application programming interface adapter submodule and a web page document parser submodule; The application programming interface adapter submodule is configured to send a query request conforming to the specification to an external data source that provides structured data services, based on the task configuration file. The web page document parser submodule is configured to load a headless browser kernel according to the task configuration file, simulate a client sending a Hypertext Transfer Protocol request to an external data source that provides web page document services, and capture the document stream responded by the server.

5. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The data conflict arbitration and structured parsing module is configured to map data records into a bipartite graph topology. The bipartite graph topology includes a set of source nodes and a set of object nodes. Each source node in the set of source nodes uniquely corresponds to an external data source, and each object node in the set of object nodes uniquely corresponds to a credit indicator object. The data conflict arbitration and structured parsing module establishes connection relationships based on the observation matrix. Only when an external data source provides a valid observation value for a credit indicator object, an undirected edge is created between the corresponding source node and the object node, and the weight of the undirected edge is set to the valid observation value.

6. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 5, characterized in that, The data conflict arbitration and structured parsing module performs iterative calculations on the bipartite graph topology, including a truth value estimation step for credit indicator objects and a source node reputation update step. In the step of estimating the true value of the credit indicator object, the data conflict arbitration and structured parsing module uses the source node reputation determined in the previous iteration as a weight to perform a weighted average calculation on the observations provided by all connected external data sources to obtain the estimated true value of each credit indicator object in the current round. In the source node reputation update step, the data conflict arbitration and structured parsing module calculates the difference between the observation provided by each external data source and the estimated true value, calculates the weighted error loss in combination with the sample variance, and maps the weighted error loss to the updated source node reputation through a logarithmic function. When the change in the reputation of all source nodes is less than the preset convergence threshold, the data conflict arbitration and structured parsing module terminates the iteration and determines the final estimated truth value as the enterprise credit feature vector.

7. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The credit risk quantification assessment module performs a feature fusion operation, concatenating the retained business context vector with the cleaned enterprise credit feature vector to construct a comprehensive risk input vector; The credit risk quantification assessment module includes a multilayer perceptron network. The multilayer perceptron network uses a risk feature weight matrix and a bias vector to perform a linear transformation on the comprehensive risk input vector, and outputs the hidden layer feature vector through a modified linear unit activation function. The credit risk quantification assessment module uses the output layer weights to map the hidden layer feature vectors into scalars, and then uses the Sigmoid function to compress the scalars to the range of zero to one, thereby obtaining the predicted default probability value.

8. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 7, characterized in that, The credit risk quantification assessment module includes a scoring conversion logic, which maps the predicted default probability value to the credit score based on a preset benchmark score and a doubling ratio factor. The scoring conversion logic calculates the ratio of the non-default probability to the predicted default probability to obtain the default odds, calculates the logarithm of the default odds, multiplies the logarithm by the doubling ratio factor, and then adds it to the base score to obtain the credit score.

9. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 1, characterized in that, The grayscale closed-loop feedback module is preset with an automatic pass threshold and an automatic rejection threshold, wherein the automatic pass threshold is greater than the automatic rejection threshold. The grayscale closed-loop feedback module determines whether the credit score is within the closed interval formed by the automatic pass threshold and the automatic rejection threshold; If the credit score is within the closed interval, the gray-scale closed-loop feedback module determines the current state as a gray-scale decision interval and calculates the decision ambiguity based on the distance between the credit score and the center score of the gray-scale interval.

10. The enterprise credit report automatic acquisition and credit business collaborative processing system according to claim 9, characterized in that, After determining that the grayscale decision interval has been entered, the grayscale closed-loop feedback module constructs the bias correction vector for the dynamic policy routing generation module. The grayscale closed-loop feedback module identifies external data sources that were not collected in the previous round of routing, calculates the activation bias increment based on the decision ambiguity and the preset retest gain coefficient, and assigns the activation bias increment to the corresponding component in the bias correction vector. The dynamic policy routing generation module receives the bias correction vector, superimposes the bias correction vector onto the internal routing bias vector, and re-executes the data collection task generation step.

Images