A data processing method, medium and device based on a quantum computer

Abstract

Embodiments of the present application provide a data processing method based on a quantum computer, a medium and equipment, the method comprising: structurally processing a plurality of accounts receivable original records to obtain a standardized object pool; removing objects with invalid rights status from the standardized object pool according to a rights status field to obtain a reserved object list; generating a forbidden set and a quadratic high penalty term according to mutual exclusion relationships in the reserved object list; calculating a single comprehensive risk cost of each object in the reserved object list; generating a sparse pairwise structure risk term for each target object pair in the reserved object list that does not belong to the forbidden set; determining a basic risk function according to the single comprehensive risk cost and each sparse pairwise structure risk term; superimposing the basic risk function, the quadratic high penalty term and various square penalty terms to obtain a total QUBO objective function; and compiling the QUBO objective function into a parameter configuration executable by a quantum processor and determining output data through the quantum processor.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, the embodiments of this application relate to a data processing method, medium, and device based on a quantum computer. Background Technology

[0002] In scenarios involving the construction of asset or project portfolios based on multi-source heterogeneous data, the objects to be processed are typically not single data records, but rather candidate datasets composed of multiple candidate records. This construction process not only needs to handle the basic attributes of the records themselves (such as amount, time, and entity identifier), but also simultaneously process status fields reflecting the validity and executability of the records (such as contract identifier, registration status, dispute status, and duplicate identifier), and output a set of selected data records that meet the requirements under various constraints such as budget, resource usage, term distribution, and concentration control.

[0003] Traditional processing methods experience a sharp increase in processor time as the number of candidate records increases. When there are hundreds of candidate records, the processor time may take several hours, making it difficult to meet the timeliness requirements of data processing.

[0004] Existing approaches attempt to model combinatorial optimization problems as QUBO (Quadratic Unconstrained Bivariate Optimization) models and process them using quantum processors. These approaches involve extracting the objective function, converting it into QUBO coefficients, and then submitting these coefficients as input to the quantum processor.

[0005] However, the above approach has at least the following technical problems:

[0006] 1. Slow processing speed

[0007] When the number of candidate records is large, the constructed QUBO model contains a huge number of secondary coupling terms. When processing such a high-density coupled model, the annealing time required by the quantum processor increases significantly, and embedding failures occur frequently due to the mismatch between the coupling relationships and the physical connectivity graph. The overall processing time is significantly increased, failing to leverage the speed advantage of the quantum processor.

[0008] II. Poor hardware compatibility

[0009] The directly calculated QUBO coefficient values ​​often exceed the physical limits allowed by the quantum processor (for example, the absolute values ​​of local fields and coupling coefficients typically do not exceed 2). This leads to numerical truncation or saturation within the processor after submission, causing the actual target function to deviate significantly from the design intent. Furthermore, the coupling relationships between different logical variables cannot match the physical coupling graph of the quantum processor, resulting in repeated failures in the embedding process or requiring extensive manual intervention, severely slowing down the processing flow.

[0010] III. Results are unavailable or unreliable

[0011] In the sampling results returned by quantum processors, there are often instances where multiple physical bits corresponding to the same logical variable have inconsistent values ​​(i.e., broken links), and many results do not meet business constraints such as budget and resource consumption. Current practices typically discard these infeasible results directly, resulting in an extremely low effective sampling rate (sometimes less than 1%), requiring thousands of repeated samplings to obtain a small number of feasible solutions; or they simply repair the inconsistent bits through voting and output them directly without subsequent constraint verification and adjustment, causing the repaired combination to still violate the preset constraints and making it unusable for downstream automated processing systems. Low utilization and unreliable output make existing solutions difficult to implement in real-world data processing tasks.

[0012] In summary, existing quantum processor-based processing schemes have significant shortcomings in terms of processing speed, hardware adaptability, and result reliability. Therefore, a method that can systematically solve the above-mentioned technical problems is needed. Summary of the Invention

[0013] The purpose of this application is to provide a data processing method, medium, and device based on a quantum computer. Addressing the slow processing speed of quantum processors, this application employs candidate reduction and conflict compilation based on right states and mutual exclusion relationships. This eliminates unusable objects and pre-generates forbidden sets and secondary high-penalty terms. Simultaneously, it generates sparse pairwise risk terms only for strongly correlated object pairs, significantly reducing the variable size and secondary coupling density of the QUBO model. This reduces the annealing time and embedding failure rate required by the quantum processor, thereby improving data processing speed. To address the issue of unusable or unreliable results, this application employs closed-loop post-processing—including link break repair, inverse embedding, constraint replay, and feasible solution recovery—to stably transform unusable physical samples returned from quantum sampling into feasible combinations that satisfy all business constraints. This significantly improves the utilization rate of quantum samples and ensures that the output results can be directly used in downstream automated decision-making systems.

[0014] In a first aspect, embodiments of this application provide a data processing method based on a quantum computer. The data processing method includes: performing structured processing on multiple original accounts receivable records to obtain a normalized object pool, wherein each object in the normalized object pool corresponds to an accounts receivable and includes attribute information, a rights status field, and an identifier field for constraint compilation of the corresponding accounts receivable; removing objects with invalid rights status from the normalized object pool at least based on the rights status field to obtain a list of retained objects; generating a forbidden set and a secondary high penalty corresponding to each mutual exclusion relationship based on the mutual exclusion relationships in the list of retained objects; and calculating the... The single-transaction comprehensive risk cost for each object in the retained object list is calculated. For each pair of target objects in the retained object list that does not belong to the forbidden set, sparse pairwise structural risk terms are generated, wherein the target object pair belongs to two objects that meet the structural association strength requirement. The basic risk function is obtained by weighted summation of each single-transaction comprehensive risk cost and each sparse pairwise structural risk term based on the decision variable, wherein the decision variable is used to characterize whether the corresponding original record of accounts receivable is selected. The value of each decision variable is determined at least based on the basic risk function and the quantum processor, and a set of data records is obtained based on whether the object corresponding to the value is selected.

[0015] In some embodiments of this application, the step of determining the value of each decision variable based at least on the basic risk function and the quantum processor, and obtaining a data record set based on whether the object corresponding to the value is selected, includes: superimposing the basic risk function, the quadratic high penalty term, and various squared penalty terms to obtain a total QUBO objective function, wherein the various squared penalty terms are used to convert business constraints into penalty terms in the QUBO model for quantum computer solution, and the various squared penalty terms are obtained by discretizing various constraints and encoding binary relaxation variables; compiling the total QUBO objective function into a parameter configuration executable by the quantum processor; submitting the parameter configuration to the quantum processor for sampling to obtain quantum samples; sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum samples to obtain the value of each decision variable, and obtaining a data record set based on whether the object corresponding to the value is selected.

[0016] This application addresses the slow processing speed of quantum processors by employing candidate reduction and conflict compilation based on right states and mutual exclusion relationships. It eliminates unusable objects and pre-generates forbidden sets and secondary high-penalty terms. Simultaneously, it generates sparse pairwise risk terms only for strongly correlated object pairs, significantly reducing the variable size and secondary coupling density of the QUBO model. This reduces the annealing time and embedding failure rate required by the quantum processor, thereby improving data processing speed. To address the issue of unusable or unreliable results, this application uses closed-loop post-processing—including link break repair, inverse embedding, constraint replay, and feasible solution recovery—to stably transform unusable physical samples returned from quantum sampling into feasible combinations that satisfy all business constraints. This significantly improves the utilization rate of quantum samples and ensures that the output results can be directly used in downstream automated decision-making systems.

[0017] In some embodiments, before superimposing the basic risk function, the secondary high penalty, and various squared penalties to obtain the total QUBO objective function, the data processing method further includes: receiving control parameters, wherein the control parameters include at least a budget cap, a capital occupation cap, a maturity exposure cap, and a group concentration cap defined by payer, supplier, group, business entity, or region; discretizing and binary slack variable encoding the budget, capital occupation, maturity exposure, and group concentration constraints according to the control parameters to generate the various squared penalties, wherein the various squared penalties include: a budget squared penalty, a capital occupation squared penalty, a maturity exposure squared penalty, and a group concentration squared penalty defined by the group exposure cap.

[0018] In some embodiments, the step of sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample to obtain the value of each decision variable includes: performing the chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample according to the budget ceiling, the capital occupation ceiling, the term exposure ceiling, and the group concentration ceiling to obtain the value of each decision variable.

[0019] In some embodiments, the step of structuring multiple original accounts receivable records to obtain a standardized object pool includes: performing entity parsing on the multiple original accounts receivable records for payers and suppliers based on an entity parsing strategy, identifying and marking records belonging to the same entity, and obtaining a set of entity parsed records. The entity parsing strategy is as follows: prioritizing the use of the unified social credit code to confirm whether payers and suppliers in different original accounts receivable records match; if the unified social credit code is missing or inconsistent, confirming whether suppliers and payers in different original accounts receivable records match based on similarity scores; the entity parsing... The data in the record set includes: payer unique identifier, supplier unique identifier, matching score, matching rule number, pending review mark, and field-level source mapping. Using the account receivable identifier as the primary key and the contract identifier, invoice identifier, and statement identifier as secondary keys, records with the same primary key or those that can be associated with the same primary key through the secondary keys are grouped together, resulting in a record set grouped by receivable object. Multiple records within each group correspond to different sources of the same accounts receivable. For each group in the record set, the following numerical field merging process is performed: extract the numerical fields of all records within the corresponding group and compare the differences... If the source values ​​are consistent, the corresponding value is directly adopted as the target value. If they are inconsistent, the target value is determined according to the source priority, record timestamp, and field confidence, and conflict markers and rules are recorded simultaneously to obtain the numerical field merging result of the corresponding group. For each group in the record set, the following status field merging process is performed: multiple status fields in each group are merged according to a preset status priority or target status rule to obtain the status field merging result of the corresponding group. The numerical field merging result corresponding to each group in the record set is combined with the status field merging result. The result is that a merged object record is generated for each group; the merged object record is converted into a normalized object in the normalized object pool according to the transformation strategy, wherein the transformation strategy includes: status standardization, calculation of actual risk exposure, and supplementation of missing fields; the normalized object is represented by at least the following information: object unique identifier, contract identifier, payer identifier, supplier identifier, group identifier, business carrier identifier, region identifier, account amount, prepayment ratio, actual risk exposure, term, term tier, risk weight, registration status, dispute status, duplicate financing status, field-level source mapping, and object-level confidence.

[0020] The embodiments of this application address the inconsistency in payer and supplier identities in multi-source data by prioritizing entity resolution based on the unified social credit code and supplementing it with name similarity matching, thus providing a reliable entity alignment foundation for subsequent concentration constraints and aggregation. A multi-level grouping strategy using account identifiers as primary keys and contracts / invoices as secondary keys accurately merges records of the same claim scattered across different systems, avoiding duplicate counting or omissions. Conflict detection and priority adjudication of numerical fields, and priority or strictest rule merging of status fields, ensure the accuracy and traceability of data fusion. Finally, a standardized object pool formed through actual risk exposure calculation, term tier division, and missing field supplementation provides high-quality and complete input data for subsequent candidate reduction, risk calculation, and constraint compilation, thereby reducing invalid variables and noise in the quantum optimization model and improving solution efficiency and engineering reliability of the results.

[0021] In some embodiments, the step of removing objects with invalid rights status from the normalized object pool based at least on the rights status field to obtain a list of retained objects includes: deleting objects from the normalized object pool whose rights status field content is irresolvable registration conflict, dispute freeze, or duplicate financing, and deleting objects whose contracts contain prohibition clauses on assignment and have no exemption.

[0022] The embodiments of this application automatically eliminate accounts receivable with irreparable defects in mandatory requirements based on the rights status field (prohibition of transfer, registration conflict, dispute, duplicate financing) before the optimization problem is constructed. This avoids introducing these invalid objects into the quantum optimization model, thereby reducing the number of decision variables and constraint complexity, and reducing the scale and secondary coupling density of the subsequent QUBO model. At the same time, by eliminating infeasible objects in advance, it effectively prevents the appearance of combinations with invalid mandatory requirements in the quantum sampling results, improves the executability and compliance of the final output combination, reduces the repair cost in the feasible solution recovery stage, and improves the overall engineering efficiency and reliability of the technical solution.

[0023] In some embodiments, generating a forbidden set and a secondary high penalty corresponding to each mutual exclusion relationship based on the mutual exclusion relationships in the reserved object list includes: traversing all object pairs in the reserved object list; if a pair of objects is confirmed to satisfy a preset mutual exclusion rule, then adding the corresponding object pair to the forbidden set and assigning a corresponding secondary high penalty coefficient to the corresponding object pair. The mutual exclusion rule includes: mutual exclusion rules related to rights conflicts or mutual exclusion rules related to compliance conflicts. The secondary high penalty in the target QUBO model is represented as: multiplying each pair of mutually exclusive objects in the forbidden set by a positive coefficient on the product term of the corresponding binary decision variable.

[0024] The embodiments of this application pre-compile the mutual exclusion relationships in the reserved object list, unifying rights conflicts (such as multiple mutually exclusive payments under the same contract) and compliance conflicts (such as the pre-compilation of the concentration limit of the same payer) into a forbidden set and generating corresponding secondary high penalty terms. This enables the quantum optimization model to avoid selecting mutually exclusive pairs simultaneously during sampling, thus eliminating the need to handle these conflicts additionally in subsequent constraint replay. At the same time, the secondary high penalty term, by setting a sufficiently large positive coefficient, ensures that the objective function value of any solution containing mutually exclusive pairs is strictly higher than that of all feasible solutions, fundamentally eliminating the possibility of business rule violations and improving the compliance of the combination and the reliability of the solution results.

[0025] In some embodiments, calculating the single comprehensive risk cost for each object in the list of retained objects includes: for the i-th object in the list of retained objects: determining the i-th expected credit loss ELi based on the i-th actual risk exposure Ei, the i-th probability of default PDi, and the i-th loss due to default LGDi; determining the i-th dilution risk DIi based on the i-th dilution risk coefficient δi and the i-th actual risk exposure Ei; determining the i-th fraud and operational risk FOi based on the i-th fraud and operational risk coefficient φi and the i-th actual risk exposure Ei; and determining the i-th mandatory requirement based on the i-th mandatory requirement compliance risk coefficient ψi and the i-th actual risk exposure Ei. Compliance risk LCI; Based on the term Ti of the i-th object and the funding matching term T^F, determine the i-th term mismatch factor μi and the i-th term mismatch risk; Based on the capital opportunity cost, capital adequacy ratio parameter CAR, the risk weight of the i-th object, and the i-th actual risk exposure Ei, determine the i-th capital occupation risk KCi; The i-th expected credit loss ELi, the i-th dilution risk DIi, the i-th fraud and operational risk FOi, the i-th mandatory compliance risk LCI, the i-th term mismatch risk, and the i-th capital occupation risk Kci are weighted and summed to obtain the single comprehensive risk cost of the i-th object.

[0026] The single-transaction comprehensive risk cost determination method in this application quantifies different dimensions of risk (credit, dilution, fraud, mandatory compliance requirements, maturity mismatch, capital occupation) in banking financing business into a unified cost value in monetary terms, and flexibly adapts to the risk preferences of different banks through configurable weighting coefficients. The mandatory compliance risk coefficient ψi dynamically depends on front-end rights status fields such as notification status, registration status, and dispute status, enabling risk pricing to reflect the effectiveness of mandatory requirements for accounts receivable in real time, thereby accurately distinguishing the quality differences of individual assets in the optimization model. This multi-dimensional weighted fusion method, compared to traditional assessment models that only consider credit risk, provides a more comprehensive cost basis that better reflects the bank's actual risk-bearing capacity for subsequent portfolio optimization, helping the quantum solver to select the portfolio with the truly lowest comprehensive risk for inclusion in the pool.

[0027] In some embodiments, generating sparse pairwise structural risk terms for each target object pair in the retained object list that does not belong to the forbidden set includes: weighting and summing the similarity or association degree of the p-th object and the q-th object on multiple preset association dimensions to obtain the structural association strength corresponding to the two object pairs, wherein the p-th object and the q-th object belong to the two objects in the retained object list, and the p-th object and the q-th object do not belong to the object pair in the forbidden set, and the multiple preset association dimensions refer to a predefined set of attributes or relationships used to measure the degree of structural association between two accounts receivable; if the structural association strength is greater than or equal to a preset sparsification threshold, then confirm that the two object pairs belong to the target object pairs and generate a secondary coupling edge; generating a pq-th sparse pairwise structural risk term corresponding to the two object pairs, wherein the sparse pairwise structural risk term is used to characterize the structural risk when the p-th object and the q-th object are selected simultaneously, and the structural risk term is used to characterize the combined risk arising from the structural association between the two accounts receivable corresponding to the p-th object and the q-th object.

[0028] The embodiments of this application generate secondary coupling edges and structural risk terms only for object pairs that do not belong to the forbidden set and whose structural correlation strength reaches a preset threshold, while ignoring object pairs with low correlation strength. This achieves a sparsity construction of the QUBO model, significantly reducing the number of quadratic terms and the model coupling density, making it adaptable to the limited physical bit connectivity and parameter range of current quantum processors. At the same time, the structural correlation strength integrates multiple business dimensions such as payer, supplier, group, region, proximity of maturity, and supply chain network correlation, which can accurately capture the combined risk increment caused by the common risk exposure of two accounts receivable. This allows the quantum optimization model to retain the ability to model major structural risks while reducing computational complexity, avoiding the loss of key risk correlations due to excessive sparsity.

[0029] In some embodiments, generating the pqth sparse pairwise structural risk term corresponding to the two object pairs includes: multiplying the structural association strength corresponding to the pth object and the qth object by the square root of the product of the actual risk exposure of the pth object and the actual risk exposure of the qth object to obtain a first result; dividing the single comprehensive risk cost of the pth object by the actual risk exposure of the pth object, and dividing the single comprehensive risk cost of the qth object by the actual risk exposure of the qth object to obtain two ratios, and then calculating the arithmetic mean of the two ratios; multiplying the first result by the arithmetic mean to obtain the pqth sparse pairwise structural risk term.

[0030] The pairwise structural risk terms in this application are calculated by multiplying the structural correlation strength, the geometric mean of risk exposures, and the arithmetic mean of the ratio of individual risk cost to risk exposure, thus achieving a unified monetary scale model for structural incremental risk and individual risk. The geometric mean of risk exposures reflects the scale effect of the portfolio, while the ratio of risk cost to exposure reflects the risk efficiency per unit exposure. Their product can reasonably quantify the additional portfolio risk arising from shared risk factors (such as same payer, same group, same region, etc.) when two assets are simultaneously included in the pool. This design maintains the additivity of the QUBO model while ensuring that the structural risk terms and the linear terms in the basic risk function can be directly superimposed. This allows quantum optimization to simultaneously weigh the quality of individual assets and the structural correlation risk between assets under the same objective function, avoiding the problems of neglecting portfolio risk or double-counting in traditional methods, and improving the risk control effect of the final pooled portfolio.

[0031] In some embodiments, the step of weighting and summing each individual comprehensive risk cost and each sparse pairwise structural risk term according to the decision variables to obtain the basic risk function includes: multiplying the individual comprehensive risk cost of each object by the corresponding object's binary decision variables and summing the results to obtain a linear part; multiplying each sparse pairwise structural risk term by the product of the binary decision variables of the corresponding two objects and summing the results to obtain a quadratic part; and adding the linear part and the quadratic part to obtain the basic risk function.

[0032] The underlying risk function in this application achieves a unified mathematical expression for the risk of individual asset quality and structural correlation between assets by superimposing the comprehensive risk cost of a single asset as a linear term (multiplied by the decision variable) and the sparse pairwise structural risk term as a quadratic term (multiplied by the product of the decision variables). The linear part ensures that the risk cost of each selected asset is accurately included, while the quadratic part calculates additional portfolio risk only for strongly correlated and non-mutually exclusive pairs of objects, avoiding the model over-density caused by full connectivity. This function serves as the superposition basis for various subsequent constraints and penalties (budget, capital, term, concentration, etc.), maintaining the additivity and convexity of the objective function. This allows quantum optimization to comprehensively weigh business benefits and compliance constraints within the same energy framework, ultimately selecting the pooled portfolio that meets both bank risk control requirements and minimizes overall risk.

[0033] In some embodiments, the step of superimposing the basic risk function, the secondary high penalty, and various squared penalties to obtain the total QUBO objective function includes: summing the basic risk function, the secondary high penalty, the budget squared penalty, the capital occupation squared penalty, the maturity exposure squared penalty, and the concentration squared penalties for each group defined by the group exposure cap to obtain the QUBO objective function.

[0034] In some embodiments, compiling the total QUBO objective function into a parameter configuration executable by the quantum processor includes: converting the linear and quadratic coefficients of the total QUBO objective function into local fields and coupling coefficients of the Ising model; normalizing the local fields and coupling coefficients according to the local field range and coupling range supported by the quantum processor to obtain normalized local fields and normalized coupling coefficients; mapping each logical variable to a set of physical bits according to the physical coupling graph of the quantum processor and setting the ferromagnetic coupling strength within the chain to form a physical bit chain configuration, wherein the physical bit chain configuration requires that the physical bit chains corresponding to different logical variables do not overlap and that there are connecting edges between chains corresponding to non-zero logical couplings; wherein the parameter configuration includes at least the normalized local fields, the normalized coupling coefficients, and the physical bit chain configuration, and the physical bit chain configuration includes physical bit chain mapping relationships and chain strengths.

[0035] Some embodiments of this application convert the linear and quadratic coefficients of the overall QUBO objective function into local field and coupling coefficients of the Ising model, and normalize them according to the local field and coupling ranges supported by the quantum processor. This ensures that the parameters can be correctly loaded by the hardware and do not exceed physical limitations. Based on the physical coupling graph, each logical variable is mapped to a set of physical bits, and the ferromagnetic coupling strength within the chain is set to form a physical bit chain configuration. At the same time, it is required that different chains do not overlap and that there are connecting edges between chains with non-zero logical coupling. This resolves the contradiction between the connection relationship between logical variables and the limited connectivity of the hardware, enabling the sparse QUBO model, which could not be embedded due to excessive coupling, to run stably on a real quantum annealing machine. This compilation step is a technical bridge connecting the abstract mathematical model and physical quantum hardware, significantly improving the engineering feasibility of quantum processors in combinatorial optimization problems.

[0036] In some embodiments, submitting the parameter configuration to the quantum processor for sampling to obtain a quantum sample includes: setting annealing parameters, wherein the annealing parameters include annealing time; and controlling the quantum processor to perform a quantum annealing sampling based on the annealing time and the parameter configuration to obtain a set of physical bit assignments as the quantum sample.

[0037] Some embodiments of this application, by configuring parameters such as annealing time, control the quantum processor to perform annealing sampling based on the quantum tunneling effect. This enables rapid exploration of low-energy states in complex combinatorial optimization problems, offering a potential speed advantage over classical algorithms (such as simulated annealing or branch-and-bound) when solving large-scale sparse QUBO models. Simultaneously, by obtaining a set of physical bit assignments through a single sampling, the probabilistic characteristics of the quantum annealing process are preserved, providing the original data foundation for subsequent multi-sample post-processing. This step connects the compiled mathematical model with the execution stage of the quantum hardware, achieving a technical closed loop from abstract optimization objectives to physical sampling, thus enabling the effective utilization of quantum computing resources.

[0038] In some embodiments, the process of sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum samples to obtain the value of each decision variable, and obtaining a data record set based on whether the object corresponding to the value is selected, includes: replaying budget, capital occupation, term exposure, group concentration, and rights status constraints for each candidate business variable combination; marking combinations that satisfy all constraints as original feasible solutions and marking combinations that violate at least one constraint as infeasible solutions; repairing each infeasible solution by iteratively flipping or removing violating variables to obtain repaired feasible solutions; merging the original feasible solutions and the repaired feasible solutions to form a feasible solution set; substituting the business variable values ​​of each feasible solution in the feasible solution set into the total QUBO objective function to calculate the objective function value; selecting the feasible solution with the smallest objective function value as the target feasible solution; marking the object corresponding to the decision variable with a value of 1 in the target feasible solution as selected, marking the object with a value of 0 as unselected, and outputting a data record set containing the selected marks.

[0039] The post-processing technique in this application verifies the feasibility of all candidate business variable combinations through constraint replay. Solutions that satisfy the constraints are marked as original feasible solutions, while infeasible solutions are repaired by iterative reversal or removal of non-compliant variables. The two types of feasible solutions are then merged to form a feasible solution set, thereby maximizing the retention of effective information in the quantum sampling results and significantly improving the utilization rate of quantum samples and the hit rate of feasible solutions. Each feasible solution is then substituted into the overall QUBO objective function to calculate the objective function value, and the solution with the smallest objective function value is selected as the final output, ensuring that the combination risk is minimized. This closed loop not only overcomes the unavailability problem caused by noise or embedded chain breaks in quantum sampling samples but also embeds the compliance of business constraints (budget, capital, timeframe, concentration, and rights status) into the recovery process. This ensures that the final output data record set not only meets all bank risk control requirements but also achieves optimality in an optimization sense, realizing a technical improvement from physical samples to executable business results.

[0040] Secondly, some embodiments of this application provide a computer program readable storage medium having a computer program stored thereon, which, when executed, can implement the method described in any one of the embodiments of the first aspect.

[0041] Thirdly, some embodiments of this application provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can implement the method as described in any one of the embodiments of the first aspect. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A quantum computer-based data processing system is provided for embodiments of this application.

[0044] Figure 2 This is a flowchart illustrating a quantum computer-based data processing method according to an embodiment of this application.

[0045] Figure 3 This application provides a schematic diagram of the electronic device composition. Detailed Implementation

[0046] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0047] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0048] Some embodiments of this application.

[0049] Please refer to Figure 1 , Figure 1 A processing system for performing a quantum computer-based data processing method is provided in some embodiments of this application. The hybrid processing includes an electronic computer 100 and a quantum annealing device 200.

[0050] In some embodiments of this application, the electronic computer 100 includes a preprocessing module 110, a compilation module 120, a conversion module 130, and a post-processing module 140, and these modules are used to perform the following operations:

[0051] The preprocessing module 110 performs the following data preprocessing operations: entity parsing (identifying the same payer or supplier), object merging (merging multiple sources of the same receivable) and status standardization (encoding notification status, registration status, dispute status, and duplicate financing status into a unified format) on the input raw accounts receivable records, generating a standardized object pool. This process utilizes the processor's parallel processing capabilities to complete string matching and numerical calculations. For example, in some embodiments of this application, the notification status indicates whether the bank has notified the payer of the transfer or pledge of accounts receivable, with values ​​and codes as follows: 0 = not notified, 1 = notified, 2 = notification rejected, 3 = notification pending confirmation. The registration status indicates whether the accounts receivable has been registered in the unified registration system for movable property financing, and whether there are any registration conflicts, with values ​​and codes as follows: 0 = not registered, 1 = registered and no conflict, 2 = prior registration exists (conflict cannot be resolved), 3 = registration information is incomplete. The dispute status indicates whether there is a dispute between the payer and the supplier that affects payment. Values ​​and codes are: 0 = No dispute, 1 = In dispute, 2 = In litigation or arbitration, 3 = Dispute resolved but payment not yet made. The duplicate financing status indicates whether the accounts receivable have been used for financing by other financial institutions. Values ​​and codes are: 0 = No duplicate financing, 1 = Suspected duplicate financing (requires verification), 2 = Duplicate financing confirmed, 3 = Duplicate financing resolved but registration not updated.

[0052] The compilation module 120 is used to perform the following constraint compilation and model building: In memory, it performs candidate reduction (removing objects with invalid rights states), conflict compilation (generating forbidden sets and secondary high penalties) on the normalized object pool, calculates individual risk costs, generates sparse pairwise risk terms (generating secondary terms only for strongly correlated object pairs), and discretizes and encodes binary slack variables for constraints such as budget, capital, time, and concentration, and finally superimposes them to form the overall QUBO objective function. This stage utilizes the processor of the electronic computer to accelerate large-scale matrix operations and sparsity processing, significantly reducing the model size and secondary coupling density.

[0053] The conversion module 130 is used to perform the following quantum parameter conversion and embedding: converting the QUBO objective function into the local field and coupling coefficients of the Ising model, normalizing them according to the local field range and coupling range supported by the quantum processor, mapping each logical variable to a set of physical bits (forming a physical bit chain) according to the physical coupling graph of the quantum processor, setting the chain strength and annealing parameters (annealing time, number of samplings, annealing schedule), and outputting parameter configurations that can be directly loaded by the quantum processor.

[0054] The post-processing module 140 is used to perform the following post-processing and result output: receiving physical bit samples (i.e. quantum samples) returned by the quantum processor 210 in the quantum annealing device 200, sequentially performing chain break repair (majority voting to merge inconsistent physical bits in the same chain), inverse embedding (mapping logical spin samples back to business variables), constraint replay (verifying constraints such as budget, capital, and concentration), and feasible solution recovery (iteratively flipping or eliminating illegal solutions), and substituting each feasible solution into the overall QUBO objective function to calculate the objective function value, selecting the optimal solution, and finally outputting a set of structured data records that can be directly loaded by the downstream automated decision-making system 300.

[0055] The quantum processor 210 (located in the quantum annealing device 200) is responsible for performing quantum sampling: loading the parameter configuration sent by the electronic computer 100 (normalized local field, coupling coefficient, physical bit chain mapping, chain strength, annealing time, number of samplings, etc.), and performing multiple annealing samples using the quantum tunneling effect, outputting a set of physical bit assignments (i.e., quantum samples) for each sample. These samples are returned to the electronic computer 100 for post-processing.

[0056] In some embodiments of this application, the automated decision-making system 300 automatically acquires a set of data records (each marked as selected or unselected) through a data receiving interface, eliminating the need for manual input. Through a built-in rule engine, it automatically generates specific execution actions or review tasks based on the business attributes of the selected records (such as amount, term, and interest rate) and system preset rules (e.g., a single transaction exceeding 1 million requires supervisor review). For cases where review is not required, it directly enters the execution process. A workflow orchestrator connects multiple actions into a complete workflow. For example, for selected assets, it automatically executes: contract generation, electronic signature, sending loan disbursement instructions, and updating ledgers, all without manual clicking or intervention. In other embodiments of this application, the automated decision-making system 300 interfaces with bank data processing systems, payment systems, and contract management systems through standardized interfaces (such as APIs and enterprise service buses) to achieve fully automated execution of operations such as contract generation, fund transfer, and accounting processing.

[0057] The structured data record set output by this application serves as the input to the automated decision-making system. Each record in this set contains fields such as: unique identifier, selection flag, risk exposure, term, interest rate, and contract template ID. The automated decision-making system reads records with the selection flag set to true, automatically triggering subsequent processing flows without requiring bank staff to review each record individually. This application, in conjunction with the automated decision-making system, forms a complete automated closed loop from raw data input to business action execution.

[0058] It should be noted that the embodiments of this application enable the quantum processor to sample quickly and stably by completing sparse compilation and hardware adaptation (normalization and embedding) in an electronic computer, avoiding problems such as long annealing time or embedding failure caused by excessively dense models or parameters exceeding limits. The closed-loop post-processing performed by the post-processing module in the electronic computer repairs unusable quantum samples into compliant and feasible output data, significantly improving sampling utilization and overall data processing speed, and finally outputting structured data that can be directly used by downstream systems.

[0059] The following is combined Figure 2 Explanation by Figure 1 Quantum-based data processing methods are performed using electronic computers and quantum annealing devices.

[0060] like Figure 2 As shown, some embodiments of this application provide a quantum computer-based data processing method, including:

[0061] S110: Multiple original accounts receivable records are structured to obtain a standardized object pool. Each object in the standardized object pool corresponds to an accounts receivable and includes the corresponding accounts receivable's attribute information, rights status field, and identifier field used for constraint compilation. For example, attribute information includes: account amount, prepayment ratio, actual risk exposure, term, term tier, risk weight, payer industry, payer region, supplier's group affiliation, and business vehicle identifier. Rights status fields include: notification status, registration status, dispute status, and duplicate financing status (standardized). Identifier fields used for constraint compilation include: payer identifier, supplier identifier, group identifier, business vehicle identifier, region identifier, contract identifier, and object unique identifier.

[0062] S120, at least based on the rights status field, remove objects with invalid rights status from the normalized object pool to obtain a list of retained objects. Examples include: objects prohibited from transfer and without exemption; objects with irresolvable prior transfer or pledge registration conflicts; objects under disputed freeze; and objects that trigger duplicate financing exclusion rules. These are all specific manifestations of invalid rights status.

[0063] S130, based on the mutual exclusion relationships in the reserved object list, generate a forbidden set and a secondary high penalty term corresponding to each mutual exclusion relationship. For example, in embodiments of this application, mutual exclusion relationships include conflicts of rights or compliance. Conflict of rights refers to mutual exclusion caused by mandatory requirements or rights status, such as the same account receivable being repeatedly financed, or multiple accounts under the same contract stipulating that only one can be selected. Compliance conflict refers to mutual exclusion caused by internal bank risk control strategies or regulatory requirements, such as if two accounts from the same payer are selected simultaneously, it will exceed the bank's concentration limit for that payer. In some embodiments of this application, a secondary high penalty term refers to: a quadratic coefficient set for a pair of mutually exclusive objects, the coefficient being large enough that the objective function value of simultaneously selecting both objects in the mutually exclusive pair is higher than the objective function value of any feasible solution that does not simultaneously select the mutually exclusive pair. For example, in some embodiments of this application, in the QUBO model, if the objects corresponding to two variables cannot be selected simultaneously, a large positive number (i.e., a penalty term) is added to their product term. Thus, when both variables are equal to 1, the overall objective function value becomes very large (i.e., very poor), and the optimization solver will automatically avoid choosing this combination.

[0064] S140, calculate the single comprehensive risk cost for each object in the list of retained objects.

[0065] S150, for each target object pair in the reserved object list that does not belong to the forbidden set, generate sparse pairwise structural risk terms, wherein the target object pair belongs to two objects that meet the structural association strength requirement.

[0066] S160, the basic risk function is obtained by weighted summation of each individual comprehensive risk cost and each sparse pairwise structural risk item based on the decision variables, wherein the decision variables are used to characterize whether the corresponding original record of accounts receivable is selected.

[0067] S170, at least based on the basic risk function and the quantum processor, determine the value of each decision variable, and obtain a set of data records based on whether the object corresponding to the value is selected.

[0068] In some embodiments of this application, S170 includes: superimposing the basic risk function, the quadratic high penalty term, and various squared penalty terms to obtain a total QUBO objective function, wherein the various squared penalty terms are used to convert business constraints into penalty terms in the QUBO model for quantum computer solution, and the various squared penalty terms are obtained by discretizing various constraints and encoding binary relaxation variables; compiling the total QUBO objective function into a parameter configuration executable by the quantum processor; submitting the parameter configuration to the quantum processor for sampling to obtain quantum samples; sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum samples to obtain the value of each decision variable, and obtaining a data record set based on whether the corresponding object is selected according to the value.

[0069] In some embodiments of this application, before the basic risk function, the secondary high penalty, and various squared penalties are superimposed to obtain the total QUBO objective function, the data processing method further includes: receiving control parameters, wherein the control parameters include at least a budget cap, a capital occupation cap, a maturity exposure cap, and a group concentration cap defined by payer, supplier, group, business entity, or region; discretizing and binary slack variable encoding the budget, capital occupation, maturity exposure, and group concentration constraints according to the control parameters to generate the various squared penalties, wherein the various squared penalties include: budget squared penalties, capital occupation squared penalties, maturity exposure squared penalties, and each group concentration squared penalty defined by the group exposure cap.

[0070] In some embodiments of this application, the step of sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample to obtain the value of each decision variable includes: based on the budget ceiling, capital occupation ceiling, time exposure ceiling, and group concentration ceiling in the control parameters, sequentially performing the chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample to obtain the value of each decision variable.

[0071] This application addresses the slow processing speed of quantum processors by employing candidate reduction and conflict compilation based on right states and mutual exclusion relationships. It eliminates unusable objects and pre-generates forbidden sets and secondary high-penalty terms. Simultaneously, it generates sparse pairwise risk terms only for strongly correlated object pairs, significantly reducing the variable size and secondary coupling density of the QUBO model. This reduces the annealing time and embedding failure rate required by the quantum processor, thereby improving data processing speed. To address the issue of unusable or unreliable results, this application uses closed-loop post-processing—including link break repair, inverse embedding, constraint replay, and feasible solution recovery—to stably transform unusable physical samples returned from quantum sampling into feasible combinations that satisfy all business constraints. This significantly improves the utilization rate of quantum samples and ensures that the output results can be directly used in downstream automated decision-making systems.

[0072] The following example illustrates... Figure 2 How each step is implemented.

[0073] It should be noted that, in some embodiments of this application, before executing S110, the quantum computer-based data processing method further includes: receiving candidate accounts receivable financing original records and control parameters. The candidate accounts receivable financing original records include at least fields for characterizing the creditor, transaction entity, amount ratio, rights status, and source information. The control parameters include at least the financing scale target value, budget band upper and lower bounds and / or budget upper limit, capital occupation original upper limit, total maturity exposure original upper limit, original exposure upper limit defined by group, and solution parameters of the target quantum processor.

[0074] It should be noted that, in some embodiments of this application, the control parameters are configured in an auditable manner to obtain the original upper limits of financing budget, capital occupation, maturity exposure, and group exposure; ΔE, ΔK, and ΔM represent the discretization units of amount, capital, and maturity exposure, respectively; λB, λK, λM, and λconc represent the squared penalty coefficients of the corresponding constraints. These parameters are obtained from the bank's risk management configuration, regulatory or internal risk weight mapping table, and target quantum processor configuration, and the field source, value range, default rules, and version number are recorded.

[0075] The business judgments involving mutual exclusion, concentration, limit, threshold, and candidate reduction involved in some embodiments of this application can be executed as a preset rule base and control parameters, and are not limited to a single fixed strategy of a certain bank. Each rule in the preset rule base includes at least a rule identifier, rule version, activation flag, scope of application, field set, judgment operator, threshold, reason code, and effective time. When compiling the prohibited set, concentration constraint, limit constraint, and sparsity threshold, the system reads the currently effective rule version and saves the rule version, hit reason code, and miss reason code in the output audit record. For example, it can be configured to enable any one or more of the following rules: concentration limit for the same payer, concentration limit for the same group, exclusion for co-selection of the same contract, or registration conflict exclusion.

[0076] In some embodiments of this application, the structured processing of multiple original accounts receivable records to obtain a standardized object pool as described in S110 includes: performing entity parsing of the multiple original accounts receivable records for payers and suppliers based on an entity parsing strategy, identifying and marking records belonging to the same entity, and obtaining an entity parsing record set. The entity parsing strategy is as follows: prioritizing the use of the unified social credit code to confirm whether the payers and suppliers in different original accounts receivable records match; if the unified social credit code is missing or inconsistent, confirming whether the suppliers and payers in different original accounts receivable records match based on a similarity score; the data packets in the entity parsing record set... This includes: unique payer identifier, unique supplier identifier, matching score, matching rule number, pending review flag, and field-level source mapping; using the account receivable identifier as the primary key and the contract identifier, invoice identifier, and statement identifier as secondary keys, records with the same primary key or those that can be associated with the same primary key through the secondary keys are grouped together to obtain a set of record groups after grouping by debt object, where multiple records in each group correspond to different sources of the same account receivable; for each group in the set of record groups, the following numerical field merging process is performed: extract the account receivable amount, prepayment ratio, term, and risk weight of all records in the corresponding group, and compare whether the values ​​of different sources are consistent. If they are consistent, directly... The corresponding value is used as the target value. If they are inconsistent, the target value is determined according to the source priority, record timestamp, and field confidence, and conflict markers and adoption rules are recorded simultaneously to obtain the numerical field merging result of the corresponding group. For each group in the record group set, the following status field merging process is performed: for notification status, registration status, dispute status, and duplicate financing status, merging is performed according to preset status priority or target status rules to obtain the status field merging result of the corresponding group. The numerical field merging result corresponding to each group in the record group set and the status field merging result are combined to generate a merged object record for each group. The conversion strategy is then applied to... The merged object records are converted into normalized objects in the normalized object pool. The conversion strategy includes: state standardization, calculation of actual risk exposure, and supplementation of missing fields (e.g., the conversion strategy includes mapping notification status, registration status, dispute status, and duplicate financing status to preset enumeration values ​​or 0 / 1 codes, calculating actual risk exposure Ei based on the account amount Ai and prepayment ratio ηi, and generating term tiers based on the term Ti). The payer's industry, region identifier loci, supplier's group affiliation gi, business carrier identifier pi, and risk weight RWi for subsequent constraint compilation are supplemented from entity parsing results, external enterprise relationship data, or internal bank configuration.Each object in the standardized AR object pool includes at least the following information: object unique identifier, contract identifier, payer identifier bi, supplier identifier si, group identifier gi, business carrier identifier pi, region identifier loci, account amount Ai, prepayment ratio ηi, actual risk exposure Ei, term Ti, term tier, risk weight RWi, standardized notification status ni, registration status regi, dispute status disi, duplicate financing status repi, field-level source mapping, and object-level confidence level. The standardized object is represented by at least the following information: object unique identifier, contract identifier, payer identifier, supplier identifier, group identifier, business carrier identifier, region identifier, account amount, prepayment ratio, actual risk exposure, term, term tier, risk weight, registration status, dispute status, duplicate financing status, field-level source mapping, and object-level confidence level.

[0077] In other words, in some embodiments of this application, by executing S110, entity parsing, merging of accounts receivable objects, and state standardization are performed on the candidate accounts receivable financing original records to generate a normalized object pool.

[0078] In some embodiments of this application, the step of removing objects with invalid rights status from the normalized object pool based at least on the rights status field to obtain a list of retained objects includes: deleting objects from the normalized object pool whose rights status field content is irresolvable registration conflict, dispute freeze, or duplicate financing, and deleting objects whose contracts contain prohibition clauses on assignment and have no exemption.

[0079] In some embodiments of this application, generating a forbidden set and a secondary high penalty corresponding to each mutually exclusive relationship based on the mutual exclusion relationships in the reserved object list includes: traversing all object pairs in the reserved object list; if a pair of objects is confirmed to satisfy a preset mutual exclusion rule, then adding the corresponding object pair to the forbidden set and assigning a corresponding secondary high penalty coefficient to the corresponding object pair. The mutual exclusion rule includes: mutual exclusion rules related to rights conflicts or mutual exclusion rules related to compliance conflicts. The secondary high penalty in the target QUBO model is represented as: for each pair of mutually exclusive objects (i,j) in the forbidden set, multiplying the product term of the binary decision variables xi and xj by a positive coefficient. For example, in some embodiments of this application, a mutual exclusion rule refers to a set of judgment conditions used to determine whether any two objects in the reserved object list cannot be selected simultaneously. Two objects are mutually exclusive when they meet any of the following conditions: Both objects have the same contract identifier, and bank policy stipulates that multiple accounts receivable under the same contract cannot be selected simultaneously; Both objects have the same payer identifier, and the sum of their risk exposures exceeds the concentration limit for that payer (pre-compiled mutual exclusion); Both objects have the same supplier identifier, and the sum of their risk exposures exceeds the total lending limit for that supplier (pre-compiled mutual exclusion); Both objects have the same group identifier, and the sum of their risk exposures exceeds the concentration limit for that group (pre-compiled mutual exclusion); Both objects have the same business vehicle identifier, and the sum of their risk exposures exceeds the limit for that business vehicle (pre-compiled mutual exclusion); Both objects have the same region identifier, and the sum of their risk exposures exceeds the concentration limit for that region (pre-compiled mutual exclusion). Of these mutually exclusive rules, the first category is a conflict of rights (mandatory requirement level), while the rest are compliance conflicts (bank policy level). In actual implementation, some of these rules can be selected and enabled based on the bank's specific risk strategy.

[0080] In some embodiments of this application, calculating the single comprehensive risk cost for each object in the list of retained objects includes: for the i-th object in the list of retained objects: determining the i-th expected credit loss ELi based on the i-th actual risk exposure Ei, the i-th probability of default PDi, and the i-th loss due to default LGDi; determining the i-th diluted risk DII based on the i-th diluted risk coefficient δi and the i-th actual risk exposure Ei; determining the i-th fraud and operational risk FOi based on the i-th fraud and operational risk coefficient φi and the i-th actual risk exposure Ei; and determining the i-th mandatory requirement based on the i-th mandatory requirement compliance risk coefficient ψi and the i-th actual risk exposure Ei. Compliance risk LCI; Based on the term Ti of the i-th object and the funding matching term T^F, determine the i-th term mismatch factor μi and the i-th term mismatch risk; Based on the capital opportunity cost, capital adequacy ratio parameter CAR, the risk weight RWi of the i-th object, and the i-th actual risk exposure Ei, determine the i-th capital occupation risk KCi; The i-th expected loss ELi, the i-th dilution risk DIi, the i-th fraud and operational risk FOi, the i-th mandatory requirement compliance risk LCI, the i-th credit term mismatch risk, and the i-th capital occupation risk Kci are weighted and summed to obtain the single comprehensive risk cost of the i-th object.

[0081] In some embodiments of this application, generating sparse pairwise structural risk terms for each target object pair in the reserved object list that does not belong to the forbidden set includes: weighted summing of the similarity or association degree of the p-th object and the q-th object across multiple preset association dimensions to obtain the structural association strength corresponding to the two object pairs, wherein the p-th object and the q-th object belong to the two objects in the reserved object list, and the p-th object and the q-th object do not belong to the object pair in the forbidden set, and the multiple preset association dimensions refer to a predefined set of dimensions used to measure the relationship between the two objects. The attributes or relationships of the structural association between accounts receivable; if the structural association strength is greater than or equal to a preset sparsity threshold, then the two object pairs are confirmed to belong to the target object pair and a secondary coupling edge is generated; a pq-th sparse pairwise structural risk term is generated corresponding to the two object pairs, wherein the pq-th sparse pairwise structural risk term is used to characterize the structural risk when the p-th object and the q-th object are selected at the same time, and the pq-th sparse pairwise structural risk term is used to characterize the combined risk arising from the structural association between the two accounts receivable corresponding to the p-th object and the q-th object.

[0082] For example, in some embodiments of this application, the similarity or correlation degree on multiple preset correlation dimensions refers to: on each preset dimension, a basic quantitative value obtained by comparing two objects is used to measure the degree of correlation between the two objects under that dimension. Similarity: Applicable to binary dimensions such as whether the payer is the same, whether the supplier is the same, etc., where the same is 1 and different is 0. Correlation degree: Applicable to continuous or network dimensions, such as the term difference which can be mapped to the [0,1] interval (the smaller the difference, the higher the correlation), and the supply chain network correlation degree is a value between [0,1]. Weighted summation is obtained by multiplying the basic quantitative values ​​of each dimension by the corresponding weights and then summing them to obtain the final structural correlation strength. In some embodiments of this application, multiple preset correlation dimensions refer to a predefined set of attributes or relationships used to measure the degree of structural correlation between two accounts receivable, such as whether the payer is the same, whether the supplier is the same, whether the group is the same, whether the region is the same, whether the term is close, whether there is an upstream or downstream relationship in the supply chain, etc.

[0083] In some embodiments of this application, generating the pqth sparse pairwise structural risk term corresponding to the two object pairs includes: multiplying the structural association strength corresponding to the pth object and the qth object by the square root of the product of the actual risk exposure of the pth object and the actual risk exposure of the qth object to obtain a first result; dividing the single comprehensive risk cost of the pth object by the actual risk exposure of the pth object, and dividing the single comprehensive risk cost of the qth object by the actual risk exposure of the qth object to obtain two ratios, and then calculating the arithmetic mean of the two ratios; multiplying the first result by the arithmetic mean to obtain the pqth sparse pairwise structural risk term.

[0084] In some embodiments of this application, the step of weighted summation of each individual comprehensive risk cost and each sparse pairwise structural risk term based on decision variables to obtain the basic risk function includes: multiplying the individual comprehensive risk cost of each object by the corresponding object's binary decision variable and summing the results to obtain a linear part; multiplying each sparse pairwise structural risk term by the product of the binary decision variables of the corresponding two objects and summing the results to obtain a quadratic part; and adding the linear part and the quadratic part to obtain the basic risk function.

[0085] In some embodiments of this application, the step of superimposing the basic risk function, the secondary high penalty, and various squared penalties to obtain the total QUBO objective function includes: summing the basic risk function, the secondary high penalty, the budget squared penalty, the capital occupation squared penalty, the maturity exposure squared penalty, and the concentration squared penalties for each group defined by the upper limit of group exposure to obtain the QUBO objective function.

[0086] For example, in some embodiments of this application, the method for determining the budget squared penalty includes: first, setting a budget discrete unit, dividing the risk exposure amount of each account receivable by this unit and rounding it to the nearest integer to obtain the budget occupancy integer for each account receivable; simultaneously, rounding the bank's total budget ceiling to the nearest integer using the same unit to obtain the budget ceiling integer. Then, using subset and dynamic programming methods, it is checked whether there exists a subset of accounts receivable such that the sum of their budget occupancy integers is exactly equal to the budget ceiling integer. If such a subset exists, an exact budget squared penalty is constructed: multiplying a penalty coefficient by the square of (the sum of the budget occupancy integers of all selected accounts receivable minus the budget ceiling integer). If such a subset does not exist, it automatically switches to budget band mode: first determining the maximum reachable lower bound integer and the minimum reachable upper bound integer, and then introducing two slack variables (each slack variable represented by a set of binary bits) to construct the upper bound squared penalty and the lower bound squared penalty. The upper bound penalty is calculated by multiplying the penalty coefficient by the square of (the sum of the integer budget occupancy of all selected accounts plus the upper bound slack variable minus the upper bound integer); the lower bound penalty is calculated by multiplying the penalty coefficient by the square of (the sum of the integer budget occupancy of all selected accounts minus the lower bound slack variable minus the lower bound integer). The method for determining the capital occupancy square penalty includes: setting a capital discrete unit. For each account receivable, first calculate its capital occupancy amount (i.e., capital adequacy ratio multiplied by risk weight and then multiplied by risk exposure), then divide by the capital discrete unit and round to the nearest integer to obtain the capital occupancy integer for each account; the bank's total capital occupancy limit is also rounded to the nearest integer using the same unit to obtain the capital limit integer. A binary slack variable (encoded as a weighted sum of binary bits) is introduced to construct the capital occupancy square penalty: the penalty coefficient multiplied by the square of (the sum of the integer capital occupancy of all selected accounts plus the slack variable minus the capital limit integer). The method for determining the maturity exposure square penalty includes: setting a maturity discrete unit. For each account receivable, first calculate its maturity exposure (i.e., remaining maturity multiplied by risk exposure), then divide by the maturity discrepancy unit and round to the nearest integer to obtain the integer maturity exposure for each account receivable; the bank's total maturity exposure cap is also rounded to the nearest integer to obtain the maturity cap. A binary slack variable is introduced to construct the maturity exposure squared penalty: the penalty coefficient is multiplied by the square of (the sum of the integer maturity exposures of all selected accounts receivable plus the slack variable minus the integer maturity cap). The method for determining the grouped concentration squared penalty includes: for each group dimension (e.g., payer, supplier, group, business entity, region), using the same discrepancy unit as the budget. For each group, divide the risk exposure amount of each account receivable within that group by the discrepancy unit and round to the nearest integer to obtain the integer budget allocation for each account receivable; the exposure cap for that group is also rounded to the nearest integer to obtain the group upper limit.Then, a binary slack variable is introduced for this group to construct the group concentration squared penalty: the penalty coefficient multiplied by the square of (the sum of the integer budget occupancy of all selected accounts within the group plus the slack variable and then minus the group upper limit integer). In some embodiments of this application, the hard upper limit class constraint is conservatively discretized.

[0087] In some embodiments of this application, compiling the total QUBO objective function into a parameter configuration executable by a quantum processor includes: converting the linear and quadratic coefficients of the total QUBO objective function into local fields and coupling coefficients of the Ising model; normalizing the local fields and coupling coefficients according to the local field range and coupling range supported by the quantum processor to obtain normalized local fields and normalized coupling coefficients; mapping each logical variable (i.e., each spin variable st obtained after the QUBO to Ising model transformation) to a set of physical bits according to the physical coupling graph of the quantum processor and setting the ferromagnetic coupling strength within the chain to form a physical bit chain configuration, wherein the physical bit chain configuration requires that the physical bit chains corresponding to different logical variables do not overlap and that there are connecting edges between chains corresponding to non-zero logical couplings; wherein the parameter configuration includes at least the normalized local fields, the normalized coupling coefficients, and the physical bit chain configuration, and the physical bit chain configuration includes physical bit chain mapping relationships and chain strengths.

[0088] In some embodiments of this application, submitting the parameter configuration to the quantum processor for sampling to obtain a quantum sample includes: setting annealing parameters, wherein the annealing parameters include annealing time; and controlling the quantum processor to perform a quantum annealing sampling according to the annealing time and the parameter configuration to obtain a set of physical bit assignments as the quantum sample.

[0089] In some embodiments of this application, the step of sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample to obtain the value of each decision variable, and finally outputting a set of marked data records based on whether the object corresponding to the value is selected, includes: replaying budget, capital occupation, term exposure, group concentration, and rights status constraints for each candidate business variable combination; marking combinations that satisfy all constraints as original feasible solutions and marking combinations that violate at least one constraint as infeasible solutions; repairing each infeasible solution by iteratively flipping or removing violating variables to obtain repaired feasible solutions; merging the original feasible solutions and the repaired feasible solutions to form a set of feasible solutions; substituting the business variable values ​​of each feasible solution in the set of feasible solutions into the total QUBO objective function to calculate the objective function value; selecting the feasible solution with the smallest objective function value as the target feasible solution; marking the object corresponding to the decision variable with a value of 1 in the target feasible solution as selected, marking the object with a value of 0 as unselected, and outputting a set of data records containing the selected marks.

[0090] The following will illustrate with specific examples. Figure 2 The implementation process of the relevant steps.

[0091] S1 performs entity parsing, account object merging, and status standardization on the candidate accounts receivable financing original records (i.e., accounts receivable original records). When a unified social credit code exists in the record, it is used as the primary key for entity parsing; when no unified social credit code exists, the payer entity and / or supplier entity identity are determined based on name normalization and fuzzy matching, and a matching score, a pending review flag, and entity parsing results are generated. Subsequently, based on account identifiers, contract identifiers, invoice identifiers, account statement identifiers, or combinations thereof, original records belonging to the same debt object are merged into normalized AR objects, and the notification status is updated. Registration status Disputed status and repeated financing status Perform standardized coding to generate object-level confidence and field-level source mapping.

[0092] In this step, bi, si, and gi represent the payer, supplier, and group or affiliate corresponding to the i-th normalized AR object, respectively; ni, regi, disi, and repi represent the notification status, registration status, dispute status, and duplicate financing verification result, respectively. The above status parameters come from the original records, registration or financing service platform, contract, and notification receipt, and are written into the source mapping along with the object-level confidence level.

[0093] S2, extract the account balance for each normalized object. Prepayment ratio Payer identification Supplier identification Group logo Business carrier identification Regional identifiers Notification status Registration status Disputed status Repeat financing status ,the term Risk weights In addition, risk parameters are defined, and a field-level source mapping is established. In this step, Ai represents the face value of the accounts receivable or the available financing balance of the i-th normalized AR object, ηi represents the prepayment ratio or credit discount rate, and Ei=ηiAi represents the actual risk exposure; pi, loci, Ti, and RWi represent the business vehicle, region, accounts receivable term, and risk weight, respectively. The units, enumeration scope, value range, and source fields of the above parameters are listed in the "Symbol, Parameter Source, and Value Configuration Table" at the end of the document.

[0094]

[0095] Calculate the first Actual risk exposure of a normalized object ,in, and Therefore In this invention This serves as a unified calculation method for budget occupancy, capital occupancy, and grouped exposure occupancy.

[0096] In one implementation, candidate reduction and rights conflict compilation are performed based on the standardized front-end state: Candidate reduction tags are generated for objects that meet the criteria of prohibition of transfer without exemption, existence of irresolvable prior transfer or pledge registration conflicts, being under disputed freeze status, or triggering duplicate financing exclusion rules, and these objects are removed from the subsequent QUBO compilation set; object pairs that, while not directly removed, have mutually exclusive rights conflicts or compliance conflicts are added to the prohibited set. And generate the corresponding conflict edges and quadratic high penalties:

[0097]

[0098] in, Indicates the conflicting objects The second-order penalty coefficient. The total number of objects retained after candidate reduction is denoted as... .

[0099] Furthermore, in the second instance of a high penalty and All are binary decision variables, taking values ​​of 0 or 1; the corresponding variable takes the value of 1 when the i-th or j-th retained object is selected into the financing portfolio, otherwise it takes the value of 0. Indices i and j traverse the set of retained objects after candidate reduction. ,in and only for those that satisfy The object generates the quadratic term; This is a prohibited set of rules, generated by rules prohibiting transfer, registration conflicts, disputed freezes, duplicate financing, or other compliance conflicts. Determined based on the lower bound rules of the penalty or the configuration file, this is used to ensure that the objective function value is higher than any feasible combination when both conflicting object pairs are selected simultaneously.

[0100] S3, calculate the comprehensive risk cost per transaction. For the first... For each retained item, first calculate the expected credit loss, dilution risk, fraud and operational risk, mandatory compliance risk, maturity mismatch risk, and capital occupation risk:

[0101]

[0102]

[0103]

[0104]

[0105] Among them, T F Indicates the funding matching period, κ cap c represents the regulatory or internal capital provision factor. cap This represents the capital opportunity cost coefficient. Further, based on weights ω1 to ω6, we obtain: In this step, PDi represents the probability of default by the payer or debtor, LGDi represents the loss due to default, δi represents the dispute or defect reduction factor, and φ... i Indicates the adjustment factor for the confirmation or notification status, ψ i This represents the adjustment factor for registration, pledging, or repeated financing verification; weights ω1 to ω6 correspond to expected credit loss, dilution risk, fraud and operational risk, mandatory compliance risk, maturity mismatch risk, and capital occupation risk, respectively. These risk parameters are generated from internal bank ratings, historical loss data, authorized external risk data, or rule mapping tables.

[0106] Specifically, It is obtained by mapping the payer's or debtor's internal rating, external credit score, delinquency migration rate or historical default sample, and the value ranges from 0 to 1; The value is determined based on guarantees, recourse rights, aging of accounts receivable, repayment history, collateral coverage, and historical loss recovery rate, and ranges from 0 to 1. This represents the expected credit loss of the i-th object. Derived from dispute status, defect type, and invoice or contract matching results, this is used to calculate dilution risk. ; It is derived from the status of confirmation of rights, effective notification, payment commitment or receipt, and is used to calculate fraud and operational risks. ; It is derived from the mapping of registration, pledge, duplicate financing verification, and mandatory requirement compliance status, and is used to calculate the mandatory requirement compliance risk. The mapping tables above all record the source field, enumeration criteria, default rule, confidence level, and effective version.

[0107]

[0108] In one implementation, the weight The coefficient is determined by historical loss data calibration, expert rule configuration, or a combination of both; among them, a compliance risk coefficient is a mandatory requirement. It is determined based on at least one of the following: the status of the prohibition on assignment clause, the notification status, the status of the assignment or pledge registration, the status of the dispute, and the status of repeated financing.

[0109] S4, after candidate reduction and right conflict compilation, for those not belonging to the forbidden set The structural association strength is calculated for object pairs, and sparse pairwise structural risk terms are generated only for object pairs that meet a threshold condition. For any two retained objects... and The structural correlation strength can be expressed as:

[0110] In this step, Γij represents the association strength between the i-th object and the j-th object, Cij represents the structural risk term generated when the two objects are added to the pool simultaneously, τ represents the sparsification threshold of the structural association strength, and ΔT represents the time proximity threshold. This indicates the strength of net relationship or trade network relationship. The system retains only object pairs that meet the configured thresholds and records the relationship source and reason code.

[0111]

[0112] in, For indicator functions, The threshold is the nearest time limit. This represents the network connectivity calculated from external supply chain relationship maps or transaction networks. Weights for related dimensions.

[0113] Based on the aforementioned structural correlation strength, sparse pairwise structural risk terms are generated according to the following formula:

[0114]

[0115] in, and All risks are uniformly expressed as risk costs in monetary terms and can be directly incorporated into the same objective function. If and only if... and At that time, a secondary coupling edge is generated for the corresponding asset pair. Structural risk items And the relationship reason code; when When, no corresponding quadratic term is generated, thus forming a sparsely coupled edge set; when In this case, the object pair is not treated as a normal structural associated edge, but its conflict edges and conflict high penalties are retained.

[0116] S5, using binary decision variables Indicates the first Whether a reserved object is included in the pool is determined by constructing a basic risk function based on the single comprehensive risk cost and sparse pairwise risk items:

[0117]

[0118] Among them, for those belonging to the forbidden set The object pair can make the corresponding And through the aforementioned high penalties for conflict

[0119] It separately indicates its compliance conflict constraints.

[0120] S6 performs discretization and reachability checks on the original budget target value. Let the budget discretization unit be... ,but:

[0121]

[0122] Here, Round represents the preset rounding rule used for the target budget point; for the budget upper limit or budget band boundary, the system adopts conservative discretization or replays the verification according to the original continuous amount during the recovery phase. Based on the discretization result, budget grid reachability detection is first performed; in one implementation, the reachable set Reach[t] can be represented by subsets and dynamic programming tables or bitmaps, and the existence of satisfying the requirements can be updated step by step. A feasible subset.

[0123] When the existence of satisfying is detected When feasible combinations are available, generate precise budget penalties:

[0124]

[0125] When the precise budget discrete point is unreachable, it automatically switches to budget band form. And recompile the budget with penalties. Specifically, the upper bound of the budget band can be written as:

[0126]

[0127] And through slack variables Compiled as:

[0128]

[0129] A budget with a lower bound can be written as:

[0130]

[0131] And through slack variables Compiled as:

[0132]

[0133] Therefore, the budget plus total penalty items are satisfied:

[0134]

[0135] in, and All use the same binary bit weight encoding as other inequality constraints.

[0136] S7 discretizes the capital occupancy constraint, maturity exposure constraint, and concentration constraint defined by the exposure cap per group. The capital occupancy constraint adopts the capital occupancy caliber. The discrete result is as follows:

[0137]

[0138] in, This indicates the original upper limit of capital occupation. The maturity exposure constraint uses the total maturity exposure as the criterion. The discrete result is as follows:

[0139]

[0140] in, This indicates the original upper limit of the total exposure over the specified period, rather than the average period threshold.

[0141] For the concentration constraint defined by the upper limit of exposure per group, let the first... The grouping function corresponding to the class grouping dimension is ,in, Payable by payee, supplier, group, business entity, or region; for any group identifier ,definition:

[0142]

[0143] The corresponding original exposure upper limit constraint is:

[0144]

[0145] After discretization, it can be written as:

[0146]

[0147] The aforementioned exposure cap for each group is used as a hard constraint on concentration; indicators such as the Herfindahl-Hirschman Index (HHI) for payers and the Herfindahl-Hirschman Index (HHI) for business entities can be used as output indicators.

[0148] S8 encodes the discretized inequality constraints using binary slack variables. For any inequality constraint... Introducing binary slack variables And order:

[0149] In this step, This represents the relaxation bit width of the m-th inequality constraint. This represents the discretized residual of the constraint. This represents the ℓth slack variable corresponding to the m-th constraint; the system automatically determines the slack bit width based on the constraint upper limit, discretization unit, and maximum candidate size, and writes the calculation process into the audit log.

[0150]

[0151] Thus, the squared penalty term is obtained:

[0152]

[0153] Corresponding to capital occupation constraints and maturity exposure constraints, we obtain:

[0154]

[0155]

[0156] The same method can be used to construct group penalties for any group exposure cap constraint:

[0157]

[0158] Among them, relaxation bit width and residual weight The remaining amount after the discrepancy should be covered.

[0159] S9, superimpose the basic risk function, conflict high penalty (i.e., quadratic high penalty), budget penalty, and various squared penalties to form a total bivariate quadratic optimization objective. Let all the summarized bivariate variables be denoted as... The overall goal can then be written as:

[0160]

[0161] Among them, only those that satisfy Upper triangular quadratic coupling term In implementation, linear terms, upper triangular quadratic terms, variable index mappings, relational reason codes, conflict reason codes, and constraint sources can be organized into interface-readable data structures.

[0162] In one implementation, the total penalty term resulting from various constraint violations is denoted as P(z), and the minimum positive violation amount is denoted for the m-th type of hard constraint. The minimum and maximum values ​​of the basic risk function in the same discrete variable space are denoted as follows: and Assuming the feasible region is non-empty, all penalty terms are non-negative, and comparisons are made using the same QUBO metric before normalization, the penalty term coefficients can satisfy the following conservative lower bound:

[0163]

[0164] This ensures that the objective value of infeasible or conflicting solutions is strictly higher than that of feasible solutions. Therefore, quantum QUBO solving is no longer simply about deriving interface parameters; it first completes the construction of a normalized AR object pool, candidate reduction, weight conflict compilation, budget reachability detection, relaxation encoding, and penalty calibration, before forming a stable sparse QUBO instance.

[0165] S10, convert the QUBO objective into Ising parameters and perform quantum embedding. Let:

[0166]

[0167] With only the upper triangular QUBO coefficients preserved, the Ising parameter satisfies:

[0168]

[0169] Furthermore, based on the upper limit of the absolute value of the local field supported by the target quantum processor... and the upper limit of the absolute value of coupling Set the normalization coefficient:

[0170]

[0171] When the corresponding denominator is 0, the corresponding ratio is not included in the minimum value calculation. The normalized parameters satisfy:

[0172]

[0173] Subsequently, based on the physical coupling diagram GH of the target quantum processor, the logical variable st is mapped to the physical bit chain Ct. It is required that the chains corresponding to different logical variables do not overlap, and the physical bit values ​​corresponding to the same logical variable are kept consistent through ferromagnetic coupling within the chain.

[0174] In one implementation, a chain term can be represented as:

[0175]

[0176] in, Represents the spin of a physical bit. Indicates chain strength.

[0177] Then set the annealing time. Number of samplings and / or annealing schedule The normalized Ising parameters and embedding configuration are then submitted to the target quantum processor to perform quantum sampling.

[0178] In this step, Hmax and Jmax represent the range of local fields and coupling coefficients supported by the target quantum processor, respectively; α represents the normalization ratio of the QUBO or Ising coefficients; G^H represents the hardware connection graph; Ct represents the physical bit chain corresponding to the logical variable; λchain represents the ferromagnetic coupling strength or chain strength within the chain; TA represents the annealing time; and R represents the number of samples. Before system submission, the normalization range, embedding configuration, chain strength value range, and anomaly handling rules are recorded.

[0179] S11: Obtain quantum sampling results and perform chain break repair, inverse embedding, and feasibility recovery. Let the quantum processor output R groups of physical samples σ^(r). First, perform majority voting, minimum energy repair, or a preset repair rule on each physical bit chain to obtain a logical spin sample s(r), then inversely map it to a binary sample z(r) and a business variable sample x(r). For each candidate sample, replay the budget, capital, deadline, group exposure limit, forbidden set F, and compliance constraints corresponding to notification status, registration status, dispute status, and repeated financing status. When a sample does not meet the business constraints, iteratively flip or remove variables according to the principles of maximum penalty contribution, minimum marginal benefit, or minimum sample frequency impact, prioritizing the resolution of variables hitting conflict edges or compliance status restrictions. If the repaired sample is still infeasible, continue performing a preset number of recovery replays. If it still cannot be recovered, mark the sample as an unrecoverable sample and record the failure reason code; it will not participate in the selection of the optimal feasible solution. Finally, the optimal feasible solution is selected based on the objective function value, constraint violation degree and / or sample frequency, and the following are output: pooled asset set, QUBO / Ising coefficient, quantum embedding parameter, annealing parameter, quantum sampling frequency, chain break repair record, constraint verification result, candidate reduction result, conflict reason code, and field-level source and confidence record.

[0180] The core innovation of this invention lies not in abstracting the bank portfolio problem into general QUBO interface parameters, but in establishing a quantum QUBO compilation and recovery closed loop oriented towards quantum processor executability: first, the original candidate records are compiled into a normalized AR object pool with rights state, compliance state, and source mapping; then, candidate reduction, conflict compilation, budget reachability detection, and penalty labeling are performed to form sparse QUBO instances that retain only valid quadratic slots; subsequently, QUBO / Ising parameter mapping, quantum embedding, and quantum sampling are completed, and the quantum samples are stably restored into executable financing portfolios through chain break repair, inverse embedding, constraint replay, and feasible solution recovery.

[0181] The following describes some technical solutions of this application with reference to several embodiments.

[0182] Example 1: The method flow of this example includes the following steps:

[0183] The first step is to receive the original records and control parameters of the candidate accounts receivable financing. Each original record should include at least... The statuses are: notification status, registration status, dispute status, and duplicate financing status.

[0184] The second step involves performing entity parsing, account object merging, and state standardization on the original records to generate a normalized AR object pool, entity primary keys, similarity scores, and matching confidence scores.

[0185] The third step involves executing candidate contract reductions based on the standardized notification, registration, dispute, and duplicate financing statuses, generating candidate contract reduction tags and a prohibited set. Conflict edges and / or high penalties.

[0186] The fourth step is to calculate the actual risk exposure for each retained object. and the overall risk cost per transaction .

[0187] Fifth step: Calculate the values ​​of objects that do not belong to the forbidden set. Only for Objects generate structural risk items Construct a sparse logic graph using relational reason codes.

[0188] Step 6: Construct the basic risk function of the combination. and high penalties for conflict .

[0189] Step 7: Use discretization units calculate and And perform budget reachability testing using bitmaps or dynamic programming.

[0190] Step 8: If the exact budget point is reachable, generate an exact budget penalty; if it is not reachable, switch to budget with constraints and rebuild the upper and lower bound penalties.

[0191] The ninth step involves discretizing and encoding binary slack variables for the capital occupation constraints, maturity exposure constraints, and exposure cap constraints for each group.

[0192] Step 10: Summarize all coefficients, assemble QUBO parameters according to the upper triangular caliber, and complete the penalty calibration in conjunction with the penalty lower limit rule.

[0193] Step 11: Map the QUBO parameters to the Ising parameters, normalize the coefficients, and complete the quantum embedding and annealing parameter settings according to the quantum hardware topology.

[0194] Step 12: Perform quantum sampling to obtain candidate spin samples and complete chain break repair, reverse embedding, compliance constraint replay and feasibility recovery.

[0195] Step 13: Output the set of assets to be added to the pool, candidate reduction results, reason codes for not being added to the pool, reason codes for conflicts, quantum sampling frequency, number of assets added to the pool in the dominant cluster, and field-level audit records.

[0196] Example 2: The following exemplarily illustrates the process of quantum QUBO solving compilation and quantum embedding provided in some embodiments of this application.

[0197] In this embodiment, the constraint compilation step first expands the conflict high penalty items formed based on the notification status, registration status, dispute status and duplicate financing status, as well as the budget penalty items, capital occupation penalty items, term exposure penalty items and group exposure upper limit penalty items into constant terms, linear terms and quadratic terms respectively, and then summarizes them into upper triangular coefficients.

[0198] In one implementation, the contribution of the precise budget penalty to the QUBO coefficient after expansion can be expressed as:

[0199]

[0200] For other squared penalties, they can also be expanded and mapped to constant terms, linear terms, and upper triangular quadratic terms. Then, the minimum positive violation for each penalty is estimated, and based on... The penalty coefficient is conservatively calibrated.

[0201] In obtaining Then, according to Perform Ising mapping to obtain Then, based on the local field range and coupling range supported by the quantum processor, the normalization coefficients are calculated. and output Next, the logical variables are embedded into the physical coupling graph of the quantum processor, a physical bit chain is assigned to each logical variable, and the chain strength is set. Annealing time Number of samplings Together with the annealing schedule parameters, a quantum QUBO solution configuration is formed.

[0202] Example 4: In some embodiments of this application, after obtaining the quantum sampling results through quantum sample post-processing and result output steps, in addition to the pooled asset set, the following artifacts are also output:

[0203] The results of constraint satisfaction include the verification results of budget constraints, capital occupation constraints, maturity exposure constraints, group exposure cap constraints, and compliance constraints corresponding to notification status, registration status, dispute status, and duplicate financing status.

[0204] Relationship reason codes and conflict reason codes are used to identify the specific reasons that trigger secondary coupling edges or prohibited edges, such as same payer, same supplier, same group, same business vehicle, same region, adjacent terms, network association, notification conflict, registration conflict, dispute conflict, and duplicate financing conflict.

[0205] Reason codes for not entering the pool are used to identify budget overruns, capital overruns, maturity exposure overruns, group exposure overruns, notification status restrictions, registration status conflicts, dispute status restrictions, duplicate financing hits, excessive marginal risk increments, or failure of feasibility recovery after chain break repairs.

[0206] Concentration metrics include the Herfindahl-Hirschman Index (HHI) for payers, the Herfindahl-Hirschman Index (HHI) for business entities, and the number of assets in the dominant concentration cluster.

[0207] The quantum sampling post-processing records include sample frequency, chain break repair log, inverse embedding results, embedding parameters, and annealing parameters.

[0208] Field-level source and confidence records. For derived fields. Each outputs its source field set and the confidence level after propagation.

[0209] Candidate reduction and object merging records are used to identify which original records are merged into normalized AR objects, which objects are removed due to front-end state rules, and the corresponding removal reasons.

[0210] In one embodiment of this application, chain-level majority voting repair is first performed on each quantum sample, and then the repaired sample is replayed with budget, capital, deadline, group exposure limit, and notification / registration / dispute / duplicate financing constraints. If the sample is still infeasible, the relevant variables are removed first according to the maximum penalty contribution, and objects that hit conflict edges or compliance status restrictions are eliminated first until a feasible solution is restored, and the repair path is recorded in the audit log.

[0211] Example 5: A small numerical example is provided below.

[0212] This embodiment provides a numerical example of four candidate assets to illustrate the complete chain of processing from a normalized AR object pool to a quantum processor.

[0213] The original records of the candidate accounts receivable are as follows:

[0214] Asset 1: A1=100, η1=0.8, E1=80; Payer b1=P1, Supplier s1=S1, Group g1=G1, Business Vehicle p1=V1, Region loc1=L1, Term T1=90, Risk Weight RW1=0.5; Notification status is effective, Registration status has no prior conflict, Dispute status has no dispute, Duplicate financing status is not hit.

[0215] Asset 2: A2=120, η2=0.7, E2=84; Payer b2=P1, Supplier s2=S2, Group g2=G1, Business Vehicle p2=V2, Region loc2=L1, Term T2=120, Risk Weight RW2=0.8; Notification status is effective, Registration status has no prior conflict, Dispute status has no dispute, Duplicate financing status is not hit.

[0216] Asset 3: A3=80, η3=0.85, E3=68; Payer b3=P2, Supplier s3=S3, Group g3=G2, Business Vehicle p3=V1, Region loc3=L2, Term T3=60, Risk Weight RW3=0.4; Notification status is effective, Registration status has no prior conflict, Dispute status has no dispute, Duplicate financing status is not hit.

[0217] Asset 4: A4=60, η4=0.75, E4=45; Payer b4=P3, Supplier s4=S4, Group g4=G3, Business Vehicle p4=V3, Region loc4=L3, Term T4=150, Risk Weight RW4=0.6; Notification status is effective, Registration status is no prior conflict, Dispute status is observation level dispute, Duplicate financing status is not hit.

[0218] During the construction phase of the normalized AR object pool (i.e., the normalized object pool), if there is a record in the original accounts receivable record that has the same contract identifier and account identifier as asset 2 and is repeatedly uploaded by different access channels, then the duplicate record will be merged into the normalized AR object corresponding to asset 2 and will not form a separate optimization variable; if there is an object that hits the rules of irresolvable prior registration conflict, dispute freeze or duplicate financing exclusion, then it will be directly removed in the candidate reduction phase and will not enter the subsequent quantum QUBO compilation.

[0219] The comprehensive risk cost for a single transaction is calculated based on preset risk parameters and weights:

[0220]

[0221] In this embodiment, none of the four normalized objects matched the front-end exclusion rule, so after candidate reduction, the four optimized variables remain, forming a forbidden set. Empty. Set the structural association threshold. Asset 1 and Asset 2 have the same payer, the same group, and the same region, therefore they meet the requirements. Generate coupling edges Structural risk items and the relationship reason codes RC_SAME_PAYER, RC_SAME_GROUP, and RC_SAME_REGION. Other assets...

[0222] If , then no corresponding quadratic term will be generated.

[0223] Set the original value of the budget target Budget Discrete Unit ,but After performing reachability testing on the discrete mesh, it was found that... Since there are no combinations in the reachable set whose exact sum is 165, the constraint automatically switches to budget band constraints. An example budget band value formula is shown below:

[0224] .

[0225] Assuming the payer group exposes the original upper limit, the payer... The upper limit is Then the combination Although the budget usage is However, its payer Group exposure as Violation of group exposure limit; combination The budget usage is It is located within the budget zone and meets the payer group exposure limit.

[0226] In this embodiment, after mapping the QUBO coefficients to the Ising parameter, the chain strength λchain, annealing time TA, and sampling number R are set for the target quantum processor. Among the candidate samples from simulated sampling or returned by the target quantum processor, after chain break repair, inverse embedding, and feasibility recovery, the candidate feasible combinations with high sample frequency and satisfying the constraints include {2,3}. The system then determines the target feasible solution based on the objective function value, constraint violation degree, and sample frequency.

[0227] The following results were generated based on this:

[0228] Candidate reduction and object merging records: The original records of Asset 2 that were uploaded repeatedly were merged into the same normalized AR object without creating additional optimization variables.

[0229] Reason codes for not being included in the pool: Asset 1 can be marked as RC_GROUP_EXPOSURE, and Asset 4 can be marked as RC_TERM_OR_MARGINAL_RISK or RC_DISPUTE_STATUS.

[0230] Number of dominant clustered assets in the pool: Since the selected asset 2 and asset 3 do not satisfy Γ23≥τ, the maximum number of connected component nodes is 1.

[0231] Constraints satisfied: Budget band constraint, payer group exposure constraint, capital occupation constraint, term exposure constraint, and notification, registration, dispute, and duplicate financing constraints are all satisfied.

[0232] Quantum sampling post-processing records include the frequency of feasible samples corresponding to the combination {2,3}, the number of chain break repairs, the variable vector after inverse embedding, and the replay results of compliance constraints.

[0233] Regarding the propagation of field-level source and confidence levels, if the confidence levels of the account balance field and the prepayment ratio field of Asset 2 are respectively... and Then, under the minimum value propagation rule:

[0234]

[0235] If further based on and calculate Then record it in the result workpiece. The corresponding source field set and the propagated data .

[0236] Table 1. Symbol, Parameter Sources, and Value Configuration Table

[0237] Some embodiments of this application provide a computer program readable storage medium having a computer program stored thereon, which, when executed, can perform the methods described in any of the above embodiments.

[0238] like Figure 3 As shown, some embodiments of this application provide an electronic device 400, which includes a memory 410, a processor 420, and a computer program stored in the memory 410 and executable on the processor 420, wherein the processor 420 executes the program to implement the method described in any of the above embodiments.

[0239] Processor 420 can process digital signals and may include various computing architectures. For example, it may be a complex instruction set computer architecture, a reduced instruction set computer architecture, or an architecture that implements multiple instruction set combinations. In some examples, processor 420 may be a microprocessor.

[0240] Memory 410 can be used to store instructions executed by processor 420 or data related to the execution of instructions. These instructions and / or data may include code used to implement some or all of the functions of one or more modules described in the embodiments of this application. The processor 420 of the embodiments of this disclosure can be used to execute the instructions in memory 410 to implement… Figure 2 The method shown. Memory 410 includes dynamic random access memory, static random access memory, flash memory, optical memory, or other memory well known to those skilled in the art.

[0241] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0242] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0243] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A data processing method based on a quantum computer, characterized in that, The data processing method includes: A standardized object pool is obtained by structuring multiple original accounts receivable records. Each object in the standardized object pool corresponds to an accounts receivable and contains the attribute information, rights status field, and identification field for constraint compilation of the corresponding accounts receivable. At least objects with invalid rights status are removed from the normalized object pool based on the rights status field to obtain a list of retained objects; Based on the mutual exclusion relationships in the reserved object list, generate a set of prohibited matches and a secondary high penalty term corresponding to each mutual exclusion relationship; Calculate the single comprehensive risk cost for each object in the list of retained objects; For each pair of target objects in the reserved object list that does not belong to the forbidden set, a sparse pairwise structural risk term is generated, wherein the pair of target objects belongs to two objects that meet the structural association strength requirement; The basic risk function is obtained by weighted summation of each individual comprehensive risk cost and each sparse pairwise structured risk item based on the decision variables. The decision variables are used to characterize whether the corresponding original record of accounts receivable is selected. The value of each decision variable is determined at least based on the basic risk function and the quantum processor, and a set of data records is obtained based on whether the object corresponding to the value is selected.

2. The data processing method as described in claim 1, characterized in that, The process of determining the value of each decision variable based at least on the basic risk function and the quantum processor, and obtaining a set of data records based on whether the object corresponding to the value is selected, includes: The total QUBO objective function is obtained by superimposing the basic risk function, the second-order high penalty term, and various squared penalty terms. The various squared penalty terms are used to convert business constraints into penalty terms in the QUBO model. The various squared penalty terms are obtained by discretizing various constraints and encoding them with binary slack variables. The total QUBO objective function is compiled into a parameter configuration executable by the quantum processor; The parameter configuration is submitted to the quantum processor for sampling to obtain quantum samples; The quantum samples are sequentially subjected to chain break repair, inverse embedding, constraint replay and feasible solution recovery to obtain the value of each decision variable, and the data record set is obtained according to whether the object corresponding to the value is selected.

3. The data processing method as described in claim 1, characterized in that, Before the data processing method involves superimposing the basic risk function, the quadratic high penalty term, and various squared penalties to obtain the overall QUBO objective function, it further includes: Receive control parameters, wherein the control parameters include at least budget caps, capital occupancy caps, maturity exposure caps, and group concentration caps defined by payer, supplier, group, business entity, or region; Based on the control parameters, the budget, capital occupation, maturity exposure, and group concentration constraints are discretized and encoded using binary slack variables to generate the various types of squared penalties. These various types of squared penalties include: budget squared penalty, capital occupation squared penalty, maturity exposure squared penalty, and group concentration squared penalties defined according to the upper limit of group exposure.

4. The data processing method as described in claim 3, characterized in that, The process of sequentially performing chain break repair, inverse embedding, constraint replay, and feasible solution recovery on the quantum sample yields the value of each decision variable, including: Based on the budget ceiling, the capital occupation ceiling, the time exposure ceiling, and the group concentration ceiling, the chain break repair, the inverse embedding, the constraint replay, and the feasible solution recovery are sequentially performed on the quantum sample to obtain the value of each decision variable.

5. The data processing method as described in claim 1, characterized in that, The process of structuring multiple original accounts receivable records to obtain a standardized object pool includes: Based on the entity parsing strategy, the multiple original accounts receivable records are parsed for the payer and supplier, and records belonging to the same entity are identified and marked to obtain an entity parsing record set. The entity parsing strategy is as follows: first, the payer and supplier in different original accounts receivable records are confirmed to match based on the unified social credit code; if the unified social credit code is missing or inconsistent, the supplier and payer in different original accounts receivable records are confirmed to match based on the similarity score. The data in the entity parsing record set includes: unique payer identifier, unique supplier identifier, matching score, matching rule number, pending review mark, and field-level source mapping. Using the account receivable identifier as the primary key and the contract identifier, invoice identifier, and statement identifier as auxiliary keys, records with the same primary key or that can be associated with the same primary key through the auxiliary keys are grouped together to obtain a set of record groups after being grouped by the creditor object. Among them, multiple records in each group correspond to different sources of the same account receivable. For each group in the record set, the following numerical field merging process is performed: extract the numerical fields of all records in the corresponding group, compare whether the values ​​from different sources are consistent, if they are consistent, directly use the corresponding value as the target value, if they are inconsistent, determine the target value according to the source priority, record timestamp and field confidence, and record the conflict mark and the adopted rule at the same time to obtain the numerical field merging result of the corresponding group. For each group in the record set, the following status field merging process is performed: multiple status fields in each group are merged according to a preset status priority or target status rule to obtain the status field merging result of the corresponding group; Combine the merged results of the numerical fields corresponding to each group in the record set with the merged results of the status fields to generate a merged object record for each group; The merged object records are converted into normalized objects in the normalized object pool according to the conversion strategy, wherein the conversion strategy includes: status standardization, calculation of actual risk exposure, and supplementation of missing fields; the normalized objects are represented by at least the following information: object unique identifier, contract identifier, payer identifier, supplier identifier, group identifier, business carrier identifier, region identifier, account amount, prepayment ratio, actual risk exposure, term, term tier, risk weight, registration status, dispute status, duplicate financing status, field-level source mapping, and object-level confidence level.

6. The data processing method as described in claim 1, characterized in that, The step of removing objects with invalid rights status from the normalized object pool based at least on the rights status field to obtain a list of retained objects includes: Remove objects from the standardized object pool whose rights status field contains irresolvable registration conflicts, disputed freezes, or duplicate financing, and remove objects whose contracts contain prohibition clauses on assignment and have no exemptions.

7. The data processing method as described in claim 6, characterized in that, The step of generating a set of prohibited matches and a secondary high penalty term corresponding to each mutual exclusion relationship based on the mutual exclusion relationships in the reserved object list includes: Traverse all object pairs in the reserved object list. If a pair of objects is confirmed to satisfy a preset mutual exclusion rule, add the corresponding object pair to the forbidden set and assign a corresponding quadratic high penalty coefficient to the corresponding object pair. The mutual exclusion rule includes: mutual exclusion rules related to rights conflicts or mutual exclusion rules related to compliance conflicts. The quadratic high penalty is expressed in the target QUBO model as: for each pair of mutually exclusive objects in the forbidden set, multiply the product term of the corresponding binary decision variable by a positive coefficient.

8. The data processing method according to any one of claims 1-7, characterized in that, The calculation of the single comprehensive risk cost for each object in the list of retained objects includes: For the i-th object in the list of reserved objects: The expected credit loss ELi is determined based on the actual risk exposure Ei, the probability of default PDi, and the loss rate of default LGDi. The dilution risk DIi of the i-th dilution risk coefficient δi and the i-th actual risk exposure Ei are determined; The fraud and operational risk FOi is determined based on the i-th fraud and operational risk coefficient φi and the i-th actual risk exposure Ei. The compliance risk LCi of the i-th mandatory requirement is determined based on the compliance risk coefficient ψi of the i-th mandatory requirement and the actual risk exposure Ei of the i-th mandatory requirement. Based on the term Ti of the i-th object and the funding matching term T^F, determine the term mismatch factor μi and the term mismatch risk of the i-th term; The capital occupation risk KCi of the i-th object is determined based on the capital opportunity cost, the capital adequacy ratio parameter CAR, the risk weight of the i-th object, and the actual risk exposure Ei of the i-th object. The single comprehensive risk cost of the i-th object is obtained by weighting and summing the i-th expected credit loss ELi, the i-th dilution risk DII, the i-th fraud and operational risk FOi, the i-th mandatory requirement compliance risk LCI, the i-th maturity mismatch risk, and the i-th capital occupation risk Kci.

9. The data processing method as described in claim 8, characterized in that, The step of generating sparse pairwise risk terms for each pair of target objects in the reserved object list that does not belong to the forbidden set includes: The similarity or association degree of the p-th object and the q-th object on multiple preset association dimensions are weighted and summed to obtain the structural association strength corresponding to the two object pairs. The p-th object and the q-th object belong to the two objects in the reserved object list, and the p-th object and the q-th object do not belong to the object pairs in the forbidden set. The multiple preset association dimensions refer to a predefined set of attributes or relationships used to measure the degree of structural association between two accounts receivable. If the structural association strength is greater than or equal to the preset sparsification threshold, then the two object pairs are confirmed to belong to the target object pair and a secondary coupling edge is generated. Generate a pq-th sparse pairwise structural risk term corresponding to the two object pairs, wherein the pq-th sparse pairwise structural risk term is used to characterize the structural risk when the p-th object and the q-th object are selected at the same time; in, The generation of the pq-th sparse pairwise structure risk term corresponding to the two object pairs includes: The first result is obtained by multiplying the structural association strength corresponding to the p-th object and the q-th object by the square root of the product of the actual risk exposure of the p-th object and the actual risk exposure of the q-th object; Divide the single comprehensive risk cost of the p-th object by the actual risk exposure of the p-th object, and divide the single comprehensive risk cost of the q-th object by the actual risk exposure of the q-th object to obtain two ratios, and then calculate the arithmetic mean of the two ratios; Multiplying the first result by the arithmetic mean yields the pq-th sparse pairwise structure risk term.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the program, it can implement the method as described in any one of claims 1-9.

Images