Method, medium, and apparatus for random selection of compliant suppliers under multiple constraints

Abstract

The application discloses a method, medium and equipment for random extraction of a compliance supplier under multiple constraints, the method comprising: in response to an extraction instruction, obtaining project and sub-project identification; screening suppliers from a database based on pre-configured extraction conditions, constructing a knowledge graph and predicting an associated risk probability using a graph neural network, and generating a compliance list after eliminating high-risk suppliers; aggregating the compliance list according to parent companies to generate a candidate list; constructing a digital twin model, loading the candidate list, extraction algorithm and verification rule to a simulation engine to perform multiple simulation extractions, updating the knowledge graph state and recalculating the risk probability after each simulation, and recording the result distribution; based on the result distribution, evaluating an expected effect indicator, if the threshold is met, performing formal extraction, otherwise adjusting the extraction parameters to simulate again until the threshold is met. The application realizes dynamic prediction of associated risks through a knowledge graph and a graph neural network, significantly improving the compliance and fairness of supplier extraction.

Description

Technical Field

[0001] This application relates to the fields of data processing and intelligent decision-making technology in government procurement or supply chain management.

[0002] Specifically, this relates to a method, medium, and equipment for randomly selecting compliant suppliers under multiple constraints. Background Technology

[0003] Government procurement is an important part of public finance expenditure. The process of randomly selecting a number of qualified suppliers for price inquiry or bidding is crucial. The fairness, compliance and traceability of the selection process are directly related to the effectiveness of public funds and social credibility.

[0004] Existing supplier selection methods primarily rely on manual or semi-manual approaches, where purchasing departments perform coarse-grained industry matching based on their own maintained, fragmented lists, followed by random selection. These methods have revealed the following prominent problems in practical operation:

[0005] First, the selection criteria are too simplistic, usually using only the industry category as a single matching dimension. There is a lack of ability to combine and screen multiple constraints such as supplier operating status, associated risks, and historical performance, which makes it difficult to guarantee compliance of the selection results.

[0006] Second, the selection mechanism lacks dynamic risk assessment methods and cannot identify potential risks that may be transmitted between suppliers due to equity, personnel and other related relationships. There is a risk that suppliers who are "compliant on the surface but high in reality" will be selected.

[0007] Third, the formulation of extraction strategies relies on the experience of the personnel in charge and lacks quantitative evaluation methods, making it impossible to achieve comprehensive optimization across multiple target dimensions such as supplier diversity, geographical coverage balance, and risk control.

[0008] Therefore, how to achieve compliant, fair, and optimal random selection of suppliers under multiple constraints is a pressing technical problem that needs to be solved. Summary of the Invention

[0009] In view of the above problems, this application provides a technical solution for the random selection of compliant suppliers under multiple constraints, in order to solve the technical problem that the existing supplier random selection method lacks a dynamic prediction and multi-objective simulation optimization mechanism for supplier-related risks, which makes it difficult to achieve the comprehensive optimality in terms of compliance, fairness and risk control of the selection results.

[0010] To achieve the above objectives, in a first aspect, this application provides a method for randomly selecting compliant suppliers under multiple constraints, the method comprising:

[0011] S1: In response to the extraction start command for the target procurement project, obtain the project identifier of the target procurement project and the sub-project identifier of the currently pending sub-project;

[0012] S2: Based on the pre-configured extraction conditions associated with the sub-project identifiers, select supplier identifiers that meet the multi-dimensional constraints from the supplier database and generate a qualified list; wherein, the multi-dimensional constraints include industry classification matching, supplier status validity, and preset supplier attribute condition matching;

[0013] In the supplier status validity verification, a knowledge graph containing supplier nodes, legal person nodes, shareholder nodes, senior executive nodes and affiliated company nodes is constructed, and a pre-trained graph neural network model is used to predict the risk transmission of the knowledge graph, outputting the dynamic association risk transmission probability of each supplier node. If the dynamic association risk transmission probability exceeds a preset risk threshold, the supplier status is determined to be invalid and it is removed from the qualified list.

[0014] S3: Aggregate the supplier identifiers in the qualified list according to their parent companies to generate a parent company candidate list, wherein multiple supplier identifiers under the same parent company correspond to one candidate qualification in the parent company candidate list;

[0015] S4: Construct a digital twin model of the current sub-project, specifically including: loading the parent company candidate list, the preset random sampling algorithm and the real-time status verification rules into the simulation engine. The simulation engine is used to perform the following simulated sampling steps: randomly select N parent companies from the parent company candidate list, then randomly select a supplier identifier from the supplier identifiers that meet the requirements of the selected parent company, and call the simulated real-time status verification.

[0016] S5: Perform multiple simulated extractions based on the digital twin model. Each simulated extraction calls the real-time state verification of the simulation and records the result distribution of each simulated extraction. After each simulated extraction, the simulation engine updates the state information of the relevant nodes in the knowledge graph according to the result of this simulated extraction, and re-calls the graph neural network model to calculate the dynamic association risk transmission probability for the weight adjustment of the next simulated extraction.

[0017] S6: Based on the distribution of the results, evaluate the expected performance indicators of the current extraction strategy. The expected performance indicators include supplier diversity index, regional coverage balance, weighted historical performance risk probability, and risk diversity balance index.

[0018] S7: If the expected effect indicator meets the preset indicator threshold, then perform formal extraction according to the current extraction strategy to generate the final extraction list of the sub-projects; if the expected effect indicator does not meet the preset indicator threshold, then adjust the extraction parameters in the current extraction strategy and return to step S5 to re-perform the simulated extraction until the preset indicator threshold is met; wherein, the extraction parameters include the number of parent companies N extracted in the simulated extraction step, the weight adjustment coefficient of the dynamic correlation risk transmission probability, and the target threshold of the risk diversity balance index;

[0019] S8: Determine whether there are any unprocessed sub-projects in the target procurement project. If so, return to step S1 to process the next sub-project. If not, end the extraction process.

[0020] Furthermore, in step S2, the step of using a pre-trained graph neural network model to predict risk transmission in the knowledge graph specifically includes:

[0021] S21: Encode the attribute information of each supplier node, legal person node, shareholder node, senior executive node, and affiliated company node in the knowledge graph into an initial feature vector;

[0022] S22: Input the initial feature vector into the pre-trained graph attention network (GAT) model. The GAT model calculates the attention coefficient between each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation of the features of the neighboring nodes based on the attention coefficient to obtain the updated feature vector of each node.

[0023] S23: Input the updated feature vector of each supplier node into the risk prediction classifier and output the dynamic correlation risk transmission probability of that supplier node;

[0024] S24: If the probability of dynamic association risk transmission exceeds the preset risk threshold, the supplier status is determined to be invalid and the supplier is removed from the qualified list.

[0025] Furthermore, step S2, which involves constructing a knowledge graph including supplier nodes, legal entity nodes, shareholder nodes, executive nodes, and affiliated company nodes, also includes the construction of regional nodes and regional association edges, specifically including:

[0026] In the knowledge graph, a regional node is constructed for each supplier node's registered region and operating region, including provincial regional nodes and municipal regional nodes;

[0027] For each supplier node, a regional association edge is constructed with its registered regional node and operating regional node, and a regional association strength weight is assigned to each regional association edge. The regional association strength weight is proportional to the proportion of the supplier's business in that region.

[0028] To construct regional adjacency edges between different city-level regional nodes belonging to the same provincial-level regional node, and to assign an adjacency attenuation coefficient to each regional adjacency edge. The adjacent attenuation coefficient It is inversely proportional to the geographical distance between the two municipal-level regional nodes;

[0029] When calculating the attention coefficient in the GAT model, the regional association strength weight of the regional association edge and the adjacent decay coefficient of the regional adjacent edge are used as adjustment factors for the attention coefficient, so that the risk transmission probability between supplier nodes in the same region or adjacent regions is suppressed, and the suppression magnitude is proportional to the product of the regional association strength weight and the adjacent decay coefficient.

[0030] Furthermore, in step S5, the simulation engine updates the state information of relevant nodes in the knowledge graph based on the results of this simulation extraction, including:

[0031] S41: The simulation engine executes the simulated extraction step to generate a set of simulated extraction results;

[0032] S42: The simulation engine calls the real-time state verification of the simulation, which includes: identifying the extracted supplier node and its associated nodes according to the set of simulation extraction results, wherein the associated nodes include legal person nodes, shareholder nodes, senior executive nodes and related company nodes associated with the extracted supplier node, and marking the state of the extracted supplier node and its associated nodes as extracted in the knowledge graph.

[0033] The step of recalculating the dynamic association risk transmission probability by re-invoking the graph neural network model specifically includes:

[0034] Based on the labeled knowledge graph, the dynamic association risk transmission probability of each unextracted supplier node is calculated; among them, the dynamic association risk transmission probability of supplier nodes that have direct or indirect association with the extracted suppliers is attenuated and adjusted according to the length of the association path and the association strength.

[0035] Furthermore, in step S5, after each simulated extraction, an interpretability output step for the extraction results is also included, specifically including:

[0036] After each simulated extraction, the simulation engine extracts the risk transmission path of the extracted supplier node and its associated nodes from the knowledge graph based on the results of this simulated extraction. The risk transmission path includes the shortest path from the extracted supplier node, through legal person nodes, shareholder nodes, senior executive nodes or related company nodes, to other unextracted supplier nodes.

[0037] For each risk transmission path, calculate the path risk contribution value, which is the product of the association strength weights of all edges on the path and the reciprocal of the path length.

[0038] The risk transmission paths are sorted from largest to smallest according to their risk contribution value, and the top K risk transmission paths are selected as key risk transmission paths, where K is a preset positive integer;

[0039] Generate an interpretability report, which includes: the identifier of the supplier selected in this simulation sampling result, the node sequence of the key risk transmission path, the path risk contribution value of each key risk transmission path, and the contribution ratio of the path risk contribution value to the probability of dynamic associated risk transmission.

[0040] The interpretability report is associated with and stored in conjunction with the results of this simulation extraction, and then output to the user interface for visualization.

[0041] Furthermore, in step S5, the step of re-invoking the graph neural network model to calculate the dynamic correlation risk transmission probability further includes the following steps:

[0042] In the knowledge graph, competitive relationship edges are constructed between parent company nodes that have historical bidding relationships, and each competitive relationship edge is assigned a competitive intensity weight, which is proportional to the number of historical biddings between the two parent companies.

[0043] After each simulated sampling, identify the parent company node to which the sampled supplier belongs and record it as the sampled parent company node;

[0044] Traverse all unextracted supplier nodes in the knowledge graph. For each unextracted supplier node, obtain its parent company node, which is denoted as a candidate parent company node.

[0045] If there is a competitive relationship between the candidate parent node and any extracted parent node, the probability of dynamic association risk transmission of the supplier nodes under the candidate parent node is adjusted upward according to the competition intensity weight of the competitive relationship edge, and the adjustment range is proportional to the competition intensity weight.

[0046] The increased probability of dynamic correlation risk transmission will be used as the basis for adjusting the weight of the supplier node in the next simulation extraction.

[0047] Furthermore, in step S2, the construction of the knowledge graph, which includes supplier nodes, legal entity nodes, shareholder nodes, executive nodes, and affiliated company nodes, also includes a supplier node importance weighting step, specifically including:

[0048] In the knowledge graph, a node importance index is calculated for each supplier node. This node importance index includes a first-dimensional index and a second-dimensional index. The first-dimensional index is the supply chain irreplaceability index, which is calculated based on the number of exclusive supply records, core component supply records, and alternative suppliers for that supplier node in historical projects. The calculation formula is as follows:

[0049] Indispensability Index = ×(Number of exclusive supply records / Total supply records)+ ×(Number of core component supply records / Total supply records)+ ×(1 / number of alternative suppliers), where, , , For the preset weighting coefficients, and =1;

[0050] The second dimension indicator is the network centrality index, which is calculated by weighting the supplier node's degree centrality, betweenness centrality, and proximity centrality in the knowledge graph. The calculation formula is as follows:

[0051] Centrality index = × degree centrality+ ×between centrality+ × is close to centrality, where , , For the preset weighting coefficients, and ;

[0052] The supply chain irreplaceability index and the network centrality index are weighted and fused to obtain the comprehensive node importance value of the supplier node. The calculation formula is as follows:

[0053] Overall node importance value = ×Irreplaceability Index+ × Centrality index, where , For the preset weighting coefficients, and ;

[0054] In step S5, when the graph neural network model is called again to calculate the probability of dynamic correlation risk transmission, the comprehensive value of node importance is used as an adjustment factor for the attention coefficient of the GAT model, so that the supplier node with the higher the comprehensive value of node importance, the greater the range of risk transmission influence on other nodes after it is extracted, and the adjustment range of the probability of dynamic correlation risk transmission is proportional to the comprehensive value of node importance.

[0055] Furthermore, in step S7, adjusting the extraction parameters in the current extraction strategy specifically includes:

[0056] When the risk diversity balance index is lower than a preset balance threshold:

[0057] If the supplier diversity index is lower than the preset diversity threshold, then the number N of parent companies extracted in the simulation extraction step is increased.

[0058] If the weighted historical performance risk probability is higher than the preset risk threshold, then the weight adjustment coefficient of the dynamic associated risk transmission probability is reduced.

[0059] Wherein, the preset balance threshold, preset diversity threshold, and preset risk threshold are all preset constants.

[0060] In a third aspect, this application provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when executed by the processor, the computer program implements the random selection method for compliant suppliers under multiple constraints as described in the first aspect of this application.

[0061] Unlike existing technologies, the above-mentioned technical solution involves a method, medium, and equipment for randomly selecting compliant suppliers under multiple constraints. The method includes: responding to a selection command and obtaining project and sub-project identifiers; filtering suppliers from a database based on pre-configured selection conditions, constructing a knowledge graph and using a graph neural network to predict the probability of associated risks, and generating a qualified list after removing high-risk suppliers; aggregating the qualified list by parent company to generate a candidate list; constructing a digital twin model, loading the candidate list, selection algorithm, and verification rules into a simulation engine to perform multiple simulated selections, updating the knowledge graph state and recalculating the risk probability after each simulation, and recording the result distribution; evaluating the expected performance indicators based on the result distribution, and if the threshold is met, performing formal selection; otherwise, adjusting the selection parameters and resimulating until the target is met. This application achieves dynamic prediction of associated risks through knowledge graphs and graph neural networks, and combines digital twins for multi-objective simulation optimization, significantly improving the compliance, fairness, and overall effectiveness of supplier selection.

[0062] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0063] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of this application and other related content, and should not be considered as limitations on this application.

[0064] In the accompanying drawings of the instruction manual:

[0065] Figure 1 A flowchart illustrating the random selection method for compliant suppliers under multiple constraints as described in the first exemplary embodiment of this application;

[0066] Figure 2 This is a schematic diagram of an electronic device according to an exemplary embodiment of this application;

[0067] The reference numerals used in the above figures are explained as follows:

[0068] 10. Electronic equipment; 101. Processor; 102. Storage medium. Detailed Implementation

[0069] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that this application can achieve, the following detailed description is provided in conjunction with the specific embodiments listed and the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this application, and are therefore only examples and should not be used to limit the scope of protection of this application.

[0070] In the first aspect, such as Figure 1 As shown, this application provides a method for randomly selecting compliant suppliers under multiple constraints, the method comprising:

[0071] S1: In response to the extraction start command for the target procurement project, obtain the project identifier of the target procurement project and the sub-project identifier of the currently pending sub-project;

[0072] S2: Based on the pre-configured extraction conditions associated with the sub-project identifiers, select supplier identifiers that meet the multi-dimensional constraints from the supplier database and generate a qualified list; wherein, the multi-dimensional constraints include industry classification matching, supplier status validity, and preset supplier attribute condition matching;

[0073] In the supplier status validity verification, a knowledge graph containing supplier nodes, legal person nodes, shareholder nodes, senior executive nodes and affiliated company nodes is constructed, and a pre-trained graph neural network model is used to predict the risk transmission of the knowledge graph, outputting the dynamic association risk transmission probability of each supplier node. If the dynamic association risk transmission probability exceeds a preset risk threshold, the supplier status is determined to be invalid and it is removed from the qualified list.

[0074] S3: Aggregate the supplier identifiers in the qualified list according to their parent companies to generate a parent company candidate list, wherein multiple supplier identifiers under the same parent company correspond to one candidate qualification in the parent company candidate list;

[0075] S4: Construct a digital twin model of the current sub-project, specifically including: loading the parent company candidate list, the preset random sampling algorithm and the real-time status verification rules into the simulation engine. The simulation engine is used to perform the following simulated sampling steps: randomly select N parent companies from the parent company candidate list, then randomly select a supplier identifier from the supplier identifiers that meet the requirements of the selected parent company, and call the simulated real-time status verification.

[0076] S5: Perform multiple simulated extractions based on the digital twin model. Each simulated extraction calls the real-time state verification of the simulation and records the result distribution of each simulated extraction. After each simulated extraction, the simulation engine updates the state information of the relevant nodes in the knowledge graph according to the result of this simulated extraction, and re-calls the graph neural network model to calculate the dynamic association risk transmission probability for the weight adjustment of the next simulated extraction.

[0077] S6: Based on the result distribution, evaluate the expected performance indicators of the current extraction strategy, including supplier diversity index, regional coverage balance, weighted historical performance risk probability, and risk diversity balance index.

[0078] S7: If the expected effect indicator meets the preset indicator threshold, then perform formal extraction according to the current extraction strategy to generate the final extraction list of the sub-projects; if the expected effect indicator does not meet the preset indicator threshold, then adjust the extraction parameters in the current extraction strategy and return to step S5 to re-perform the simulated extraction until the preset indicator threshold is met; wherein, the extraction parameters include the number of parent companies N extracted in the simulated extraction step, the weight adjustment coefficient of the dynamic correlation risk transmission probability, and the target threshold of the risk diversity balance index;

[0079] S8: Determine whether there are any unprocessed sub-projects in the target procurement project. If so, return to step S1 to process the next sub-project. If not, end the extraction process.

[0080] In this embodiment, the target procurement project refers to a complete procurement business document initiated by the procuring entity. A single project can be broken down into multiple business items, corresponding to several sub-projects. Each sub-item has independent industry, qualification, and regional extraction constraints.

[0081] Sub-project identifiers are unique codes for each business item under a project, used to bind the specific extraction rules, industry classifications, qualification thresholds, exclusion lists of suppliers, and other constraints for that item.

[0082] Pre-configured extraction conditions refer to the set of filtering rules stored in the system in advance as metadata. These rules include multi-dimensional constraints such as industry classification codes, supplier operating status, place of origin, enterprise size, special qualifications, and project-specific exclusion lists. Rules can be added or adjusted without modifying the underlying code.

[0083] The supplier database refers to a unified central supplier resource database. The data source is synchronized with the business information of the tax system. Suppliers from all regions are uniformly entered into the database, eliminating the problem of data fragmentation across departments.

[0084] Multi-dimensional constraints refer to a three-tiered progressive screening rule. The first tier is industry classification matching, which relies on standardized tax industry codes to achieve precise matching of subdivided industries. The second tier is supplier status verification, which includes verification of business status, blacklist, dishonesty, and associated risks. The third tier is preset supplier attribute conditions, which include registered region, business scale, special qualifications, and local / out-of-town location attributes.

[0085] Knowledge graphs are structured networks with suppliers, legal entities, shareholders, senior executives, and affiliated companies as entity nodes, and business relationships such as equity, employment, cooperation, region, and competition as related edges. They are used to uncover the risks of implicit relationships within enterprises.

[0086] Graph neural network models refer to pre-trained graph attention network (GAT) deep learning models that automatically calculate the probability of risk transmission between enterprises based on the features of graph nodes and associated edges.

[0087] The probability of dynamic associated risk transmission is a quantitative indicator that represents the likelihood that a supplier may be linked to procurement dishonesty or operational abnormalities due to related relationships such as legal persons, shareholders, equity, and cooperative enterprises.

[0088] Parent company aggregation refers to using the parent company of a corporate group as the aggregation dimension. Multiple compliant businesses under the same parent company occupy only one candidate quota, thus avoiding a single group monopolizing the quota.

[0089] A digital twin model is a virtual simulation model built for supplier extraction business. It maps the real candidate pool, extraction algorithm, and compliance verification rules to the simulation engine to realize a full-process virtual simulation before the formal extraction.

[0090] The simulation engine refers to the computing module that carries the simulation logic of digital twins, which can perform multiple rounds of simulation extraction, real-time verification, graph status update, and indicator statistics in batches.

[0091] The random selection algorithm refers to a uniform and unbiased random shuffling algorithm, which ensures that the probability of selecting candidate objects is equal. The random seed is persistently stored to meet the audit reproducibility requirements.

[0092] Real-time status verification rules refer to the four-layer verification logic of the tax system, which includes real-time verification of four items: business opening status, tax arrears and liabilities, normal operation, and not yet closed. The interface communication failure has a three-time exponential backoff retry mechanism.

[0093] The expected results indicators refer to four types of quantitative evaluation indicators, namely supplier diversity index, regional coverage balance, weighted historical performance risk probability, and risk diversity balance index, which are used to comprehensively evaluate the rationality of the sampling plan.

[0094] Extraction parameters refer to adjustable variables used for simulation iterative optimization, including the number of parent companies N extracted in a single run, the risk transmission probability weight adjustment coefficient, and the target threshold of the risk diversity balance index.

[0095] The working principle of this embodiment is as follows:

[0096] In step S1, after the extraction command is triggered on the consultation program operation page, the system verifies that the project's pre-extraction status is pending extraction, generates a globally unique extraction serial number to be stored in the audit log, synchronously reads the global identifier of the current procurement project, traverses all sub-projects under the project, and locks the identifier of the first sub-project that has not been extracted as the object of this processing.

[0097] In step S2, the system retrieves the pre-configured filtering rules based on the sub-project identifier, traverses all normally operating businesses in the central supplier database, and matches industry, status, and attribute constraints layer by layer. In the status verification stage, a multi-level entity association knowledge graph is built, and all node features of the graph are input into a pre-trained graph neural network to automatically calculate the dynamic association risk transmission probability of each supplier. Suppliers whose dynamic association risk transmission probability exceeds the preset risk threshold are judged as invalid and directly removed from the preliminary compliance list, thus blocking the hidden association procurement risk from the source.

[0098] Step S2 describes a two-level storage reuse architecture for the pre-configured extraction conditions, enabling flexible rule configuration and project isolation. The specific architecture logic is as follows: The system pre-builds a global pre-configured condition template, storing general multi-dimensional constraint rules such as industry classification, qualification thresholds, regional constraints, and blacklist filtering. This allows for direct reuse by multiple different target procurement projects. When any target procurement project is bound to a corresponding pre-configured extraction condition, the system automatically generates a condition snapshot specific to that sub-project and stores it independently. The snapshot completely preserves all constraint parameters of the current version. Subsequent adjustments to the global pre-configured condition template by maintenance personnel, such as adding, modifying, or deleting, only apply to newly created procurement projects. Historical projects with existing condition snapshots and projects currently executing extraction processes are unaffected by template changes, achieving complete isolation of extraction condition logic between different projects. This eliminates the need to repeatedly configure the same business rules, reducing rule maintenance costs.

[0099] In step S3, all businesses in the preliminary compliance list are matched with their parent company entities. Multiple compliance business records under the same parent company are merged. Each parent company retains only a single candidate qualification, generating a parent company candidate list without duplicate groups. This mechanism ensures that small and medium-sized suppliers have an equal chance of being selected.

[0100] After generating the candidate parent company list in step S3, the system executes the branch determination logic based on the number of candidate parent companies, setting differentiated business processing paths for different numbers of candidates:

[0101] In the first scenario, if the number of candidate entities in the parent company candidate list is less than 1, the pre-screening of this sub-project is determined to have failed, and the process will automatically jump to the review process for cancellation of the failed extraction. All cancellation operations and failure reason records must be reviewed and confirmed by the administrator before being stored in the audit log.

[0102] In the second scenario, if the number of parent company candidate entities is greater than or equal to 1 but less than 3, the system opens an external supplier supplementation channel. Business operators can manually enter the identifiers of compliant external suppliers that are not stored in the supplier database to expand the parent company's candidate list. If the operator does not supplement the candidate entities to 3 but still confirms to continue the selection, the system will forcibly open the reason entry module. The written reasons filled in by the operator will be permanently stored and bound to the selection transaction record as the basis for audit verification.

[0103] In the third scenario, if the number of parent company candidate entities is greater than or equal to 3, no external supplementation is required. The process can directly proceed to step S4 to build the digital twin model of the current sub-project and perform simulated extraction.

[0104] In step S4, the parent company candidate list, fixed random sampling algorithm, and tax real-time verification rules are uniformly loaded into the simulation engine. The simulation engine replicates the real sampling logic: first, N parent companies are randomly selected from the parent company candidate pool, then a single business name is randomly selected from the compliant business names under each parent company, and the simulated real-time verification is run to completely replicate the entire chain of formal sampling.

[0105] In step S5, the simulation engine repeatedly performs simulated extraction and retains the complete results of each round. After each round of simulated extraction is completed, the status of the extracted suppliers and related legal persons, shareholders, executives and related companies nodes in the knowledge graph is automatically updated and marked as extracted. Then, the graph neural network model is called again to recalculate the dynamic association risk transmission probability of the remaining suppliers based on the updated graph. The new calculation results are used as the basis for weight adjustment in the next round of simulated extraction to dynamically correct the extraction priority of candidate suppliers.

[0106] In step S6, based on the distribution of multi-round simulation results, four comprehensive evaluation indicators are calculated. The advantages and disadvantages of the current extraction strategy are quantitatively evaluated from four dimensions: supplier category richness, regional distribution balance, historical performance risk average, and the degree of balance between risk and diversity, so as to form a standardized quantitative evaluation result.

[0107] In step S7, the four indicators—supplier diversity index, geographical coverage balance, weighted historical performance risk probability, and risk diversity balance index—are compared with their corresponding preset thresholds. If all indicators meet the standards, the current extraction strategy is directly reused to execute the formal extraction, and the final extraction list for the current sub-project is output. If any indicator fails to meet the preset threshold, the corresponding extraction parameters are modified, and the simulation is restarted in step S5, iterating until all indicators meet the standards. During the formal extraction process, random seeds, screening logs, tax verification records, and replacement operation records are simultaneously retained to form tamper-proof audit data. After the formal extraction is completed, a tiered replacement strategy is implemented for unqualified suppliers: suppliers are first re-extracted within the corresponding parent company; if all suppliers in the parent company fail verification, a new parent company is re-extracted from the parent company's candidate pool to complete the replacement.

[0108] In step S7, when the final list of sub-projects is generated through formal extraction according to the compliance extraction strategy, a real-time status verification process of the actual tax system is executed simultaneously. This differs from the simulated real-time verification within the simulation engine; the actual verification interfaces with external tax business interfaces, and the verification dimensions include four aspects: supplier's business operation status, debt status, operating status, and closure status. If communication fails at the interface, the system automatically performs three retry operations, with the retry interval using an exponential backoff rule, set to 1 second, 2 seconds, and 4 seconds respectively.

[0109] If a selected supplier identifier is determined to be non-compliant through real tax verification, the system executes a tiered replacement strategy: The first tier is a priority replacement strategy, which randomly selects one supplier identifier from all compliant supplier identifiers of the parent company to which the supplier identifier belongs, and initiates real-time tax verification again; if all supplier identifiers under the parent company fail verification, the second tier is a fallback replacement strategy, which randomly selects one parent company from the complete list of parent company candidates, and then selects a single supplier from the supplier identifiers under that parent company and repeats the tax verification until a compliant supplier is obtained.

[0110] In step S8, after the current sub-item is extracted, the system checks whether there are any unprocessed sub-items under the project. If there are, it returns to step S1 to process the next sub-item. If all sub-items are extracted, the overall status of the project is changed. If all sub-items are successfully extracted, they are transferred to the pending registration status. If any sub-item is extracted but failed, it is transferred to the pending supplementation status. The system simultaneously judges the project budget amount. If the budget exceeds the preset amount, the consultation announcement release process is automatically triggered.

[0111] After extracting all sub-projects for a single project, the system determines the overall project status based on the extraction results of each sub-project. Simultaneously, it performs standardized fault semantic classification on all sub-projects that failed extraction, categorizing fault types into three types: system communication failure, insufficient number of compliant supplier identifiers, and no matching supplier for the industry category. Each type of fault is assigned a unique fault code and written to the audit log. If any sub-project within a project fails to be extracted, the entire project transitions to a dedicated "pending supplementation" state. This state supports project state machine re-entry logic: operators can enter the "pending supplementation" state and expand the candidate pool by adding compliant supplier identifiers through the external supplier supplementation channel. After supplementation is completed, there is no need to recreate the project; the extraction process for this sub-project can be directly restarted without reconfiguring all pre-configured extraction conditions. This significantly improves the fault tolerance of the business process and prevents the entire procurement project from being terminated due to a temporary shortage of suppliers.

[0112] This embodiment establishes a full-chain intelligent extraction framework adaptable to multi-item procurement projects. It leverages a unified supplier database to address the data fragmentation and information asynchrony issues inherent in traditional models. Through knowledge graphs and graph neural networks, it uncovers hidden risks associated with enterprise relationships, compensating for the inability of manual screening to identify risks associated with chain operations. A parent company aggregation mechanism restricts duplicate entries from a single group, balancing extraction opportunities for large, medium, and small enterprises. A digital twin simulation module performs multiple rounds of virtual simulations before the formal extraction, automatically optimizing extraction parameters using four-dimensional quantitative indicators, avoiding the time costs and compliance risks associated with offline manual trial and error. Structured logs are maintained for every step of the process, random seeds are persistently stored, and the extraction logic is fully reproducible, meeting the traceability requirements of government procurement audits. Furthermore, configurable screening rules can quickly adapt to business changes without modifying the underlying code, comprehensively balancing extraction fairness, procurement compliance, business adaptability, and system auditability.

[0113] In addition, this method is equipped with a complete, tamper-proof audit and traceability mechanism covering all operation nodes from extraction initiation to result archiving:

[0114] First, a persistent encrypted storage mechanism for random seeds is implemented. Each simulated extraction and formal random extraction process generates a unique random seed, which is encrypted and stored in the database. After auditors retrieve the random seed, they can fully reproduce all the operational logic of multi-dimensional screening, random shuffling, and supplier extraction, making the extraction process reproducible.

[0115] Second, the system retains structured logs for all operations. The system generates structured logs for all operations, including loading of extraction conditions, filtering of multi-dimensional constraints, calling tax interfaces, random shuffling calculations, replacement of unqualified suppliers, and supplementation of external suppliers. The logs are stored on an immutable storage medium and include operation time, operator, operation parameters, and operation results, which meet the requirements of fiscal procurement audit and verification.

[0116] Third, the multi-industry classification parallel screening mechanism allows multiple sets of independent industry classification constraints to be configured for the same target procurement project and to perform screening operations in parallel. The supplier identifiers obtained from each group of screenings are stored in isolation according to the classification groups. After all extractions are completed, all group results are automatically merged and duplicate supplier identifiers are removed, adapting to complex procurement scenarios where multiple categories and multiple sub-industries are extracted simultaneously.

[0117] In some embodiments, step S2, which involves using a pre-trained graph neural network model to predict risk transmission in the knowledge graph, specifically includes:

[0118] S21: Encode the attribute information of each supplier node, legal person node, shareholder node, senior executive node, and affiliated company node in the knowledge graph into an initial feature vector;

[0119] S22: Input the initial feature vector into the pre-trained graph attention network (GAT) model. The GAT model calculates the attention coefficient between each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation of the features of the neighboring nodes based on the attention coefficient to obtain the updated feature vector of each node.

[0120] S23: Input the updated feature vector of each supplier node into the risk prediction classifier and output the dynamic correlation risk transmission probability of that supplier node;

[0121] S24: If the probability of dynamic association risk transmission exceeds the preset risk threshold, the supplier status is determined to be invalid and the supplier is removed from the qualified list.

[0122] In this embodiment, node attribute information refers to the static and dynamic data of each entity node in the graph. Supplier nodes include years of operation, qualifications, performance records, and industry classification; legal person nodes, shareholder nodes, and senior executive nodes include years of service, shareholding ratio, and number of affiliated companies; affiliated company nodes include cooperation duration, transaction frequency, and equity ratio.

[0123] The initial feature vector refers to the numerical vector that converts the node attributes of text, numerical values, and classification categories into a uniform-dimensional numerical vector through standardized encoding. It is the input data that the graph neural network can recognize.

[0124] Graph Attention Network (GAT) refers to a pre-trained deep learning model with a core multi-head attention mechanism. It can autonomously identify the weights of different associated edges on the risk of nodes, which is different from graph convolutional networks with fixed weights.

[0125] Multi-head attention mechanism refers to the parallel operation of multiple independent attention computing units, which calculate the importance of the relationship between nodes from different correlation dimensions such as equity, position, and cooperation, and then integrate the multi-dimensional results;

[0126] The attention coefficient is a quantified value that represents the strength of the influence of a neighboring node on the risk status of the current target supplier node.

[0127] The updated feature vector refers to a new vector generated by the target node after fusing its original attributes with the weighted features of all neighboring nodes, which fully carries the risk information of the enterprise itself and its upstream and downstream related entities;

[0128] A risk prediction classifier is a binary classification model built on a training dataset. It takes the supplier's updated feature vector as input and outputs a dynamic correlation risk transmission probability value in the 0-1 range.

[0129] The preset risk threshold refers to a critical value set in advance by business supervision. If the value is exceeded, the supplier is deemed to have too high a risk of association and is prohibited from entering the candidate pool.

[0130] The working principle of this embodiment is as follows:

[0131] In step S21, full-dimensional attribute data is extracted for all entity nodes in the knowledge graph. Text, numbers, and classification labels are uniformly converted into fixed-dimensional initial feature vectors through standardized coding rules, eliminating the obstacle that unstructured data cannot be processed, and realizing the digitization of graph entity information.

[0132] In step S22, all initial feature vectors are input into the pre-trained GAT model. The model enables a multi-head attention mechanism. Each attention unit independently traverses all neighboring edges of a node and calculates the attention coefficient corresponding to each edge. This coefficient represents the weight of the influence of the corresponding associated entity on the target supplier's risk. Then, using the attention coefficient as the weight, the feature vectors of all neighboring nodes are weighted and summed, and the original features of the node itself are superimposed to generate an updated feature vector that integrates multiple associated information, fully absorbing the chain risk information brought by legal persons, shareholders, and cooperative enterprises.

[0133] In step S23, the updated feature vectors of all supplier nodes are extracted individually and input into the risk prediction classifier in batches. The classifier, based on the risk feature patterns learned during the training phase, outputs the dynamic correlation risk transmission probability corresponding to each supplier, intuitively quantifying the possibility of procurement risks caused by the correlation relationship for the enterprise.

[0134] In step S24, the calculated dynamic correlation risk transmission probability is compared with a preset risk threshold. If the dynamic correlation risk transmission probability exceeds the preset risk threshold, it means that the supplier has a high hidden chain risk. The system automatically removes the supplier's identifier from the preliminary compliance list to prevent high-risk suppliers from entering the subsequent extraction stage.

[0135] This embodiment employs a GAT (Gaussian Attention Network) multi-head attention network to replace traditional fixed-rule screening. It can autonomously distinguish the strength of risk impact from different related entities, significantly improving the accuracy of identifying hidden risk linkages compared to manually formulated static risk rules. It can capture hidden risk chains such as multi-layered indirect shareholding and cross-company employment. Multi-dimensional parallel attention calculation enhances model robustness, avoiding misjudgments caused by a single association dimension. Standardized automatic elimination is achieved based on quantified risk probabilities, requiring no manual intervention throughout the process. This unifies risk assessment criteria, eliminates unfairness caused by subjective judgments by personnel, and improves the overall compliance level of procurement operations from the screening source.

[0136] In some embodiments, the construction of a knowledge graph including supplier nodes, legal entity nodes, shareholder nodes, executive nodes, and affiliated company nodes in step S2 further includes the construction of regional nodes and regional association edges, specifically including:

[0137] In the knowledge graph, a regional node is constructed for each supplier node's registered region and operating region, including provincial regional nodes and municipal regional nodes;

[0138] For each supplier node, a regional association edge is constructed with its registered regional node and operating regional node, and a regional association strength weight is assigned to each regional association edge. The regional association strength weight is proportional to the proportion of the supplier's business in that region.

[0139] To construct regional adjacency edges between different city-level regional nodes belonging to the same provincial-level regional node, and to assign an adjacency attenuation coefficient to each regional adjacency edge. The adjacent attenuation coefficient It is inversely proportional to the geographical distance between the two municipal-level regional nodes;

[0140] When calculating the attention coefficient in the GAT model, the regional association strength weight of the regional association edge and the adjacent decay coefficient of the regional adjacent edge are used as adjustment factors for the attention coefficient, so that the risk transmission probability between supplier nodes in the same region or adjacent regions is suppressed, and the suppression magnitude is proportional to the product of the regional association strength weight and the adjacent decay coefficient.

[0141] In this embodiment, a geographic node refers to a newly added geographic dimension entity in the knowledge graph, which is divided into provincial geographic nodes and municipal geographic nodes, corresponding to the supplier's registered address and actual business location.

[0142] Geographic association edges refer to the association lines established between supplier nodes and nodes in their registered and operating regions, used to bind the geographical operating attributes of enterprises;

[0143] The weight of regional association strength refers to the value assigned to the regional association edge, which is directly proportional to the proportion of the supplier's annual business revenue in that region. The higher the proportion of business revenue, the greater the weight value.

[0144] Geographical adjacency edges refer to the connecting lines established between different municipal-level geographic nodes under the same provincial node, representing the geographical adjacency relationship between cities;

[0145] Adjacent attenuation coefficient This refers to the specific adjustment value for adjacent geographical areas; the greater the geographical distance between two city-level nodes, the more significant the adjustment. The smaller the value, the closer the geographical distance. The larger the value;

[0146] Attention coefficient adjustment factor refers to the correction parameter embedded in the GAT model calculation, which is used to correct the risk transmission weight between suppliers in the same region and adjacent regions;

[0147] The risk transmission probability suppression magnitude refers to the correction value for reducing the risk transmission probability for suppliers in the same or adjacent regions. The suppression magnitude is equal to the product of the regional correlation strength weight and the adjacent attenuation coefficient.

[0148] This embodiment adds hierarchical regional entity nodes in addition to the original enterprise-related nodes. The system reads the business registration address and actual operating address of each supplier, matches them with corresponding city-level and provincial-level regional nodes, and builds bidirectional regional association edges. Based on the supplier's annual business revenue share in each region, a corresponding regional association strength weight is assigned. For all city-level regional nodes within the same provincial region, pairwise regional adjacency edges are constructed, and an adjacency attenuation coefficient is assigned based on the straight-line geographical distance between cities. The closer the distance The higher the value, the farther the distance. The lower the value.

[0149] When calculating the attention coefficient between nodes, the GAT model simultaneously reads the weight of regional association strength and the attenuation coefficient of regional adjacent edges as adjustment factors. If two suppliers belong to the same region or adjacent regions, the attention coefficient of risk transmission between them will be reduced synchronously. The reduction is obtained by multiplying the two parameters, thereby suppressing the risk transmission value between enterprises in the same region and avoiding misjudgment of batch supplier risks due to regional common policies and industry environment.

[0150] This embodiment supplements the knowledge graph with a geographical dimension of association logic, resolving the misjudgment problem of traditional models uniformly classifying suppliers in the same region as having high association risk. It dynamically adjusts the regional risk suppression level based on business revenue and geographical distance, aligning with the actual operational distribution characteristics of enterprises and refining the granularity of risk prediction. By adjusting the GAT attention coefficient through adjustment factors, the risk calculation logic is optimized without altering the main model structure. This allows the model to adapt to procurement businesses in different regions without retraining, balancing the accuracy of risk identification with the cost of model iteration.

[0151] In this embodiment, in step S5, the simulation engine updates the state information of relevant nodes in the knowledge graph based on the results of this simulation extraction, including:

[0152] S41: The simulation engine executes the simulated extraction step to generate a set of simulated extraction results;

[0153] S42: The simulation engine calls the real-time state verification of the simulation, which includes: identifying the extracted supplier node and its associated nodes according to the set of simulation extraction results, wherein the associated nodes include legal person nodes, shareholder nodes, senior executive nodes and related company nodes associated with the extracted supplier node, and marking the state of the extracted supplier node and its associated nodes as extracted in the knowledge graph.

[0154] The step of recalculating the dynamic association risk transmission probability by re-invoking the graph neural network model specifically includes:

[0155] Based on the labeled knowledge graph, the dynamic association risk transmission probability of each unextracted supplier node is calculated; among them, the dynamic association risk transmission probability of supplier nodes that have direct or indirect association with the extracted suppliers is attenuated and adjusted according to the length of the association path and the association strength.

[0156] In this embodiment, the pseudo-real-time status verification refers to the virtual verification module within the simulation engine that replicates the verification logic of the tax system. It does not require actual calls to external interfaces and only relies on locally synchronized supplier snapshot data to complete the verification of the four statuses: opening, liabilities, operation, and closing.

[0157] The extracted status markers refer to the built-in status fields of knowledge graph nodes. After the simulation is completed, the selected supplier and all its associated entity nodes are marked with a unique marker, which serves as the basic identifier for the next round of simulation risk calculation.

[0158] The associated path refers to the combination of entity connections in the graph that extend from the extracted supplier nodes to other unextracted supplier nodes, including multi-layered links such as direct equity, indirect shareholding, and cross-enterprise employment;

[0159] The path length refers to the number of entity nodes between the extracted suppliers and the target supplier. The more nodes there are, the larger the path length value.

[0160] Association strength refers to the weight value inherent in each associated edge within the path, including shareholding ratio, length of service, scale of business cooperation, etc.

[0161] Risk attenuation adjustment refers to reducing the dynamic correlation risk transmission probability of unselected suppliers that have a related link with the selected suppliers. The longer the path and the lower the correlation strength, the greater the attenuation reduction.

[0162] The working principle of this embodiment is as follows:

[0163] In step S41, the simulation engine performs a complete round of simulation extraction, outputting the selected parent company and corresponding business name combination to form a complete simulation extraction result set.

[0164] In step S42, a simulated real-time status verification is initiated. All selected vendors in the result set are traversed, and their associated legal entities, shareholders, senior executives, and affiliated companies are traced layer by layer. The status fields of all these entity nodes are modified in the knowledge graph, uniformly marked as extracted, and the marking operation audit log is retained. After marking is completed, the system re-invokes the graph neural network model, traversing all unextracted supplier nodes in the database and retrieving all association paths between each node and all extracted suppliers. For each association path, an attenuation correction value is calculated based on the path length and the association strength of each edge within the path. The more nodes the path passes through and the lower the weight of the associated edges, the larger the attenuation correction value, and the greater the reduction in the probability of dynamic association risk transmission for the corresponding supplier. After completing the probability correction for all nodes, the risk values ​​of all suppliers are updated as the weight input basis for the next round of simulated extraction.

[0165] This embodiment achieves synchronized updates between simulation extraction and knowledge graph status. Each round of virtual extraction realistically simulates the business scenario of "changes in the associated risks of remaining entities after some suppliers are shortlisted," ensuring that the simulation results closely align with real procurement business logic. By applying layered attenuation corrections based on the length and strength of the association path, it accurately reproduces the objective law that risk transmission gradually weakens as the link extends, avoiding the simulation model's overestimation of the risks of distant associated enterprises. This significantly improves the authenticity and reference value of the simulation evaluation results and reduces compliance deviations in formal extraction.

[0166] In some embodiments, after each simulated extraction in step S5, an interpretability output step for the extraction results is further included, specifically including:

[0167] After each simulated extraction, the simulation engine extracts the risk transmission path of the extracted supplier node and its associated nodes from the knowledge graph based on the results of this simulated extraction. The risk transmission path includes the shortest path from the extracted supplier node, through legal person nodes, shareholder nodes, senior executive nodes or associated company nodes, to other unextracted supplier nodes.

[0168] For each risk transmission path, calculate the path risk contribution value, which is the product of the association strength weights of all edges on the path and the reciprocal of the path length.

[0169] The risk transmission paths are sorted from largest to smallest according to their risk contribution value, and the top K risk transmission paths are selected as key risk transmission paths, where K is a preset positive integer;

[0170] Generate an interpretability report, which includes: the identifier of the supplier selected in this simulation sampling result, the node sequence of the key risk transmission path, the path risk contribution value of each key risk transmission path, and the contribution ratio of the path risk contribution value to the probability of dynamic associated risk transmission.

[0171] The interpretability report is associated with and stored in conjunction with the results of this simulation extraction, and then output to the user interface for visualization.

[0172] In this embodiment, the shortest path of risk transmission refers to the association link in the updated and labeled knowledge graph that starts from the extracted supplier node and connects to any unextracted supplier node, passing through the fewest number of entity nodes. The link only allows four types of entities—legal persons, shareholders, senior executives, and affiliated companies—as intermediate transmission nodes, filtering out irrelevant regions and competitive auxiliary nodes.

[0173] The path risk contribution value refers to the weight of the impact of a single risk link on the risk probability of the target supplier. The path risk contribution value is obtained by multiplying the weights of all associated edges in the link and dividing by the length of the path nodes. The higher the value, the stronger the risk transmission capability of the link.

[0174] The preset positive integer K refers to the configurable display threshold used to filter high-risk core links. It is customized by the procurement supervisor according to the audit granularity, and the common values ​​are 5, 10, etc.

[0175] The key risk transmission path refers to the top K high-impact transmission paths extracted after all risk links are sorted in descending order of path risk contribution value. These are the core sources of increased risk associated with unextracted suppliers.

[0176] Interpretable reports refer to structured and standardized output documents that carry comprehensive risk tracing data extracted from a single round of simulation, eliminating the algorithmic black box of graph neural network simulation.

[0177] Contribution percentage refers to the percentage of the risk contribution value of a single critical path to the total contribution value of all transmission paths of the target supplier, which intuitively distinguishes the primary and secondary impacts of each risk link;

[0178] Visualization refers to the front-end interface using a network diagram and numerical bar chart for simultaneous rendering. The graph nodes distinguish between extracted and unextracted suppliers, and the link thickness matches the path risk contribution value.

[0179] The working principle of this embodiment is as follows:

[0180] After each round of simulation extraction is completed and the knowledge graph has completed the marking of extracted nodes, the simulation engine starts the full-link path retrieval. Starting from all the selected supplier nodes in this round, it traverses all unextracted suppliers in the graph and retrieves the connection path only through the legal person, shareholder, senior executive, and affiliated company entity nodes, and selects the shortest transmission link between each group of supply and demand nodes as the basic analysis object.

[0181] For each shortest path retrieved, the pre-defined association strength weight of each associated edge within the path is extracted. All weights are multiplied consecutively to obtain the total link weight, which is then divided by the number of nodes contained in the path to calculate the path risk contribution value specific to that path. After completing the numerical calculation for all paths, they are sorted from highest to lowest according to their path risk contribution values. The top K links in the sorted results are selected as key risk transmission paths, focusing on high-risk transmission channels.

[0182] The system generates a standardized and interpretable report based on the selected critical paths. The report fully records the unique identifier of the shortlisted suppliers in this simulation, the complete node flow sequence of each critical path, the risk contribution value of each path, and the contribution percentage of the path to the overall associated risk probability of the target supplier. The report is bound to the original results of this round of simulation extraction and stored in the database. The front-end business interface loads the report data synchronously and completes the visualization display with network diagrams and numerical charts, allowing operators and auditors to intuitively view the risk transmission logic.

[0183] The above solution adds comprehensive risk tracing capabilities to the simulation extraction process, breaking the algorithmic black box problem of graph neural network simulation. All changes in supplier risk probability can be traced back to specific enterprise-related links. By quantifying the contribution value of path risks and screening core risk channels, equity and employment relationships that cause high supplier-related risks can be quickly identified, facilitating manual intervention to adjust the extraction strategy. The combination of standardized reporting and visualization lowers the barrier for auditors and procurement personnel to understand complex relationship networks. End-to-end risk data is persistently stored, further improving the audit traceability system for procurement business and significantly enhancing the transparency and regulatory credibility of the extraction process.

[0184] In some embodiments, step S5, the re-invocation of the graph neural network model to calculate the dynamic association risk transmission probability, further includes the following steps:

[0185] In the knowledge graph, competitive relationship edges are constructed between parent company nodes that have historical bidding relationships, and each competitive relationship edge is assigned a competitive intensity weight, which is proportional to the number of historical biddings between the two parent companies.

[0186] After each simulated sampling, identify the parent company node to which the sampled supplier belongs and record it as the sampled parent company node;

[0187] Traverse all unextracted supplier nodes in the knowledge graph. For each unextracted supplier node, obtain its parent company node, which is denoted as a candidate parent company node.

[0188] If there is a competitive relationship between the candidate parent node and any extracted parent node, the probability of dynamic association risk transmission of the supplier nodes under the candidate parent node is adjusted upward according to the competition intensity weight of the competitive relationship edge, and the adjustment range is proportional to the competition intensity weight.

[0189] The increased probability of dynamic correlation risk transmission will be used as the basis for adjusting the weight of the supplier node in the next simulation extraction.

[0190] In this embodiment, the parent company node competitive relationship edge refers to the newly added enterprise group dimension association connection in the knowledge graph. It is only established between two parent company entities that have multiple historical bidding intersections. It is different from equity and employment-related risk association edges and independently depicts market competition business relationships.

[0191] The competition intensity weight is a numerical parameter specific to the competitive relationship. It is directly proportional to the number of times the two parent companies have participated in the same procurement project in the past. The more historical bidding overlaps, the higher the competition intensity weight value.

[0192] The extracted parent company nodes refer to the parent company entity nodes of suppliers that have successfully entered the shortlist in this round of simulation extraction, and are used as the benchmark for judging competition risk.

[0193] Candidate parent company nodes refer to all parent company entity nodes in the knowledge graph that have not generated shortlisted records, and all their subordinate suppliers are candidates to be extracted.

[0194] The upward adjustment of the risk transmission probability refers to the increase in the probability of risk transmission dynamically associated with suppliers under the parent company with strong competitive relationships. The upward adjustment is directly proportional to the competition intensity weight.

[0195] The weight adjustment is based on the risk probability value after the competition relationship correction, which is directly input into the GAT model as the node feature correction parameter for the next round of simulation, thereby raising the avoidance priority of highly competitive suppliers.

[0196] The working principle of this embodiment is as follows:

[0197] The system pre-traverses the bidding archives of historical procurement projects, identifies parent company pairs that have multiple bidding records, constructs competitive relationship edges between the corresponding parent company nodes in the knowledge graph, and assigns a competition intensity weight based on the number of times the two bid together in history. The more bidding overlaps, the greater the weight value.

[0198] After each round of simulated extraction and marking of extracted nodes, the system extracts the parent companies of all selected suppliers in this round and marks them uniformly as extracted parent company nodes. Then, it iterates through all unextracted suppliers, reverse-matching their respective parent companies to obtain candidate parent company nodes. For any pairing of candidate parent companies with extracted parent companies, it searches for competitive edges between them. If a competitive relationship exists, it reads the corresponding competition intensity weight and increases the dynamic association risk transmission probability of all subordinate suppliers of that candidate parent company according to a proportional rule; the higher the competition intensity, the greater the probability increase.

[0199] After completing all competitive relationship correction calculations, the updated risk probability values ​​are persistently written to the graph nodes as weight correction inputs for the GAT model calculations in the next round of simulation extraction, thereby enabling the simulation process to automatically avoid homogeneous competitive supplier combinations.

[0200] This embodiment integrates market competition business logic into knowledge graph risk calculation, overcoming the shortcomings of traditional risk models that only focus on equity and personnel relationships while ignoring the risk of market competition imbalances. By quantifying the degree of competition among enterprises through competition intensity weights, it automatically increases the risk probability of competing suppliers in the same industry, making the simulation extraction more inclined to select differentiated industries and differentiated competitors, avoiding the concentration of extraction results on a few leading competitors, and optimizing supplier diversity. The competitive relationship edges are constructed independently of the original risk chain, without requiring changes to the existing GAT model's basic computational logic, lightweightly expanding the model's business adaptability and aligning with the regulatory requirements for balanced selection of diverse suppliers in government procurement.

[0201] In some embodiments, step S2, the construction of a knowledge graph including supplier nodes, legal entity nodes, shareholder nodes, executive nodes, and affiliated company nodes, further includes a supplier node importance weighting step, specifically including:

[0202] In the knowledge graph, a node importance index is calculated for each supplier node. This node importance index includes a first-dimensional index and a second-dimensional index. The first-dimensional index is the supply chain irreplaceability index, which is calculated based on the number of exclusive supply records, core component supply records, and alternative suppliers for that supplier node in historical projects. The calculation formula is as follows:

[0203] Indispensability Index = ×(Number of exclusive supply records / Total supply records)+ ×(Number of core component supply records / Total supply records)+ ×(1 / number of alternative suppliers), where, , , For the preset weighting coefficients, and =1;

[0204] The second dimension indicator is the network centrality index, which is calculated by weighting the supplier node's degree centrality, betweenness centrality, and proximity centrality in the knowledge graph. The calculation formula is as follows:

[0205] Centrality index = × degree centrality+ ×between centrality+ × is close to centrality, where , , For the preset weighting coefficients, and ;

[0206] The supply chain irreplaceability index and the network centrality index are weighted and fused to obtain the comprehensive node importance value of the supplier node. The calculation formula is as follows:

[0207] Overall node importance value = ×Irreplaceability Index+ × Centrality index, where , For the preset weighting coefficients, and ;

[0208] In step S5, when the graph neural network model is called again to calculate the probability of dynamic correlation risk transmission, the comprehensive value of node importance is used as an adjustment factor for the attention coefficient of the GAT model, so that the supplier node with the higher the comprehensive value of node importance, the greater the range of risk transmission influence on other nodes after it is extracted, and the adjustment range of the probability of dynamic correlation risk transmission is proportional to the comprehensive value of node importance.

[0209] In this embodiment, the supply chain irreplaceability index is a quantitative indicator that measures the degree of irreplaceability of a single supplier in a specific procurement category. It is calculated based on three types of business data: exclusive supply records, core component supply records, and the number of market alternative manufacturers.

[0210] , , It refers to the fixed weight coefficients of the sub-items of the supply chain irreplaceability index, the sum of which is always equal to 1, and is preset by the procurement supervisor based on the importance of the category;

[0211] The network centrality index is a comprehensive indicator that characterizes the pivotal position of a supplier in the enterprise network graph, integrating three classic network topology indicators.

[0212] Degree centrality refers to the total number of edges directly connected to a single supplier node within a knowledge graph, reflecting the scale of the entities directly associated with the enterprise.

[0213] Betweenness centrality refers to the frequency with which a node acts as a relay link between other entities, representing the intermediary hub role of an enterprise;

[0214] Closeness centrality refers to the average shortest path length from a node to all other nodes in the graph; the smaller the value, the stronger the connectivity.

[0215] , , It refers to the weighting coefficient of network centrality, the sum of which is always equal to 1, used to balance the calculation proportion of the three types of topology indicators;

[0216] The node importance composite value is a composite indicator that integrates supply chain value and network hub status, and is used to correct the risk transmission strength of the GAT model;

[0217] η and θ are the weights of the two types of exponents, and their sum is always equal to 1;

[0218] The attention coefficient adjustment factor refers to embedding the comprehensive value of node importance into the GAT calculation, which amplifies the scope of the chain risk transmission after the core supplier is extracted. The higher the importance, the greater the correction of the risk probability of other related suppliers.

[0219] This embodiment adds supplier node importance quantification calculation logic in addition to basic entity nodes and regional nodes. The system calculates two main dimensions of indicators for each supplier node, as detailed below:

[0220] The first dimension, the supply chain irreplaceability index, is calculated by retrieving the supplier's historical number of exclusive supply projects, core component supply records, and the total number of alternative suppliers in the market, and then substituting them into a weighted formula. The higher the proportion of exclusive supply and core supply, and the fewer alternative manufacturers, the higher the index value.

[0221] The second dimension, the network centrality index, extracts node degree centrality, betweenness centrality, and proximity centrality, respectively, and combines them with preset weighting coefficients to obtain a comprehensive topological index, representing the enterprise's hub level in the network. The two indices are then combined... , The weighted fusion yields a comprehensive value for node importance, which fully considers both the actual value of the supplier's delivery and the influence of the associated network.

[0222] When recalculating the probability of dynamic correlation risk transmission using the GAT model, the comprehensive value of node importance is used as an attention coefficient adjustment factor in the calculation. If a highly important supplier is selected in this round of simulation, the model will simultaneously amplify its risk transmission impact on all related nodes. The higher the comprehensive value of node importance, the greater the increase in the risk probability of other related suppliers, thereby reflecting the chain supply chain risks brought about by the inclusion of core suppliers.

[0223] This embodiment distinguishes between ordinary suppliers and core key suppliers in the supply chain, addressing the issue that a unified risk transmission standard cannot be applied to exclusive or core suppliers. It quantifies supplier importance from two dimensions: business supply value and network topology, with calculation logic closely aligning with real-world procurement supply chain operations. By embedding comprehensive importance values ​​into the GAT model calculation, the risk transmission intensity can be dynamically adjusted without retraining the model. Simulation extraction proactively avoids selecting multiple high-hub, highly irreplaceable suppliers simultaneously, reducing the significant procurement risks of supply chain disruptions in a single product category and related chain reactions of breaches of trust, thus enhancing the overall risk management capability of the extraction strategy.

[0224] In some embodiments, step S7, adjusting the extraction parameters in the current extraction strategy, specifically includes:

[0225] When the risk diversity balance index is lower than a preset balance threshold:

[0226] If the supplier diversity index is lower than the preset diversity threshold, then the number N of parent companies extracted in the simulation extraction step is increased.

[0227] If the weighted historical performance risk probability is higher than the preset risk threshold, then the weight adjustment coefficient of the dynamic associated risk transmission probability is reduced.

[0228] Wherein, the preset balance threshold, preset diversity threshold, and preset risk threshold are all preset constants.

[0229] In this embodiment, the risk diversity balance index is a comprehensive evaluation index that balances the diversity of supplier categories and the overall level of performance risk. The lower the value, the worse the balance of the sampling scheme.

[0230] The preset balance threshold refers to the critical value of the balance index that is pre-configured in the system, which serves as the overall judgment criterion for simulation iterative optimization;

[0231] The supplier diversity threshold refers to the lower limit of the diversity index. A value below this value indicates that the industry, group, and regional distribution of the selected suppliers are too concentrated.

[0232] The number of parent companies N to be selected refers to the number of group entities randomly selected from the candidate list of parent companies in a single simulation sampling stage, which directly determines the base number of shortlisted suppliers.

[0233] The weighted historical performance risk probability refers to the overall risk average obtained by weighting the historical default and overdue records of all simulated shortlisted suppliers.

[0234] The preset risk threshold refers to the upper limit of the overall performance risk. If the average risk of the simulation results exceeds this value, it means that the overall supplier performance reliability of the extracted scheme is insufficient.

[0235] The dynamic correlation risk transmission probability weight adjustment coefficient is a global correction parameter used by the GAT model to calculate the risk probability. Reducing the coefficient will decrease the constraint of correlation risk on the extraction weight and relax the restrictions on suppliers with high correlation risk.

[0236] The working principle of this embodiment is as follows:

[0237] The system first determines the risk diversity balance index of the current simulation results. If the risk diversity balance index is lower than the preset balance threshold, it means that the current extraction strategy has defects such as insufficient diversity or high overall performance risk, and enters the sub-scenario parameter adjustment branch.

[0238] The first scenario is that the supplier diversity index is lower than the preset diversity threshold, which indicates that the supplier groups, industries and regions in the sampling results are highly homogeneous. The system automatically increases the number of parent companies N in a single simulation sampling to expand the selection base of candidate groups and improve the richness of supplier distribution in the next round of simulation.

[0239] The second scenario: If the weighted historical performance risk probability is higher than the preset risk threshold, it indicates that the overall number of historical defaults and overdue records of the shortlisted suppliers in this round of simulation is too high. The system automatically reduces the weight adjustment coefficient of the probability of dynamic correlation risk transmission, reduces the constraint ratio of correlation risk values ​​in the weight calculation, and appropriately relaxes the screening restrictions for suppliers with mild correlation risks, thus balancing the overall performance risk and the range of suppliers that can be selected.

[0240] After the parameters are adjusted, the process jumps back to step S5 to re-execute multiple rounds of digital twin simulation, iteratively evaluating the indicators and adjusting the parameters until all four expected effect indicators meet the corresponding preset thresholds before the formal extraction process can be executed.

[0241] This embodiment sets up hierarchical, quantifiable parameter adaptive optimization rules, eliminating the need for repeated manual debugging of simulation parameters and achieving fully automated iterative optimization of the extraction strategy, significantly reducing the maintenance and operational costs of the procurement system. Dedicated parameter tuning logic is matched to address two core deficiencies: insufficient diversity and high overall performance risk, accurately correcting the shortcomings of the extraction scheme and balancing the two core requirements of supplier richness and procurement risk control. All judgment thresholds can be customized according to different procurement categories and regulatory requirements, adapting to differentiated business scenarios such as small-scale consulting procurement and large-scale engineering project procurement, improving the versatility and adaptability of the entire extraction method.

[0242] In some embodiments, this application provides a standardized extraction implementation logic that combines constraint weighting, post-real-time verification, and sequential inheritance. The specific execution flow is as follows:

[0243] First, obtain all compliant supplier identifiers to form a complete supplier set. Each supplier identifier is uniquely linked to a corresponding parent company identifier p(d); the total number k of compliant supplier identifiers under each parent company j is counted. j Set the number of target parent companies to be extracted in a single extraction to M. If the target extraction number M is greater than the total number of parent companies in the candidate parent company list, then adjust the extraction number to the total number of parent companies and give a system prompt.

[0244] Secondly, a uniform unbiased random shuffle algorithm (Fisher-Yates shuffle algorithm) is used to perform a global random rearrangement of the supplier set D, generating an ordered random candidate queue. The random seed generated during this shuffle is stored synchronously and persistently for subsequent audit tracing and complete reproduction of the extraction process.

[0245] Then, the mechanical energy sequentially traverses and completes the initial extraction of parent companies to remove duplicates. Specifically, this includes: initializing an empty set of extracted parent companies S and a temporary candidate list C; traversing each supplier identifier in the order of the random candidate queue Q; if the parent company to which the current supplier identifier belongs is already stored in the set of extracted parent companies S, then skipping the supplier identifier; otherwise, adding the supplier identifier to the temporary candidate list C, and simultaneously writing its parent company to the set of extracted parent companies S; continuing to traverse until the number of suppliers in the temporary candidate list C reaches the target extraction number M, or terminating the initial extraction after completely traversing the random candidate queue Q.

[0246] The above scheme is based on the number k of compliant supplier identifiers under each parent company. jFor weighted sampling without replacement, the probability that parent company j is the first parent company to be sampled is equal to k. j / N; For any two different parent companies j and l, if the total number of compliant suppliers under parent company j is k j Greater than the total number of compliant suppliers under the parent company l k l If the probability of parent company j being selected is strictly higher than that of parent company l, then this weighted fair selection scheme does not require pre-calculation or storage of various parent company weight parameters. It can achieve a fair selection mechanism that matches the probability of group size matching through a single shuffle and linear traversal operation with a time complexity of O(N).

[0247] This embodiment only calls the external tax system interface to perform real-time business status verification for supplier identifiers in the temporary candidate list C. The verification mode supports serial, concurrent, or batch interface calls. Supplier identifiers that meet the compliance requirements are included in the final selection list R, while supplier identifiers that fail the verification are marked as invalid and removed from the temporary candidate list.

[0248] For each supplier identifier that fails verification, the same random candidate queue Q generated in the previous steps is used to search backward from the current traversal cursor position. The first supplier identifier that simultaneously meets the following conditions—not marked as failed, not included in the final extraction list, and whose parent company is not in the already extracted parent company set S—is selected as the replacement target. The replacement supplier identifier is included in the candidate and the real-time status verification step is re-executed. The replacement and verification operations are repeated until the number of suppliers in the final extraction list R reaches the target extraction number M or the random candidate queue Q is completely traversed.

[0249] In this embodiment, the replacement process reuses the initial random candidate queue and a unified parent company deduplication rule. The entire process does not re-execute random shuffling or initiate a second overall extraction, thus fully preserving the weighted probability distribution structure formed in the initial extraction and ensuring the fairness of the random extraction. At the same time, the entire scheme only initiates external interface verification for shortlisted candidate suppliers. The total number of tax interface calls is equal to the target extraction quantity M plus the number of replacements caused by verification failures. The total number of interface requests is decoupled from the total number of suppliers N in the supplier database, significantly reducing the frequency of cross-system interface interactions and the overall extraction computation time.

[0250] In a second aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for randomly selecting compliant suppliers under multiple constraints as described in the first aspect of this application.

[0251] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0252] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.

[0253] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of this application are intended to include these and any other suitable types of memory.

[0254] like Figure 2 As shown, in a third aspect, this application provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the random selection method for compliant suppliers under multiple constraints as described in the first aspect of this application.

[0255] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may use at least one of the following: circuit, single or multiple application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), central processing units (CPUs), controllers, microcontrollers, and microprocessors, thereby enabling the processor to perform some or all of the steps, or any combination thereof, of the random selection method for compliant suppliers under multiple constraints described in the various embodiments of this application.

[0256] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A method for randomly selecting compliant suppliers under multiple constraints, characterized in that, The method includes: S1: In response to the extraction start command for the target procurement project, obtain the project identifier of the target procurement project and the sub-project identifier of the currently pending sub-project; S2: Based on the pre-configured extraction conditions associated with the sub-project identifiers, select supplier identifiers that meet the multi-dimensional constraints from the supplier database and generate a qualified list; wherein, the multi-dimensional constraints include industry classification matching, supplier status validity, and preset supplier attribute condition matching; In the supplier status validity verification, a knowledge graph containing supplier nodes, legal person nodes, shareholder nodes, senior executive nodes and affiliated company nodes is constructed, and a pre-trained graph neural network model is used to predict the risk transmission of the knowledge graph, outputting the dynamic association risk transmission probability of each supplier node. If the dynamic association risk transmission probability exceeds a preset risk threshold, the supplier status is determined to be invalid and it is removed from the qualified list. S3: Aggregate the supplier identifiers in the qualified list according to their parent companies to generate a parent company candidate list, wherein multiple supplier identifiers under the same parent company correspond to one candidate qualification in the parent company candidate list; S4: Construct a digital twin model of the current sub-project, specifically including: loading the parent company candidate list, the preset random sampling algorithm and the real-time status verification rules into the simulation engine. The simulation engine is used to perform the following simulated sampling steps: randomly select N parent companies from the parent company candidate list, then randomly select a supplier identifier from the supplier identifiers that meet the requirements of the selected parent company, and call the simulated real-time status verification. S5: Perform multiple simulated extractions based on the digital twin model. Each simulated extraction calls the real-time state verification of the simulation and records the result distribution of each simulated extraction. After each simulated extraction, the simulation engine updates the state information of relevant nodes in the knowledge graph according to the result of this simulated extraction, and re-calls the graph neural network model to calculate the dynamic association risk transmission probability for weight adjustment in the next simulated extraction. S6: Based on the distribution of the results, evaluate the expected performance indicators of the current extraction strategy. The expected performance indicators include supplier diversity index, regional coverage balance, weighted historical performance risk probability, and risk diversity balance index. S7: If the expected effect indicator meets the preset indicator threshold, then perform formal extraction according to the current extraction strategy to generate the final extraction list of the sub-projects; if the expected effect indicator does not meet the preset indicator threshold, then adjust the extraction parameters in the current extraction strategy and return to step S5 to re-perform the simulated extraction until the preset indicator threshold is met; wherein, the extraction parameters include the number of parent companies N extracted in the simulated extraction step, the weight adjustment coefficient of the dynamic correlation risk transmission probability, and the target threshold of the risk diversity balance index; S8: Determine whether there are any unprocessed sub-projects in the target procurement project. If so, return to step S1 to process the next sub-project. If not, end the extraction process.

2. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 1, characterized in that, In step S2, the step of using a pre-trained graph neural network model to predict risk transmission in the knowledge graph specifically includes: S21: Encode the attribute information of each supplier node, legal person node, shareholder node, senior executive node, and affiliated company node in the knowledge graph into an initial feature vector; S22: Input the initial feature vector into the pre-trained graph attention network (GAT) model. The GAT model calculates the attention coefficient between each node and its neighboring nodes through a multi-head attention mechanism, and performs weighted aggregation of the features of the neighboring nodes based on the attention coefficient to obtain the updated feature vector of each node. S23: Input the updated feature vector of each supplier node into the risk prediction classifier and output the dynamic correlation risk transmission probability of that supplier node; S24: If the probability of dynamic association risk transmission exceeds the preset risk threshold, the supplier status is determined to be invalid and the supplier is removed from the qualified list.

3. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 2, characterized in that, Step S2 describes constructing a knowledge graph that includes supplier nodes, legal entity nodes, shareholder nodes, executive nodes, and affiliated company nodes. It also includes constructing regional nodes and regionally related edges, specifically: In the knowledge graph, a regional node is constructed for each supplier node's registered region and operating region, and the regional nodes include provincial regional nodes and municipal regional nodes; For each supplier node, a regional association edge is constructed with its registered regional node and operating regional node, and a regional association strength weight is assigned to each regional association edge. The regional association strength weight is proportional to the proportion of the supplier's business in that region. To construct regional adjacency edges between different city-level regional nodes belonging to the same provincial-level regional node, and to assign an adjacency attenuation coefficient to each regional adjacency edge. The adjacent attenuation coefficient It is inversely proportional to the geographical distance between the two municipal-level regional nodes; When calculating the attention coefficient in the GAT model, the regional association strength weight of the regional association edge and the adjacent decay coefficient of the regional adjacent edge are used as adjustment factors for the attention coefficient, so that the risk transmission probability between supplier nodes in the same region or adjacent regions is suppressed, and the suppression magnitude is proportional to the product of the regional association strength weight and the adjacent decay coefficient.

4. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 1, characterized in that, In step S5, the simulation engine updates the state information of relevant nodes in the knowledge graph based on the results of this simulation extraction, including: S41: The simulation engine executes the simulated extraction step to generate a set of simulated extraction results; S42: The simulation engine calls the real-time state verification of the simulation, which includes: identifying the extracted supplier node and its associated nodes according to the set of simulation extraction results, wherein the associated nodes include legal person nodes, shareholder nodes, senior executive nodes and related company nodes associated with the extracted supplier node, and marking the state of the extracted supplier node and its associated nodes as extracted in the knowledge graph. The step of recalculating the dynamic association risk transmission probability by re-invoking the graph neural network model specifically includes: Based on the labeled knowledge graph, the dynamic association risk transmission probability of each unextracted supplier node is calculated; among them, the dynamic association risk transmission probability of supplier nodes that have direct or indirect association with the extracted suppliers is attenuated and adjusted according to the length of the association path and the association strength.

5. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 4, characterized in that, In step S5, after each simulated extraction, there is also a step of outputting the interpretability of the extraction results, specifically including: After each simulated extraction, the simulation engine extracts the risk transmission path of the extracted supplier node and its associated nodes from the knowledge graph based on the results of this simulated extraction. The risk transmission path includes the shortest path from the extracted supplier node, through legal person nodes, shareholder nodes, senior executive nodes or related company nodes, to other unextracted supplier nodes. For each risk transmission path, calculate the path risk contribution value, which is the product of the association strength weights of all edges on the path and the reciprocal of the path length. The risk transmission paths are sorted from largest to smallest according to their risk contribution value, and the top K risk transmission paths are selected as key risk transmission paths, where K is a preset positive integer; Generate an interpretability report, which includes: the identifier of the supplier selected in this simulation sampling result, the node sequence of the key risk transmission path, the path risk contribution value of each key risk transmission path, and the contribution ratio of the path risk contribution value to the probability of dynamic associated risk transmission. The interpretability report is associated with and stored in conjunction with the results of this simulation extraction, and then output to the user interface for visualization.

6. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 4, characterized in that, In step S5, the step of re-invoking the graph neural network model to calculate the dynamic correlation risk transmission probability further includes the following steps: In the knowledge graph, competitive relationship edges are constructed between parent company nodes that have historical bidding relationships, and each competitive relationship edge is assigned a competitive intensity weight, which is proportional to the number of historical biddings between the two parent companies. After each simulated sampling, identify the parent company node to which the sampled supplier belongs and record it as the sampled parent company node; Traverse all unextracted supplier nodes in the knowledge graph. For each unextracted supplier node, obtain its parent company node, which is denoted as a candidate parent company node. If there is a competitive relationship between the candidate parent node and any extracted parent node, the probability of dynamic association risk transmission of the supplier nodes under the candidate parent node is adjusted upward according to the competition intensity weight of the competitive relationship edge, and the adjustment range is proportional to the competition intensity weight. The increased probability of dynamic correlation risk transmission will be used as the basis for adjusting the weight of the supplier node in the next simulation extraction.

7. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 2, characterized in that, In step S2, the construction of a knowledge graph including supplier nodes, legal entity nodes, shareholder nodes, senior management nodes, and affiliated company nodes also includes a supplier node importance weighting step, specifically including: In the knowledge graph, a node importance index is calculated for each supplier node. This node importance index includes a first-dimensional index and a second-dimensional index. The first-dimensional index is the supply chain irreplaceability index, which is calculated based on the number of exclusive supply records, core component supply records, and alternative suppliers for that supplier node in historical projects. The calculation formula is as follows: Indispensability Index = ×(Number of exclusive supply records / Total supply records)+ ×(Number of core component supply records / Total supply records)+ ×(1 / number of alternative suppliers), where, , , For the preset weighting coefficients, and =1; The second dimension indicator is the network centrality index, which is calculated by weighting the supplier node's degree centrality, betweenness centrality, and proximity centrality in the knowledge graph. The calculation formula is as follows: Centrality index = × degree centrality+ ×between centrality+ × is close to centrality, where , , For the preset weighting coefficients, and ; The supply chain irreplaceability index and the network centrality index are weighted and fused to obtain the comprehensive node importance value of the supplier node. The calculation formula is as follows: Overall node importance value = ×Irreplaceability Index+ × Centrality index, where , For the preset weighting coefficients, and ; In step S5, when the graph neural network model is called again to calculate the probability of dynamic correlation risk transmission, the comprehensive value of node importance is used as an adjustment factor for the attention coefficient of the GAT model, so that the supplier node with the higher the comprehensive value of node importance, the greater the range of risk transmission influence on other nodes after it is extracted, and the adjustment range of the probability of dynamic correlation risk transmission is proportional to the comprehensive value of node importance.

8. The method for randomly selecting compliant suppliers under multiple constraints as described in claim 1, characterized in that, In step S7, adjusting the extraction parameters in the current extraction strategy specifically includes: When the risk diversity balance index is lower than a preset balance threshold: If the supplier diversity index is lower than the preset diversity threshold, then the number N of parent companies extracted in the simulation extraction step is increased. If the weighted historical performance risk probability is higher than the preset risk threshold, then the weight adjustment coefficient of the dynamic associated risk transmission probability is reduced. Wherein, the preset balance threshold, preset diversity threshold, and preset risk threshold are all preset constants.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for randomly selecting compliant suppliers under multiple constraints as described in any one of claims 1 to 8.

10. An electronic device having a computer program stored thereon, characterized in that, It includes a processor and a storage medium, on which a computer program is stored, which, when executed by the processor, implements the random selection method for compliant suppliers under multiple constraints as described in any one of claims 1 to 8.

Images