An inter-enterprise matching management method and management system based on a knowledge graph

Abstract

The present application relates to the technical field of enterprise service digitization and intelligent management, and particularly relates to a method and system for inter-enterprise matching management based on a knowledge graph. The present application gathers and cleans multi-source data such as industry and commerce, capacity, equipment, products / processes, qualifications, policies, geography, risks, and visits, and constructs an enterprise knowledge graph according to chain constraints; parses the demand into a demand subgraph, and carries out two-channel matching (graph path scoring and RAG semantic scoring) with a candidate enterprise supply subgraph, and performs consistency arbitration by a rule layer and a learning layer. The system generates a traceable evidence chain and a hierarchical confidence, supports event-driven incremental recalculation, and cockpit / mobile terminal closed-loop learning, thereby improving the accuracy, explainability, and timeliness of the matching.

Description

Technical Field

[0001] This invention relates to the field of enterprise service digitalization and intelligent management technology, and in particular to an inter-enterprise matching management method and management system based on knowledge graph. Background Technology

[0002] Against the backdrop of industrial digitalization and regional industrial chain upgrading, the supply and demand matching, collaborative manufacturing, and cross-regional support between governments and industrial parks, and between leading enterprises and supporting enterprises, increasingly rely on unified modeling and intelligent retrieval of multi-source data such as enterprise capabilities, processes, qualifications, production capacity, and delivery time. Knowledge graphs, due to their ability to express entity relationships and their queryability across multi-source heterogeneous data, have been widely used in scenarios such as supply and demand retrieval, supplier management, industrial chain analysis, and information dissemination. Existing solutions can mostly complete the basic closed loop of "data aggregation, graph construction, and semantic retrieval / recommendation," but in complex inter-enterprise matching tasks, they still suffer from problems such as insufficient expression of chain position constraints, limited interpretability of matching results, slow response to dynamic changes, and a lack of business-oriented closed-loop calibration.

[0003] For example, Chinese patent CN112541072B discloses a supply and demand information recommendation method based on knowledge graphs. This method improves retrieval speed and accuracy by extracting entities and relationships from supply and demand texts, constructing a supply and demand information graph, and retrieving matching items from a graph database, thus generating a recommendation list. However, it primarily focuses on supply and demand retrieval based on history and semantics, with limited attention to the integration of constraints at the enterprise's production factor level and the evidentiary interpretability of the output. Similarly, Chinese patent CN113836310B proposes a knowledge graph-driven industrial product supply chain management method: it extracts entities based on inquiries, relies on existing graphs to retrieve and match products, and pushes business opportunities, improving the matching efficiency on the procurement side. However, it still focuses on product / customer similarity and historical statistics, failing to systematically address the issues of connection and traceability under multiple constraints related to enterprise capabilities, processes, and supply chain positions.

[0004] On the graph construction side, Chinese patent CN114610898A presents a method for constructing a supply chain operation knowledge graph, emphasizing multi-source data extraction, knowledge fusion, graph database storage, and semantic retrieval and precise push services, providing a general technical path of "ontology, data, and retrieval" for subsequent applications; however, its main goal is to support operational retrieval and knowledge services, without addressing chain position constraints and evidence chain management when matching between enterprises. Regarding supply chain knowledge fusion, Chinese patent CN113157940A improves the accuracy of graph entities by performing entity disambiguation and referential resolution on semi-structured and unstructured data, providing a foundation for entity-level consistency in supply chain scenarios, but it does not address the propagation and interpretative output of field-level confidence during the matching stage.

[0005] Regarding semantic association and supply chain reasoning, Chinese patent CN116502807B calculates sentence vector similarity between projects and items based on a technology knowledge graph, expands the technology-item relationship and generates new triples for enterprise production decisions, representing the idea of ​​enriching the supply chain graph with semantic computing; however, such solutions mostly focus on "semantic expansion and association discovery", and there is little discussion on the joint arbitration of alignment matching and rule constraints of "demand subgraphs and supply subgraphs".

[0006] Regarding dynamic maintenance, Chinese patent CN109597855A proposes a pipeline mechanism for constructing and incrementally updating domain knowledge graphs from a big data-driven perspective; Chinese patent CN118626811A explicitly proposes incremental updates by monitoring data change events in an industry chain scenario, updating only changed nodes and edges, and combining real-time data stream access to reduce maintenance costs. These technologies can shorten the update window, but they mostly remain at the level of "incremental updates of the graph ontology or embedded elements," lacking specific descriptions of event-driven local recalculation and online calibration loops for "matching scores and candidate sets."

[0007] Furthermore, applications for supplier management are relatively mature. Chinese patent CN112256887A uses text analysis, knowledge graphs, and graph databases for comprehensive management and risk control of suppliers within power grid companies, emphasizing the integration of internal and external data and the calculation of risk indicators. This is helpful for supplier profiling and monitoring, but it belongs to a different problem domain than online matching and interpretable report generation for "cross-enterprise supply and demand docking / matching".

[0008] In summary, existing technologies can construct knowledge graphs from multi-source data into industry chain / supply chain knowledge graphs, perform semantic retrieval, and provide a certain degree of recommendation or risk control. However, they still have common shortcomings in the specific task of "inter-enterprise matching management": First, the matching stage lacks a unified expression and constraint solution for multiple constraints such as enterprise capabilities, processes, qualifications, production capacity / delivery time, geographical layout, and policy adaptation, making it difficult to reflect the consistency of an enterprise's position in the industry chain. Second, it mainly relies on similarity or rule-based retrieval, lacking dual-channel verification and consistency arbitration that simultaneously utilizes graph path evidence and unstructured text semantics, which can easily lead to semantic illusions or unauthorized matching. Third, it lacks field-level evidence chains and confidence propagation mechanisms, making it difficult to support audit traceability and expert verification. Fourth, although incremental updates exist, it is rare to see a closed-loop process oriented towards business, integrating "event monitoring, affected subgraph location, topology pruning, partial recalculation of matching, and online updating of thresholds / weights." These shortcomings mean that in a volatile industry environment, there is still room for improvement in the accuracy, interpretability, and timeliness of enterprise matching. Summary of the Invention

[0009] The technical objective of this invention is to provide a knowledge graph-based inter-enterprise matching management method and system. Addressing issues such as insufficient expression of chain position constraints, weak interpretability of results, and slow response to dynamic changes in supply and demand matching and supplier selection, this invention constructs an integrated model of "enterprise, capability or process, and industrial chain link," combining graph path and text semantic dual-channel matching. It then performs consistency arbitration using rules such as reachability, taboo edges, and chain position order to generate field-level evidence chains and confidence levels. Combined with event-driven subgraph incremental recalculation and cockpit / mobile terminal feedback write-back, this achieves highly accurate, interpretable, and low-latency online matching and continuous optimization of enterprise matching relationships.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A knowledge graph-based inter-enterprise matching management method includes the following steps:

[0012] S1. Obtain relevant enterprise information from multi-source data, and use the composite key of unified social credit code and name standardization to perform entity parsing and disambiguation on enterprise identifiers, forming a cleaned and standardized basic enterprise database.

[0013] S2. Graph Modeling and Construction: Constructing an enterprise knowledge graph with chain position constraints;

[0014] S3. Subgraph Location: Extract structured fields and text from the target demand to obtain the demand subgraph corresponding to the demand; aggregate potential supply-side enterprise elements according to chain position constraints to locate the supply subgraph.

[0015] S4, Dual-channel matching and consistency arbitration:

[0016] S4.1 Perform meta-path search with chain position constraints on the knowledge graph and score the paths to obtain graph path scores;

[0017] S4.2. Using the demand description as the query and the unstructured files of candidate companies as the corpus, perform a generative process to enhance retrieval and obtain semantic scores.

[0018] S4.3. The graph path score and semantic score are jointly assessed and conflict detected by the consistency arbitrator. The verification is performed based on reachability, taboo edges, and chain position order rules. The comprehensive score and ranking of the candidate enterprise set are output, and the candidates that do not meet the consistency conditions are rejected.

[0019] S5. Evidence Chain and Confidence Propagation: For each candidate company, generate an evidence chain subgraph containing entities, relationships, data sources, and timestamps, and propagate and normalize the field-level confidence scores involved in the matching from bottom to top to form an interpretable matching report.

[0020] As a preferred option, the multi-source data in step S1 includes business registration, finance and production capacity, equipment and production lines, products and processes, qualification certification, policy guidelines, geographical and manufacturing layout, risk and public opinion, and visit records.

[0021] And / or, in step S2, the construction of the enterprise knowledge graph is based on the upper-level ontology of the industrial chain, defining enterprise entities, capability or process entities, industrial chain link entities and the relationships between them, and expanding to include relationship types such as products or parts, equipment or production lines, capacity and delivery time, qualifications and compliance, geographical and manufacturing layout, policy adaptability, and risk signals.

[0022] As a preferred approach, the entity resolution and disambiguation in step S1 adopts a multi-feature voting mechanism based on lateral evidence features such as code consistency, name similarity, legal representative or equity or number of insured persons, and graph community consistency, and uses the stability of connected components as the final decision criterion.

[0023] As a preferred option, the path scoring in step S4.1 should include at least path length penalty, key node confidence weight, capacity and delivery time penalty, and geographical and manufacturing layout penalty, and set hard constraint filters for qualification or policy requirements.

[0024] As a preferred approach, the generative processing for enhanced retrieval in step S4.2 expands the query by using a thesaurus, process equivalence group and component-to-machine relationship dictionary during the recall phase, and sets high weight factors for necessary qualifications and compliance clauses during the fine ranking phase.

[0025] As a preferred option, the consistency arbiter in step S4.3 includes a rule layer and a learning layer; the rule layer is responsible for reachability verification, taboo edge removal, and chain position order verification; the learning layer performs feature concatenation on the graph path score and semantic score based on historical successful and failed samples and outputs the calibrated comprehensive score; when any hard constraint of the rule layer is not satisfied, the candidate is directly rejected.

[0026] As a preferred method, the method also includes step S6: when events such as business registration changes, capacity and delivery date adjustments, production line migration, policy updates, risk sentiment and new records from visits are detected, the subgraph related to the event is located, topology pruning and local recalculation are performed, and the graph path score, semantic score, comprehensive score and evidence chain are updated online.

[0027] And / or, the method also includes step S7: writing back the matching results, transaction or landing feedback and expert revisions through the cockpit and mobile assistant, and updating the meta-path weight, consistency arbitration threshold and entity alignment rules online.

[0028] As a preferred option, each triple in the evidence chain includes the source type, collection time, extraction method, and field-level confidence level. The field-level confidence level is calculated based on the source credibility weight and time decay, and is weighted and aggregated at the path and candidate levels for report display and audit traceability.

[0029] As a preferred approach, event-driven incremental maintenance listens through an event subscription interface, determines the local recalculation range based on the multi-hop neighborhood of the affected entity and the whitelist of relationship types, and the default recalculation neighborhood is one to three hops; when pruning fails or dependency conflicts occur, a full recalculation rollback strategy is triggered.

[0030] As a preferred option, the cockpit is used for batch task and strategy template publishing, while the mobile assistant is used for text, voice and image collection during visits and one-click write-back. Both ends share multi-tenant and hierarchical permission control policies.

[0031] Furthermore, the present invention also provides an inter-enterprise matching management system based on knowledge graphs, comprising:

[0032] The data aggregation and entity alignment module is used to execute step S1.

[0033] The graph modeling and construction module is used to perform step S2 and persist the enterprise knowledge graph with chain position constraints in the graph database.

[0034] The subgraph positioning module is used to generate demand subgraphs and supply subgraphs based on demand and candidate supply.

[0035] The matching engine module includes a graph path scoring unit and a semantic scoring unit, which are used to obtain graph path scores and semantic scores, respectively.

[0036] The consistency arbitration module is used to perform joint confidence assessment and conflict detection on the two types of scores. It performs rule verification based on reachability, taboo edges and chain position order and outputs comprehensive scores and rankings.

[0037] The evidence chain and confidence propagation module is used to generate evidence chain subgraphs and perform field-level confidence propagation and aggregation, outputting an interpretable matching report.

[0038] The event-driven incremental maintenance module is used for event listening, subgraph location, topology pruning, and local recalculation.

[0039] The cockpit linkage module and mobile assistant module are used for result push, feedback write-back, and online parameter updates.

[0040] As a preferred option, the graph path scoring unit supports at least the following meta-path sets: enterprise to capability to chain position, enterprise to product to capability to qualification, enterprise to equipment or production line to capacity to delivery date, and provides path length penalties and chain position hard constraints.

[0041] As a preferred embodiment, the consistency arbitration module includes a rule layer and a learning layer; the rule layer performs reachability verification, taboo edge removal, and chain position order verification; the learning layer calibrates and fuses the graph path score and semantic score, and rejects candidates that fail the rule layer verification.

[0042] As a preferred approach, the evidence chain and confidence propagation module records the source type, collection time, and extraction method for each field, and calculates the field-level confidence based on source weight and time decay. Subsequently, it aggregates the data at the path and candidate levels for report display, auditing, and backtracking.

[0043] As a preferred option, the event-driven incremental maintenance module includes an event subscription submodule, an affected subgraph location submodule, a topology pruning submodule, and a local recalculation submodule; when the affected range crosses a predetermined multi-hop neighborhood or a dependency conflict occurs, a full recalculation rollback process is triggered.

[0044] As a preferred option, the cockpit linkage module and the mobile assistant module support multi-tenant isolation, hierarchical access control and audit logs, and use transaction or landing feedback and expert revisions to update meta-path weights, consistency arbitration thresholds and entity alignment rules online.

[0045] Furthermore, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method.

[0046] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0047] Compared with existing technologies, this invention achieves a synergistic improvement in accuracy, interpretability, and timeliness within the same framework: Under chain position constraints, graph metapath scoring and text semantic retrieval enhancement are generated and integrated in parallel, and consistency arbitration is performed through rules such as reachability, taboo edges, and chain position order, significantly suppressing semantic illusions and unauthorized matching, thus improving accuracy and reliability; Field-level evidence chains and confidence levels are introduced, making each match traceable to specific data sources and timestamps, facilitating auditing, expert verification, and rapid error correction; Event-driven subgraph positioning and topology pruning are adopted, performing only local recalculation on affected areas, significantly reducing computational overhead and shortening the delay from data change to result update; Through cockpit and mobile terminal write-back, transactions and expert feedback are transformed into online updates of path weights and arbitration thresholds, forming a closed loop of "matching, verification, and calibration," enabling the system to maintain stable, accurate, interpretable, and low-latency inter-enterprise matching capabilities even in scenarios with frequent changes in enterprise status and high data noise. Attached Figure Description

[0048] Figure 1 The system architecture is shown in the specific embodiments of the present invention.

[0049] Figure 2 This is a flowchart of the method of the present invention.

[0050] Figure 3 This is a flowchart of step S4 of the present invention. Detailed Implementation

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0052] I. Nouns and Conventions

[0053] A business entity refers to a legal person or business entity whose primary key is the Unified Social Credit Code. A capability or process entity refers to reusable manufacturing capabilities, key processes, or testing capabilities. A link in the industrial chain, also known as a chain position, refers to a standardized location within a specific industrial chain, such as raw materials, components, complete machines, or after-sales service. A chain of evidence refers to a traceable subgraph composed of data items involved in the matching determination, their source, and timestamps. An event refers to a change that triggers the recalculation of the graph and the matching process, including at least business registration changes, production line relocation, adjustments to capacity and delivery dates, changes in policy interpretation, risk-related public opinion, and new additions from site visits.

[0054] II. System Overall Architecture

[0055] like Figure 1As shown, the system adopts a layered microservice architecture, with a typical deployment including the following layers: data aggregation and cleaning layer, entity parsing and graph construction layer, matching and arbitration layer, evidence and incremental layer, application layer, and auxiliary facilities layer. The data aggregation and cleaning layer is used for data access, extraction, transformation, loading, data quality verification, and anomaly handling, and can be implemented using streaming and batch job engines. The entity parsing and graph construction layer includes a graph database, text extraction service, and ontology management module, used for entity parsing and disambiguation, text element extraction, and graph writing. The matching and arbitration layer includes a graph path matching engine, a semantic retrieval enhanced matching engine, and a consistency arbitrator, used to output candidate ranking and conclusions of approval, rejection, or pending review. The evidence and incremental layer includes an evidence chain generator, a field-level confidence propagator, and an event subscription and incremental recalculation scheduler, used for interpretation and timeliness updates. The application layer includes a dashboard and a mobile intelligent assistant. The dashboard is used for batch tasks, strategy template publishing, and audit export, while the mobile app is used for text, voice, and image collection during site visits and one-click write-back. Both ends share multi-tenant isolation and hierarchical permission control. Supporting facilities include at least vector retrieval, full-text retrieval, message queues, configuration center, and audit logs, used for retrieval performance, event distribution, parameter governance, and traceability auditing.

[0056] III. Example 1: Specific implementation of the method steps (e.g.) Figure 2 (As shown)

[0057] S1 Data Aggregation and Entity Analysis

[0058] 1) Data access and layering

[0059] A compliant access mechanism is established for nine types of data sources, including at least business registration, finance and production capacity, equipment and production lines, products and processes, qualification certifications, policy interpretations, geographical and manufacturing layout, risk and public opinion, and on-site visit records. Access methods support batch import, real-time API collection, message subscription, database change capture, and mobile collection. All raw data first enters the original destination area for read-only storage, then enters the transition processing area to complete column name and type alignment and basic validation, and finally enters the main data area to form detailed and summary data for parsing. The format of each source field is included in version control; destructive changes require gray-scale verification before switching.

[0060] 2) Standardization and Cleaning

[0061] Unified character encoding and standards were implemented, removing invisible characters and differences between full-width and half-width characters. Punctuation and capitalization were standardized, and time fields were standardized to a standard time format. Redundant suffixes were normalized for company names, misspellings and simplified / traditional character mappings were performed, and punctuation was standardized, creating a standard name and alias table and recording its source and time. Registered addresses and production or manufacturing addresses were hierarchically segmented and geocoded, generating latitude, longitude, and geographic grid codes. Manufacturing layouts used independent structured records, supporting multiple factories and multiple industrial parks. Industry codes were mapped to the latest national standards; products and components were unified into an internal category tree and aligned with the component-to-finisher mapping table; processes and capabilities were aligned to a standard terminology library. Production capacity used recent-period statistical values ​​while retaining dispersion; delivery time was standardized to days and supported minimum, regular, and peak commitment levels. Certificates and policy clauses were uniformly numbered, typed, issued by the issuing authority, effective date, expiration date, and status; policy definitions were extracted by clause number and necessary condition tags. Risk and public opinion elements were structured and enumerated, including severity level and handling status. Personal information recorded during visits is anonymized and made visible according to a strategy, and the original text is stored in controlled storage while retaining fingerprints.

[0062] 3) Text Structure Extraction

[0063] For unstructured texts such as business scope, product catalogs, process specifications, equipment lists, quality inspection or certification reports, bidding announcements, visit minutes, image text recognition results, and speech-to-text results, a hybrid process combining rule templates, domain dictionaries, and named entity recognition and relation extraction models is used to extract product models, materials, key dimensions, capabilities or process points, equipment models and key process windows, production capacity and delivery date commitments, certificate numbers and clause numbers, and geographical and manufacturing layout descriptions. Each extraction result must include at least field values, source, extraction method, and initial confidence level. Results below the threshold are marked as pending verification and sent to the manual verification platform.

[0064] 4) Entity resolution and disambiguation (composite key, multi-feature voting, community consensus)

[0065] The Unified Social Credit Code is used as the primary key. When it is missing or questionable, a composite key consisting of the name standardization result, administrative division code, and legal representative or website domain name is used as the candidate blocking key. Candidate generation rules include similar name fingerprints, differences in the first character and length of the name within a threshold, consistency of the geographical grid code prefix of the registered address, consistency of the website domain name, and identical or homophonous legal representatives. A multi-feature voting mechanism is used for candidate pairs, including at least code consistency, name similarity, lateral evidence features such as legal representative, equity, or number of insured persons, and graph community consistency.

[0066] Name similarity can be calculated using the standardized edit distance, denoted as . :

[0067] ,

[0068] Parameter definition: , The standardized string for the name; To edit distance; , The length of the string; The larger the value, the more similar the two numbers are.

[0069] Consistency of time series evidence, such as the number of insured persons, can be expressed using an exponential decay method, denoted as . :

[0070] ,

[0071] Parameter definition: , This is a sequence vector of the number of insured persons by month or quarter; It is a 2-norm; This is the scale parameter.

[0072] Graph community consistency can be measured using a two-hop internal connectivity penalty form, denoted as . :

[0073] ,

[0074] Parameter definition: , This represents the nodes of candidate companies in the subgraph representing the relationship between "equity, senior management, upstream and downstream suppliers, and geographical industrial park". This represents the number of hops in the shortest path. The attenuation scale.

[0075] The sum of scores for multi-feature voting is denoted as :

[0076] ,

[0077] Parameter definition: The characteristic number; For the first Each feature weight; This corresponds to a similarity or consistency score; The feature threshold; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0078] The following decision-making rules can be adopted for merging, splitting, and reviewing decisions based on the aggregated voting scores:

[0079] ,

[0080] Parameter definition: This is the merging threshold; The separation threshold; .

[0081] The final decision adopts the connectivity component stability criterion: nodes that meet the merging conditions are merged into enterprise clusters. If the merging results in multiple valid unified social credit codes within the same cluster, or significantly damages the original supporting network community structure, the merging is rolled back and the conflict is marked for manual adjudication; the adjudication results are written to the audit log.

[0082] 5) Base library generation and field liveness rules

[0083] A unique enterprise identifier is generated for each enterprise cluster to form the basic enterprise database. Field values ​​are selected based on a survival rule that prioritizes source, time freshness, and method reliability. Authoritative business registration authorities are given priority, followed by competent authorities, third-party authorities, enterprise self-reporting, and on-site records. Structured data retrieval is given priority, followed by manual verification, model extraction, image text recognition, or speech-to-text. More recent data has higher weight. Registered addresses and production addresses are managed concurrently. Qualifications are deduplicated by certificate number and their validity period is verified. Production capacity is taken from recent period statistics and the fluctuation range is retained. Old values ​​and reasons for overwriting are written to a conflict log.

[0084] 6) Quality control and abnormal handling

[0085] Perform integrity, legality, uniqueness, referential integrity and enumeration legality checks, and generate verification work orders for anomalies such as sudden capacity changes, coordinate drift without relocation events, certificates nearing or expiring, inconsistencies between equipment or production lines and capacity, and policy conflicts. The verification results are written back to the base database and events are triggered.

[0086] S2 map modeling and construction

[0087] 1) Body design and chain system

[0088] The top-level concepts of the graph should at least include enterprises, capabilities or processes, products or components, equipment or production lines, capacity and delivery time, qualifications and compliance, policy provisions, risk events, geographical regions, supply chain links, and evidence and events. To support chain-point constraints, a standardized set of chain points is established, and a chain-point reachability matrix and allowed adjacent chain-point pairs are maintained as the basic configuration for subsequent meta-path search and consistency verification. Chain-point reachability can be represented by matrix elements, denoted as […]. :

[0089] ;

[0090] Parameter definition: , For chain bit index; It is a chain reachability matrix; Setting it to 1 indicates that it is allowed from the chain bit. To the chain position Jump or link.

[0091] 2) Graph Pattern and Attribute Specification

[0092] A property graph model is used, with unified naming and attribute constraints for nodes and relationships. Node types include at least enterprise, capability or process, product or component, equipment or production line, capacity and delivery time, certificate and compliance, policy clause, risk event, geographical region, chain position, evidence and event. Enterprise nodes include a unique system identifier, a unique constraint of the unified social credit code, standard name, alias set, industry code, status, creation time, update time, and tenant identifier. Capability or process nodes include name, category, parameter mode, and applicable product types. Product or component nodes include category, model, key specifications, and chain position code. Equipment or production line nodes include type, model, factory location, supported process set, and rated capacity. Capacity and delivery time nodes include cycle, capacity value, unit, delivery days, effective and expiration time, etc. Certificate nodes include a unique constraint of certificate number, certificate type, issuing authority, effective and expiration dates, and status. Policy clause nodes include clause code, summary, adaptation tag, geographical scope, and effective and expiration dates, etc. Risk event nodes include type, severity, occurrence time, status, and source, etc. Geographic region nodes include administrative division codes, names, center point coordinates, geographic grid prefixes, and boundary references. Link nodes include unique link code constraints, sequential indexes, upstream and downstream sets, etc. Evidence nodes include source type, source identifier, collection time, extraction method, confidence level, original reference location, and operator, etc. Event nodes include event type, associated entity references, occurrence time, and payload references, etc. Correspondence is established between enterprises and capabilities, products, equipment, qualifications, policies, risks, geography, and links; demand relationships are established between products and capabilities; attribution relationships are established between production capacity and delivery time and enterprises or equipment; evidence is linked to nodes or relationships, enabling node-level and relationship-level traceability. All nodes and edges uniformly carry source type, source identifier, collection time, version number, tenant identifier, and most recent event identifier to support traceability, version replay, and multi-tenant governance.

[0093] 3) Data Plotting and Incremental Construction

[0094] Full batch construction reads standard data from the enterprise master data and evidence repository, writes nodes and relationships idempotently centered on the enterprise, and simultaneously generates evidence nodes to form a minimum evidence chain. Incremental updates locate affected entities through event subscription, perform partial updates of nodes and relationships according to event type, and write evidence and event records. Chain position determination adopts a combination of rules and models. The rules are based on the scoring thresholds of the dominant product chain position code, process mapping, and equipment type mapping. Conflicts are marked as pending review. After a chain position change, the path cache related to the chain position is refreshed. Performance is improved through indexes and constraints, and replayable auditing is achieved through a dual time dimension of construction version number and event time. After construction, graph quality rules are executed and repairable issues are automatically fixed, while non-decision-making issues are pushed to the manual workbench.

[0095] S3 Subgraph Positioning

[0096] The goal of S3 is to transform a single set of requirements into a computable graph query, producing a set of demand subgraphs and a set of supply subgraphs. The inputs are the graph and the description of the current requirement; the outputs are the demand subgraph, the set of supply subgraphs, and a candidate list. The requirement profile can be uniformly represented as a set of requirement points. :

[0097] ;

[0098] Parameter definition: For the first One demand point; As weight; For constraint type, This indicates a hard constraint. Indicates soft constraints; This represents the number of demand points.

[0099] After demand analysis, the execution unit and specifications are standardized, terminology is standardized, and category tree mapping is performed. Hard constraints and soft constraints are separated, labeled, and weighted to form a demand subgraph as a query blueprint. The supply-side initial screening performs hard filtering on enterprises, including chain position compatibility, valid qualifications, policy conditions met, geographical scope satisfied, production capacity and delivery time satisfied, and exclusion of forbidden edges and major risks. For each enterprise that passes the initial screening, a one- to two-hop supply subgraph is constructed, including at least chain position, main product, capacity or process, equipment and production line, near-cycle production capacity and delivery time, valid qualifications and evidence, policy tags, geographical nodes, and risk events. Basic features and evidence citations are calculated to provide input for S4's dual-channel matching and arbitration.

[0100] S4 Dual-Channel Matching and Consistency Arbitration

[0101] like Figure 3As shown, the inputs to S4 include a demand subgraph, a supply subgraph set, a chain reachability matrix, a meta-path whitelist, a corporate text corpus, a thesaurus of synonyms and equivalent processes, a list of qualifications and policy hard constraints, a list of forbidden edges, and thresholds and parameters. The output is a Top-K list of candidate companies, including a comprehensive score, graph path score, semantic score, consistency verification conclusion, key evidence chains, and reasons for rejection or pending review.

[0102] 1) Map Path Scoring Channel

[0103] Capability or process coverage can be denoted as :

[0104] ,

[0105] Parameter definition: Candidate companies; It is a collection of a company's capabilities or processes; Indicate demand points Whether it falls under the category of skill-based requirements; the rest is the same as before.

[0106] Capacity fit can be denoted as :

[0107] ;

[0108] Parameter definition: This refers to the company's available production capacity within the validity period; To meet demand and production capacity.

[0109] Delivery time matching can be denoted as :

[0110] ;

[0111] Parameter definition: We commit to a delivery time of several days for businesses; This refers to the required delivery time in days; The attenuation scale.

[0112] Geographic time-limited penalties can be recorded as :

[0113] ;

[0114] Parameter definition: The distance from the manufacturing point to the demand location for the enterprise; For the scale.

[0115] Under the gating of chain position constraints, forbidden edges, and hard filtering by qualification policies, a single hit path The path score can be denoted as :

[0116] ;

[0117] Parameter definition: Indicates whether the path satisfies the hard constraints; As a weighting factor; For penalty weighting; This represents the number of nodes in the path. Points will be deducted for risk. , , , This refers to the corresponding indicator for that company.

[0118] Enterprise path aggregation can be denoted as :

[0119] ;

[0120] Parameter definition: For enterprises The former High-resolution path; For diversity reward coefficient; This is the set of metapath types for these paths; The type distribution entropy.

[0121] 2) Semantic retrieval enhancement channel

[0122] The candidate enterprise text corpus is cleaned and segmented, and inverted and vector indexes are built. The query is expanded using synonyms, process equivalence groups, and a component-to-whole-machine dictionary. After recall and reordering, the satisfaction level of each segment is evaluated. The weighted summary of semantic satisfaction can be denoted as... :

[0123] ;

[0124] Parameter definition: For semantic evaluators to enterprises At the point of demand Satisfaction confidence; The weight of the demand point.

[0125] 3) Consistency Arbitration

[0126] The difference between the two channels can be denoted as :

[0127] ;

[0128] Parameter definition: The difference between graph path segmentation and semantic segmentation is used for conflict detection.

[0129] Learning layer fusion calibration can be performed using logistic regression, outputting a comprehensive score or success probability. :

[0130] ;

[0131] Parameter definition: ; These are model parameters; For evidence coverage; As an indicator of the freshness of evidence; Points will be deducted for risk.

[0132] The rule layer first performs reachability checks, taboo edge removal, chain order checks, qualification and policy hard constraint checks, and key field validity checks. If any hard constraint is not met, the application is rejected directly. The learning layer then processes the candidate outputs that have passed the rule layer. And provide a graded conclusion such as recommended, pending review, or not recommended; when When the conflict threshold is exceeded, mark the conflict as pending review and output the gap diagnosis.

[0133] S5 Evidence Chain and Confidence Propagation

[0134] The goal of S5 is to structure, traceably display, and quantify the confidence level of the matching criteria in S4, forming a three-level confidence propagation system encompassing fields, paths, and candidates. Evidence nodes must at least record the source type, source identifier, collection time, extraction method, original confidence level, and visibility level, and be linked to nodes or relationships through evidence pointing relationships.

[0135] The confidence level of a single piece of evidence can be normalized by source weight, method weight, and time decay, denoted as . :

[0136] ;

[0137] Parameter definition: Source weight; Weights for extraction methods; Time weighting; Original confidence level; .

[0138] Time weights can be expressed in an exponentially decaying form, denoted as . :

[0139] ;

[0140] Parameter definition: This refers to the number of days since the evidence was collected. The attenuation scale.

[0141] Multi-evidence fusion for the same field can adopt the non-overlapping probability form of independent corroboration, with field-level confidence denoted as... :

[0142] ;

[0143] Parameter definition: To support fields The collection of evidence; Believe each piece of evidence; When evidence is contradictory and cannot be explained by time changes, the field is downgraded and marked as conflict pending verification; the field corresponding to expired or invalid certificates is directly marked as unavailable.

[0144] Path-level confidence adopts a conservative form of weighted aggregation of the lower bound of hard constraint fields and general fields, denoted as... :

[0145] ;

[0146] Parameter definition: For hard constraint fields; A general collection of fields; , Confidence for the field; This represents the lower limit percentage. For general field weights and satisfying .

[0147] Candidate-level comprehensive confidence and fusion score linkage, denoted as :

[0148] ;

[0149] Parameter definition: The arbitration output is a fusion score; For path confidence summarization; For evidence coverage; To adjust the index.

[0150] Based on this, the system generates an interpretable report, which includes at least a comprehensive score, a graph path score, a semantic score, a comprehensive confidence level, a critical path and evidence catalog, and a conflict and gap list, and supports audit traceability and export.

[0151] S6 event-driven incremental maintenance

[0152] When events such as business registration changes, capacity and delivery date adjustments, production line relocations, policy updates, risk-related public opinion events, new entries from site visits, and text data entry or manual verification are detected, the system listens for these events through event subscriptions and performs impact analysis, subgraph location, topology pruning, and partial recalculation. Affected subgraphs can be identified by the entity linked to the event. of The skip neighborhood is defined as follows: :

[0153] ;

[0154] Parameter definition: The shortest path hop count in the graph; The radius of the neighborhood is preferably 1 to 3; The set of affected nodes.

[0155] The system in Within the scope, expired certificates, illegal chain connections, and forbidden edges are removed. The graph path score, semantic score, fusion score, and evidence chain of the affected candidates are recalculated, and the cache is precisely invalidated. A full recalculation rollback strategy is triggered when pruning fails or dependency conflicts occur. The recalculation results are published using a shadow index and atomic version number switching method to avoid intermediate states affecting online services.

[0156] S7 Closed-Loop Learning and Collaborative Release

[0157] The system writes back matching results, transaction or implementation feedback, expert revisions, manual verification, and user behavior logs to update meta-path weights, consistency arbitration thresholds, and entity alignment rules online. It also performs version control on synonyms, process equivalence groups, component-to-machine dictionaries, and strategy templates. Meta-path weights can be updated using an exponential sliding update method, denoted as... :

[0158] ;

[0159] Parameter definition: For the first Online weighting; The suggested weights are derived from the attribution of recent successful samples; For smoothing coefficients, Larger updates are more robust. Strategy and model releases employ canary releases and rollback mechanisms to ensure online stability and consistent output.

[0160] Application Example: Inter-enterprise matching and closed-loop verification of new energy component supply needs

[0161] The municipal industrial and information technology department spearheaded a special action to "cultivate specialized, refined, and innovative industries and connect them with the industrial chain," establishing a knowledge graph of enterprises and launching an inter-enterprise matching management system for the new energy vehicle and energy storage industrial chains. The pilot program covered three districts and counties and six industrial parks, including 12,460 enterprises, of which 7,930 were manufacturing companies. Data sources accessed by the system included: business registration databases, output and capacity reports from tax or statistical perspectives, park equipment registration and production line ledgers, enterprise product catalogs and process descriptions, quality system and industry certification databases, policy clause databases, park geographical and logistics timeliness data, judicial and penalty data, public opinion data, and data recorded during mobile site visits. To evaluate the technical effectiveness of this invention, a real-world task was selected for comparative testing, and the results were tagged as "actual connection results and expert review conclusions."

[0162] 1. Requirements, tasks, and target chain

[0163] The demander is an energy storage system integrator, posting a supplier selection request for "aluminum alloy battery enclosures and endplate components." Key requirements include: 6xxx series aluminum alloy material; extrusion, welding, and finishing capabilities; monthly production capacity of no less than 8,000 sets; standard delivery time of no more than 21 days; ISO9001 certification required, with preference given to IATF16949 certification; production location within a 250km logistics radius; and exclusion of companies with significant quality penalties or breaches of trust in the past year. The system maps this demand to a set of demand points. ,in The hard constraints include "ISO9001 valid", "chain position is a component or its adjacent compatible", "capacity and delivery time thresholds", and "risk taboo edge filtering", etc.; the soft constraints include "IATF16949", "shorter delivery time", "closer distance", and "more sufficient evidence", etc. The target chain position is set to "component".

[0164] 2. Method Comparison and Test Setup

[0165] Three sets of plans were set up for comparison:

[0166] Group A (this invention): Employs "knowledge graph chain position constraints, graph path scoring, semantic retrieval enhancement scoring, consistency arbitration, evidence chain and belief propagation, event incremental maintenance, and closed-loop learning".

[0167] Group B (Graph Single-Channel Comparison): Only knowledge graph meta-path retrieval and path scoring are used, without enabling semantic retrieval enhancement channels and consistency arbitration.

[0168] Group C (Semantic Single-Channel Comparison): Only performs semantic retrieval and large model evaluation on unstructured enterprise files, without enabling graph path reasoning and reachability verification with chain position constraints.

[0169] Evaluation metrics include: Top-10 accuracy (the percentage of Top-10 companies deemed by experts to "meet hard constraints and recommended for matching"), hard constraint false pass rate (the percentage of companies that do not meet hard constraints but still make it into the Top-10), average decision-making time, evidence availability rate (the percentage of Top-10 companies that can output an "evidence chain diagram"), and ranking stabilization recovery time after event changes (the time from the event to the updated availability). Additionally, the "final matching result" is recorded: the number of companies that entered into substantive negotiations and signed a framework agreement.

[0170] 3. Key Calculation Process

[0171] The system first performs entity resolution and disambiguation in S1, and then uses the similarity of company names to perform similarity analysis.

[0172] ,

[0173] And combined with the consistency of the lateral sequence

[0174] ,

[0175] Consistency with the community

[0176] ,

[0177] Multi-feature voting merging is used to avoid duplicate candidates, fragmented evidence, or erroneous exclusion caused by "duplicate entities" of the same enterprise under different system definitions.

[0178] S2 establishes the chain reachability matrix Link reachability via

[0179] ;

[0180] Constrain meta-path search to prevent structural errors such as "raw material companies being recommended as component suppliers without proper authorization".

[0181] The S4 graph channel provides information on each candidate company. Computing capacity coverage, production capacity matching, delivery time matching, geographical penalties and risk deductions, and for each path calculate

[0182] ;

[0183] Further aggregation yields enterprise path graphs.

[0184] ;

[0185] The semantic channel evaluates the satisfaction level of each slice of evidence item by item and summarizes them.

[0186] ;

[0187] The difference between the two channels is

[0188] ;

[0189] Ultimately, the arbitration output is integrated through rule-based prioritization and learning-calibration.

[0190] ;

[0191] In S5, evidence chains and confidence propagation are generated, and field confidence is...

[0192] ;

[0193] and

[0194] ;

[0195] Aggregation yields candidate-level confidence.

[0196] ;

[0197] The final output is "Top-10, Pass / Reject / Pending Review, Evidence Chain Subgraph, Gap Diagnosis".

[0198] 4. Results and Comparison

[0199] In this task, the expert group verified all candidates and identified 18 companies that "met all hard constraints and were recommended for integration." The results of the three sets of solutions are shown in the table below.

[0200] Table 1 Comparison of matching effects for required tasks (Top-10 level)

[0201]

[0202] As can be seen, the proposed solution improves the Top-10 accuracy by approximately 33% compared to Group B and by approximately 60% compared to Group C. The false pass rate of hard constraints is significantly reduced, mainly because the rule layer reachability, chain position order, taboo edges, and qualification validity checks of the proposed solution directly reject candidates that "semantically appear to match but structurally do not." Simultaneously, the evidence chain generation and confidence propagation of the proposed solution enable all Top-10 results to output auditable evidence chain subgraphs, significantly improving decision interpretability and review efficiency.

[0203] 5. Incremental maintenance effect under event triggering

[0204] On the 12th day after the requirement was released, the system received two types of events: first, the IATF16949 certificate status of candidate company E07 changed to expire; second, candidate company E03 added a new production line, increasing capacity and updating delivery commitments. The system mapped these events to event entities. ,according to The subgraph affected by skip neighborhood localization:

[0205] ;

[0206] Pick This invention performs partial recalculation only on affected fields, paths, and candidate sets, and uses shadow indexes to atomically switch and publish results. Actual testing showed that the time from event enqueuing to the visibility of rankings, evidence chains, and comprehensive scores was 58 seconds; if a full recalculation strategy were used, the average time measured under equivalent hardware resources would be 19 minutes and 40 seconds. Therefore, this invention significantly reduces maintenance costs and improves result timeliness through event-driven pruning and partial recalculation.

[0207] 6. Conclusion on Technical Effectiveness

[0208] Through the above application examples, under the same data source and candidate scale conditions, this invention can significantly improve the accuracy of inter-enterprise matching and reduce the false pass rate of hard constraints by simultaneously using graph path reasoning and semantic evidence evaluation to form a consistent arbitration, while satisfying chain position constraints and hard constraint filtering. It can also achieve interpretable and auditable output for each candidate through evidence chain and belief propagation, thereby improving the efficiency of expert review and implementation. Furthermore, it can achieve rapid updates from seconds to minutes through incremental maintenance under high-frequency events such as certificate expiration and capacity changes, avoiding the high resource consumption and service jitter caused by full recalculation. Thus, it achieves comprehensive and repeatable technical effects in terms of accuracy, interpretability, and timeliness.

[0209] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A knowledge graph-based inter-enterprise matching management method, characterized in that, Includes the following steps: S1. Obtain relevant enterprise information from multi-source data, and use the composite key of unified social credit code and name standardization to perform entity parsing and disambiguation on enterprise identifiers, forming a cleaned and standardized basic enterprise database. S2. Construct an enterprise knowledge graph with chain position constraints; In step S2, the enterprise knowledge graph is constructed based on the upper-level ontology of the industry chain, defining enterprise entities, capability or process entities, and entities in the industry chain links, as well as the relationships between them. This is further expanded to include relationship types such as products or components, equipment or production lines, capacity and delivery time, qualifications and compliance, geographical and manufacturing layout, policy adaptability, and risk signals. To support chain position constraints, a standardized set of chain positions is established, and a chain position reachability matrix and allowed adjacent chain position pairs are maintained as the basic configuration for subsequent meta-path search and consistency verification. Chain position reachability is represented by matrix elements, denoted as... : ； Parameter definitions: i and j are chain indexes; A is the chain reachability matrix; A ij ∈{0,1}, taking 1 indicates that a jump or connection from link i to link j is allowed; S3. Extract structured fields and text from the target requirements to obtain a requirement subgraph corresponding to the requirements; aggregate potential supply-side enterprise elements according to chain position constraints to locate the supply subgraph. S4, Dual-channel matching and consistency arbitration: S4.1 Perform meta-path search with chain position constraints on the knowledge graph and score the paths to obtain graph path scores; S4.

2. Using the demand description as the query and the unstructured files of candidate companies as the corpus, perform a generative process to enhance retrieval and obtain semantic scores. S4.

3. The graph path score and semantic score are jointly assessed and conflict detected by the consistency arbitrator. The verification is performed based on reachability, taboo edges, and chain position order rules. The comprehensive score and ranking of the candidate enterprise set are output, and the candidates that do not meet the consistency conditions are rejected. The path scoring in step S4.1 includes at least path length penalty, key node confidence weight, capacity and delivery time penalty, and geographical and manufacturing layout penalty, and sets hard constraint filtering for qualification or policy requirements; In step S4.2, the enhanced generative processing for retrieval employs query expansion during the recall phase, including a thesaurus, a price group of processes, and a component-to-machine relationship dictionary, and sets high-weight factors for necessary qualifications and compliance clauses during the fine-ranking phase. S5. For each candidate company, generate an evidence chain subgraph containing entities, relationships, data sources, and timestamps, and propagate and normalize the field-level confidence scores involved in the matching from bottom to top to form an interpretable matching report.

2. The method according to claim 1, characterized in that, The multi-source data in step S1 includes business registration, finance and production capacity, equipment and production lines, products and processes, qualification certifications, policy guidelines, geographical and manufacturing layout, risk and public opinion, and visit records.

3. The method according to claim 1, characterized in that, In step S1, entity resolution and disambiguation employ a multi-feature voting mechanism that includes code consistency, name similarity, lateral evidence features such as legal representative, equity or number of insured persons, and graph community consistency, and uses the stability of connected components as the final decision criterion.

4. The method according to claim 1, characterized in that, In step S4.3, the consistency arbiter includes a rule layer and a learning layer. The rule layer is responsible for reachability verification, taboo edge removal, and chain position order verification. The learning layer performs feature concatenation on the graph path score and semantic score based on historical successful and failed samples and outputs a calibrated comprehensive score. Candidates are directly rejected when any hard constraint of the rule layer is not met.

5. The method according to claim 1, characterized in that, The method also includes step S6: when an event such as business registration change, capacity and delivery date adjustment, production line migration, policy update, risk sentiment and newly added events recorded during visits are detected, the subgraph related to the event is located, topology pruning and local recalculation are performed, and the graph path score, semantic score, comprehensive score and evidence chain are updated online. And / or, the method also includes step S7: writing back the matching results, transaction or landing feedback and expert revisions through the cockpit and mobile assistant, and updating the meta-path weight, consistency arbitration threshold and entity alignment rules online.

6. The method according to claim 5, characterized in that, Each triple in the chain of evidence includes the source type, collection time, extraction method, and field-level confidence level. The field-level confidence level is calculated based on the source credibility weight and time decay, and is weighted and aggregated at the path and candidate levels for report display and audit traceability. And / or, event-driven incremental maintenance listens through an event subscription interface, determines the local recalculation range based on the multi-hop neighborhood of the affected entity and the whitelist of relationship types, and the default recalculation neighborhood is one to three hops; when pruning fails or dependency conflicts occur, a full recalculation rollback strategy is triggered. And / or, the cockpit is used for batch task and strategy template publishing, and the mobile assistant is used for text, voice and image collection during visits and one-click write-back, with both ends sharing multi-tenant and hierarchical permission control policies.

7. A knowledge graph-based inter-enterprise matching management system, characterized in that, The system is used to implement the method according to any one of claims 1-6, comprising: The data aggregation and entity alignment module is used to perform step S1; The graph modeling and construction module is used to perform step S2 and persist the enterprise knowledge graph with chain position constraints in the graph database. The subgraph positioning module is used to generate demand subgraphs and supply subgraphs based on demand and candidate supply. The matching engine module includes a graph path scoring unit and a semantic scoring unit, which are used to obtain graph path scores and semantic scores, respectively. The consistency arbitration module is used to perform joint confidence assessment and conflict detection on the two types of scores. It performs rule verification based on reachability, taboo edges and chain position order and outputs comprehensive scores and rankings. The evidence chain and confidence propagation module is used to generate evidence chain subgraphs and perform field-level confidence propagation and aggregation, outputting an interpretable matching report. The event-driven incremental maintenance module is used for event listening, subgraph location, topology pruning, and local recalculation. The cockpit linkage module and mobile assistant module are used for result push, feedback write-back, and online parameter updates.

8. The system according to claim 7, characterized in that, The graph path scoring unit supports at least the following meta-path sets: enterprise to capability to chain position, enterprise to product to capability to qualification, enterprise to equipment or production line to capacity to delivery date, and provides path length penalty and chain position hard constraint. And / or, the consistency arbitration module includes a rule layer and a learning layer; the rule layer performs reachability verification, taboo edge removal, and chain bit order verification; The learning layer calibrates and fuses the graph path score and semantic score, and rejects the output of candidates that fail the rule layer's verification. And / or, the evidence chain and confidence propagation module records the source type, collection time and extraction method for each field, and calculates the field-level confidence based on the source weight and time decay, and then aggregates it at the path and candidate levels for report display, auditing and backtracking; And / or, the event-driven incremental maintenance module includes an event subscription submodule, an affected subgraph location submodule, a topology pruning submodule, and a local recalculation submodule; When the affected area spans a predetermined multi-hop neighborhood or a dependency conflict occurs, a full recalculation rollback process is triggered. And / or, the cockpit linkage module and mobile assistant module support multi-tenant isolation, hierarchical access control and audit logs, and use transaction or landing feedback and expert revisions to update meta-path weights, consistency arbitration thresholds and entity alignment rules online.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method of any one of claims 1-6.

Images