Supplier combination recommendation method, medium and device based on knowledge graph

Abstract

This application provides a knowledge graph-based supplier combination recommendation method, medium, and device, belonging to the field of data processing technology. The method includes: acquiring a set of procurement demand information; filtering a set of candidate suppliers that meet the procurement demand information set and the matching degree between each candidate supplier and each procurement item demand from a preset supplier knowledge graph; forming multiple candidate supplier combinations based on the procurement item demand met by each candidate supplier; identifying whether there is collaborative information between the candidate suppliers, and calculating the collaborative weight between the corresponding suppliers based on the collaborative information; calculating a combination score for each candidate supplier combination based on the matching degree of the candidate suppliers and the collaborative weight; and generating supplier combination recommendation information from the final candidate supplier combinations based on the combination score. This application can improve the accuracy of supplier combination recommendations.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, medium, and apparatus for recommending supplier combinations based on knowledge graphs. Background Technology

[0002] With the rapid development of the market economy and the continuous expansion of enterprise operations, the complexity of enterprise procurement is increasing daily, the types of goods procured are becoming more diverse, and supplier resources are experiencing explosive growth. Especially in the procurement of standardized goods (such as office supplies and computer accessories), enterprises need to select high-quality partners from a vast number of suppliers. Procurement efficiency and the scientific nature of decision-making directly affect enterprise operating costs, supply chain stability, and core competitiveness.

[0003] As supplier networks continue to expand and become more complex, enterprise supplier recommendation methods are increasingly focusing on constructing supplier knowledge graphs based on transaction data and evaluation metrics. These knowledge graphs are then used to analyze the matching between each supplier and current procurement needs, resulting in supplier recommendations. However, enterprises often require more than one supplier, and cooperation between suppliers can impact actual procurement efficiency. While existing supplier recommendation methods are effective for individual suppliers, they still exhibit significant discrepancies between recommendations and actual needs for procurement requiring collaboration from multiple suppliers, thus requiring further improvement. Summary of the Invention

[0004] The purpose of this application is to provide a knowledge graph-based supplier portfolio recommendation method, medium, and device to solve at least one of the above-mentioned technical problems.

[0005] In a first aspect, this application provides a supplier combination recommendation method based on knowledge graphs, the method comprising:

[0006] Obtain a collection of procurement demand information;

[0007] From a preset supplier knowledge graph, a set of candidate suppliers that meet one or more procurement item requirements of the procurement requirement information set are selected, along with the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement.

[0008] Based on the procurement requirements satisfied by each candidate supplier, multiple candidate supplier combinations are formed, and each candidate supplier combination satisfies all procurement requirements in the procurement requirements information set.

[0009] Identify whether there is any collaboration information among the candidate suppliers, and calculate the collaboration weight between the corresponding suppliers based on the collaboration information;

[0010] A combined score for each candidate supplier combination is calculated based on the matching degree of the candidate suppliers and the collaborative weight.

[0011] Supplier portfolio recommendation information is generated from the final candidate supplier portfolio based on the combined score.

[0012] Optionally, calculating the collaboration weights between the corresponding suppliers based on the collaboration information includes:

[0013] Using suppliers as nodes and collaboration information between suppliers as edges of the nodes, the edge weights are calculated based on the collaboration information to construct a supplier collaboration graph.

[0014] Based on the supplier collaboration diagram and the procurement demand information set, a supplier group with collaborative relationships is generated;

[0015] The collaboration weights between suppliers are calculated based on the supplier collaboration graph and the supplier groups.

[0016] Optionally, generating supplier groups with collaborative relationships based on the supplier collaboration graph and the procurement demand information set includes:

[0017] Based on the procurement demand information set, the edge weights in the supplier collaboration graph are corrected to obtain the supplier corrected collaboration graph.

[0018] The nodes are aggregated according to the supplier modified collaboration graph to form a supplier group.

[0019] Optionally, the step of aggregating nodes according to the supplier modified collaboration graph to form a supplier group includes:

[0020] Step 1: Treat each node in the supplier modification collaboration graph as an independent community;

[0021] Step 2: Calculate the modularity gain brought about by a node in the supplier modified collaboration graph moving to its neighboring community;

[0022] Step 3: If the modularity gain is greater than the preset gain threshold, move the node to its neighboring community; otherwise, refuse to move the node to its neighboring community, thus forming a new community distribution.

[0023] Repeat steps 2 and 3 iteratively until no module gain exceeds the gain threshold or the number of iterations reaches the preset iteration threshold.

[0024] Suppliers in the latest community distribution are identified as the corresponding supplier groups.

[0025] Optionally, the collaboration information includes one or more collaboration events between suppliers, each collaboration event including a corresponding collaboration project, collaboration time, and event intensity value; the step of calculating the edge weight based on the collaboration information includes:

[0026] The initial weights between suppliers for each collaborative event are calculated based on the collaborative projects and event intensity values ​​between suppliers.

[0027] Calculate the corresponding weight decay value based on the collaboration time in the collaboration event;

[0028] The edge weights between suppliers are calculated based on the initial weight of each collaborative event and the weight decay value.

[0029] Optionally, calculating the combined score for each candidate supplier combination based on the matching degree of the candidate suppliers and the collaborative weight includes:

[0030] Identify the importance of candidate suppliers within the set of procurement demand information;

[0031] A base score for each candidate supplier combination is calculated based on the matching degree and the importance.

[0032] Identify whether any candidate suppliers in the candidate supplier portfolio belong to the same supplier group. If so, weight the portfolio's basic score based on the collaborative weight to obtain the portfolio score of the candidate supplier portfolio.

[0033] Optionally, the step of filtering a set of candidate suppliers from a preset supplier knowledge graph that meet one or more procurement item requirements of the procurement demand information set, and the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement, includes:

[0034] From the preset supplier knowledge graph, a set of candidate suppliers that meet one or more procurement item requirements that meet the procurement requirement information set are selected to form a candidate supplier subgraph;

[0035] An initial score vector is extracted from each candidate supplier from the candidate supplier subgraph, where each parameter in the initial score vector represents an evaluation metric for one dimension of the procurement requirement;

[0036] Calculate the dimension weight of each dimension in the initial score vector based on the requirements for each type of purchased item;

[0037] The matching degree is calculated based on the dimension weights and the initial score vector.

[0038] Optionally, extracting the initial score vector for each candidate supplier from the candidate supplier subgraph includes:

[0039] Based on the candidate supplier subgraph, an N×N adjacency matrix A and an N×F feature matrix X are constructed, where N is the total number of nodes in the candidate supplier subgraph, F is the initial feature dimension of each node, and the elements a in the adjacency matrix are... ij Used to reflect the connection relationship between node i and node j in the candidate supplier subgraph;

[0040] The adjacency matrix A and the feature matrix F are input into a preset graph neural network model for nonlinear transformation processing, and the initial score vector of each candidate supplier is output.

[0041] In a second aspect, this application provides a computer-readable storage medium storing executable instructions that, when executed by a processor, cause the processor to perform the method described in any embodiment of this application.

[0042] In a third aspect, this application provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the method as described in any embodiment of this application.

[0043] The knowledge graph-based supplier combination recommendation method, medium, and device in this application, after obtaining suppliers that meet procurement requirements based on the supplier knowledge graph, further analyzes the collaboration information between various candidate suppliers. Based on this collaboration information, it analyzes the advantages and disadvantages of various candidate supplier combinations and assigns collaboration weights. Based on these collaboration weights, it calculates a combination score for each supplier combination and finally provides recommendation information. This application can improve the accuracy of supplier combination recommendations, making the provided supplier combinations more closely aligned with procurement needs. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0045] Figure 1 This is a flowchart illustrating a knowledge graph-based supplier combination recommendation method in one embodiment.

[0046] Figure 2 This is a schematic diagram of a process for calculating the collaboration weights between corresponding suppliers based on the collaboration information in one embodiment;

[0047] Figure 3 This is a schematic diagram of the process for calculating edge weights based on the cooperation information in one embodiment;

[0048] Figure 4 This is a flowchart illustrating the process of selecting a set of candidate suppliers from a preset supplier knowledge graph that meet one or more procurement item requirements from the procurement requirement information set, and the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement, in one embodiment.

[0049] Figure 5 This is a schematic diagram of the structure of an electronic device in one embodiment. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0052] For example, the terms "first," "second," etc., used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element.

[0053] For example, the terms "comprising" or "including" used in this application indicate the presence of features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0054] This application provides a knowledge graph-based supplier combination recommendation method, such as... Figure 1 As shown, the method includes:

[0055] Step 110: Obtain the set of procurement demand information.

[0056] In this embodiment, the procurement demand information set refers to the sum of all information required by an enterprise to complete a procurement task. It can include various types of information such as basic information about the procured items, procurement constraints, and core procurement objectives. Specifically, the basic information of each procured item, along with its corresponding procurement constraints and core procurement objectives, constitutes a procurement item demand. That is, the procurement demand information set can contain one or more procurement item demands. Basic item information can include fundamental attributes such as the specifications, quantity, and category of the corresponding item, ensuring clear matching criteria for supplier selection. Procurement constraints can include hard constraints such as delivery time, budget limits, and compliance requirements for the corresponding item or the entire procurement demand, excluding suppliers who do not meet the basic conditions. Core procurement objectives can be, for example, cost priority, quality priority, or delivery priority, providing weighting guidance for subsequent matching degree calculations and combined scoring.

[0057] For example, a manufacturing company needs to purchase 100 laptops, 20 laser printers, and 50 boxes of office paper for production. Its procurement requirements include: 1. Laptops - Quantity 100 units - Specifications {Laptops must have i7 processor / 16GB RAM}; 2. Printers - Quantity 20 units - Specifications {Supports duplex printing and color printing}; 3. Office paper - Quantity 50 boxes - Specifications {A4 paper}. The total budget for the above procurement is ≤500,000 RMB, and the delivery deadline is within 30 days. The core objective is: quality first, followed by cost control.

[0058] Specifically, procurement needs can be extracted from standardized forms within the enterprise's internal procurement management system, forming a collection of procurement requirements information. These forms include mandatory fields (item name, specifications, quantity, delivery date) and optional fields (budget, special compliance requirements, target priority). After procurement personnel fill out the forms online, structured requirement data is automatically generated. For non-standardized needs (such as customized accessories), attachments (technical parameter documents) can be uploaded, and key parameters can be extracted and added to the procurement requirements using OCR technology. During the process of filling out and uploading the standardized procurement forms, further information verification and standardization can be performed. Electronic devices can have built-in verification rules for procurement needs, such as "the budget amount must match the procurement quantity × average market price (deviation not exceeding ±30%)" and "the delivery period must be ≥ the industry average delivery cycle (e.g., laptop procurement delivery cycle ≥ 7 days)." If the filled-out or uploaded standardized procurement form does not meet these verification rules, the relevant procurement personnel will be prompted to correct or review it. Simultaneously, unstructured descriptions in the requirements (e.g., "high-quality printer") are standardized into quantifiable indicators (e.g., "printer mean time between failures ≥ 1000 hours").

[0059] The standardized procurement forms can be categorized and stored according to requirements. For example, the collected requirement information can be stored in a relational database according to item type (item_info), constraint type (constraint_info), target type (target_info), etc., and a unique requirement ID can be generated to facilitate subsequent querying and association with supplier knowledge graph.

[0060] Step 120: Select a set of candidate suppliers from the preset supplier knowledge graph that meet one or more procurement item requirements of the procurement requirement information set, and determine the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement.

[0061] In this embodiment, the supplier knowledge graph is a pre-constructed structured knowledge graph centered on "entity-relationship-attribute". Entities include suppliers, purchased items, historical orders, and certifications; relationships include supply relationships between suppliers and items, certification relationships between suppliers and certifications, and historical cooperation relationships between suppliers and enterprises / suppliers; attributes include supplier delivery cycles and item specifications. For example, the knowledge graph of supplier A records: node: Supplier A; entity attributes: delivery cycle 10 days, laptop supply qualification, 2023 cooperation satisfaction score 92; relationship with the laptop entity: "Supply (supports i7 processor / 16GB memory model)".

[0062] The candidate supplier set represents a group of suppliers selected from the knowledge graph that meet the requirement of at least one item in the procurement needs. For example, for a procurement need of a laptop and a printer, the candidate set might include supplier A, which supplies only laptops; supplier B, which supplies only printers; and supplier C, which supplies both items.

[0063] Matching score indicates the degree to which a single candidate supplier matches the demand for one or more items. A higher matching score indicates a higher degree of matching. For example, supplier A has a matching score of 90 for "laptop requirements" (completely matching specifications, delivery cycle of 10 to 30 days, historical satisfaction score of 92), but a matching score of 0 for "printer requirements" (no printer supply qualifications).

[0064] Specifically, for each type of procurement item requirement in the procurement demand information set, the supplier knowledge graph is used to filter and query suppliers that meet the requirements, which are then selected as candidate suppliers. All the candidate suppliers retrieved form the candidate supplier set.

[0065] Upon obtaining each candidate supplier, the matching degree between the supplier and the corresponding procurement item requirements is analyzed based on the supplier's knowledge graph. Optionally, the electronic equipment pre-sets multiple matching degree evaluation dimensions, which may include one or more of the following: specification matching, delivery time suitability, quality compliance, and compliance. By analyzing the dimension matching values ​​between the candidate supplier and each evaluation dimension, the matching degree between the candidate supplier and the corresponding item can be calculated based on these dimension matching values. The matching degree can be a weighted sum of the matching values ​​of the above-mentioned dimensions. The weights of each evaluation dimension can be fixed or determined based on the procurement item requirements. The more important the relevant evaluation dimension under the procurement item requirements, the higher its corresponding weight. The dimension matching value is quantified by analyzing the supplier's historical performance and requirements under the corresponding dimension. The higher the dimension matching value, the better the relevant supplier can meet the procurement requirements under that dimension.

[0066] In one embodiment, algorithms such as cosine similarity can be used to calculate the information in each dimension (such as item requirements, delivery cycle, historical quality pass rate, etc.) into vectors. Based on these vectors and the relevant vectors in the knowledge graph of the candidate suppliers (such as the initial score vector below), pre-similarity calculations are performed to obtain the corresponding dimension matching values.

[0067] Step 130: Based on the procurement item requirements satisfied by each candidate supplier, multiple candidate supplier combinations are formed, and each candidate supplier combination satisfies all procurement item requirements in the procurement requirement information set.

[0068] In this embodiment, a candidate supplier combination refers to a group consisting of at least two candidate suppliers, and this group covers all procurement item requirements in the procurement requirement information set, meaning that there is at least one supplier that can supply each item. For example, for the requirement to purchase a laptop + printer, the candidate combinations may include: Supplier A (for laptops) + Supplier B (for printers), or Supplier A (for laptops) + Supplier D (for printers).

[0069] The electronic device has pre-defined procurement rules, including basic rules and optimization rules for generating combinations. The basic rules may include those that need to cover all procured items and ensure that there is no competitive conflict between suppliers within the combination. The optimization rules may include rules that prioritize suppliers with fewer suppliers and rules that avoid dependence on a single supplier.

[0070] Based on established procurement rules, electronic equipment manufacturers create a subset of candidate suppliers for each type of procured item. By arranging and combining candidate suppliers from different subsets, they form combinations that can cover all procured items. Understandably, there are various ways to combine these candidate suppliers.

[0071] Step 140: Identify whether there is any collaboration information among the candidate suppliers, and calculate the collaboration weight between the corresponding suppliers based on the collaboration information.

[0072] In this embodiment, the electronic device further includes collaboration information between suppliers, which refers to cooperation data generated between candidate suppliers in historical procurement activities. Each instance of cooperation data constitutes a collaboration event. Each collaboration event includes a collaboration project, collaboration time, and event intensity value. A collaboration project can represent multiple suppliers working together to complete a single procurement, or collaboration between multiple suppliers themselves. It is the specific business scenario that generates a collaborative relationship between suppliers, including joint order fulfillment, joint R&D, and supporting supply. Different types of collaboration projects contribute differently to the degree of supplier collaboration. The event intensity value is a quantitative score of the completion quality, collaboration depth, and outcome value of a single collaboration event, comprehensively evaluated by the procuring party based on key indicators during the collaboration process (such as performance qualification rate, delivery timeliness, collaboration satisfaction, and collaboration amount).

[0073] For example, if supplier A supplies motherboards and supplier B supplies casings, and the two suppliers previously collaborated on an order, then a collaboration event between supplier A and supplier B can be recorded, indicating the existence of collaboration information. For this collaboration event, the corresponding collaboration time is also recorded, and based on the specific collaboration event and collaboration satisfaction / effect, a corresponding event intensity value is calculated. Collaboration satisfaction and effect can be evaluated by relevant employees or calculated automatically by the system. The larger the amount involved in the collaboration event between suppliers, the higher the collaboration satisfaction, and the better the collaboration effect, the higher the corresponding event intensity value.

[0074] Collaboration weight is an indicator used to quantify the degree of collaboration between candidate suppliers. A higher collaboration weight indicates closer collaboration and lower risk. The magnitude of the collaboration weight is positively correlated with the number of collaboration events and the intensity of each event. Furthermore, the closer the collaboration event's time is to the current moment, the higher the collaboration weight. The collaboration weight between suppliers with no collaboration information can be set to 0 or 1 by default, while the collaboration weight between suppliers with collaboration information will be higher than 0 or 1. For example, the collaboration weight between supplier A and C is 0.8, and the collaboration weight between supplier A and B is 0 (no historical collaboration).

[0075] Specifically, historical collaboration data between suppliers can be extracted from the supplier knowledge graph to form a collaboration information database. The database can then be used to identify whether there is collaboration information between candidate suppliers. Based on the extracted collaboration information, a collaboration weight can be calculated. If no collaboration information exists, a fixed collaboration weight can be assigned by default.

[0076] Step 150: Calculate the combined score for each candidate supplier combination based on the matching degree of the candidate suppliers and the collaborative weight.

[0077] The portfolio score is a quantitative metric used to measure the overall quality of a portfolio of candidate suppliers; a higher score indicates a better portfolio. For example, if portfolio {A,C} has a portfolio score of 92 and portfolio {B,C} has a portfolio score of 85, then {A,C} is better. The portfolio score is positively correlated with the matching degree and the synergy weight; the higher the values ​​of both, the higher the portfolio score.

[0078] Optionally, any suitable calculation rule can be used to calculate the combined score based on the matching degree and the synergy weight. For example, the combined score can be a weighted sum of the matching degree and the corresponding synergy weight for each candidate supplier; or a basic combined score can be calculated first from the corresponding matching degree, and then the combined score of the candidate supplier combination can be calculated based on the basic combined score and the synergy weight. For example, a final weight can be calculated based on the synergy weight between each candidate supplier, and the value obtained by multiplying the final weight by the basic combined score is used as the combined score.

[0079] In one embodiment, the average weight of a candidate supplier can be obtained by weighting the synergistic weights between the candidate supplier and each other candidate supplier in the portfolio. The matching degree of each candidate supplier is then weighted and summed with its average weight to obtain the portfolio score.

[0080] Step 160: Generate supplier combination recommendation information from the final candidate supplier combinations based on the combined score.

[0081] In this embodiment, the electronic device can select one or more candidate supplier combinations with the highest combined scores as supplier combination recommendation information. The supplier combination recommendation information includes one or more supplier combinations ultimately recommended to the purchasing personnel. Furthermore, it may include the reasons for the recommendation, as well as additional information such as the corresponding expected procurement cost, delivery cycle, and risk point warnings. For example, the supplier combination recommendation information may suggest the optimal combination as {Supplier A + Supplier C} (score 100 points), and the alternative combination as {Supplier C} (score 92 points), along with the procurement cost, delivery cycle, and risk point warnings for each combination.

[0082] The knowledge graph-based supplier portfolio recommendation method in this application firstly filters candidate suppliers using a pre-defined supplier knowledge graph and provides the matching degree between candidate suppliers and procurement needs. Compared with traditional manual screening, this method can quickly locate and quantify suppliers with supply capabilities. Secondly, during the supplier portfolio recommendation process, the method further analyzes the collaboration information between suppliers, quantifies the historical collaboration information between suppliers into collaboration weights, and incorporates these collaboration weights into the portfolio score. This ensures that the score not only reflects the capabilities of individual suppliers but also the degree of collaboration within the portfolio, making the recommended supplier portfolio more closely aligned with procurement needs.

[0083] In one embodiment, prior to step 120, a step of constructing a supplier knowledge graph is included, which includes: collecting multi-source data of suppliers; extracting entities and semantic relationships from the multi-source data; importing the entities and semantic relationships into a graph database; and constructing a supplier knowledge graph.

[0084] In this embodiment, the collected multi-source data includes one or more of the following: internal enterprise historical purchase order data, supplier basic qualification data, product attribute data, historical collaboration record data, and external third-party supplier credit rating data, industry policy data, network information data, environmental certification data, etc. This data may be pre-processed data.

[0085] The preprocessing process may include missing value imputation, outlier removal, deduplication, and format standardization. Missing value imputation can be performed using K-nearest neighbor imputation or mean imputation, while outlier removal can be achieved using boxplot analysis or DBSCAN clustering algorithms. Entities can be extracted using BERT-based entity recognition algorithms, and semantic relationships can be extracted using relation extraction algorithms based on dependency parsing. Entities include supplier entities, purchased item entities, qualification certification entities, and historical cooperation project entities; semantic relationships include supply relationships, qualification association relationships, collaboration relationships, and policy compliance relationships. The graph database can be based on Neo4j, ArangoDB, or NebulaGraph. The knowledge graph supports incremental updates of entity attributes and semantic relationships, and the resulting structured data can be in the format of "entity-attribute-value" triples.

[0086] For entity extraction, the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers) can be used to perform semantic understanding on the collected multi-source data, identifying preset entity types (such as suppliers, purchased items, and certifications) and their attributes. For example, from the text data "XX Technology Co., Ltd. (Unified Social Credit Code: 91110105XXXXXXXXX) supplies XXMateBook14 laptops, which have ISO9001 quality certification", the supplier entity "XX Technology Co., Ltd." (attribute: Unified Social Credit Code = 91110105XXXXXXXXX), the purchased item entity "XXMateBook14 laptops", and the certification entity "ISO9001 quality certification" can be identified.

[0087] To address semantic relationships, the semantic connections between entities are identified by analyzing the dependency relationships (such as subject-verb and verb-object) of words in a sentence. For example, in the sentence "XX Technology Co., Ltd. supplies Huawei MateBook 14 laptops," dependency parsing identifies "supply" as the core verb, "XX Technology Co., Ltd." as the subject (supplier entity), and "Huawei MateBook 14 laptops" as the object (purchased item entity), thereby extracting the "supply relationship" (supplier → purchased item).

[0088] For example, a supplier entity includes one or more attributes such as supplier ID, name, unified social credit code, registered capital, establishment date, business scope, delivery cycle, quality pass rate, and credit rating; a purchased item entity includes one or more attributes such as item ID, name, brand, specifications, unit price, category, and environmental protection level; a qualification certification entity includes one or more attributes such as certification ID, certification name, issuing authority, validity period, and certification level; and a historical cooperation project entity includes one or more attributes such as project ID, project name, cooperation time, cooperation amount, and cooperation satisfaction.

[0089] Supply relationships can be described as: Supplier entity → Purchased item entity (e.g., "S001→G003", describing "S001 supplies G003 laptops"); Qualification relationships can be described as: Supplier entity → Qualification certification entity (e.g., "S001→C001", describing "S001 has C001 ISO9001 certification"); Collaboration relationships can be described as: Supplier entity → Supplier entity (e.g., "S001→S005", describing "S001 and S005 have historical collaboration records"); Policy compliance relationships can be described as: Supplier entity → Industry policy entity (e.g., "S001→P001", describing "S001 complies with P001 green procurement policy").

[0090] Furthermore, for the constructed supplier knowledge graph, a corresponding supplier profile is set for each supplier node. The supplier profile is the core of the knowledge graph node. After multi-dimensional feature extraction and quantification, each supplier generates a formalized profile vector, for example, supplier A=[price:85, quality:92, delivery:88, service:79,...]. This vector is directly stored in the graph database as the attribute set of the supplier entity node. When a new procurement requirement is generated, the system first filters out relevant candidate supplier nodes based on the requirement, reads the attribute set containing its profile vector, and performs subsequent analysis.

[0091] In one embodiment, such as Figure 2 As shown, the calculation of the collaboration weights between the corresponding suppliers based on the collaboration information includes:

[0092] Step 210: Using suppliers as nodes and collaboration information between suppliers as edges of the nodes, calculate the edge weights based on the collaboration information to construct a supplier collaboration graph.

[0093] In this embodiment, the supplier collaboration graph is an undirected graph structure with suppliers as nodes, collaboration information between suppliers as edges, and the degree of collaboration as edge weights. It is used to visualize and quantify the collaborative relationships between suppliers. The edge weight refers to the quantified value corresponding to the edge connecting two nodes in the supplier collaboration graph, used to measure the strength of collaboration between suppliers. It is calculated comprehensively from the collaboration information between suppliers (such as frequency, quality, and timeliness); the larger the value, the closer the collaboration between the two suppliers. For each collaboration event, the electronic device can evaluate its assessment value in various dimensions, then sum the weighted values ​​of each dimension to obtain the assessment value of each collaboration event, and finally sum the weighted values ​​of each collaboration event to obtain the corresponding edge weight. The dimensions for evaluating each collaboration event may include collaboration quality, cost, delivery, service, environmental compliance, qualification compliance, and regional adaptability, etc., and the corresponding weights may be correction factors such as quality correction factor, cost correction factor, delivery correction factor, compliance correction factor, and regional adaptability correction factor. The values ​​of these correction factors used in calculating the edge weights can be default values. The weights of each collaborative event can be determined based on factors such as the occurrence of the collaborative event and the importance of the collaborative event.

[0094] The supplier nodes in the supplier collaboration graph can be extracted from the entire supplier knowledge graph, making the constructed graph more comprehensive. Alternatively, only candidate supplier nodes from the supplier knowledge graph can be used as nodes in the supplier collaboration graph to reduce graph redundancy. Node attributes in the supplier collaboration graph can include supplier ID (primary key), supplier name, core supply category (e.g., electronic equipment, office supplies), whether basic procurement requirements are met (yes / no), supplier profile, etc., and attribute data can be directly extracted from the supplier knowledge graph. When two supplier nodes have at least one shared performance record or a complementary resource collaboration record (e.g., supplier A supplies CPUs, supplier B supplies motherboards, and they were previously supplied for the same order), an undirected edge connecting the two nodes is created. Each edge is associated with corresponding collaboration information metadata, forming an edge attribute record, which may include, for example, the number of shared performances, average collaboration satisfaction, most recent collaboration time, and category of collaborative items, facilitating subsequent edge weight calculation and traceability.

[0095] Step 220: Generate a group of suppliers with collaborative relationships based on the supplier collaboration graph and the procurement demand information set.

[0096] In this embodiment, a supplier group refers to a collection of suppliers with close internal collaboration relationships (i.e., a community or cluster in graph theory). The frequency and quality of collaboration among suppliers within a group are higher than those outside the group. For example, when purchasing "laptops + printers + office paper," suppliers S001 (laptop supplier), S002 (printer supplier), and S003 (office paper supplier) in the supplier collaboration graph have multiple co-fulfillment records, forming a supplier group G1={S001,S002,S003}; while S004 (laptop supplier) has no collaboration records with other suppliers and does not form a group.

[0097] Optionally, a community detection algorithm (such as a hierarchical multi-objective community detection algorithm) can be used to cluster the nodes in the supplier collaboration graph to form closely cooperating groups. By setting thresholds such as group size and collaboration density, the node groups that meet the requirements can be selected as supplier groups.

[0098] For example, the Louvain algorithm for community discovery, which employs modular optimization, can efficiently identify communities in a graph with tightly connected internal structures and sparse external connections, and supports large-scale node clustering. The minimum community size can be set to 2, the modularity function threshold to 0.3 (≥0.3 indicates a significant community structure), and the maximum number of iterations to 10. Iteration stops when the modularity value changes by ≤0.01 for two consecutive iterations. During node clustering, node and edge weights are input into the Louvain algorithm. The algorithm optimizes the modularity value to divide the subgraph into multiple non-overlapping communities. After obtaining and validating the supplier groups, the final output is a list of supplier groups that meet the threshold requirements. Each group is labeled with its group ID, included suppliers, collaboration density, demand coverage, core collaboration scenarios, and other group information.

[0099] Furthermore, the edge weights in the supplier collaboration graph are modified based on the procurement demand information set to obtain a modified supplier collaboration graph; node aggregation is performed based on the modified supplier collaboration graph to form a supplier group.

[0100] This allows for the preset use of multiple correction factors, which may include one or more of the following: quality correction factor, cost correction factor, delivery correction factor, compliance correction factor, and regional adaptation correction factor. Each correction factor corresponds to a correction for one dimension (such as quality, cost, delivery, service, environmental compliance, qualification compliance, and regional adaptation) in the procurement requirement information set.

[0101] For example, the values ​​of the correction factors for each of the aforementioned dimensions can be set to a default value. The system can assess the importance of each dimension based on the procurement requirement information set and adjust the values ​​of the corresponding correction factors accordingly. For instance, if the analysis indicates that cost is the priority in this procurement, the cost correction factor can be increased. After adjusting the correction factors for each dimension, the original correction factors are replaced with the adjusted ones, and the edge weights are recalculated as the corrected edge weights, thus forming the supplier correction collaboration graph. For example, the adjusted correction factors and the values ​​used in the original edge weight calculation are weighted and summed to obtain the evaluation value of each collaboration event. Then, the evaluation values ​​of each collaboration event are weighted and summed again, and the final value is used as the corrected edge weight.

[0102] After obtaining the supplier modified collaboration graph, the nodes in the supplier modified collaboration graph are aggregated based on the community detection algorithm described above to form supplier groups.

[0103] In one embodiment, node aggregation is performed based on the supplier modified collaboration graph to form supplier groups, including: Step 1, treating each node in the supplier modified collaboration graph as an independent community; Step 2, calculating the modularity gain brought about by moving a node in the supplier modified collaboration graph to its neighboring community; Step 3, if the modularity gain is greater than a preset gain threshold, moving the node to its neighboring community is executed; otherwise, moving the node to its neighboring community is rejected, forming a new community distribution; Steps 2 and 3 are repeated iteratively until no modularity gain exceeds the gain threshold or the number of iterations reaches a preset iteration threshold; the supplier in the latest community distribution is taken as the corresponding supplier group.

[0104] In this embodiment, the module degree gain Q refers to the change in the modularity value of the entire supplier modification collaboration graph after a node is moved from its current community to another community. It is a core indicator for measuring the degree of community structure optimization. A Q value greater than the gain threshold indicates that moving nodes can improve the tightness of the community structure.

[0105] The neighbor community can be retrieved by querying all connected edges of the current node (with corrected edge weights > 0) using the Cypher statement, obtaining the community identifier of the connected nodes, and then deduplicating them to obtain a list of neighbor communities. The community discovery algorithm model iterates through each supplier node and calculates the modularity gain resulting from moving it to the community of its neighbors. Q. The calculation of Q fully considers edge weights; edges with higher weights contribute more to modularity. The gain threshold can be a preset, suitable value, such as 0. For supplier node i, moving it to a neighboring community would generate the maximum positive gain. If Q is selected, a move is executed. This process is repeated iteratively until no further moves can increase the overall modularity.

[0106] Specifically, the following can be extracted: the sum of corrected edge weights ka between the current node and all nodes in its neighboring community; the sum of all corrected edge weights kb between the current node and all nodes in its neighboring community; the sum of all corrected edge weights kc between all nodes in its neighboring community; the sum of all corrected edge weights kd across the entire corrected collaboration graph; and the sum of all corrected edge weights ke between all nodes in its neighboring community (including both internal and external communities). Optionally, the weights after each adjustment... Q can be calculated using the following formula 1:

[0107] Q = (kc + ka) / kd - [(ke + kb) / kd] 2 -(ka / kd) 2 -(kb / kd) 2 -(ke / kd) 2(Formula 1)

[0108] For the newly aggregated communities, each community is used as a new supernode. The edge weight between supernodes in the new graph is equal to the sum of the edge weights between corresponding communities in the original graph. Edges within a community are transformed into self-loops of the supernode, with a weight equal to the sum of the edge weights within that community. Using the aggregated new graph as input, the above iterative process is repeated to form a hierarchical community structure until the modularity can no longer be improved.

[0109] After completing the iterations, the algorithm outputs a final list of community partitions, i.e., a list of supplier groups, denoted as communities=[C1,C2,C3,...]. Each community Ci is a list containing multiple supplier IDs, representing a highly collaborative group. This result is directly used for the weighted recommendation score of subsequent supplier combinations. For example, for suppliers belonging to the same community Ci, their final combination score will receive an added bonus based on the average edge weight of that community.

[0110] Step 230: Calculate the collaboration weights between suppliers based on the supplier collaboration graph and the supplier group.

[0111] The collaboration weight between suppliers within the same supplier group is higher than that between suppliers in different supplier groups or not within supplier groups. Specifically, for supplier pairs within a group, the collaboration weight is calculated by combining the group's collaboration density, the importance of nodes in the group (such as degree centrality), and the edge weights in the collaboration graph. The larger the variable weight, the higher the collaboration density, and the higher the importance, the greater the collaboration weight. For supplier pairs outside a group (no collaboration record or not forming a group), the initial edge weights are retained or basic weights are assigned, and the collaboration weights of all candidate supplier pairs are finally output.

[0112] In one embodiment, a group collaboration gain coefficient x1 and a node importance coefficient x2 can be introduced, with the collaboration weight q = x1 × x2 × q0. q0 represents the original edge weight or the modified edge weight between two supplier nodes (i.e., the edge weight in the supplier-modified collaboration graph). The group collaboration gain coefficient can be calculated based on the group collaboration density, for example, x1 = min(collaboration density × a0, a1), where a0 and a1 are preset density coefficient and upper limit values ​​of the gain coefficient, respectively. For example, a0 = 0.8, a1 = 0.5. If the group collaboration density is 0.67, x1 = 0.67 × 0.8 ≈ 0.536, then the upper limit of 0.5 is taken.

[0113] The importance coefficient of a node can be calculated using degree centrality, for example, x2=b / (k0-1). b is the degree of the node, and k0 is the community size. For example, in community G1 with a size of 3, node S001 has 2 edges in the community, and its corresponding x2=2 / (3-1)=1.

[0114] In one embodiment, the collaborative weight qi of supplier i is calculated as u1 × avg_c1 + u2 × avg_c2. Here, c1 represents the average weight of all edge weights within the supplier group to which supplier i belongs, c2 represents the average weight of all edge weights of supplier i's direct partners, and u1 and u2 are preset harmonic parameters.

[0115] This embodiment transforms collaboration information into a graph structure, filters effective collaboration groups, and optimizes the weight calculation model, so that the final collaboration weight can more realistically and accurately reflect the collaboration adaptability of suppliers in specific procurement scenarios, thereby improving the rationality and reliability of subsequent supplier combination recommendations.

[0116] In one embodiment, the collaboration information includes one or more collaboration events between suppliers, each collaboration event containing a corresponding collaboration project, collaboration time, and event intensity value. For example... Figure 3 As shown, calculating the edge weights based on the cooperation information includes:

[0117] Step 310: Calculate the initial weight between suppliers for each collaborative event based on the collaborative project and event intensity value between suppliers.

[0118] In this embodiment, the project type of collaboration projects between suppliers can be assigned a corresponding project type weight. Multiplying this project type weight by the event intensity value yields an initial weight between suppliers. Project types can include joint performance, supporting supply, and joint R&D, etc. Different collaboration project types do not necessarily have the same type weight, and each type involves different evaluation dimensions, each with a different proportion in the project type weight. For example, joint performance involves evaluation dimensions such as performance qualification rate (40%), delivery timeliness (30%), and satisfaction (30%); supporting supply involves dimensions such as component qualification rate (50%), delivery cycle compliance rate (30%), and adaptability (20%); joint R&D involves evaluation dimensions such as results transformation rate (40%), R&D cycle (30%), and technological innovation (30%). After quantifying the values ​​under each evaluation dimension, a weighted sum is obtained based on the corresponding proportion to obtain the corresponding project type weight. Similarly, the event intensity value can also be calculated in the same way. Multiply the project type weight by the event intensity value, and use the resulting value as the initial weight.

[0119] Step 320: Calculate the corresponding weight decay value based on the collaboration time in the collaboration event.

[0120] Step 330: Calculate the edge weights between suppliers based on the initial weights of each collaborative event and the weight decay values.

[0121] In this embodiment, the edge weight can be a weighted sum of the initial weight and the corresponding weight decay value for each collaborative event. The weight decay value can be determined based on the collaboration time; the further the collaboration time is from the current time, the smaller the weight decay value, and the smaller its contribution to the edge weight. The weight decay value can be calculated based on an exponential decay model or a linear decay model. For example, the weight decay value sd = 1 - e (-k1×t1) Alternatively, sd = min(t1 / t2, 1). Here, k1 can be a preset decay rate parameter, t1 represents the time interval between the corresponding collaboration time and the current time, and t2 is the corresponding time interval threshold.

[0122] In this embodiment, during the calculation of edge weights, collaboration information is decomposed into individual collaboration events. By combining the collaboration project type weight and event intensity value, differentiated assignment of values ​​is achieved for different types of collaboration, and quality-based quantification is implemented for collaborations of the same type. This avoids the value distortion caused by traditional one-size-fits-all weight calculations. Furthermore, by introducing a decay model to deduct weights from outdated collaboration events, the problem of traditional weight calculations ignoring time factors and relying on stale data is solved. This ensures that the final calculated edge weights comprehensively consider both the overall calculation logic and its rationality, guaranteeing the usability and comparability of the weights.

[0123] Furthermore, the calculation of the corrected edge weights in the aforementioned supplier-corrected collaboration graph may include: calculating the initial weights between suppliers corresponding to each collaboration event based on the collaboration items and event intensity values ​​of the collaboration events between suppliers; correcting the initial weights based on the procurement demand information set to obtain corrected weights; calculating the corresponding weight decay value based on the collaboration time in the collaboration event; and calculating the corrected edge weights between suppliers based on the corrected weights of each collaboration event and the weight decay value.

[0124] The process of correcting the initial weights based on the procurement information set can also be used to correct the proportion of the evaluation dimensions in the above-mentioned collaborative project types. Based on the corrected proportions, the corresponding corrected project type weights are obtained, and then the corresponding corrected weights are calculated based on the corrected project type weights and the event intensity values.

[0125] In one embodiment, calculating the combined score for each candidate supplier combination based on the matching degree of the candidate suppliers and the synergy weight includes: identifying the importance of the candidate suppliers in the procurement demand information set; calculating the combined basic score for each candidate supplier combination based on the matching degree and the importance; identifying whether there are candidate suppliers in the candidate supplier combination that are in the same supplier group, and if so, weighting the combined basic score based on the synergy weight to obtain the combined score of the candidate supplier combination.

[0126] In this embodiment, the supplier's importance reflects the core contribution of a single supplier to meeting the procurement demand information set. Importance can be determined by one or more factors, such as the importance of the corresponding supply category in the procurement demand and the supplier's irreplaceability. A higher importance indicates a greater influence of the supplier on the procurement demand. For example, supply categories can be divided into core categories, secondary categories, and auxiliary categories. Irreplaceability is calculated based on dimensions such as the supplier's product's exclusivity in the category supply, technological compatibility, and delivery assurance capabilities related to the procurement demand. The electronic device calculates the corresponding irreplaceability quantification value based on the supplier's knowledge graph information, and multiplies the two values ​​by the importance of the supplier's products in the procurement demand to obtain the corresponding importance.

[0127] For each supplier, its importance and matching degree are multiplied to obtain a corresponding supplier score. The supplier scores of each candidate supplier in the candidate supplier combination are summed to obtain the combination's base score. Based on the supplier groups to which each candidate supplier in the combination belongs, a weighted value for the candidate supplier combination can be calculated. This weighted value is multiplied by the combination's base score to obtain the combination score. Alternatively, the collaboration weight for each candidate supplier in the combination can be calculated separately, and this collaboration weight is summed with the corresponding calculated supplier score to obtain the combination score. It is understandable that this weighted value / collaboration weight is positively correlated with the closeness of connections within the candidate supplier combination's supplier group and the (corrected) edge weights between the supplier nodes in the candidate supplier combination calculated above. When no candidate supplier combination belongs to the same supplier group, the combination's base score can be directly used as the final combination score.

[0128] This application deeply integrates demand adaptability and collaboration, enabling the output score to directly guide the selection of the optimal supplier combination, thereby improving the feasibility and decision-making value of the entire supplier recommendation method.

[0129] In one embodiment, such as Figure 4 As shown, the step of filtering a set of candidate suppliers from a preset supplier knowledge graph that meets one or more procurement item requirements of the procurement demand information set, and the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement, includes:

[0130] Step 410: Select a set of candidate suppliers from the preset supplier knowledge graph that meet one or more procurement item requirements of the procurement requirement information set, and form a candidate supplier subgraph.

[0131] The candidate supplier subgraph is a subset of the supplier knowledge graph, containing only supplier nodes, corresponding item category nodes, qualification nodes, and related edges from the candidate supplier set. Entities and relationships irrelevant to the procurement needs are removed, resulting in a more streamlined and focused structure. By transforming procurement needs into screening criteria, supplier and related information relevant to the needs is accurately extracted from the complete knowledge graph, forming a subgraph focused on candidate items. This improves screening efficiency and provides structured data support for subsequent multi-dimensional evaluation.

[0132] Step 420: Extract the initial score vector for each candidate supplier from the candidate supplier subgraph.

[0133] The initial score vector is a vector data quantifying the performance of candidate suppliers across various evaluation dimensions of the procurement requirements. Each parameter corresponds to a standardized score for one evaluation dimension. Each parameter in the initial score vector represents an evaluation indicator for one dimension of the procurement requirements. Evaluation dimensions represent core indicators measuring the suitability of suppliers for the procured goods, covering core dimensions such as supply capacity, compliance, technological adaptability, and service assurance, and are strongly correlated with the procurement requirements. Optionally, this initial score vector can be extracted based on the aforementioned supplier profile, or obtained by fusing its own supplier profile with its network context.

[0134] Specifically, based on the candidate supplier subgraph, an N×N adjacency matrix A and an N×F feature matrix X are constructed, where N is the total number of nodes in the candidate supplier subgraph, F is the initial feature dimension of each node, and the element a in the adjacency matrix... ij The adjacency matrix A and the feature matrix F are used to reflect the connection relationship between nodes i and j in the candidate supplier subgraph. The adjacency matrix A and the feature matrix F are input into a preset graph neural network model for nonlinear transformation processing, and the initial score vector of each candidate supplier is output.

[0135] The adjacency matrix A is a square matrix representing the connectivity relationships between all nodes in the candidate supplier subgraph, where N is the total number of nodes in the subgraph, and element a ij (The element in the i-th row and j-th column of matrix A) represents the connection strength between node i and node j. The feature matrix X refers to the matrix that quantizes the attribute features of all nodes in the subgraph, F is the initial feature dimension of each node, and the elements x... if (The element in the i-th row and f-th column of matrix X) represents the standardized score of node i on the f-th feature dimension. This standardized score can be the numerical value of the corresponding feature dimension described by the corresponding supplier profile. The initial feature vector of each supplier in feature matrix X can be obtained from its supplier profile.

[0136] By loading a pre-set graph neural network model and training parameters, the adjacency matrix and feature matrix are preprocessed, such as normalized and dimension adapted, and then used as input to the graph neural network model. Then, the convolutional layer in the graph neural network model aggregates the features of the node itself and its neighboring nodes, and performs non-linear transformation through the activation function to output a high-dimensional vector that is mapped to the evaluation dimension, thus obtaining a standardized initial score vector.

[0137] For each node in the candidate supplier subgraph, each layer of the graph neural network (GNN) model performs neighbor sampling, neighbor information aggregation, and self-feature update. For neighbor sampling: a certain number of direct neighbor nodes (e.g., first-order neighbors) are randomly sampled for each target node (e.g., supplier A). For neighbor information aggregation: the current feature vectors of the sampled neighbor nodes are aggregated using an aggregation function (e.g., mean pooling, max pooling) to generate a neighbor information vector representing the local network structure of the target node. For self-feature update: the target node's current feature vector is concatenated or added to the neighbor information vector obtained in the previous step, and then passed through a fully connected neural network layer and subjected to a non-linear transformation using the ReLU activation function to obtain the new feature vector output by the target node at this GNN layer. This process is stacked multiple times (e.g., 2 layers) so that the final feature vector of each node can capture the information of its multi-hop neighbors, ultimately yielding the initial score vector.

[0138] For example, for supplier node S, its initial feature vector h_s consists of its profile [price: 85, quality: 92, delivery: 88, etc.]. Its neighbors include the product G it supplies (features are product attributes) and another supplier P with which it has a cooperative relationship (features are the profile of P). In the first layer of the GNN, node S aggregates the features from G and P, generating an aggregated vector. Then, h_s is combined with this aggregated vector, and through neural network updates, the first layer output feature h_s of S is obtained. 1 This new feature already includes information about S's direct partners. In the second-layer GNN, S will further aggregate new features from its neighbors (whose features have also been updated; for example, P's features include information about P's partners). After the second-layer update, the final output feature h_s of S is... 2 This not only includes its own profile but also indirectly includes information about its partners' partners, thus achieving the capture of complex relationships. This final h_s 2 It is the initial score vector of node S, which is a fusion of its own profile and its network context.

[0139] Step 430: Calculate the dimension weight of each dimension in the initial score vector based on the requirements of each type of purchased item.

[0140] Dimension weights refer to the proportion of importance of each evaluation dimension in the matching degree calculation, which is determined by the core objectives of the procurement needs. For example, if the procurement needs are "high-precision sensors and installation and commissioning services", the dimension weights are allocated as follows: technical compatibility (0.4), qualification compliance (0.3), category matching degree (0.15), delivery cycle compliance (0.1), and service support degree (0.05).

[0141] Optionally, the procurement requirement text can be parsed using natural language processing to identify core target keywords (such as high precision, urgent delivery, cost control, etc.). Based on the identified core target keywords and the requirement text, semantic analysis can be performed to quantify the importance of each dimension and obtain the dimension weights. For example, the analytic hierarchy process (AHP) can be used to calculate the dimension weights.

[0142] Step 440: Calculate the matching degree based on the dimension weights and the initial score vector.

[0143] Given the calculated dimension weights and the initial score vector, the parameters in the initial score vector can be weighted and summed with their corresponding dimension weights, and the resulting value can be used as the matching degree.

[0144] In one embodiment, a computer-readable storage medium is provided having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the steps in the above method embodiments.

[0145] In one embodiment, an electronic device is also provided, including one or more processors; and a memory storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the steps in the above method embodiments. The electronic device may be a device that deploys the aforementioned large language model and / or knowledge graph, such as a backend server for a platform that provides knowledge-based question answering for enterprise personnel.

[0146] In one embodiment, a computer-readable storage medium is provided having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the steps in the above method embodiments.

[0147] In one embodiment, an electronic device is also provided, including one or more processors; and a memory storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the steps in the above-described method embodiments. The electronic device may be a device that deploys the aforementioned related algorithm models and / or knowledge graphs, such as a backend server for a platform that provides enterprise personnel with supplier combination recommendations.

[0148] In one embodiment, such as Figure 5 The diagram illustrates the structure of an electronic device used to implement an embodiment of this application. The electronic device includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 502 or a program loaded from a storage portion 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0149] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 510 as needed so that computer programs read from it can be installed into storage section 508 as needed.

[0150] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer-readable medium carrying instructions that, in such embodiments, can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the instructions are executed by central processing unit (CPU) 501, the various method steps described in this application are performed.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0152] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A supplier combination recommendation method based on knowledge graphs, characterized in that, The method includes: Obtain a collection of procurement demand information; From a preset supplier knowledge graph, a set of candidate suppliers that meet one or more procurement item requirements of the procurement requirement information set are selected, along with the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement. Based on the procurement requirements satisfied by each candidate supplier, multiple candidate supplier combinations are formed, and each candidate supplier combination satisfies all procurement requirements in the procurement requirements information set. Identify whether there is any collaboration information among the candidate suppliers, and calculate the collaboration weight between the corresponding suppliers based on the collaboration information; A combined score for each candidate supplier combination is calculated based on the matching degree of the candidate suppliers and the collaborative weight. Supplier combination recommendation information is generated from the final candidate supplier combination based on the combined score; The step of calculating the collaboration weights between suppliers based on the collaboration information includes: using suppliers as nodes and collaboration information between suppliers as edges of the nodes, calculating edge weights based on the collaboration information, constructing a supplier collaboration graph, generating supplier groups with collaborative relationships based on the supplier collaboration graph and the procurement demand information set, and calculating the collaboration weights between suppliers based on the supplier collaboration graph and the supplier groups. The step of calculating the combined score of each candidate supplier combination based on the matching degree of the candidate suppliers and the collaboration weight includes: identifying the importance of the candidate suppliers in the procurement demand information set; calculating the basic combined score of each candidate supplier combination based on the matching degree and the importance; identifying whether there are candidate suppliers in the candidate supplier combination that are in the same supplier group; if so, weighting the basic combined score based on the collaboration weight to obtain the combined score of the candidate supplier combination.

2. The method according to claim 1, characterized in that, The generation of supplier groups with collaborative relationships based on the supplier collaboration graph and the procurement demand information set includes: Based on the procurement demand information set, the edge weights in the supplier collaboration graph are corrected to obtain the supplier corrected collaboration graph. The nodes are aggregated according to the supplier modified collaboration graph to form a supplier group.

3. The method according to claim 2, characterized in that, The step of aggregating nodes according to the supplier modified collaboration graph to form a supplier group includes: Step 1: Treat each node in the supplier modification collaboration graph as an independent community; Step 2: Calculate the modularity gain brought about by a node in the supplier modified collaboration graph moving to its neighboring community; Step 3: If the modularity gain is greater than the preset gain threshold, move the node to its neighboring community; otherwise, refuse to move the node to its neighboring community, thus forming a new community distribution. Repeat steps 2 and 3 iteratively until no module gain exceeds the gain threshold or the number of iterations reaches the preset iteration threshold. Suppliers in the latest community distribution are identified as the corresponding supplier groups.

4. The method according to claim 1, characterized in that, The collaboration information includes one or more collaboration events between suppliers, and each collaboration event includes a corresponding collaboration project, collaboration time, and event intensity value; The step of calculating edge weights based on the cooperation information includes: The initial weights between suppliers for each collaborative event are calculated based on the collaborative projects and event intensity values ​​between suppliers. Calculate the corresponding weight decay value based on the collaboration time in the collaboration event; The edge weights between suppliers are calculated based on the initial weight of each collaborative event and the weight decay value.

5. The method according to any one of claims 1 to 4, characterized in that, The step of selecting a set of candidate suppliers from a preset supplier knowledge graph that meets one or more procurement item requirements of the procurement demand information set, and the matching degree between each candidate supplier in the candidate supplier set and each procurement item requirement, includes: From the preset supplier knowledge graph, a set of candidate suppliers that meet one or more procurement item requirements that meet the procurement requirement information set are selected to form a candidate supplier subgraph; An initial score vector is extracted from each candidate supplier from the candidate supplier subgraph, where each parameter in the initial score vector represents an evaluation metric for one dimension of the procurement requirement; Calculate the dimension weight of each dimension in the initial score vector based on the requirements for each type of purchased item; The matching degree is calculated based on the dimension weights and the initial score vector.

6. The method according to claim 5, characterized in that, The step of extracting the initial score vector for each candidate supplier from the candidate supplier subgraph includes: Based on the candidate supplier subgraph, an N×N adjacency matrix A and an N×F feature matrix X are constructed, where N is the total number of nodes in the candidate supplier subgraph, F is the initial feature dimension of each node, and the elements a in the adjacency matrix are... ij Used to reflect the connection relationship between node i and node j in the candidate supplier subgraph; The adjacency matrix A and the feature matrix F are input into a preset graph neural network model for nonlinear transformation processing, and the initial score vector of each candidate supplier is output.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores executable instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the method as described in any one of claims 1 to 6.

Images