Supply chain multi-level relation identification method based on graph neural network

Abstract

The invention discloses a supply chain multi-level relation identification method based on a graph neural network. The method comprises the following steps of S1, collecting supply chain data and performing preprocessing to generate an original graph structure; s2, constructing a three-dimensional node feature vector according to the enterprise industry type, the legal person property and the transaction frequency; s3, extracting in-degree, out-degree, path depth and direction of the nodes, generating hierarchical position codes, and splicing the hierarchical position codes with the feature vectors to obtain a structural hierarchical embedded vector set; s4, inputting a Geom-GCN model, selecting neighbor nodes based on graph distance and direction consistency, and generating a structure perception embedding representation set; s5, constructing a node pair input classification network, and outputting a relation type label; s6, complementing the supply edge shielded by the agent node according to the label, and constructing a repaired graph structure; and S7, executing path hop count and upstream role verification, and generating a multi-level relation identification result. According to the invention, accurate identification of the supply chain multi-level relationship and efficient repair of the path structure are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of supply chain relationship identification, and in particular to a supply chain multi-level relationship identification method based on a graph neural network. BACKGROUND

[0002] In the process of supply chain management, there are various relationship structures such as multi-level transactions, processing, agency and distribution between enterprises. Traditional supply chain relationship identification methods mostly rely on rule engines, expert annotations or statistical methods based on field matching, which are difficult to effectively model the complex structure path and hierarchical relationship between nodes, especially when there are actual transaction behaviors such as agency, transit and cross-layer jumping. The traditional method cannot accurately identify the real supply path between enterprises, resulting in incomplete supply chain structure graph and missing or incorrect key path.

[0003] In recent years, graph neural networks have shown superior modeling capabilities in structured graph data learning tasks, but mainstream models mostly use fixed adjacency matrix construction methods, without considering the hierarchical position, path direction and role attribute of nodes in the actual supply chain structure, which can easily cause inaccurate relationship classification, node role confusion and path misjudgment. When existing models aggregate neighbor node information, they lack consistent judgment of path direction and structure semantics, cannot effectively exclude reverse path interference, and cannot identify the real supply path hidden by proxy nodes.

[0004] For the complex relationship scenarios such as multi-level structure, agency behavior and path jumping in the supply chain, there is still a lack of a graph neural network identification method that combines graph structure distance, node role attribute and path direction relationship. The existing technology has deficiencies in processing accuracy and structure explanation ability in terms of path completion, upstream role judgment and multi-class supply relationship classification. Therefore, how to provide a supply chain multi-level relationship identification method based on a graph neural network is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a supply chain multi-level relationship identification method based on a graph neural network. The present application uses structure level embedding modeling, path direction consistency judgment and relationship classification mechanism to automatically identify the supply, agency, cooperation and other relationships between enterprises, and has the advantages of accurate structure identification, controllable path completion and fine node role discrimination.

[0006] According to the supply chain multi-level relationship identification method based on the graph neural network of the embodiment of the present application, the following steps are included: S1, collect the data of the supply chain, and pre-process to build the original supply chain graph structure; S2, generate a three-dimensional node feature vector according to the industry type, legal nature and historical transaction frequency of the enterprise; S3. Construct a hierarchical position code based on the in-degree, out-degree, shortest path depth and path direction identifier of each node in the original supply chain graph structure. Concatenate the hierarchical position code with the 3D node feature vector to generate a set of structural hierarchical embedding vectors. S4. Input the set of structure-level embedding vectors into the Geom-GCN model, perform a neighbor selection operation based on graph distance, use the structural orientation consistency judgment to filter out the set of aggregated nodes that match the supply path direction, apply attention weighting operation to the set of aggregated nodes, and generate a set of structure-aware embedding representations. S5. Construct a node pair input set based on the structure-aware embedding representation set, input it into a multi-class classification network, and output the supply chain relationship type label for each pair of nodes; S6. Complete the path of the original graph structure according to the supply chain relationship type label, insert the real supply edge that is hidden by the agent node, and build the graph structure after structural repair. S7. Perform path hop count verification and upstream role consistency verification on the graph structure after structural repair to generate the final multi-level supply chain relationship identification result.

[0007] Optionally, the supply chain data in step S1 includes enterprise identifier, industry type, legal entity nature, contract amount, transaction frequency, and logistics direction; the preprocessing includes filling missing values, normalizing fields, aligning time, and unifying node identifiers in the supply chain data.

[0008] Optionally, the construction of hierarchical location coding in step S3 specifically includes: calculating the in-degree of each node in the original supply chain graph structure as the first coding component; calculating the out-degree of each node as the second coding component; calculating the shortest path step length from the node to the upstream node based on the shortest path algorithm in the graph as the third coding component; extracting the direction identifiers of the outgoing and incoming edges of the node, generating path direction identifier codes according to the logistics flow direction as the fourth coding component; and concatenating the four components of in-degree, out-degree, shortest path depth, and path direction identifier to form a hierarchical location coding used to describe the hierarchical information of the node structure.

[0009] Optionally, step S3, which involves concatenating the hierarchical position encoding with the three-dimensional node feature vector, includes: connecting the three-dimensional node feature vector of each node with the corresponding hierarchical position encoding vector in the dimensional direction to generate a structural hierarchical embedding vector containing node attributes and structural position features, and summarizing the structural hierarchical embedding vectors of all nodes to form a set of structural hierarchical embedding vectors.

[0010] Optionally, the process of generating the structure-aware embedding representation set in step S4 includes the following operations: The set of structural hierarchy embedding vectors is input into the Geom-GCN model. For each target node in the original supply chain graph structure, the graph distance calculation operation is performed to obtain the shortest path hop count from the target node to all other nodes. A graph distance mapping table is generated to record the hop count relationship between each potential neighbor node and the target node. Select all nodes with a hop count of less than or equal to two from the mapping table to form an initial neighbor candidate set. Only retain one-hop and two-hop nodes, and remove high-order nodes in the graph structure that are too far away and may cause feature interference. Read the logistics direction field of the edge between each node and the target node in the initial neighbor candidate set, extract the shortest path depth dimension value corresponding to the embedding vector of the neighbor node structure hierarchy, and calculate the difference with the same dimension value of the target node. When the difference is positive and the logistics direction field indicates that the information flow is from the neighbor node to the target node, it is determined that the neighbor node is upstream of the target node in the structural hierarchy and has the same supply direction. The neighbor node is retained in the set of aggregation nodes with the same direction, and the other neighbor nodes do not participate in information aggregation. For each node in the set of nodes with consistent orientation, the semantic proximity score between the node in the set of nodes with consistent orientation and the target node is calculated based on the spatial position distribution of the node embedding vector. The semantic proximity score represents the degree of consistency of the structural semantics of the two nodes in the embedding space. The semantic proximity scores of all aggregated nodes are normalized to form the attention weight set corresponding to the target node. The contribution of different aggregated nodes to the representation of the target node is controlled by the weights. The embedding vectors of all aggregate nodes are weighted and summed according to the normalized weights to obtain the structure-aware embedding vector of the current target node. The perception embedding vector integrates the features of this node, the structural hierarchy information, and the upstream and downstream semantic features of the direction-consistent aggregate substructure. The structure-aware embedding vectors generated from all target nodes are combined sequentially to form a structure-aware embedding representation set.

[0011] Optionally, the multi-class classification network processing in step S5 includes: For any two enterprise nodes in the original supply chain graph that are connected by a structural path, extract the embedding vectors of the two enterprise nodes in the set of structure-aware embedding representations, and denote them as the first node vector and the second node vector, respectively. The first node vector and the second node vector are concatenated in sequence to construct the node pair input vector, and the node pair input vector is input into the multilayer feedforward neural network. The multilayer feedforward neural network includes a set of input layers with fixed input dimensions, two or more fully connected transformation layers with nonlinear activation functions, and a classification layer consisting of a set of output nodes. In each fully connected transformation layer, matrix transformation and activation function calculation operations are performed on the input node pair vector to generate intermediate feature vectors of fixed dimensions. Several output nodes are set in the classification layer, each output node corresponds to a predefined supply chain relationship type. Linear transformation and normalization are performed on the intermediate feature vector to generate a set of numerical scores corresponding to each relationship type. Based on the index number corresponding to the output node with the highest score in the numerical score set, determine the supply chain relationship type label to which the current node belongs.

[0012] Optionally, the supply chain relationship types include direct raw material supply, tiered supply, agency supply, production collaboration and distribution, and the rules for determining different types of relationships include: For direct supply relationships of raw materials, the criteria for identifying output nodes are: there are direct transaction records between node pairs, the contract amount is higher than the set threshold, the transaction frequency is greater than the set number of times, and there are no transit nodes in the corresponding logistics path; For the hierarchical supply relationship type, the identification criteria for output nodes are: there is a two-hop path connection between node pairs in the supply chain graph structure, the intermediate nodes in the path do not participate in the contract signing record, and the logistics direction remains consistent in the path; For the agency supply relationship type, the identification criteria for the output node are: there are transaction records between the node pairs, and there is an actual supply path with more than three hops in the graph structure, where the intermediate node of the path has transaction contract records with several other nodes, but the logistics direction field does not point to the target node. For production collaboration relationships, the criteria for identifying output nodes are: the industry type of the node pair is the same, the overlap ratio of contract performance cycles is higher than the set threshold, and the proportion of transaction amount to the total transaction amount of the node is lower than the set threshold. For distribution relationship types, the criteria for identifying output nodes are: the out-degree of the identified node in the node pair is greater than its in-degree, and the node belongs to the retail or terminal category.

[0013] Optionally, step S6 involves performing path completion on the original graph structure, including identifying node pairs classified as agent supply relationships, tracing indirect upstream supply paths, and inserting real supply edges that directly connect node pairs into the graph structure.

[0014] Optionally, the path hop count verification in step S7 includes recalculating the shortest path hop count for each supply relationship node pair in the repaired graph structure, and determining whether the recalculated shortest path hop count is lower than the path hop count in the original graph structure, thus verifying whether the path completion operation effectively shortens the supply path length.

[0015] Optionally, the upstream role consistency verification in step S7 includes: for each upstream node in the repaired path, based on the industry type dimension and legal person nature dimension in the three-dimensional node feature vector, combined with the path depth component and path direction identifier in the structural hierarchy embedding vector, determining whether the node has the upstream enterprise characteristics required by the corresponding supply relationship type; when the node attributes do not meet the upstream role requirements of the identified relationship type, the corresponding path edge of the node is removed from the graph structure.

[0016] The beneficial effects of this invention are: (1) This invention constructs a structural hierarchy embedding vector that integrates node attributes and structural positions, and introduces a geometric neighbor selection mechanism based on graph distance and path direction to achieve accurate modeling of upstream and downstream structural relationships in the supply chain graph, identify proxy occlusion paths and skip supply paths, and improve the integrity and recognition depth of the graph structure.

[0017] (2) In this invention, a path direction consistency screening rule and an attention weighted aggregation mechanism are set in the graph neural network to ensure that the information aggregation process is carried out only between nodes with correct semantic direction, thereby avoiding the interference of reverse paths and non-real adjacency relationships on the embedded representation and improving the accuracy of relationship recognition.

[0018] (3) This invention constructs a multi-relationship classification network, and classifies and distinguishes relationship types such as direct supply of raw materials, hierarchical transfer supply, agency supply, production cooperation and distribution based on node attributes, path depth and relationship structure differences. Combined with path completion and upstream role consistency verification, it realizes dynamic repair of supply chain structure and multi-level semantic output. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Fig. 1 This is an overall flowchart of a supply chain multi-level relationship identification method based on graph neural networks proposed in this invention; Fig. 2 This is a diagram of the Geom-GCN directional consistency neighbor filtering and attention aggregation structure of a multi-level supply chain relationship identification method based on graph neural networks proposed in this invention. Fig. 3 This is a multi-class relationship classification network structure diagram of a multi-level relationship identification method for supply chain based on graph neural networks proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figs. 1-3 A method for identifying multi-level relationships in a supply chain based on graph neural networks includes the following steps: S1. Collect supply chain data and preprocess it to construct the original supply chain diagram structure; S2. Generate a three-dimensional node feature vector based on the enterprise's industry type, legal entity nature, and historical transaction frequency; S3. Construct a hierarchical position code based on the in-degree, out-degree, shortest path depth and path direction identifier of each node in the original supply chain graph structure. Concatenate the hierarchical position code with the 3D node feature vector to generate a set of structural hierarchical embedding vectors. S4. Input the set of structure-level embedding vectors into the Geom-GCN model, perform a neighbor selection operation based on graph distance, use the structural orientation consistency judgment to filter out the set of aggregated nodes that match the supply path direction, apply attention weighting operation to the set of aggregated nodes, and generate a set of structure-aware embedding representations. S5. Construct a node pair input set based on the structure-aware embedding representation set, input it into a multi-class classification network, and output the supply chain relationship type label for each pair of nodes; S6. Complete the path of the original graph structure according to the supply chain relationship type label, insert the real supply edge that is hidden by the agent node, and build the graph structure after structural repair; Specifically: after the supply chain relationship type label is generated, identify the node pairs that belong to the agent supply relationship, determine the actual supply node position by tracing the multi-hop path of the node pair in the original graph structure, insert the direct edge connecting the supply node and the target node in the graph, complete the path completion, and generate the multi-level supply chain graph structure after structural repair. S7. Perform path hop count verification and upstream role consistency verification on the graph structure after structural repair to generate the final multi-level supply chain relationship identification result.

[0023] In this embodiment, the supply chain data in step S1 includes enterprise identifier, industry type, legal entity nature, contract amount, transaction frequency, and logistics direction; the preprocessing includes filling missing values, normalizing fields, aligning time, and unifying node identifiers in the supply chain data.

[0024] In this embodiment, the construction of hierarchical position coding in step S3 specifically includes: calculating the in-degree of each node in the original supply chain graph structure as the first coding component; calculating the out-degree of each node as the second coding component; calculating the shortest path step length from the node to the upstream node based on the shortest path algorithm in the graph as the third coding component; extracting the direction identifiers of the outgoing and incoming edges of the node, and generating path direction identifier codes according to the logistics flow direction as the fourth coding component; and concatenating the four components of in-degree, out-degree, shortest path depth, and path direction identifier to form a hierarchical position coding used to describe the hierarchical information of the node structure.

[0025] In this embodiment, step S3, which involves concatenating the hierarchical position encoding with the three-dimensional node feature vector, includes: connecting the three-dimensional node feature vector of each node with the corresponding hierarchical position encoding vector in the dimensional direction to generate a structural hierarchical embedding vector containing node attributes and structural position features, and summarizing the structural hierarchical embedding vectors of all nodes to form a set of structural hierarchical embedding vectors.

[0026] In this embodiment, the process of generating the structure-aware embedding representation set in step S4 includes the following operations: The hierarchical embedding vector set is input into the Geom-GCN model. For each target node in the original supply chain graph structure, a graph distance calculation operation is performed to obtain the shortest path hop count from the target node to all other nodes, generating a graph distance mapping table that records the hop count relationship between each potential neighbor node and the target node. The graph distance calculation operation is based on the node connection relationship in the original supply chain graph structure. For each target node, a breadth-first traversal algorithm is performed, starting from the target node and traversing its reachable node set layer by layer. In each traversal, the hop count is incremented by one, recording the minimum hop count from the target node to each neighbor node. During the traversal, a graph distance mapping table is maintained. The key in the mapping table is the identifier of the potential neighbor node, and the value is the corresponding shortest hop count. After the traversal is completed, the mapping table fully describes the path hop count distribution between the target node and all reachable nodes.

[0027] In the mapping table, all nodes with a hop count of two or less are selected to form an initial neighbor candidate set. Only one-hop and two-hop nodes are retained, and high-order nodes that are too far away in the graph structure and may cause feature interference are removed. A one-hop node is a node in the graph structure that has a direct directed edge connection to the target node, representing an upstream or downstream enterprise with a direct transaction or cooperation relationship with the target node. A two-hop node is a node in the graph structure that can only be reached from the target node through an intermediate node, usually reflecting indirect supply, sub-supply, or agency cooperation relationships. When constructing the initial neighbor candidate set, only nodes with a graph distance of one or two are retained to ensure that the selected neighbor nodes have a clear upstream and downstream path connection with the target node in the supply chain structure and have a strong structural correlation.

[0028] Read the logistics direction field of the edge between each node and the target node in the initial neighbor candidate set, extract the shortest path depth dimension value corresponding to the embedding vector of the neighbor node structure hierarchy, and calculate the difference with the same dimension value of the target node. When the difference is positive and the logistics direction field indicates that the information flow is from the neighbor node to the target node, it is determined that the neighbor node is upstream of the target node in the structural hierarchy and has the same supply direction. The neighbor node is retained in the set of aggregation nodes with the same direction, and the other neighbor nodes do not participate in information aggregation. For each node in the set of nodes with consistent orientation, the semantic proximity score between the node in the set of nodes with consistent orientation and the target node is calculated based on the spatial position distribution of the node embedding vector. The semantic proximity score represents the degree of consistency of the structural semantics of the two nodes in the embedding space. The semantic proximity scores of all aggregate nodes are normalized to form the attention weight set corresponding to the target node. The contribution of different aggregate nodes to the target node representation is controlled by the weights. Specifically, when performing the aggregation operation of the graph neural network, the target node does not treat the features of all neighboring nodes equally and average them. Instead, it assigns a real value as an attention weight based on the semantic proximity score calculated between each aggregate node and the target node. The attention weight is transformed into a proportional coefficient between 0 and 1 through a normalization mechanism and is used to weight the embedded feature vectors of the aggregate nodes. The higher the weight of the aggregate node, the greater the proportion of its feature information in the final representation of the target node, and thus the stronger its influence in the embedding update process. Conversely, the lower the weight of the node, the weaker its influence. This mechanism ensures that the target node pays more attention to neighboring nodes that are more consistent with the structural direction and semantic features, effectively improving the discriminativeness of the embedding representation and the ability to express upstream and downstream relationships. The embedding vectors of all aggregate nodes are weighted and summed according to the normalized weights to obtain the structure-aware embedding vector of the current target node. The perception embedding vector integrates the features of this node, the structural hierarchy information, and the upstream and downstream semantic features of the direction-consistent aggregate substructure. The structure-aware embedding vectors generated from all target nodes are combined sequentially to form a structure-aware embedding representation set.

[0029] In this embodiment, the processing procedure of the multi-class classification network in step S5 includes: For any two enterprise nodes in the original supply chain graph that are connected by a structural path, extract the embedding vectors of the two enterprise nodes in the set of structure-aware embedding representations, and denote them as the first node vector and the second node vector, respectively. The first node vector and the second node vector are concatenated in sequence to construct the node pair input vector, and the node pair input vector is input into the multilayer feedforward neural network. The multilayer feedforward neural network includes a set of input layers with fixed input dimensions, two or more fully connected transformation layers with nonlinear activation functions, and a classification layer consisting of a set of output nodes. In each fully connected transformation layer, matrix transformation and activation function calculation operations are performed on the input node pair vector to generate a fixed-dimensional intermediate feature vector. Specifically, the steps are as follows: First, matrix multiplication is performed on the input node pair vector and the weight matrix of the current fully connected layer, and the corresponding bias vector is superimposed to obtain a linear transformation result. Then, the linear transformation result is input into a nonlinear activation function, usually ReLU or LeakyReLU, to enhance the network's ability to express different feature intervals, and the intermediate feature vector after nonlinear mapping is output. This process is repeated in each layer, so that the node pair vector gradually completes semantic reconstruction and relation boundary adjustment in the multi-layer feature space, providing a stable high-dimensional feature foundation for the final relation type classification output.

[0030] Several output nodes are set in the classification layer, each corresponding to a predefined supply chain relationship type. A linear transformation and normalization process is performed on the intermediate feature vectors to generate a set of numerical scores corresponding to each relationship type. Specifically, the process of performing the linear transformation and normalization on the intermediate feature vectors includes the following operations: First, a linear matrix multiplication operation is performed between the intermediate feature vector output from the previous fully connected layer and the weight matrix of the classification layer, and a bias term is added to obtain the raw response score of each output node. Next, these raw scores are used as input to the multi-classification objective and normalized using the Softmax function to calculate the normalized probability value of the relationship type corresponding to each output node. The Softmax function transforms the response value distribution of all output nodes into a probability distribution with a sum of 1, representing the probability that the current node belongs to each supply chain relationship type. This processing method enables the classification layer to score multiple predefined relationship types in parallel, providing a standardized and comparable numerical basis for the final relationship type judgment.

[0031] Based on the index number corresponding to the output node with the highest score in the numerical score set, determine the supply chain relationship type label to which the current node belongs.

[0032] In this embodiment, the supply chain relationship types include direct raw material supply, tiered transfer supply, agency supply, production collaboration and distribution. The rules for determining different types of relationships include: For direct supply relationships of raw materials, the criteria for identifying output nodes are: there are direct transaction records between node pairs, the contract amount is higher than the set threshold, the transaction frequency is greater than the set number of times, and there are no transit nodes in the corresponding logistics path; For the hierarchical supply relationship type, the identification criteria for output nodes are: there is a two-hop path connection between node pairs in the supply chain graph structure, the intermediate nodes in the path do not participate in the contract signing record, and the logistics direction remains consistent in the path; For the agency supply relationship type, the identification criteria for the output node are: there are transaction records between the node pairs, and there is an actual supply path with more than three hops in the graph structure, where the intermediate node of the path has transaction contract records with several other nodes, but the logistics direction field does not point to the target node. For production collaboration relationships, the criteria for identifying output nodes are: the industry type of the node pair is the same, the overlap ratio of contract performance cycles is higher than the set threshold, and the proportion of transaction amount to the total transaction amount of the node is lower than the set threshold. For distribution relationship types, the criteria for identifying output nodes are: the out-degree of the identified node in the node pair is greater than its in-degree, and the node belongs to the retail or terminal category.

[0033] In this embodiment, step S6 involves performing path completion on the original graph structure, including identifying node pairs classified as agent supply relationships, tracing indirect upstream supply paths, and inserting real supply edges that directly connect node pairs into the graph structure.

[0034] In this embodiment, the path hop count verification in step S7 includes recalculating the shortest path hop count for each supply relationship node pair in the repaired graph structure, and determining whether the recalculated shortest path hop count is lower than the path hop count in the original graph structure, thus verifying whether the path completion operation effectively shortens the supply path length.

[0035] In this embodiment, the upstream role consistency verification in step S7 includes: for each upstream node in the repaired path, based on the industry type dimension and legal person nature dimension in the three-dimensional node feature vector, combined with the path depth component and path direction identifier in the structural hierarchy embedding vector, determining whether the node has the upstream enterprise characteristics required by the corresponding supply relationship type; when the node attributes do not meet the upstream role requirements of the identified relationship type, the corresponding path edge of the node is removed from the graph structure.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to the supply network identification and management task of a manufacturing enterprise group. The enterprise has multiple production and processing subsidiaries and a large number of supply chain transaction relationships. However, due to the lack of historical data records, some actual supply paths are obscured by intermediate nodes such as agents and transshipment companies, resulting in a serious distortion of the supply network diagram structure constructed in the internal management system. This makes it impossible to accurately identify the real supply and demand relationship and multi-level structure between enterprises, thereby affecting the identification of upstream suppliers, the calculation of dependence on key materials, and the identification of risk nodes.

[0037] During implementation, the historical supply chain transaction data of the enterprise was first cleaned and preprocessed to extract information such as enterprise identifier, industry type, legal entity nature, contract amount, transaction frequency, and logistics direction. Missing fields were filled and the data was normalized to a unified format. Then, an initial supply chain graph structure was constructed with each enterprise as a node and the existing contract and logistics transaction relationships as edges. For each node, a three-dimensional node feature vector was generated based on its industry type, legal entity nature, and historical transaction frequency. For each edge, a one-dimensional attribute vector containing amount, frequency, and logistics direction was generated. A hierarchical position code was constructed based on the node's in-degree, out-degree, shortest path depth, and path direction, and concatenated with the three-dimensional node feature vector to obtain a set of structural hierarchical embedding vectors.

[0038] The structural hierarchy embedding vector is input into the Geom-GCN model. The model automatically performs graph distance-based aggregated neighbor filtering on the target node, retaining only nodes with a hop count of less than or equal to 2 that are structurally consistent. Attention weighting is applied to these nodes, and a set of node pairs is constructed based on the aggregated structure-aware embedding representation. This set is then input into a multi-level classification network, which predicts the relationship type of all node pairs. The output includes five types of relationship labels: direct supply of raw materials, tiered transfer supply, agency supply, production collaboration, and distribution. Based on this, path completion and edge insertion operations are performed on the original graph structure according to the recognition results to reconstruct the real multi-level supply chain structure. Path hop count verification and upstream role consistency verification are then performed.

[0039] This invention identified 86 actual supply paths obscured by proxy nodes in the dataset. The average number of compressed hops was 1.72, and the proportion of true one-hop supply relationships increased by 14.2%. By comparing with the relationship structure output by the company's original system, it was found that this method could identify 73 multi-level dependencies that were missed by the original system, including 17 hierarchical supply chain paths for key raw materials. The recovery rate of the overall supply network was improved by 12.6%. The table below shows the comparison of the identification performance of the method based on this invention and the traditional rule engine method under several key indicators, which fully verifies the identification accuracy and structure repair capability of this method in the actual supply network.

[0040] Table 1: Performance Comparison of Different Methods in Identifying Multi-Level Relationships in Supply Chains

[0041] Based on the comparison data of the multi-level supply chain relationship identification performance shown in the table, it can be seen that the method of this invention outperforms the traditional rule-based method in several key indicators. First, in terms of the accuracy of identifying real supply paths, the method of this invention reaches 91.8%, which is 15.3% higher than the traditional method. This indicates that by introducing the Geom-GCN structure and the structure-aware embedding vector modeling method, the real supply relationships that are obscured by proxy relationships or transfer paths can be identified more accurately. Second, in terms of obscured path restoration, the method of this invention successfully identified and completed 86 actual supply paths that were not explicitly reflected in the original graph structure, while the traditional rule-based method only restored 29, an improvement of 196.6%, which fully demonstrates the effectiveness of the graph structure path completion mechanism.

[0042] Furthermore, the average path hop count compression value reached 1.72, far exceeding the 0.84 of the traditional method. This indicates that through path completion and real edge insertion, the supply paths in the supply chain network are more compact, and the identified supply and demand relationships are closer to the actual structure. In terms of the accuracy of identifying hierarchical transfer relationships, this method reached 89.1%, which is 18.5% higher than the traditional method, demonstrating its advantage in identifying multi-hop connection relationships. In addition, the pass rate of upstream role consistency verification also reached 93.4%, far exceeding the 79.8% of the traditional method. This reflects that this method is more rigorous in role attribute matching logic and graph structure semantic modeling, and can effectively avoid structural misidentification.

[0043] Finally, in terms of overall network structure recovery rate, this method improves by 12.6% compared to the original graph structure, which means that not only are more real paths identified, but also the hierarchical coherence and topological accuracy of the network are effectively enhanced. The above data comprehensively verify the reliability and practical value of this invention in the task of identifying multi-level relationships in the supply chain.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying multi-level relationships in a supply chain based on graph neural networks, characterized in that, Includes the following steps: S1. Collect supply chain data and preprocess it to construct the original supply chain diagram structure; S2. Generate a three-dimensional node feature vector based on the enterprise's industry type, legal entity nature, and historical transaction frequency; S3. Construct a hierarchical position code based on the in-degree, out-degree, shortest path depth and path direction identifier of each node in the original supply chain graph structure. Concatenate the hierarchical position code with the 3D node feature vector to generate a set of structural hierarchical embedding vectors. S4. Input the set of structure-level embedding vectors into the Geom-GCN model, perform a neighbor selection operation based on graph distance, use the structural orientation consistency judgment to filter out the set of aggregated nodes that match the supply path direction, apply attention weighting operation to the set of aggregated nodes, and generate a set of structure-aware embedding representations. S5. Construct a node pair input set based on the structure-aware embedding representation set, input it into a multi-class classification network, and output the supply chain relationship type label for each pair of nodes; S6. Complete the path of the original graph structure according to the supply chain relationship type label, insert the real supply edge that is hidden by the agent node, and build the graph structure after structural repair. S7. Perform path hop count verification and upstream role consistency verification on the graph structure after structural repair to generate the final multi-level supply chain relationship identification result.

2. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 1, characterized in that, The supply chain data in step S1 includes enterprise identification, industry type, legal entity nature, contract amount, transaction frequency, and logistics direction; the preprocessing includes filling missing values, normalizing fields, aligning time, and unifying node identification in the supply chain data.

3. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 2, characterized in that, The construction of hierarchical location coding in step S3 specifically includes: calculating the in-degree of each node in the original supply chain graph structure as the first coding component; calculating the out-degree of each node as the second coding component; calculating the shortest path step length from the node to the upstream node based on the shortest path algorithm in the graph as the third coding component; extracting the direction identifiers of the outgoing and incoming edges of the node, and generating path direction identifier codes according to the logistics flow direction as the fourth coding component; and concatenating the four components of in-degree, out-degree, shortest path depth, and path direction identifier to form a hierarchical location coding used to describe the hierarchical information of the node structure.

4. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 3, characterized in that, Step S3, which involves concatenating the hierarchical position encoding with the three-dimensional node feature vector, includes: connecting the three-dimensional node feature vector of each node with the corresponding hierarchical position encoding vector in the dimensional direction to generate a structural hierarchical embedding vector containing node attributes and structural position features, and summarizing the structural hierarchical embedding vectors of all nodes to form a set of structural hierarchical embedding vectors.

5. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 4, characterized in that, The process of generating the structure-aware embedded representation set in step S4 includes the following operations: The set of structural hierarchy embedding vectors is input into the Geom-GCN model. For each target node in the original supply chain graph structure, the graph distance calculation operation is performed to obtain the shortest path hop count from the target node to all other nodes. A graph distance mapping table is generated to record the hop count relationship between each potential neighbor node and the target node. Select all nodes with a hop count of less than or equal to two from the mapping table to form an initial neighbor candidate set. Only retain one-hop and two-hop nodes, and remove high-order nodes in the graph structure that are too far away and may cause feature interference. Read the logistics direction field of the edge between each node and the target node in the initial neighbor candidate set, extract the shortest path depth dimension value corresponding to the embedding vector of the neighbor node structure hierarchy, and calculate the difference with the same dimension value of the target node. When the difference is positive and the logistics direction field indicates that the information flow is from the neighbor node to the target node, it is determined that the neighbor node is upstream of the target node in the structural hierarchy and has the same supply direction. The neighbor node is retained in the set of aggregation nodes with the same direction, and the other neighbor nodes do not participate in information aggregation. For each node in the set of nodes with consistent orientation, the semantic proximity score between the node in the set of nodes with consistent orientation and the target node is calculated based on the spatial position distribution of the node embedding vector. The semantic proximity score represents the degree of consistency of the structural semantics of the two nodes in the embedding space. The semantic proximity scores of all aggregated nodes are normalized to form the attention weight set corresponding to the target node. The contribution of different aggregated nodes to the representation of the target node is controlled by the weights. The embedding vectors of all aggregate nodes are weighted and summed according to the normalized weights to obtain the structure-aware embedding vector of the current target node. The perception embedding vector integrates the features of this node, the structural hierarchy information, and the upstream and downstream semantic features of the direction-consistent aggregate substructure. The structure-aware embedding vectors generated from all target nodes are combined sequentially to form a structure-aware embedding representation set.

6. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 5, characterized in that, The multi-class classification network processing in step S5 includes: For any two enterprise nodes in the original supply chain graph that are connected by a structural path, extract the embedding vectors of the two enterprise nodes in the set of structure-aware embedding representations, and denote them as the first node vector and the second node vector, respectively. The first node vector and the second node vector are concatenated in sequence to construct the node pair input vector, and the node pair input vector is input into the multilayer feedforward neural network. The multilayer feedforward neural network includes a set of input layers with fixed input dimensions, two or more fully connected transformation layers with nonlinear activation functions, and a classification layer consisting of a set of output nodes. In each fully connected transformation layer, matrix transformation and activation function calculation operations are performed on the input node pair vector to generate intermediate feature vectors of fixed dimensions. Several output nodes are set in the classification layer, each output node corresponds to a predefined supply chain relationship type. Linear transformation and normalization are performed on the intermediate feature vector to generate a set of numerical scores corresponding to each relationship type. Based on the index number corresponding to the output node with the highest score in the numerical score set, determine the supply chain relationship type label to which the current node belongs.

7. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 6, characterized in that, The supply chain relationship types include direct raw material supply, tiered supply, agency supply, production collaboration and distribution. The rules for determining different types of relationships include: For direct supply relationships of raw materials, the criteria for identifying output nodes are: there are direct transaction records between node pairs, the contract amount is higher than the set threshold, the transaction frequency is greater than the set number of times, and there are no transit nodes in the corresponding logistics path; For the hierarchical supply relationship type, the identification criteria for output nodes are: there is a two-hop path connection between node pairs in the supply chain graph structure, the intermediate nodes in the path do not participate in the contract signing record, and the logistics direction remains consistent in the path; For the agency supply relationship type, the identification criteria for the output node are: there are transaction records between the node pairs, and there is an actual supply path with more than three hops in the graph structure, where the intermediate node of the path has transaction contract records with several other nodes, but the logistics direction field does not point to the target node. For production collaboration relationships, the criteria for identifying output nodes are: the industry type of the node pair is the same, the overlap ratio of contract performance cycles is higher than the set threshold, and the proportion of transaction amount to the total transaction amount of the node is lower than the set threshold. For distribution relationship types, the criteria for identifying output nodes are: the out-degree of the identified node in the node pair is greater than its in-degree, and the node belongs to the retail or terminal category.

8. The method for identifying multi-level relationships in a supply chain based on graph neural networks according to claim 7, characterized in that, In step S6, path completion is performed on the original graph structure, including identifying node pairs classified as agent supply relationships, tracing indirect upstream supply paths, and inserting real supply edges that directly connect node pairs into the graph structure.

9. A method for identifying multi-level relationships in a supply chain based on a graph neural network according to claim 8, characterized in that, The path hop count verification in step S7 includes recalculating the shortest path hop count for each supply relationship node pair in the repaired graph structure, and determining whether the recalculated shortest path hop count is lower than the path hop count in the original graph structure, thus verifying whether the path completion operation effectively shortens the supply path length.

10. A method for identifying multi-level relationships in a supply chain based on a graph neural network according to claim 9, characterized in that, The upstream role consistency verification in step S7 includes: for each upstream node in the repaired path, based on the industry type dimension and legal person nature dimension in the three-dimensional node feature vector, combined with the path depth component and path direction identifier in the structural hierarchy embedding vector, determining whether the node has the upstream enterprise characteristics required by the corresponding supply relationship type; when the node attributes do not meet the upstream role requirements of the identified relationship type, the corresponding path edge of the node is removed from the graph structure.

Images