Hktgnn-based industrial chain risk assessment and prediction model and method and medium

Abstract

The application discloses an industry chain risk assessment and prediction model based on HKTGNN, comprising a data acquisition module, a preprocessing module, a graph embedding encoder GEE module, a center-based cross-domain adaptive feature completer CCAFC module, a center-based domain adaptive message passing DCAMP module and a domain transferable classifier DTC module. The application also provides an industry chain risk assessment and prediction method. Through hierarchical processing based on a hierarchical structure, the application converts a traditional constructed directed heterogeneous graph into a product network homogeneous directed graph with a smaller graph scale, reduces the calculation time and memory consumption of a single graph, and effectively saves the calculation resources. Through feature embedding and feature supplementing, the performance of the model is improved. Meanwhile, based on the center of the field difference feature supplement and the message passing, the data starvation problem is solved.

Description

Technical Field

[0001] This invention relates to the field of supply chain graph data mining technology, specifically to a supply chain risk assessment and prediction model, method, and medium based on hierarchical knowledge transferable graph neural network (HKTGNN). Background Technology

[0002] The increasing complexity of supply chains, coupled with the growing number of companies, presents challenges to assessing supply chain risks. These challenges are compounded by the digital revolution in traditional business processes and supply chain connections. Investors and suppliers have recognized that these issues may offer future competitive advantages. Beyond earlier methods based on graph theory and complexity, machine learning approaches have been proposed in recent years to perform link prediction or clustering of different labeled networks to build supply chain risk assessment models. These methods are gaining popularity among researchers using knowledge graphs (KG) or graph neural networks (GNN) for knowledge reasoning.

[0003] Because the regulatory bodies in a supply chain include individuals and suppliers, the complexity of the supply chain and the specific market objectives of suppliers make them reluctant to disclose all data. Therefore, typical data collection methods often produce only partial graphical information. Existing work in the field of machine learning has primarily focused on link prediction to overcome these difficulties. In addition to using topology-based frameworks, knowledge graphs, and graph neural networks for link prediction, Britrup et al. proposed the "Digital Supply Chain Risk Monitoring" (DSCRS) framework in 2022. These methods leverage machine learning to mine data, improve supply chain data visibility, and enable relevant professionals to extract useful information from the supply chain, such as hidden supply and component relationships, to enhance supply chain visibility. However, link prediction methods do not fundamentally solve the problem of collecting supply chain data from its source.

[0004] Furthermore, enhancing the visibility of supply chain relationships has not yet successfully addressed the issues of data scarcity and distribution transfer. On the one hand, due to observational limitations, incomplete domain knowledge, and distribution transfer problems, the collected graphs often contain incomplete node features and unusable labels, leading to data hunger in various machine learning applications. On the other hand, non-listed companies are not obligated to publish their financial statements, while listed companies may lack financial statements for specific quarters, resulting in missing data for certain nodes. Both types of companies are important for financial risk assessment. This information is more about the specific details of nodes in the supply chain, and mining graph data solely from the perspective of link prediction is insufficient.

[0005] Furthermore, the supply chain includes multiple supply chains with only a single product, resulting in a very large volume of graph data. The numerous nodes, edges, and feature attributes in the supply chain network require and consume significant computing resources, often leading to memory shortages. How to retain as much information as possible in the graph while minimizing computing resource usage has become one of the challenges in practical applications. Summary of the Invention

[0006] To address the problems of the existing technologies, this invention provides a supply chain risk assessment and prediction model based on HKTGNN. It utilizes data from the supply chain that is easier to collect and more in line with the real world, combined with a hierarchical structure concept to construct a homogeneous directed graph of product networks with a smaller graph size, saving computational resources. Furthermore, it uses a centrality-based domain difference knowledge transfer module to implement feature completers for data-biased nodes and intra-domain / cross-domain message passing through CB-Graph, alleviating the data hunger problem.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A supply chain risk assessment and prediction model based on HKTGNN includes:

[0009] The data acquisition module is used to acquire industry chain datasets based on graph neural networks, which include company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links.

[0010] The preprocessing module is used to construct an industry chain network based on the acquired industry chain dataset through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain.

[0011] The Graph Embedding Encoder (GEE) module is used to embed the single product network graph of each product and its carried feature information into the homogeneous directed graph of the product network using graph representation learning methods. The feature information of each node consists of graph embedding features, financial statement features, and financial features.

[0012] The centrality-based cross-domain adaptive feature complementer (CCAFC) module is used to transfer the feature information of the complete nodes in the embedded homogeneous directed graph of the product network to the corresponding biased nodes according to the defined complete nodes and biased nodes, thereby completing the feature information supplementation of the biased nodes.

[0013] The centrality-based domain adaptive message passing (DCAMP) module is used for cross-domain message passing between complete and biased nodes in a homogeneous directed graph of a product network. It projects the source node's characteristic information onto the target node's domain by calculating a domain difference ratio factor to achieve cross-domain message passing.

[0014] The Domain Transferable Classifier (DTC) module is used to perform binary classification on the embedded nodes of the single product network graph corresponding to each product based on message passing in the homogeneous directed graph of the product network. This results in the normal or risk label of the corresponding product. Then, the impact of the product on the risk of the industrial chain is determined by analyzing the specific position of the product in the industrial chain, thereby realizing the assessment and prediction of industrial chain risk.

[0015] Specifically, a complete node is a node with complete industry chain data, while a biased node is a node with missing industry chain data.

[0016] Specifically, each node in the homogeneous directed graph of the product network is a single product network graph carrying characteristic information, and each edge is a link from one product to another.

[0017] Specifically, the graph embedding encoder (GEE) module uses graph representation learning methods to embed the single-product network graph of each product and its carried feature information into the homogeneous directed graph of the product network as follows:

[0018] Define the network graph of the i-th single product. Node set node

[0019] The graph embedding layer and its embedding information can be obtained using the following formula:

[0020]

[0021] In the above formula, Indicates the in-degree of a node. φ represents the out-degree of a node. j This indicates the node type, including four types: product nodes, listed company nodes in a single product supply chain, enterprise nodes invested in by listed companies in a single product supply chain, and source enterprise nodes from which listed companies in a single product supply chain receive supply or investment. (SPB) j Represents the shortest path to a node.

[0022] A top Laplacian vector is introduced for each node. The largest eigenvalue is the j-th value of the top Laplacian vector, from which the eigenvalues ​​of each node are obtained.

[0023]

[0024] In the above formula, yes The top Laplacian vector;

[0025] Then, the feature vectors of the industry chain network and each node are fed into the graph isomorphic network GIN for training, and finally the embedding vector of each node in the homogeneous directed graph of the product network is obtained, forming a graph CB-Graph with product-to-product links as edges and single product network graphs as nodes.

[0026] Specifically, the CB-Graph is represented as follows:

[0027] Where V represents the set of nodes in the graph, including complete nodes and biased nodes, and node v ij ∈V=V complete ∪V biased E represents the edge set in the graph; This represents the feature information attributes of all nodes in the graph, and the feature value vector of each node. Represents observable feature information attributes. This represents the unobservable feature information attribute that needs to be completed; Y represents the risk label, which marks a risky company as risky when it exists; ψ is the CB-Graph label, used to label the node attributes in the graph, when node v ij When the node is complete, ψ = 1; when node v ij When there are biased nodes, ψ = 0.

[0028] Specifically, the centrality-based cross-domain adaptive feature complementer (CCAFC) module transfers the feature information of complete nodes in the embedded homogeneous directed graph of the product network to the corresponding biased nodes. The process of supplementing the feature information of biased nodes is as follows:

[0029] Define a complete node set V complete and the biased node set V biased V complete ∪V bia d =V supply ,

[0030] Will Feature information knowledge transfer Complete the supplementation of feature information in the middle, so that Become form And use the hyperparameter K to limit the number of iterations for each transformation;

[0031] For each biased node Its unobservable vector After the supplement is completed, it will be:

[0032]

[0033] Where f(·) is the neighbor importance factor, This represents the visible input of the i-th node. ω represents the visible input of the j-th node. b and ω c This represents the final self-learning matrix. It is an unobservable vector The calibration variable eliminates domain differences:

[0034]

[0035] In the above formula, and Here is the self-learning parameter matrix; σ is the softsign activation function; It is a centrality-based transform-domain difference, where It is the feature information attribute of the i-th complete node. It is the feature information attribute of the j-th biased node. For the centrality of the i-th complete node, The centrality of the j-th biased node;

[0036] Then, eigenvector centrality is used as the calculation method. After a biased node is supplemented, the next iteration uses the supplemented biased node as the information provider to achieve the goal of...

[0037] In this process, using breadth-first search (BFS) results in:

[0038]

[0039]

[0040] In the above formula, a cb The influence weights of centrality are represented by σ1, which represents the PReLU activation function.

[0041] The algorithm converges through iteration when the number of iterations reaches the maximum value K or when all nodes are completed.

[0042] During the iteration process, a distribution consistency loss function is introduced.

[0043]

[0044] Modify the centrality.

[0045] Specifically, the centrality-based domain adaptive message passing (DCAMP) module calculates a domain difference ratio factor to project the source node feature information of the message to the domain where the target node resides, thus achieving cross-domain message passing.

[0046] For a given edge e i,j =(v i ,v j The message passing equation is:

[0047]

[0048] In the above formula, h j This represents the message passing tensor of node j. This represents the feature information calibrated using the domain difference factor; It is the neighbor importance function:

[0049]

[0050] In the above formula, σ² is the leaky ReLU activation function. This represents the self-learning matrix that needs to be obtained. This represents the weight term; therefore, the distribution shift variable Δ i,j :

[0051]

[0052] In the above formula, σ3 is the softsign activation function, ψ(v i ) represents the CB-Graph label, h i The tensor representing node i used for message passing; ΔX C It is a centrality-based transform-domain difference:

[0053]

[0054] In the above formula, It is the feature information attribute of the i-th complete node. It is the feature information attribute of the j-th biased node. For the centrality of the i-th complete node, The centrality of the j-th biased node;

[0055] This leverages centrality-based domain differences to achieve cross-domain message passing.

[0056] Furthermore, this invention also provides a supply chain risk assessment and prediction method based on HKTGNN, comprising the following steps:

[0057] S10. Obtain the industry chain dataset based on the graph neural network, which includes company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links.

[0058] S20. Based on the obtained industry chain dataset, construct an industry chain network through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain.

[0059] S30. Using graph representation learning methods, the single product network graph of each product and its carried feature information are embedded into the homogeneous directed graph of the product network. The feature information of each node is composed of graph embedding features, financial statement features, and financial features.

[0060] S40. Based on the defined complete nodes and biased nodes, transfer the feature information of the complete nodes in the embedded product network homogeneous directed graph to the corresponding biased nodes to complete the feature information supplementation of the biased nodes.

[0061] S50. Based on the message passing in the homogeneous directed graph of the product network, complete the binary classification of the embedded nodes of the single product network graph corresponding to each product, obtain the normal label or risk label of the corresponding product, and then determine its impact on the risk of the industrial chain by analyzing the specific position of the product in the industrial chain, and obtain the score result of the risk assessment of the industrial chain, thereby realizing the assessment and prediction of the risk of the industrial chain.

[0062] In the cross-domain message passing between complete nodes and biased nodes in a homogeneous directed graph of a product network, the source node feature information of the message is projected onto the domain of the target node by calculating the domain difference ratio factor to achieve cross-domain message passing.

[0063] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described supply chain risk assessment and prediction method based on HKTGNN.

[0064] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it outputs the results of supply chain risk assessment and prediction using the aforementioned supply chain risk assessment and prediction model based on HKTGNN.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) This invention constructs the concept of CB-Graph through convenient product-to-product and product-to-enterprise connections, focusing supply chain risk assessment on products within the supply chain rather than scattered enterprises, thus avoiding the problem of low data quality caused by low enterprise data visibility. In supply chain data acquisition and processing, hierarchical processing based on a hierarchical structure transforms the traditionally constructed directed heterogeneous graph into a homogeneous directed graph of product networks with a smaller graph size, reducing the computation time and memory consumption of a single graph and effectively saving computing resources; and through feature embedding and feature supplementation, the performance of the model is improved. At the same time, based on domain-differentiated feature supplementation and message passing centrality, the data hunger problem is solved. Experiments on real-world datasets demonstrate that the model of this invention is better than other models.

[0067] (2) The present invention builds a model based on investment or supply to company links, company to product links and product to product links. These links are easier to collect, effectively reducing the difficulty of data collection and avoiding the problem of collecting real company relationships due to real factors.

[0068] (3) This invention effectively distinguishes biased data nodes by configuring complete nodes and biased nodes in CB-Graph, and realizes feature information supplementation and intra-domain / cross-domain message passing through a module based on centrality-based domain difference knowledge transfer, which effectively alleviates the problem of data hunger. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the model structure of an embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of the hierarchical structure of the industrial chain network in an embodiment of the present invention.

[0071] Figure 3 This is a schematic diagram of the method flow of an embodiment of the present invention.

[0072] Figure 4 This is a schematic diagram of the hyperparameter test results in an embodiment of the present invention. Detailed Implementation

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0074] Example

[0075] like Figures 1 to 3As shown, the supply chain risk assessment and prediction model based on HKTGNN includes a data acquisition module, a preprocessing module, a graph embedding encoder (GEE) module, a centrality-based cross-domain adaptive feature complementer (CCAFC) module, a centrality-based domain adaptive message passing (DCAMP) module, and a domain transferable classifier (DTC) module.

[0076] The data acquisition module is used to acquire industry chain datasets based on graph neural networks, which include company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links.

[0077] The preprocessing module is used to construct an industry chain network based on the acquired industry chain dataset through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain.

[0078] The Graph Embedding Encoder (GEE) module is used to embed the single product network graph of each product and its carried feature information into the homogeneous directed graph of the product network using graph representation learning methods. The feature information of each node consists of graph embedding features, financial statement features, and financial features.

[0079] The centrality-based cross-domain adaptive feature complementer (CCAFC) module is used to transfer the feature information of the complete nodes in the embedded homogeneous directed graph of the product network to the corresponding biased nodes according to the defined complete nodes and biased nodes, thereby completing the feature information supplementation of the biased nodes.

[0080] The DCAMP module, based on centrality, is used to project the characteristic information of the source node of the message to the domain of the target node in cross-domain message passing between complete nodes and biased nodes in a homogeneous directed graph of a product network by calculating the domain difference ratio factor.

[0081] The Domain Transferable Classifier (DTC) module is used to perform binary classification on the embedded nodes of the single product network graph corresponding to each product based on message passing in the homogeneous directed graph of the product network. This results in the normal or risk label of the corresponding product. Then, the impact of the product on the risk of the industrial chain is determined by analyzing the specific position of the product in the industrial chain, and the result score of the industrial chain risk assessment is obtained, thereby realizing the assessment and prediction of industrial chain risk.

[0082] Regarding data acquisition, this invention conducts preliminary comprehensive analysis of various data in the industry chain and finds that only three types of directed links are easier to collect: investment or supply to company links, company to product links, and product to product links. Therefore, the modeling of this invention is based on these three types of links, avoiding the problem of difficulty in collecting potentially hidden real company relationships, and is more suitable for real-world situations.

[0083] Related knowledge application instructions:

[0084] Graph Neural Networks (GNNs) are a method for learning graph representations, and they are commonly used for semi-supervised node classification tasks. For a given graph... As the input dataset, V is the number of nodes in the set graph; E is the set of edges in the graph; Represents the characteristic attributes of a node; The labels representing the nodes used for classification are used in mainstream graph neural networks, which employ multiple layers of graph convolutional layers and message-based aggregation to learn node representations. The updated formula for each layer is as follows:

[0085]

[0086] Where h (l) Let M represent the node representation learning vector of the l-th layer of the GNNs; M represents the message equation for aggregating neighbor features; and U represents the update equation.

[0087] This invention models supply chain graph data obtained for supply chain risk assessment tasks, distinguishing between complete and biased nodes. This data contains a complex variety of node (type) and edge (type) data, whose feature attributes are crucial for representation learning. However, in the real world, collecting these feature attributes often presents challenges, such as the inability to obtain complete or partial financial statements. This necessitates defining a new label to aid model training. We define nodes with complete data as complete nodes and nodes with biased data (missing data) as biased nodes. Besides the lack of financial information for non-listed companies and the absence of financial statements for listed companies, there are also cases where structural deficiencies in the supply chain graph data construction lead to biased embedding vectors. For example, some products lack company-product links, or we cannot collect links to certain investments or supplies to companies from our sources. This is termed "completeness deficiency" in supply chain risk assessment, and we label these as biased nodes, assigning them a value of 0 and marking them as risk. Similarly, nodes in representation learning that contain missing financial statement information are also labeled as biased nodes.

[0088] CB-Graph is a graph-structured data structure designed in this invention, which can be used to train the hierarchical knowledge-transferable graph neural network HKTGNN designed in this invention, and is represented as follows:

[0089] Where V represents the set of nodes in the graph, including complete nodes and biased nodes, and node v ij ∈V=V complete ∪V biased E represents the edge set in the graph; This represents the feature information attributes of all nodes in the graph, and the feature value vector of each node. Represents observable feature information attributes. ψ represents the unobservable feature information attribute that needs to be completed; Y represents the risk label, which marks node i as risky when a risky company exists; ψ is the CB-Graph label, used to label the node attributes in the graph, when node v ij When the node is complete, ψ = 1; when node v ij When there are biased nodes, ψ = 0.

[0090] The following is a further explanation of each module:

[0091] Regarding the preprocessing module, previous data mining methods based on GNNs typically construct a large graph containing all company and product nodes. Because different types of nodes (products, companies, investments, suppliers, etc.) contain different types of investment and supply relationships, this often results in a directed heterogeneous graph with a large number of nodes. This invention cleverly utilizes the concept of hierarchical structures, constructing a directed isogeneous graph with fewer nodes through layered processing and graph embedding. Defined as a heterogeneous diagram of the entire industry chain network; As a heterogeneous graph of the industrial chain network built around a single product, that is, a single-product network graph for each product; As the only homogeneous directed graph in the product network, it is used as the bottom layer for training hierarchical structures, such as Figure 2 As shown.

[0092] We construct a real-world dataset representing a specific industry chain (such as the integrated circuit supply chain) using data provided by a listed company. After layering and graph embedding, for a given product CB-Graph... For us, the representation of the learned feature X E GEE from a hierarchical structure, with dimensions of Financial Feature X B z-score normalization from enterprise data and one-hot encoding of product business information, with dimensions of Financial Statement Feature X F These are the z-score normalized results of corporate financial statements published on Wind and RESSET, with dimensions of [missing information]. For each representation learning node i, we take each product node The characteristics of its financial statements are as follows: in This represents the average of the financial statements of all companies in node i. We use Tongdaxin's stock rating to define the risk label of listed companies: 0-80 points are set as medium or high risk; 80-100 points are set as low or no risk. For a product, the risk of its industrial chain network is determined by the risk label of the companies in it. For the CB label of a node, we mainly define the following situations as biased: (1) lack of investment or supply relationship with non-listed companies or natural persons; (2) missing or incomplete financial statements of listed companies; (3) inability to obtain relevant industrial chain companies for the corresponding product. When a company in a product has any of the situations (1) or (2), it is marked as biased. It can be seen that the characteristic attribute values ​​of biased product nodes do not need to be processed to 0.

[0093] Regarding the Graph Embedding Encoder (GEE) module, after the industry chain network is constructed, representation learning needs to begin. The encoder embeds the single-product network graph of each product into the homogeneous directed graph of the industry chain network's product network, forming a new, easily trainable homogeneous directed graph. Each node is a single-product network graph carrying feature information, and each edge is a product-to-product link connecting nodes. The implementation process of the GEE module designed in this invention is as follows.

[0094] First, the goal of graph embedding is to embed single-product supply chain network information into the characteristic attributes of nodes in the product network.

[0095] Define the network graph of the i-th single product. Node set Take the node with no attributes The topology information considered mainly includes:

[0096] 1) The degree of each node, in terms of Indicate the in-degree of each node, in... This represents the out-degree of each node;

[0097] 2) Node type φ for each node j The present invention includes four types: product nodes, listed company nodes in a single product industry chain, enterprise nodes invested by listed companies in a single product industry chain, and source enterprise nodes from which listed companies in a single product industry chain obtain supply or investment.

[0098] 3) Shortest path SPB for each node j Because the scale of the graph is reduced in the industrial chain network graph constructed in this invention, the time cost and memory cost of processing the shortest path are greatly reduced.

[0099] Using the above information and the following function formula, obtain the graph embedding layer and its embedding information:

[0100]

[0101] This also introduces the top Laplacian vector for each node. For Each node in node The largest eigenvalue is the j-th value of the top Laplacian vector, from which the eigenvalues ​​of each node are obtained.

[0102]

[0103] In the above formula, yes The top Laplacian vector;

[0104] Then, the feature vectors of the industry chain network and each node are fed into the graph isomorphic network GIN for training, and finally the embedding vector of each node in the homogeneous directed graph of the product network is obtained, forming a graph CB-Graph with product-to-product links as edges and single product network graphs as nodes.

[0105] The centrality-based cross-domain adaptive feature complementer module CCAFC needs to complete the cross-domain feature complementation for all biased nodes. First, define a complete node set V. complete and the biased node set V biased V complete ∪V biased =V supply At the start of module execution, there is Will Feature information knowledge transfer Complete the supplementation of feature information in the middle, so that Become form A hyperparameter K is used to limit the number of iterations for each transformation.

[0106] For each incomplete biased node Its unobservable vector After the supplement is completed, it will be:

[0107]

[0108] Where f(·) is the neighbor importance factor, This represents the visible input of the i-th node. ω represents the visible input of the j-th node. b and ω c This represents the final self-learning matrix. It is an unobservable vector The calibration variable eliminates domain differences:

[0109]

[0110] In the above formula, and Here is the self-learning parameter matrix; σ is the softsign activation function; It is a centrality-based transform-domain difference, where It is the feature information attribute of the i-th complete node. This is the feature information attribute of the j-th biased node; considering that the graph embedding vector is used as the visible feature attribute in this invention, which only contains the topological information of the single-product supply chain network, we improve the algorithm by using centrality weighting based on the idea of ​​weighted average, where... For the centrality of the i-th complete node, Let be the centrality of the j-th biased node.

[0111] Then, eigenvector centrality is used as the calculation method. After a biased node is supplemented, the next iteration uses the supplemented biased node as the information provider to achieve the goal of...

[0112] In this process, using breadth-first search (BFS) results in:

[0113]

[0114]

[0115] In the above formula, a cb This represents the influence weights based on centrality, and σ1 represents the PReLU activation function.

[0116] The algorithm converges through iteration when the number of iterations reaches the maximum value K or when all nodes are completed.

[0117] To ensure that the aforementioned equations It is fully utilized by introducing a distribution consistency loss function during the iteration process.

[0118]

[0119] The centrality was modified. Subsequent verification showed that this design performed better than the original formula on supply chain network data.

[0120] Regarding the DCAMP module based on centrality for domain adaptive message passing, after feature supplementation, it is necessary to pass messages on the isomorphic graph with attributes to finally complete binary classification. For message passing on the CB-Graph, there are four specific cases: (1) message passing between complete nodes and biased nodes; (2) message passing between biased nodes; (3) message passing between biased nodes and complete nodes; (4) message passing between complete nodes. Among them, cases (1) and (3) are Out-of-Distribution (OOD) cross-domain message passing; cases (2) and (4) are intra-distribution (ID) domain message passing. Cross-domain OOD message passing may be noisy and will reduce model performance. Therefore, this DCAMP module is used to solve this problem: by calculating the domain difference ratio factor, the OOD feature information of the source node of the message passing is projected to the domain where the target node is located to realize cross-domain message passing. This DCAMP module mainly considers two factors: one is the bidirectional domain difference based on centrality, and the other is the importance of neighbors.

[0121] For a given edge e i,j =(v i ,v j The message passing equation is:

[0122]

[0123] In the above formula, h j This represents the message passing tensor of node j. This represents the feature information calibrated using the domain difference factor; It is the neighbor importance function:

[0124]

[0125] In the above formula, σ² is the leaky ReLU activation function. This represents the self-learning matrix that needs to be obtained. This represents the weight term; therefore, the distribution shift variable Δ i,j :

[0126]

[0127] In the above formula, σ3 is the softsign activation function, ψ(v i ) represents the CB-Graph label, h i The tensor representing node i used for message passing; ΔX C It is a centrality-based transform-domain difference:

[0128]

[0129] In the above formula, It is the feature information attribute of the i-th complete node. It is the feature information attribute of the j-th biased node. For the centrality of the i-th complete node, The centrality of the j-th biased node;

[0130] This leverages centrality-based domain differences to achieve cross-domain message passing.

[0131] The Domain Transferable Classifier (DTC) module uses the DTC algorithm to implement the output.

[0132] In another embodiment, such as Figure 3 As shown, a supply chain risk assessment and prediction method based on HKTGNN is also provided, including the following steps:

[0133] S10. Obtain the industry chain dataset based on the graph neural network, which includes company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links.

[0134] S20. Based on the obtained industry chain dataset, construct an industry chain network through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain.

[0135] S30. Using graph representation learning methods, the single product network graph of each product and its carried feature information are embedded into the homogeneous directed graph of the product network. The feature information of each node is composed of graph embedding features, financial statement features, and financial features.

[0136] S40. Based on the defined complete nodes and biased nodes, transfer the feature information of the complete nodes in the embedded product network homogeneous directed graph to the corresponding biased nodes to complete the feature information supplementation of the biased nodes.

[0137] S50. Based on the message transmission in the homogeneous directed graph of the product network, complete the binary classification of nodes with normal labels and risk labels, and obtain the score for the risk assessment of the industrial chain.

[0138] In the cross-domain message passing between complete nodes and biased nodes in a homogeneous directed graph of a product network, the source node feature information of the message is projected onto the domain of the target node by calculating the domain difference ratio factor to achieve cross-domain message passing.

[0139] In another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described supply chain risk assessment and prediction method based on HKTGNN.

[0140] In another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, outputs the results of supply chain risk assessment and prediction using the aforementioned supply chain risk assessment and prediction model based on HKTGNN.

[0141] The aforementioned computer storage medium may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0142] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0143] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0144] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0145] The following experiment will introduce the supply chain risk assessment and prediction model and method based on HKTGNN.

[0146] I. Dataset

[0147] First, relevant data on the integrated circuit industry chain were collected in advance, saving time and effort for the data acquisition module. Within the integrated circuit industry chain, we collected data on 430 products, 1732 listed companies, and 46273 related investment / supply enterprise or individual nodes. The supply chain network for a single product contains 1875 product-to-product links and 0 to 3137 company-to-product links, with a total of 53438 links across the entire industry chain network. The adjacency matrix formed by the graph data is very sparse.

[0148] The dataset used by HKTGNN was created using the CB-Graph method, resulting in a homogeneous graph containing 430 product nodes, 1875 edges, and each node having a 99-dimensional feature attribute vector, labeled with risk labels and CB labels.

[0149] Unlike the dataset used in the hierarchical model, we need to tailor it to be compatible with the baseline. As shown in Table 1, we use four types of datasets to test our baseline. We present the betweenness centrality, degree centrality, and power-law centrality for each dataset to demonstrate the effectiveness of our model in the comparison experiments.

[0150] The main tasks for preparing the dataset are as follows:

[0151] 1) We set listed companies as the baseline and other models used in the experiment. It's important to note that since a product's supply chain involves multiple listed companies, the risk of each listed company directly impacts the risk of the product's supply chain. Furthermore, the definition of CB-Graph is more lenient than that of a hierarchical model, only considering the lack of financial data. Specifically, even if the baseline and other models achieve the same scores as our model, their actual supply chain risk assessment will still be poor due to the dispersed nature of companies throughout the supply chain.

[0152] 2) Since there are no direct connections between companies in our supply chain network, we only created two types of datasets: ① When product-to-product links exist, all listed companies in product A will be fully connected to all listed companies in product B, or all listed companies in product A will point to all listed companies in product B. ② When product-to-product links exist, listed companies in product A will be randomly connected to listed companies in product B with a certain degree of randomness. For each random connection graph in ②, we will generate ten instances, and the final comparison results will be the average, as shown in Table 2.

[0153] 3) The risk labels are created in the same way as the risk labels in this invention's model. For the characteristic attributes of each node, the financial data comes from the financial statements of each listed company. However, the operational information comes from the average financial information of all products related to the listed company.

[0154]

[0155] Table 1 Baseline Dataset

[0156] Regarding the explanation in Table 1, the dataset is divided into four groups: (1) FJG: fully connected graph; (2) 25%G: graph with 25% random connectivity; (3) 50%G: graph with 50% random connectivity; (4) 75%G: graph with 75% random connectivity. The specific methods for generating connections are provided in the Data Generation section below. We compare betweenness centrality, degree centrality, power-law centrality, and [missing information] to show that our random dataset covers a wide range of degrees and shortest paths, which better demonstrates the effectiveness of our model comparison experiments. Since random connections may cause some nodes to become isolated points, we will remove them to avoid affecting baseline message passing.

[0157] II. Baseline Model

[0158] In this experiment, we mainly used the following models and other models with the following hyperparameters for comparative experiments: (1) MLP, a neural network with multiple hidden layers. (2) GGNN, a graph neural network that uses gated recurrent units and modern optimization techniques, and then extends to the output sequence. (3) GraphSage and (4) GraphSage2, both of which are graph neural network models based on neighbor aggregation for learning graph node representations. (5) KTGNN is a knowledge-transferable graph neural network that can cope with data hunger and distribution drift. (6) HKTGNN proposed in this invention.

[0159] III. Training Details

[0160] In our tests, we found that setting λ=2 performed best for KTGNN; the learning rate and other hyperparameters were the same as in our paper. All other baselines were tuned for best performance on the industry chain dataset. For our model HKTGNN, we set λ to 2.5, γ (the weights of the KL loss) to 2, and K to 1. We set the hyperparameters for the other modules as follows: learning rate lr to 3e-2, dataset split ratio of 7:1:2, and random splitting. The remaining hyperparameters were the same as for KTGNN. We performed ten Monte Carlo experiments on each dataset, with 300 iterations per dataset. Our tests were conducted on NVIDIA GeForce RTX 3090 (32GB) hardware and Python 3.10 as the experimental environment.

[0161] IV. Experimental Results

[0162] We will present our experimental results in two aspects: evaluation scores and ablation studies.

[0163] 4.1 Test Score

[0164] Regarding Table 2, the explanation includes F1 scores and AUC scores. We compared a total of five datasets mentioned above: (1) the supply chain graph dataset used by HKTGNN; (2) the supply chain fully connected graph dataset used by the baseline model used in the comparison; (3) the supply chain graph dataset generated with 25% randomness by the baseline model used in the comparison; (4) the supply chain graph dataset generated with 50% randomness by the baseline model used in the comparison; and (5) the supply chain graph dataset generated with 75% randomness by the baseline model used in the comparison. The scores of the baseline model used in the comparison on datasets (2), (3), (4), and (5) are the average of the ten datasets.

[0165]

[0166] Table 2 Test results for each baseline

[0167] To compare the performance of different models, we selected the F1 score and AUC score, commonly used in binary classification, as evaluation metrics. As shown in Table 2, compared with the traditional baseline model and the relatively new financial risk assessment model KTGNN, the HKTGNN of this invention significantly improves both the F1 score and AUC score on real-world supply chain datasets. This indicates that the model of this invention performs better in real-world scenarios and can effectively address the data scarcity problem in supply chain risk assessment.

[0168] As mentioned earlier, our model performs better on binary classification tasks and supply chain risk assessment models; directly assessing the risk of a product-centric supply chain network is more effective than assessing the risk of companies dispersed throughout the network. This is because the risk of a single-product supply chain network is determined by the risk of all companies within that single-product network. By directly assessing the risk of the entire network, we can obtain a more comprehensive and accurate risk assessment.

[0169] The dataset information shown and analyzed in Table 1 demonstrates the completeness of our dataset. Our dataset encompasses supply chains of different industries and scales, allowing us to test the generalizability of our model. Therefore, the results of our comparative experiments are highly reliable.

[0170] 4.2 Depth Testing

[0171] A) Dataset partitioning test

[0172] This section primarily compares the impact of our data splitting on model performance, where λ = 2.5. As shown in Table 3, it is intuitively clear that when we choose a data split of 7:1:2, the average F1 score is the highest; the smaller the variance of the AUC score, the more stable the model. Overall, the 7:1:2 splitting rule is more reasonable. Our experiments will use a data splitting ratio of 7:1:2 and randomize the data.

[0173] 5:1:3 80.00±3.00 88.74±1.25 5:3:1 80.15±1.64 87.31±2.37 3:1:1 80.45±3.09 87.17±1.80 7:2:1 80.45±3.09 88.17±2.14 7:1:2 80.75±1.56 88.37±1.27

[0174] Table 3 Performance comparison of the HKTGNN model on different dataset subdivision ratios

[0175] B) Hyperparameter testing

[0176] This section primarily compares the impact of hyperparameter λ selection on model performance. The dataset is divided into a 7:1:2 ratio. We use images to illustrate our hyperparameter selection strategy, such as... Figure 4 As shown, when λ = 2.5 is chosen, our model performance reaches its highest level, with a high mean F1 score and a low variance; the mean AUC score is also high, and the variances are similar, indicating good risk assessment capability.

[0177] C) Ablation test

[0178] This section primarily compares the performance of the HKTGNN model of this invention with the baseline model KTGNN. We embedded 99-dimensional feature attributes into KTGNN and input them into the model, then conducted experiments on a supply chain graph. We used the same dataset as HKTGNN to ensure that the algorithm scores could be compared on the same dataset. We also used F1 and AUC scores, and the results are shown in Table 4. Our design of the HKTGNN model makes it more accurate and better able to distinguish between positive and negative samples in supply chain risk assessment.

[0179] KTGNN 75.82±2.56 83.44±2.38 HKTGNN 80.75±1.56 88.37±1.27

[0180] Table 4 Comparison of the performance of HKTGNN and KTGNN

[0181] In summary, the supply chain risk assessment and prediction model based on HKTGNN proposed in this invention constructs a CB-Graph concept through convenient product-to-product and product-to-enterprise connections. This focuses supply chain risk assessment on products within the supply chain rather than scattered enterprises, avoiding the problem of low data quality due to low visibility of enterprise data. Within the supply chain, a hierarchical structure transforms a directed heterogeneous graph into a directed isogeneous graph, reducing the computation time and memory consumption of individual graphs. Feature embedding and feature supplementation improve model performance. Furthermore, based on domain-specific feature supplementation and message passing centrality, the data hunger problem is addressed. Experiments on real-world datasets demonstrate that the model of this invention outperforms other models.

[0182] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes made based on the design principles of the present invention, or any non-creative modifications made thereon, shall fall within the scope of protection of the present invention.

Claims

1. A supply chain risk assessment and prediction model based on HKTGNN, characterized in that, include: The data acquisition module is used to acquire industry chain datasets based on graph neural networks, which include company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links. The preprocessing module is used to construct an industry chain network based on the acquired industry chain dataset through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain. The Graph Embedding Encoder (GEE) module is used to embed the single product network graph of each product and its carried feature information into the homogeneous directed graph of the product network using graph representation learning methods. The feature information of each node consists of graph embedding features, financial statement features, and financial features. The centrality-based cross-domain adaptive feature complementer (CCAFC) module is used to transfer the feature information of the complete nodes in the embedded homogeneous directed graph of the product network to the corresponding biased nodes according to the defined complete nodes and biased nodes, thereby completing the feature information supplementation of the biased nodes. The centrality-based domain adaptive message passing (DCAMP) module is used for cross-domain message passing between complete and biased nodes in a homogeneous directed graph of a product network. It projects the source node's characteristic information onto the target node's domain by calculating a domain difference ratio factor to achieve cross-domain message passing. The Domain Transferable Classifier (DTC) module is used to perform binary classification on the embedded nodes of the single product network graph corresponding to each product based on message passing in the homogeneous directed graph of the product network. This results in the normal or risk label of the corresponding product. Then, the impact of the product on the risk of the industrial chain is determined by analyzing the specific position of the product in the industrial chain, thereby realizing the assessment and prediction of industrial chain risk.

2. The supply chain risk assessment and prediction model based on HKTGNN according to claim 1, characterized in that, A complete node is a node with complete industry chain data, while a biased node is a node with missing industry chain data.

3. The supply chain risk assessment and prediction model based on HKTGNN according to claim 1, characterized in that, Each node in the homogeneous directed graph of the product network is a single product network graph carrying characteristic information, and each edge is a link from one product to another.

4. The supply chain risk assessment and prediction model based on HKTGNN according to claim 3, characterized in that, The graph embedding encoder (GEE) module uses graph representation learning to embed the single-product network graph of each product and its carried feature information into the homogeneous directed graph of the product network. The process is as follows: Define the network graph of the i-th single product. Node set V i supply ,node The graph embedding layer and its embedding information can be obtained using the following formula: In the above formula, Indicates the in-degree of a node. φ represents the out-degree of a node. j This indicates the node type, including four types: product nodes, listed company nodes in a single product supply chain, enterprise nodes invested in by listed companies in a single product supply chain, and source enterprise nodes from which listed companies in a single product supply chain receive supply or investment. (SPB) j Represents the shortest path to a node. A top Laplacian vector is introduced for each node. The largest eigenvalue is the j-th value of the top Laplacian vector, from which the eigenvalues ​​of each node are obtained. In the above formula, yes The top Laplacian vector; Then, the feature vectors of the industry chain network and each node are fed into the graph isomorphic network GIN for training, and finally the embedding vector of each node in the homogeneous directed graph of the product network is obtained, forming a graph CB-Graph with product-to-product links as edges and single product network graphs as nodes.

5. The supply chain risk assessment and prediction model based on HKTGNN according to claim 4, characterized in that, The CB-Graph is represented as follows: Where V represents the set of nodes in the graph, including complete nodes and biased nodes, and node v ij ∈V=V complete ∪V biased E represents the edge set in the graph; This represents the feature information attributes of all nodes in the graph, and the feature value vector of each node. Represents observable feature information attributes. This represents the unobservable feature information attribute that needs to be completed; Y represents the risk label, which marks a risky company as risky when it exists; ψ is the CB-Graph label, used to label the node attributes in the graph, when node v ij When the node is complete, ψ = 1; when node v ij When there are biased nodes, ψ = 0.

6. The supply chain risk assessment and prediction model based on HKTGNN according to claim 5, characterized in that, The centrality-based cross-domain adaptive feature complementer (CCAFC) module transfers the feature information of complete nodes in the embedded homogeneous directed graph of the product network to the corresponding biased nodes. The process of supplementing the feature information of biased nodes is as follows: Define a complete node set V complete and the biased node set V biased V complete ∪V biased =V supply , Will Feature information knowledge transfer Complete the supplementation of feature information in the middle, so that Become form And use the hyperparameter K to limit the number of iterations for each transformation; For each biased node Its unobservable vector After the supplement is completed, it will be: Where f(·) is the neighbor importance factor, This represents the visible input of the i-th node. ω represents the visible input of the j-th node. b and ω c This represents the final self-learning matrix. It is an unobservable vector The calibration variable eliminates domain differences: In the above formula, and Here is the self-learning parameter matrix; σ is the softsign activation function; It is a centrality-based transform-domain difference, where It is the feature information attribute of the i-th complete node. It is the feature information attribute of the j-th biased node. For the centrality of the i-th complete node, The centrality of the j-th biased node; Then, eigenvector centrality is used as the calculation method. After a biased node is supplemented, the next iteration uses the supplemented biased node as the information provider to achieve the goal of... In this process, using breadth-first search (BFS) results in: In the above formula, a cb This represents the influence weights based on centrality, and σ1 represents the PReLU activation function. The algorithm converges through iteration when the number of iterations reaches the maximum value K or when all nodes are completed. During the iteration process, a distribution consistency loss function is introduced. Modify the centrality.

7. The supply chain risk assessment and prediction model based on HKTGNN according to claim 6, characterized in that, The centrality-based domain adaptive message passing (DCAMP) module achieves cross-domain message passing by calculating the domain difference ratio factor and projecting the source node feature information of the message to the domain of the target node. For a given edge e i,j =(v i ,v j The message passing equation is: In the above formula, h j This represents the message passing tensor of node j. This represents the feature information calibrated using the domain difference factor; It is the neighbor importance function: In the above formula, σ² is the leaky ReLU activation function. This represents the self-learning matrix that needs to be obtained. This represents the weight term; therefore, the distribution shift variable Δ i,j : In the above formula, σ3 is the softsign activation function, ψ(v i ) represents the CB-Graph label, h i The tensor representing node i used for message passing; ΔX C It is a centrality-based transform-domain difference: In the above formula, It is the feature information attribute of the i-th complete node. It is the feature information attribute of the j-th biased node. For the centrality of the i-th complete node, The centrality of the j-th biased node; This leverages centrality-based domain differences to achieve cross-domain message passing.

8. A supply chain risk assessment and prediction method based on HKTGNN, characterized in that, Includes the following steps: S10. Obtain the industry chain dataset based on the graph neural network, which includes company nodes, product nodes, investment / supply nodes, investment / supply to company links, product to product links, and company to product links. S20. Based on the obtained industry chain dataset, construct an industry chain network through hierarchical processing. The industry chain network contains a unique homogeneous directed graph of product networks and a single product network graph based on each product in the industry chain. S30. Using graph representation learning methods, embed the single product network graph of each product and its carried feature information into the homogeneous directed graph of the product network. The feature information of each node formed consists of graph embedding features, financial statement features, and financial features. S40. Based on the defined complete nodes and biased nodes, transfer the feature information of the complete nodes in the embedded product network homogeneous directed graph to the corresponding biased nodes to complete the feature information supplementation of the biased nodes. S50. Based on the message passing in the homogeneous directed graph of the product network, complete the binary classification of the embedded nodes of the single product network graph corresponding to each product, obtain the normal label or risk label of the corresponding product, and then determine its impact on the risk of the industrial chain by analyzing the specific position of the product in the industrial chain, and obtain the score result of the risk assessment of the industrial chain, thereby realizing the assessment and prediction of the risk of the industrial chain. In the cross-domain message passing between complete nodes and biased nodes in a homogeneous directed graph of a product network, the source node feature information of the message is projected onto the domain of the target node by calculating the domain difference ratio factor to achieve cross-domain message passing.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the supply chain risk assessment and prediction method based on HKTGNN as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it outputs the results of supply chain risk assessment and prediction using the supply chain risk assessment and prediction model based on HKTGNN as described in any one of claims 1-7.

Images