Industry chain risk assessment model and method based on graph neural network, and medium

Abstract

This invention discloses a supply chain risk assessment model and method based on graph neural networks, including a heterogeneous graph construction module, a company graph generation module, a node risk assessment module, and a supply chain risk assessment module. The heterogeneous graph construction module constructs a heterogeneous graph of the supply chain, transforming supply chain risk assessment into a risk assessment problem of the graph and its nodes, thus characterizing the risk transmission relationships. The company graph generation module efficiently learns various relationships between company nodes based on upstream and downstream supply, competition, and company investment and financing relationships, constructing a supply chain company node association graph based on multiple correlations. The node risk assessment module achieves node-level risk assessment based on neighborhood sampling and attention mechanisms. The supply chain risk assessment module performs efficient risk assessment of the supply chain based on the node risk status and node location characteristics. This invention achieves efficient transformation and accurate assessment of supply chain risk problems.

Description

Technical Field

[0001] This invention relates to the field of supply chain graph data mining technology, specifically to a supply chain risk assessment model, method, and medium based on graph neural networks. Background Technology

[0002] An industrial chain, as a chain-like relationship formed between enterprises based on certain technological and economic connections and according to specific logical and spatiotemporal arrangements, reflects the social division of labor in product production. As a tightly linked chain of enterprises, the industrial chain is characterized by mutual influence and risk transmission. When a risk occurs in any link of the industrial chain, it will directly affect the resilience and stability of the entire chain. Therefore, timely identification and response to risks are key factors in ensuring economic sustainability and robustness. Industrial chain risk assessment, as a crucial strategic tool and an important component of economic and social development, helps to improve economic resilience, sustainability, and stability, and safeguard social welfare by identifying the sources of risk and the risk status of each node through timely and accurate risk assessment.

[0003] Existing risk assessment methods, such as expert evaluation, decision tree methods, and deep neural network methods, are primarily geared towards supply chain risk assessment. They can be categorized into qualitative and quantitative methods based on the assessment approach. Qualitative methods are mainly used for risk identification, analyzing and judging risk conditions based on the knowledge and experience of the assessors. Quantitative assessment methods quantify corresponding risk factors, combining the probability of risk occurrence and its consequences to obtain a risk value. With the development of artificial intelligence theory, risk assessment models built using machine learning methods have improved the accuracy of supply chain risk assessment. However, the industrial chain is more complex than the supply chain, containing multiple nodes with diverse relationships between them. Therefore, simply using supply chain risk assessment methods to quantify industrial chain risks can lead to problems such as missing relationships and poor feature extraction. Furthermore, risks are contagious; when a node in the industrial chain experiences risk, its related nodes are often affected. Existing risk assessment methods neglect the characterization of the internal structure of the upstream and downstream of the industrial chain, resulting in a lack of assessment of the internal risk transmission and its impact. Furthermore, existing assessment methods often rely on large amounts of data and information to evaluate risks. However, obtaining complete and accurate supply chain data is extremely difficult in reality. The resulting incomplete data and missing connections between internal nodes can lead to biases and inaccuracies in the assessment results. These limitations make it difficult to apply existing methods effectively to supply chain risk assessment and early warning.

[0004] Furthermore, due to the current lack of supply and competition relationships within the industry chain, direct connections cannot be established between company nodes. Therefore, risk assessment of the industry chain requires evaluating node risks based on the relationships between company nodes and their individual characteristics. Thus, the core issues in solving the risk assessment of nodes and the entire industry chain are: how to connect company nodes based on the established heterogeneous graph, what methods to use to extract risk characteristics from industry chain nodes, and how to assess the risk status based on the extracted risk characteristics. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a supply chain risk assessment model based on graph neural networks. It generates a relationship graph of company nodes in the supply chain through connection structures of "company-product-product-company" and "company-product-company", and introduces company investment and financing relationships to obtain a group of company association graphs based on different connection relationships. It combines node neighborhood sampling and graph attention mechanism to assess the risk of each node. Based on the risk status of the nodes and the structural characteristics of the nodes, it realizes the risk level assessment of the heterogeneous graph of the supply chain.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A supply chain risk assessment model based on graph neural networks includes a heterogeneous graph construction module, a company graph generation module, a node risk assessment module, and a supply chain risk assessment module, wherein:

[0008] The heterogeneous graph construction module is used to construct a heterogeneous graph of the industrial chain by collecting industrial chain data, utilizing the upstream and downstream relationships of products in the industrial chain, the production relationship of companies to products, and the financial attributes of companies. It includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships, and company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes.

[0009] The company graph generation module is used to transform the heterogeneous graph of the industrial chain into a company node supply isomorphic graph reflecting the supply relationship between company nodes and a company node product competition isomorphic graph reflecting the competitive relationship between companies producing the same product. It also combines the investment and financing relationship between company nodes within the industrial chain to generate a company node investment and financing isomorphic graph, thus forming an industrial chain company node association graph.

[0010] The node risk assessment module is used to sample and adaptively aggregate the node neighborhoods of different graphs in the industry chain company node association graph through a neighborhood sampling mechanism, and to perform binary risk level classification on the node embedding vectors of neighborhood sampling and aggregation through a relation-level attention mechanism, so as to assess the risk status of each company node and generate node-level risk level embedding vectors.

[0011] The supply chain risk assessment module is used to fuse the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company's nodes, and then use a supply chain risk assessment algorithm with an attention mechanism to quantify the supply chain risk into a three-dimensional risk level, thereby realizing multi-level risk assessment of the supply chain.

[0012] Specifically, the heterogeneous graph of the industrial chain is represented as G(V i E i ,K,δ),

[0013] Among them, V i Let E be a set of nodes. i Let K be the set of edges, K be the set of risk labels for company nodes, and δ be the risk level label for the industry chain; let V be the set of nodes. i It includes two node types: product nodes and company nodes. p V represents the set of product nodes. c Represents the set of company nodes; edge set E i It contains three edge types Γ e ={E cp E pp E pc}, E cp E represents the set of edges connecting company nodes and product nodes. pp E represents the set of edges connecting product nodes. pc V represents the set of edges connecting product nodes and company nodes; V is the set of company nodes. c Each company node v(i) contains a financial feature vector m. v(i) and structural features p v(i) The risk label set K = {0,1} contains two labels representing Safe and Risk, respectively. Each company node v(i) is associated with a node risk label y. v(i) ∈K is associated; the risk level label of the industrial chain δ={0,1,2} contains three levels of risk labels, representing low risk, medium risk and high risk respectively.

[0014] Specifically, the company nodes are selected from listed companies in the industry chain, and financial attributes are extracted from the financial statements of the listed companies as the embedded financial feature vectors of the company nodes.

[0015] Specifically, the company graph generation module transforms the heterogeneous industry chain graph into a company node supply isomorphic graph reflecting the supply relationships between company nodes and a company node product competition isomorphic graph reflecting the competitive relationships between companies producing the same product. The process of generating a company node investment and financing isomorphic graph by combining the investment and financing relationships between company nodes within the industry chain is as follows:

[0016] Define the meta-path of "company-product-company" and transform the heterogeneous graph of the industry chain to generate a company node supply isomorphic graph G1 that only contains company nodes and the connection relationships between company nodes;

[0017] Define the meta-path of "company-product-product-company" and transform the heterogeneous graph of the industry chain to generate G2, which contains only company nodes and the connection relationships between company nodes and product competition isomorphic graphs.

[0018] Based on the relationships between three types of nodes in the industry chain data—listed companies and investors within the chain, and other companies outside the chain—investment and financing relationships between company nodes within the industry chain are configured. These include investment and financing relationships where different company nodes have the same investor or other companies outside the chain, investment and financing relationships where different company nodes invest in the same company, and investment and financing relationships where one company node invests in another company node.

[0019] Based on the configured investment and financing relationships between company nodes within the industry chain, the relationships between company nodes within the industry chain are obtained, generating a company node investment and financing isomorphic graph G3 that only contains company nodes and their connection relationships.

[0020] Specifically, the industry chain company node association diagram is represented as G. j , j = {1, 2, 3}.

[0021] Specifically, the node risk assessment module samples and adaptively aggregates the node neighborhoods of different graphs in the industry chain company node association graph through a neighborhood sampling mechanism, and performs binary risk level classification on the node embedding vectors sampled and aggregated through a relation-level attention mechanism.

[0022] Based on the financial feature vector m of company node v(i) v(i) Configure the sampling neighborhood and sample the feature information of the neighboring nodes of the company node v(i);

[0023] The mean aggregation method is used to obtain the feature vector of the company node v(i) that has aggregated the information of neighbor nodes at the kth layer. The aggregated features of the neighbor nodes at the k-1th layer are obtained by aggregating the embedded neighbor nodes.

[0024] The aggregation features of the neighboring nodes at layer k are concatenated with the feature information of company node v(i) at layer k-1 to obtain the feature information of company node v(i) at layer k. Then, through normalization and layer-by-layer aggregation, the node embedding vector of company node v(i) in the corresponding graph of the industry chain company node association group is obtained.

[0025] Attention coefficients are configured for three isomorphic graphs of the node association graph of the industry chain companies, and node embedding vectors obtained under different graphs are processed. As input, it is trained and learned through a relation-level attention mechanism to obtain the attention coefficients under the corresponding graph;

[0026] Based on the obtained attention coefficients and node embedding vectors under each isomorphic graph Calculate and obtain the comprehensive feature vector Z of company node v(i). v(i) ;

[0027] The comprehensive feature vector Z of company node v(i) v(i) The input classifier performs binary risk level classification, and the risk status assessment of the company node is completed based on the classification results. The classifier is represented as...

[0028] risk(v(i))=σ(wZ v(i) +b)

[0029] In the above formula, risk represents the output result of risk binary classification, σ represents the activation function, w is the weight coefficient, and b is the bias coefficient;

[0030] The loss function L used in the risk assessment of company nodes is... NRE Represented as:

[0031]

[0032] In the above formula, y v(i) This represents the label value of the company node, y' v(i) This represents the predicted value for the company's nodes.

[0033] Specifically, the results of the binary risk level classification are encoded to generate a risk level embedding vector rEmb(v(i)) for the company node.

[0034] Specifically, the process by which the supply chain risk assessment module fuses the risk level embedding vector output by the node risk assessment module with the structural features of the company nodes, and then quantifies the supply chain risk into a ternary risk level using a supply chain risk assessment algorithm incorporating an attention mechanism, is as follows:

[0035] Structural features of company node v(i) are extracted based on its position in the industry chain node association graph. The node degree (deg) of company node v(i) is then used to determine its structural characteristics. i This is transformed into a learnable structural embedding vector p of the embedding layer. v(i) =Position_layer(deg i And encode the structural embedding vector sEmb(v(i)) = Emb_Ecd(W) of the company node v(i). p p v(i) ), where W pIndicates the weight coefficients of the node structure;

[0036] The structural embedding vector and the risk level embedding vector are concatenated and their dimensions are reduced to obtain the feature representation vector S(v(i)) = MLP([sEmb(v(i))||rEmb(v(i))]);

[0037] The feature representation vector is embedded into the graph to obtain the industry chain company node relationship graph G. j Feature representation vector Where N is the total number of company nodes in a homogeneous graph;

[0038] After aggregating the feature representation vectors of each isomorphic graph by mean-averaging, the result is input into a fully connected neural network with a softmax activation function to quantify the ternary risk level and output the multi-level risk assessment result of the industrial chain.

[0039] In the above formula, W g and b g These represent the weight coefficients and bias coefficients of the fully connected neural network, respectively; the loss function L used in the fully connected neural network is... ICRE Represented as:

[0040] L ICRE =Loss(f(G) j ;y' v(i) );δ)

[0041] In the above formula, y' v(i) δ represents the predicted value of the company node, f represents the risk level label of the industrial chain, and f represents the prediction function for industrial chain risk assessment.

[0042] Furthermore, the present invention also provides a supply chain risk assessment method based on graph neural networks, comprising the following steps:

[0043] S10. Collect supply chain data and construct a supply chain heterogeneous graph through the upstream and downstream relationships of products in the supply chain, the production relationship of companies to products, and the financial attributes of companies. This graph includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships. Company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes.

[0044] S20. Transform the heterogeneous graph of the industrial chain into a supply isomorphic graph of company nodes that reflects the supply relationship between company nodes and a product competition isomorphic graph of company nodes that reflects the competitive relationship between companies producing the same product. Combine the investment and financing relationship between company nodes in the industrial chain to generate an investment and financing isomorphic graph of company nodes, thus forming an industrial chain company node association graph.

[0045] S30. The neighborhood sampling mechanism is used to sample and adaptively aggregate the nodes of different graphs in the industry chain company node association graph. The node embedding vectors of the neighborhood sampling aggregation are classified into binary risk levels through the relation-level attention mechanism to realize the risk status assessment of each company node and generate node-level risk level embedding vectors.

[0046] S40. After fusing the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company's nodes, the risk level of the industrial chain is quantified by a three-element risk assessment algorithm that combines the attention mechanism, thereby realizing multi-level risk assessment of the industrial chain.

[0047] 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 method based on graph neural networks.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention achieves efficient transformation and accurate assessment of supply chain risk issues through a heterogeneous graph construction module, a company graph generation module, a node risk assessment module based on neighborhood sampling and attention mechanisms, and a supply chain risk assessment module for overall supply chain risk assessment. Specifically, it cleverly utilizes the "product-company" relationship and company financial information to construct a heterogeneous supply chain graph, transforming supply chain risk assessment into a risk assessment problem of the graph and its nodes, thus characterizing the risk transmission relationship. Simultaneously, a company graph generation module based on upstream and downstream supply, competition, and company investment and financing relationships is designed. It utilizes the connection relationships between heterogeneous nodes in the graph to efficiently learn various relationships between company nodes, constructing a supply chain company node association graph based on multiple relationships. Furthermore, a node risk assessment module based on neighborhood sampling and attention mechanisms is proposed. This module samples and aggregates node neighborhood features and adaptively aggregates node features under different graph relationships, achieving node-level risk assessment based on node risk characteristics. Finally, a supply chain risk assessment module for overall supply chain risk assessment is designed to efficiently assess the supply chain risk based on node risk status and node location characteristics. Experiments on real-world datasets demonstrate that the GANS model of this invention performs significantly well in the task of supply chain risk assessment. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the model structure of an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of constructing a heterogeneous industry chain graph in an embodiment of the present invention.

[0052] Figure 3 This is a flowchart illustrating the process of the company diagram generation module based on meta-paths in an embodiment of the present invention.

[0053] Figure 4 This is a schematic diagram of the company diagram generation module based on investment and financing relationships in an embodiment of the present invention.

[0054] Figure 5 This is a schematic diagram of the method flow of an embodiment of the present invention.

[0055] Figure 6 This is a comparison chart of the accuracy of the test experiments in an embodiment of the present invention.

[0056] Figure 7 This is a line graph comparing the AUC scores of the test experiments in this embodiment of the invention. Detailed Implementation

[0057] 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.

[0058] Example

[0059] like Figures 1 to 4 As shown, this supply chain risk assessment model based on graph neural networks includes a heterogeneous graph construction module, a company graph generation module, a node risk assessment module, and a supply chain risk assessment module, wherein:

[0060] The heterogeneous graph construction module is used to construct a heterogeneous graph of the industrial chain by collecting industrial chain data, utilizing the upstream and downstream relationships of products in the industrial chain, the production relationship of companies to products, and the financial attributes of companies. It includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships, and company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes.

[0061] The company graph generation module is used to transform the heterogeneous graph of the industrial chain into a company node supply isomorphic graph reflecting the supply relationship between company nodes and a company node product competition isomorphic graph reflecting the competitive relationship between companies producing the same product. It also combines the investment and financing relationship between company nodes within the industrial chain to generate a company node investment and financing isomorphic graph, thus forming an industrial chain company node association graph.

[0062] The node risk assessment module is used to sample and adaptively aggregate the node neighborhoods of different graphs in the industry chain company node association graph through a neighborhood sampling mechanism, and to perform binary risk level classification on the node embedding vectors of neighborhood sampling and aggregation through a relation-level attention mechanism, so as to assess the risk status of each company node and generate node-level risk level embedding vectors.

[0063] The supply chain risk assessment module is used to fuse the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company's nodes, and then use a supply chain risk assessment algorithm with an attention mechanism to quantify the supply chain risk into a three-dimensional risk level, thereby realizing multi-level risk assessment of the supply chain.

[0064] Application instructions for related knowledge:

[0065] Graph neural networks (GNNs) can learn from graph-structured data, extract and discover data features, and apply them to tasks such as classification and prediction. Currently, GNNs can be broadly divided into two categories: spectral methods and non-spectral methods. Spectral methods focus on the spectral representation of the graph and are widely used in node classification tasks. A typical example is GCN, which learns the representation of a node in its one-step neighborhood by restricting filters. Non-spectral methods, on the other hand, directly define convolutions on the graph. For example, GraphSAGE operates on connected neighbors in the graph, aggregating neighborhood features with a specified stride through a neural network aggregator. Most of these methods are only applicable to homogeneous networks and cannot handle heterogeneous networks with various node and edge types. In methods for processing heterogeneous graphs, many studies use meta-paths to model the graph network, setting meta-paths based on specific tasks and learning the features of neighboring nodes based on these meta-paths. However, the choice of meta-paths directly affects the effectiveness of risk assessment. Most of the above research methods only focus on the data attributes of nodes, lacking consideration for node location attributes. In supply chain risk assessment, however, location attributes have a significant impact on the risk assessment results.

[0066] Attention mechanisms are a simple method for encoding sequential data based on the importance score of each element. By introducing attention mechanisms, more critical information for the current task can be focused from a large amount of information, solving the problem of information overload and improving the efficiency and accuracy of task processing. In graph attention mechanism research, GAT introduces attention mechanisms into graph convolutional networks, using hidden self-attention layers on graph structure data to implicitly assign different weights to different neighboring nodes, learning node features. In attention networks for heterogeneous graphs, HAN et al. introduced a hierarchical attention mechanism based on heterogeneous graphs, dividing it into node-level attention and semantic-level attention. Attention weights are assigned based on different neighboring nodes and different meta-paths, respectively, and node features are aggregated based on these weights. The above studies only consider introducing attention mechanisms into a single graph and mainly focus on the application of node neighborhood attention. However, when facing complex relationships in the industry chain, requiring the processing of multiple relationship graphs, and lacking corresponding strategies for feature aggregation and risk assessment based on node features under multiple graph relationships, these approaches are insufficient.

[0067] A supply chain is a complex network composed of different links such as raw material suppliers, manufacturers, and distributors, characterized by interdependence and interconnectivity. These characteristics are related to heterogeneous graphs. As a graph data structure, heterogeneous graphs contain multiple types of nodes and edges, allowing each node to have different attributes and characteristics. They enable the modeling of complex relationships and the analysis of complex networks. In a supply chain, different enterprises and products can be viewed as different types of nodes, and each node can represent a different enterprise or organization with its own characteristics. Various types of relationships also exist between nodes, such as supply relationships and production relationships. Therefore, by constructing a supply chain heterogeneous graph, node attributes and complex relationships between nodes can be better integrated into supply chain risk assessment, enabling a more comprehensive and accurate analysis and modeling of the complex relationships and diversity within the supply chain, thus achieving more efficient supply chain risk assessment.

[0068] To address the problems of existing technologies and more effectively assess supply chain risks, this invention designs a heterogeneous graph construction module. By collecting supply chain data, it constructs a heterogeneous supply chain graph for specific application industries (such as the integrated circuit industry mentioned in subsequent experiments). This graph includes two types of nodes: product nodes and company nodes. It establishes connections between nodes through upstream and downstream product relationships, company production relationships, and company financial attributes, thereby characterizing the risk transmission between nodes. Thus, the risk assessment problem of the supply chain is transformed into a risk assessment problem of the graph and its nodes.

[0069] Specifically, in constructing the heterogeneous graph, it consists of two types of nodes: product nodes and company nodes. Products are associated when there is a direct material relationship. Company nodes are associated with product nodes through the production relationship between the company and the product. The company nodes represent the main manufacturers of the products. Due to the lack of non-listed company attributes, we selected listed companies in the industry chain as the company nodes for the products, and extracted attributes from the financial statements of these listed companies as embeddings for the company nodes. This constructs the industry chain heterogeneous graph, as shown below. Figure 2 As shown.

[0070] For the massive heterogeneous graph of the industrial chain, the risk issue is the risk assessment of each node in the heterogeneous graph and the industrial chain as a whole.

[0071] By defining parameters, the heterogeneous graph of the industry chain is represented as G(V). i E i ,K,δ), where V i Let E be a set of nodes. i Let K be the set of edges, K be the set of risk labels for company nodes, and δ be the risk level label for the industry chain; let V be the set of nodes. iIt includes two node types: product nodes and company nodes. p V represents the set of product nodes. c Represents the set of company nodes; edge set E i It contains three edge types Γ e ={E cp E pp E pc}, E cp E represents the set of edges connecting company nodes and product nodes. pp E represents the set of edges connecting product nodes. pc V represents the set of edges connecting product nodes and company nodes; V is the set of company nodes. c Each company node v(i) contains a financial feature vector m. v(i) and structural features p v(i) The risk label set K = {0,1} contains two labels representing Safe and Risk, respectively. Each company node v(i) is associated with a node risk label y. v(i) ∈K is associated; the risk level label of the industrial chain δ={0,1,2} contains three levels of risk labels, representing low risk, medium risk and high risk respectively.

[0072] Supply chain risk assessment comprehensively evaluates the risk level and importance of each module within the supply chain. Our supply chain heterogeneity diagram effectively reflects the risk factors at each node and demonstrates the transmission of risk within the supply chain. This helps to fill in missing node relationships and reflect the interconnections between company nodes, thereby improving the accuracy of node risk assessment. Subsequently, based on the transformed relationship diagram, the structural characteristics of the nodes are extracted, and through a comprehensive consideration of both, the overall risk level of the supply chain is finally obtained, enhancing the credibility of the risk assessment.

[0073] Accordingly, the optimization objective for supply chain risk assessment can be set as follows: conduct a binary risk assessment of the company nodes within the supply chain, and then quantify the overall risk of the supply chain into a ternary risk level based on the results of the node risk assessment. Therefore, the optimization objective of the GANS model of this invention can be expressed in the following two ways:

[0074] (1) Node risk assessment

[0075]

[0076] In the above formula, L NRE The loss function y represents the node risk assessment. v(i) The label value of the node, y' v(i) Represents the predicted value of a node;

[0077] (2) Supply Chain Risk Assessment

[0078] L ICRE =Loss(f(G) j ;y' v(i) );δ)

[0079] In the above formula, L ICRE The loss function y' represents the overall risk assessment of the industrial chain. v(i) δ represents the predicted value of the node, f represents the risk level label of the industrial chain, and f represents the prediction function for industrial chain risk assessment.

[0080] Regarding the company graph generation module, it is designed to address the problem of difficulty in accurately extracting node features due to the lack of direct relationships between company nodes in heterogeneous industry chain graphs. It can realize a company node graph generation method based on associated meta-paths and company investment and financing relationships.

[0081] First, based on the characteristics of the constructed heterogeneous industry chain graph G and the requirements of industry chain risk assessment, meta-paths of "company-product-company" and "company-product-product-company" are selected. According to the selected meta-paths, the heterogeneous industry chain graph is transformed to generate a company node supply isomorphic graph G1 and a company node product competition isomorphic graph G2, which only contain company nodes and their connection relationships. Figure 3 As shown.

[0082] The meta-path p can be obtained by connecting the heterogeneous edges in the heterogeneous graph G of the industrial chain, i.e. Where t l ∈Γ e This represents the type of the l-th edge in the meta-path. Based on the association between nodes via edges, it can be found between node v1 and node v2. l+1 Define a composite relationship between them. Given a composite relation R or a sequence of edge types {t1,t2,...,t} l}, then the adjacency matrix A of the metapath p. p It can be represented by the multiplication of adjacency matrices.

[0083] In the process of risk assessment, product nodes, as related nodes, do not possess multiple risk factors and risk attributes. Furthermore, in reality, direct supply and competition relationships between companies are often lacking. Therefore, we use meta-paths of "company-product-product-company" and "company-product-company" to transform the heterogeneous graph into two isomorphic graphs containing only company nodes. These graphs respectively depict the supply relationships between company nodes in the industry chain and the competition relationships between companies producing the same product.

[0084] Finally, we obtain the adjacency matrix A under the two relationships. CPC and A CPPC :

[0085]

[0086] Furthermore, the relationships between companies within a supply chain are often complex and diverse, with companies establishing connections through other factors. Therefore, the company diagram generation module of this invention adds investment and financing relationships to the existing supply and competition relationships.

[0087] Based on the investment and equity data of listed companies obtained from the industry chain data collection, this analysis identifies three node types: listed companies within the industry chain, investors, and other companies outside the chain. Based on the relationships between these three types of nodes, investment and financing relationships between company nodes within the industry chain are configured. These include investment and financing relationships where different company nodes have the same investor or have investments from other companies outside the chain, investment and financing relationships where different company nodes invest in the same company, and investment and financing relationships where one company node invests in another company node. Then, based on the configured investment and financing relationships between company nodes within the industry chain, the relationships between company nodes within the industry chain are obtained, generating a company node investment and financing isomorphism graph G3 containing only company nodes and their connections, such as... Figure 4 As shown.

[0088] This yields a node association diagram of the industry chain companies, denoted as G. j , j={1,2,3}, represent the company relationship diagrams obtained under supply, competition and investment relationships, respectively.

[0089] Regarding the node risk assessment module, it is used to assess the node association group diagram G of the industry chain companies obtained above. j To assess the risks of company nodes within the industry chain, this invention proposes a node risk assessment method that combines node neighborhood sampling and attention mechanisms.

[0090] First, the feature information of neighboring nodes is sampled. This is based on the financial feature vector m of company node v(i). v(i) Normalization yields the feature parameters x of node v(i). v(i) The sampling depth is defined by parameter K, and the sampling neighborhood of node v(i) is defined as N(i), which is expressed as N(i) = NeighborSample(v(i)).

[0091] Subsequently, the node features are aggregated, and the feature vector of the node's neighbor node information aggregated at the k-th layer is obtained by sampling mean aggregation. Represented as in Let v(j) be the feature vector obtained by neighborhood sampling aggregation at layer k-1. The aggregated features of the neighbors at layer k are obtained by aggregating the embeddings of the neighboring nodes at layer k-1. Then Concatenated with the feature vector embedding of node v(i) at layer k(1), it is represented as:

[0092]

[0093] In the above formula, σ represents the nonlinear transformation function, W k This represents the weight of the k-th layer.

[0094] Concatenating the features of node v(i) at the kth layer yields the feature vector. The feature embedding of node v(i) in the relation graph is finally obtained through normalization and hierarchical aggregation.

[0095] The aforementioned neighborhood sampling effectively addresses the low feature learning efficiency caused by numerous connecting edges in nodes within a single relationship graph. However, when faced with multiple company association graphs extracted using various semantic information, semantically specific node embeddings can only reflect a node from one perspective. To learn more comprehensive node embeddings, we need to integrate multiple semantic information. To address this issue, we designed a relationship-level attention mechanism based on three established company node relationship graphs. By assigning attention coefficients to the relationship graphs generated for each relationship, we learn the importance of different relationship graphs. We use the node embeddings obtained from neighborhood sampling aggregation under different relationship graphs as input, and combine them with the attention mechanism to obtain the attention coefficients for different relationship graphs.

[0096]

[0097] Among them, atten graph To generate corresponding attention coefficients for three different relationship graphs

[0098] To obtain the attention coefficients corresponding to the three different relationship graphs, we first embed the node embedding vectors under different relationships. A nonlinear transformation is performed. Then, the nodes in the relation graph G are obtained by multiplying the activation function with the learnable transformation parameter r. j The next scalar is then used, followed by a weighted sum and average of all the scalars at the nodes in the graph, to obtain the relation graph G. j weight

[0099]

[0100] Where W and b represent the corresponding weights and biases.

[0101] After determining the importance of each relation, they are normalized using the softmax function to obtain different relation graphs G. j Attention coefficient

[0102]

[0103] The learned attention coefficients and feature embedding vectors of nodes under various relationships are used as inputs. Through weighted summation, the comprehensive feature vector Z of company node v(i) is calculated. v(i) ;

[0104]

[0105] The comprehensive feature vector of the nodes is used as the input to the classifier. After weighted summation, a bias unit is added, and the final output value is obtained after passing through the activation function.

[0106] risk(v(i))=σ(wZ v(i) +b)

[0107] In the above formula, risk represents the output result of risk binary classification, σ represents the activation function, w is the weight coefficient, and b is the bias coefficient;

[0108] The company nodes in the industry chain are classified into binary risk levels based on the output value of the final layer, and the classification results are encoded as the risk level embedding vector rEmb(v(i)) of the node.

[0109] The supply chain risk assessment module is used to conduct an overall risk assessment of the supply chain based on the assessment results of the node risk assessment module and the structural characteristics of the company's nodes. This invention proposes a supply chain-wide risk assessment method that incorporates an attention mechanism.

[0110] First, based on the position of the company node in the industry chain node association graph, the structural feature attributes of company node v(i) are extracted, and these attributes are transformed into learnable embedding vectors. The node degree of each node v(i) is expressed as deg. i This means that the degree of a node is used to transform it into a learnable structural embedding vector p for the embedding layer. v(i) =Position_layer(deg i And encode the structural embedding vector sEmb(v(i)) = Emb_Ecd(W) of the company node v(i). p p v(i) ), where W p This represents the weight coefficient of the node structure.

[0111] Subsequently, the obtained structural embedding vector and risk level embedding vector are concatenated and their dimensionality reduced to obtain the feature representation vector S(v(i)) = MLP([sEmb(v(i))||rEmb(v(i))]) of company node v(i).

[0112] Based on the feature representation vector of company node v(i), a weighted sum is performed to obtain the industry chain company node association graph G. j Feature representation vector Represented as:

[0113] Where N is the total number of company nodes in a homogeneous graph.

[0114] After aggregating the feature representation vectors of each isomorphic graph by mean-averaging, the result is input into a fully connected neural network with a softmax activation function to quantify the ternary risk level and output the multi-level risk assessment result of the industrial chain.

[0115] In the above formula, W g and b g These represent the weight coefficients and bias coefficients of a fully connected neural network, respectively.

[0116] This enables risk assessment of the entire industry chain.

[0117] In another embodiment, such as Figure 5 As shown, a supply chain risk assessment method based on graph neural networks is also provided, including the following steps:

[0118] S10. Collect supply chain data and construct a supply chain heterogeneous graph through the upstream and downstream relationships of products in the supply chain, the production relationship of companies to products, and the financial attributes of companies. This graph includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships. Company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes.

[0119] S20. Transform the heterogeneous graph of the industrial chain into a supply isomorphic graph of company nodes that reflects the supply relationship between company nodes and a product competition isomorphic graph of company nodes that reflects the competitive relationship between companies producing the same product. Combine the investment and financing relationship between company nodes in the industrial chain to generate an investment and financing isomorphic graph of company nodes, thus forming an industrial chain company node association graph.

[0120] S30. The neighborhood sampling mechanism is used to sample and adaptively aggregate the nodes of different graphs in the industry chain company node association graph. The node embedding vectors of the neighborhood sampling aggregation are classified into binary risk levels through the relation-level attention mechanism to realize the risk status assessment of each company node and generate node-level risk level embedding vectors.

[0121] S40. After fusing the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company's nodes, the risk level of the industrial chain is quantified by a three-element risk assessment algorithm that combines the attention mechanism, thereby realizing multi-level risk assessment of the industrial chain.

[0122] 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 method based on graph neural networks.

[0123] 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.

[0124] 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.

[0125] 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.

[0126] 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).

[0127] The following experiment will introduce the supply chain risk assessment model and method based on graph neural networks.

[0128] I. Dataset

[0129] The supply chain data used in this experiment comes from real data of 1,732 listed companies in China from 2018 to 2023, including financial statements, supply chain product association diagrams, tables of major product manufacturers, and a listed company rating table. This includes 1,732 company nodes, 430 product nodes, 18 risk factors, and 36,372 company risk scores. The overall risk event span of the supply chain covers 21 quarters. The financial statements, supply chain product association diagrams, and tables of major product manufacturers were provided by Wind (a leading Chinese financial database). Based on the product association table and the product company association table, a relationship adjacency matrix table was generated after preprocessing. The listed company risk rating table was provided by Tongdaxin (a leading securities information platform). This table uses financial indicators as the primary scoring basis to assess the risk status of companies. Based on the risk scores and expert opinions, the risk status of listed companies was divided into two categories: companies with scores in the range [80, 100] were considered safe and coded as 0; companies with scores in the range [0, 79] were considered risky and coded as 1. Meanwhile, based on the company's node risk labels, the overall risk of the integrated circuit industry chain is divided into three ranges: [0,0.6], [0.6,0.8], and [0.8,1], which correspond to low risk, medium risk, and high risk of the industry chain, respectively.

[0130] II. Experimental Parameter Settings

[0131] This experiment is based on a real-world industry chain dataset, and different experimental configurations were designed. We designed experiments with different training ratios (the percentage of nodes in the training set), ranging from 20% to 60%. For each training ratio, we partitioned the dataset into three different random partitions and applied ten random seeds for random initialization of model parameters, with each model undergoing 30 trials. For the financial feature attributes of the dataset we used, we imputed missing values ​​using average interpolation. To ensure fairness, we performed hyperparameter search on all models, with each model having the same search space. We selected the number of training epochs from {200, 400, 600}, used the Adam optimizer for all experiments, and selected the learning rate in {1e-2, 1e-3, 1e-4} and the weight decay coefficient in {1e-4, 1e-3, 5e-4}. All training and validation processes for all models were performed on an NVIDIA GeForce RTX 3090 (24GB) GPU.

[0132] III. Baseline Method

[0133] The GANS model of this invention is compared with three baseline graph neural network models (GCN, GAT, GraphSAGE) and a GNN model (HAN) suitable for heterogeneous graphs.

[0134] GCN: A semi-supervised graph convolutional network designed for homogeneous graphs. Based on a graph of company node relationships obtained through a company node graph generation module, it uses the adjacency relationships between companies as input to perform a risk assessment task on company nodes and records the best performance. GAT: A semi-supervised neural network considering attention mechanisms on homogeneous graphs. Based on a graph of company node relationships obtained through a company node graph generation module, it uses the adjacency relationships between companies as input to perform a risk assessment task on company nodes and records the best performance. GraphSAGE: A semi-supervised neural network considering sampling neighbor nodes on homogeneous graphs. Based on a graph of company node relationships obtained through a company node graph generation module, it uses the adjacency relationships between companies as input to perform a risk assessment task on company nodes and records the best performance. HAN: A semi-supervised graph neural network applied to heterogeneous graphs, employing a multi-layer attention mechanism. By setting different meta-paths, heterogeneous graphs are transformed into different homogeneous graphs, and information is embedded into nodes through node-level attention and semantic-level attention. We use the heterogeneous graph of the industry chain as input to the model, perform risk assessment on company nodes, and record the best performance.

[0135] IV. Experimental Analysis

[0136] The model was trained on training sets with different proportions. Each model was tested for 200, 400, and 600 rounds respectively. The experiments were divided into two groups according to the characteristics of the baseline method. The experimental results are shown in Table 1.

[0137]

[0138] Table 1. Comparison of five models. Experimental results.

[0139] Through the above experiments, on the integrated circuit industry chain dataset, (1) GANS outperforms other models in all evaluation metrics under three training ratios, with its prediction accuracy being at least 4% higher than other GNN models. Furthermore, in terms of F1 score, the GANS model is at least 6% higher than other GNN models. Simultaneously, in terms of AUC score, GANS is at least 3% higher than other GNN models. (2) As Figure 6 and Figure 7 As shown in the test set line graph, GANS outperforms existing methods in terms of test set accuracy and AUC score. These results demonstrate that the GANS model of this invention is superior to existing GNN models in both accuracy and effectiveness when assessing the risk status of company nodes in the industry chain.

[0140] Based on the results of the GANS node-level risk assessment and the structural characteristics of the nodes, the risk weights of the company's nodes were obtained. The risk label values ​​of the company's nodes were then weighted and summed to obtain the overall risk score and label of the industry chain for the corresponding quarter. In addition, due to the lack of overall industry chain risk data, we selected the accuracy index to evaluate the experimental results. The results are shown in Table 2.

[0141] Method Accuracy GCN 76.2％ GraphSAGE 80.9％ GAT 80.9％ HAN 90.5％ GANS 95.3％

[0142] Table 2 Comparison of Overall Risk Assessment of the Industrial Chain

[0143] The experimental results show that the GANS model achieved an accuracy rate of 95.3% in assessing the risks corresponding to the integrated circuit industry chain over 21 quarters, which is at least 5% higher than other GNN models. Therefore, the GANS model of this invention can accurately assess the overall risk of the industry chain by starting from the risks of nodes within the industry chain and considering the risk status and structural characteristics of companies within the industry chain.

[0144] V. Ablation Experiment

[0145] Based on the various modules of the model proposed in this invention, four variants of GANS were designed for ablation experiments. GANS\metapath represents removing the company node graph generation module based on metapath from GANS; GANS\stock represents removing the company node graph generation module based on company investment relationships from GANS; GANS\sample represents removing the node neighborhood sampling module from GANS, using only traditional graph convolution to aggregate node neighbor information; and GANS\attention represents removing the node feature fusion module based on attention mechanisms under multi-graph relationships from GANS, only averaging node features and using fully connected layers for node classification. In our experiments, we also selected accuracy, F1 score, and AUC score as performance metrics. Then, we conducted ablation experiments to demonstrate the effectiveness of each component in our model. The results are shown in Table 3.

[0146]

[0147] Table 3 Ablation Experiment Results

[0148] Through the above experiments, all variants showed a certain degree of deterioration in accuracy, binary F1 score, and AUC score. The following conclusions were drawn: (1) After removing the meta-path-based company graph node generation module and the investment relationship-based graph node generation module, respectively, the model's performance on all three indicators decreased, proving that the associations between companies obtained through the company graph generation module are effective for accurately assessing the risk of company nodes in the industry chain. (2) After removing neighborhood node sampling, the model showed the most significant decrease in the three indicators, proving that neighborhood sampling can effectively capture the effective connections between company nodes when assessing the risk of industry chain nodes, thereby achieving efficient risk feature learning. (3) After removing the attention-based node feature fusion module, the model showed a significant decrease in the scores of the evaluation indicators, proving the effectiveness of the attention mechanism in fusing node risk features under multi-graph relationships. Overall, compared with all variants, the entire GANS achieved the best results.

[0149] In summary, the GANS (Graph Neural Network-based Supply Chain Risk Assessment) model proposed in this invention achieves efficient transformation and accurate assessment of supply chain risk issues through a heterogeneous graph construction module, a company graph generation module, a node risk assessment module based on neighborhood sampling and attention mechanisms, and a supply chain risk assessment module. Experiments based on a real-world integrated circuit supply chain dataset demonstrate that the GANS model of this invention performs significantly well in supply chain risk assessment tasks.

[0150] 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 device based on graph neural networks, characterized in that, It includes a heterogeneous graph construction module, a company graph generation module, a node risk assessment module, and a supply chain risk assessment module, among which: The heterogeneous graph construction module is used to construct a heterogeneous graph of the industrial chain by collecting industrial chain data, utilizing the upstream and downstream relationships of products in the industrial chain, the production relationship of companies to products, and the financial attributes of companies. It includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships, and company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes. The company graph generation module is used to transform the heterogeneous graph of the industrial chain into a company node supply isomorphic graph reflecting the supply relationship between company nodes and a company node product competition isomorphic graph reflecting the competitive relationship between companies producing the same product. It also combines the investment and financing relationship between company nodes within the industrial chain to generate a company node investment and financing isomorphic graph, thus forming an industrial chain company node association graph. The node risk assessment module is used to sample and adaptively aggregate the node neighborhoods of different graphs in the industry chain company node association graph through a neighborhood sampling mechanism, and to perform binary risk level classification on the node embedding vectors of neighborhood sampling and aggregation through a relation-level attention mechanism, so as to assess the risk status of each company node and generate node-level risk level embedding vectors. The supply chain risk assessment module is used to fuse the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company's nodes, and then use the supply chain risk assessment algorithm with attention mechanism to quantify the supply chain risk in three elements, so as to realize the multi-level risk assessment of the supply chain. The node risk assessment module uses a neighborhood sampling mechanism to sample and adaptively aggregate the node neighborhoods of different graphs in the industry chain company node association graph, and then uses a relation-level attention mechanism to perform binary risk level classification on the node embedding vectors of the neighborhood sampling and aggregation. Based on the financial feature vector m of company node v(i) v(i) Configure the sampling neighborhood and sample the feature information of the neighboring nodes of the company node v(i); The mean aggregation method is used to obtain the feature vector of the company node v(i) that has aggregated the information of neighbor nodes at the kth layer. The aggregated features of the neighbor nodes at the k-1th layer are obtained by aggregating the embedded neighbor nodes. The aggregation features of the neighboring nodes at layer k are concatenated with the feature information of company node v(i) at layer k-1 to obtain the feature information of company node v(i) at layer k. Then, through normalization and layer-by-layer aggregation, the node embedding vector of company node v(i) in the corresponding graph of the industry chain company node association group is obtained. ; Attention coefficients are configured for three isomorphic graphs of the node association graph of the industry chain companies, and node embedding vectors obtained under different graphs are processed. As input, it is trained and learned through a relation-level attention mechanism to obtain the attention coefficients under the corresponding graph; Based on the obtained attention coefficients and node embedding vectors under each isomorphic graph Calculate the comprehensive feature vector of company node v(i). ; The comprehensive feature vector of company node v(i) The input classifier performs binary risk level classification, and the risk status assessment of the company node is completed based on the classification results. The classifier is represented as... In the above formula, risk represents the output result of the risk binary classification. Let w represent the activation function, w be the weight coefficient, and b be the bias coefficient; The loss function L used in the risk assessment of company nodes is... NRE Represented as: In the above formula, This represents the label value of the company node. This represents the predicted value of a company node, where N is the total number of company nodes in a homogeneous graph.

2. The supply chain risk assessment device based on graph neural networks according to claim 1, characterized in that, The heterogeneity graph of the industry chain is represented as follows: , Among them, V i Let E be a set of nodes. i Let K be the set of edges, and K be the set of risk labels for company nodes. Risk level labels for the industrial chain; node set V i It includes two node types: product nodes and company nodes. p V represents the set of product nodes. c Represents the set of company nodes; edge set E i It contains three edge types E cp E represents the set of edges connecting company nodes and product nodes. pp E represents the set of edges connecting product nodes. pc V represents the set of edges connecting product nodes and company nodes; V is the set of company nodes. c Each company node v(i) contains a financial feature vector m. v(i) and structural features p v(i) The risk label set K={0,1} contains two labels representing Safe and Risk, respectively. Each company node v(i) is associated with a node risk label. Related; Risk level label of the industrial chain It includes three levels of risk labels, representing low risk, medium risk, and high risk respectively.

3. The supply chain risk assessment device based on graph neural networks according to claim 2, characterized in that, The company nodes are selected from listed companies in the industry chain, and financial attributes are extracted from the financial statements of the listed companies as the embedded financial feature vectors of the company nodes.

4. The supply chain risk assessment device based on graph neural networks according to claim 1, characterized in that, The company graph generation module transforms the heterogeneous industry chain graph into a company node supply isomorphic graph reflecting the supply relationships between company nodes and a company node product competition isomorphic graph reflecting the competitive relationships between companies producing the same product. The process of generating a company node investment and financing isomorphic graph by combining the investment and financing relationships between company nodes within the industry chain is as follows: Define the meta-path of "company-product-company" and transform the heterogeneous graph of the industry chain to generate a company node supply isomorphic graph G1 that only contains company nodes and the connection relationships between company nodes; Define the meta-path of "company-product-product-company" and transform the heterogeneous graph of the industry chain to generate G2, which contains only company nodes and the connection relationships between company nodes and product competition. Based on the relationships between three types of nodes in the industry chain data—listed companies and investors within the chain, and other companies outside the chain—investment and financing relationships between company nodes within the industry chain are configured. These include investment and financing relationships where different company nodes have the same investor or other companies outside the chain, investment and financing relationships where different company nodes invest in the same company, and investment and financing relationships where one company node invests in another company node. Based on the configured investment and financing relationships between company nodes within the industry chain, the relationships between company nodes within the industry chain are obtained, generating a company node investment and financing isomorphic graph G3 that only contains company nodes and their connection relationships.

5. The supply chain risk assessment device based on graph neural networks according to claim 4, characterized in that, The industry chain company node association diagram is represented as G. j , j={1,2,3}.

6. The supply chain risk assessment device based on graph neural networks according to claim 5, characterized in that, The results of the binary risk level classification are encoded to generate risk level embedding vectors for company nodes. .

7. The supply chain risk assessment device based on graph neural networks according to claim 6, characterized in that, The process by which the supply chain risk assessment module fuses the risk level embedding vector output by the node risk assessment module with the structural characteristics of the company nodes, and then uses an attention-based supply chain risk assessment algorithm to quantify the supply chain risk into a ternary risk level is as follows: Structural features of company node v(i) are extracted based on its position in the industry chain node association graph. The node degree (deg) of company node v(i) is then used to determine its structural characteristics. i This transforms it into a learnable structural embedding vector for the embedding layer. And the structural embedding vector of the company node v(i) is encoded. W p Indicates the weight coefficients of the node structure; The structural embedding vector and the risk level embedding vector are concatenated and their dimensions are reduced to obtain the feature representation vector S(v(i))=MLP([sEmb(v(i))||rEmb(v(i))]); The feature representation vector is embedded into the graph to obtain the industry chain company node relationship graph G. j Feature representation vector , where N is the total number of company nodes in an isomorphic graph; After aggregating the feature representation vectors of each isomorphic graph by mean-averaging, the result is input into a fully connected neural network with a softmax activation function to quantify the ternary risk level and output the multi-level risk assessment result of the industrial chain. ; In the above formula, W g and b g These represent the weight coefficients and bias coefficients of the fully connected neural network, respectively; the loss function L used in the fully connected neural network is... ICRE Represented as: In the above formula, This represents the predicted value for the company's nodes. This indicates the risk level label of the industrial chain. This represents the prediction function for supply chain risk assessment.

8. A supply chain risk assessment method based on graph neural networks, implemented by the supply chain risk assessment device based on graph neural networks as described in claim 1, characterized in that, Includes the following steps: S10. Collect supply chain data and construct a supply chain heterogeneous graph through the upstream and downstream relationships of products in the supply chain, the production relationship of companies to products, and the financial attributes of companies. This graph includes product nodes and company nodes. Product nodes are linked to each other based on direct material relationships. Company nodes are linked to product nodes through the production relationship of companies to products. The financial attributes of companies are embedded information of company nodes. S20. Transform the heterogeneous graph of the industrial chain into a supply isomorphic graph of company nodes that reflects the supply relationship between company nodes and a product competition isomorphic graph of company nodes that reflects the competitive relationship between companies producing the same product. Combine the investment and financing relationship between company nodes in the industrial chain to generate an investment and financing isomorphic graph of company nodes, thus forming an industrial chain company node association graph. S30. The neighborhood sampling mechanism is used to sample and adaptively aggregate the nodes of different graphs in the industry chain company node association graph. The node embedding vectors of the neighborhood sampling aggregation are classified into binary risk levels through the relation-level attention mechanism to realize the risk status assessment of each company node and generate node-level risk level embedding vectors. S40. After fusing the output risk level embedding vector with the structural features of the company node, the risk level of the industrial chain is quantified into a three-element risk level by combining the attention mechanism of the industrial chain risk assessment algorithm, so as to realize the multi-level risk assessment of the industrial chain.

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 method based on graph neural networks as described in claim 8.

Images