Enterprise credit risk assessment method based on cascaded hypergraph neural network
The enterprise credit risk assessment method is constructed through cascading hypergraph neural network, which solves the problem of existing methods ignoring multi-level interaction relationships and achieves more accurate credit risk assessment.
Patent Information
- Application Number
- CN202311196234.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-18
- Publication Date
- 2025-08-08
- Estimated Expiration
- 2043-09-18
AI Technical Summary
The existing corporate credit risk assessment methods ignore multi-level interactions between enterprises, resulting in limited model performance, especially underperforming data.
The cascading hypergraph neural network is adopted to build equity, supply chain, attribute hypergraph, combined with knowledge graph embedding and hypergraph attention network, learn the multi-level topological structure information of the enterprise, design attention mechanisms related to hyper-edge types, and enhance enterprise representation learning.
Effectively capture rich paired and high-order topological information between enterprises, improving the accuracy and applicability of credit risk assessments, especially showing the most advanced performance on real-world datasets.
Smart Images
Figure CN117350845B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of enterprise credit risk assessment, and in particular to an enterprise credit risk assessment method based on knowledge graph embedding and hypergraph neural network. Background Art
[0002] Corporate credit risk assessment is a key technology for investment decision-making and financial regulation. Corporate credit ratings issued by professional rating agencies such as Standard & Poor's are widely used by many financial investment and regulatory agencies. However, credit assessments by traditional rating agencies typically rely on expert experience and extensive data on financial and non-financial indicators, which is time-consuming and expensive. In recent years, some researchers have attempted to improve the efficiency of corporate credit risk assessment using statistical machine learning methods, such as random forests or support vector machines. These works typically model corporate credit risk assessment as a classification task based on corporate characteristics. These characteristics are typically derived from a system of indicators composed of financial and non-financial information, such as the quick ratio or employee quality. However, many small and medium-sized enterprises face data gaps in their financial and non-financial information, which limits the applicability and accuracy of the models. Furthermore, these methods ignore the interactions between enterprises (such as investment or transaction relationships), which contain valuable information for corporate credit risk assessment.
[0003] To address this issue, some recent graph embedding-based enterprise credit risk assessment methods have enhanced the learning of enterprise features by modeling the interactions between enterprises. However, these methods focus only on pairwise interactions between enterprises and typically only utilize one type of interaction relationship. This hinders the performance improvement of these methods for the following reasons:
[0004] In many financial scenarios, the interactions between enterprises are much more complex than pairwise relationships. For example, supply chains have been shown to be very important for enterprise risk assessment. However, for enterprises in a supply chain, it is unwise to model their interactions using simple edges connecting pairwise objects.
[0005] Interactions between enterprises are often multi-layered, such as pairwise and high-order. However, existing methods typically only consider one type of interaction, making it difficult to effectively enrich enterprise feature representations with topological information at different granularity levels. In the real world, interactions between enterprises can be sparse (i.e., isolated nodes may exist in the enterprise graph). In such cases, these nodes may only have limited topological information, or even no information at all, thus impacting model performance. Summary of the Invention
[0006] The technical problems to be solved by the present invention are:
[0007] In order to avoid the shortcomings of the existing technology, the present invention provides an enterprise credit risk assessment method based on a cascaded hypergraph neural network, which enhances enterprise representation learning by topological structures based on different granularities, including knowledge graphs of pairwise relationships and hypergraphs of high-order relationships.
[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0009] A method for assessing corporate credit risk based on a cascaded hypergraph neural network, comprising:
[0010] Obtain corporate credit-related financial data and build a credit risk assessment indicator system;
[0011] Obtain enterprise interaction relationship data and build the enterprise's equity knowledge graph, supply chain knowledge graph, equity hypergraph, supply chain hypergraph, and attribute hypergraph;
[0012] A cascaded hypergraph neural network model is designed to obtain enterprise representations. The cascaded hypergraph neural network model includes: utilizing a knowledge graph embedding network to obtain equity knowledge graph and supply chain knowledge graph embeddings; using equity and supply chain hypergraph attention networks in combination with knowledge graph embeddings to learn comprehensive enterprise representations; employing an attributed hypergraph attention network with a hyperedge type-related attention mechanism to enhance the learning of enterprise representations based on the obtained comprehensive enterprise representations to obtain a final enterprise representation; the initial features of each enterprise used in the knowledge graph embedding network are indicators included in a credit risk assessment indicator system;
[0013] Use enterprise representation to classify risks and assess enterprise credit risk.
[0014] A further technical solution of the present invention: The equity knowledge graph is specifically as follows: for equity relations, a triple is constructed with the shareholder enterprise as the head entity, the shareholding as the relationship, and the invested enterprise as the tail entity.
[0015] A further technical solution of the present invention: The supply chain knowledge graph specifically includes suppliers and customers, and triples are constructed with the enterprise as the head entity or tail entity, and the supplier or customer as the corresponding relationship.
[0016] A further technical solution of the present invention: The equity hypergraph is constructed by first converting each node v i ∈V are set as the central node, find the node v that invests in the enterprise i All enterprises in the equity knowledge graph are mapped to the equity knowledge graph, that is, find the enterprises with v iis the head entity in all triples of the tail entity; then add a hyperedge containing all these head entities and the central node; repeat this process until all n nodes are set as central nodes and n hyperedges are added to the hypergraph; use w(e) to represent the weight of the hyperedge e and initialize w(e) to 1; finally, we get the equity hypergraph G Q ={V Q ,E Q ,W Q}, where V Q is the set of nodes, E Q is the set of hyperedges.
[0017] A further technical solution of the present invention: The supply chain hypergraph is constructed by first converting each node v i ∈V are set as central nodes, and the enterprises involved in the transaction in their supply chain are found; then a hyperedge is added that includes all these enterprise nodes and the central node; this process is repeated until all n nodes are set as central nodes; finally, the supply chain hypergraph G is obtained S ={V S ,E s ,W S}, where V S is the set of nodes, E S is the set of hyperedges.
[0018] A further technical solution of the present invention: The attribute hypergraph is constructed by dividing enterprise nodes of the same industry into the same hyperedge; dividing enterprise nodes of the same tax status into the same hyperedge; and recording the attribute hypergraph as Where W A is a diagonal matrix representing the hyperedge weights.
[0019] A further technical solution of the present invention is that the knowledge graph embedding network is a TransH model, the purpose of which is to learn relationship-specific enterprise identifiers based on the equity knowledge graph and the supply chain knowledge graph, respectively. For each enterprise's entity node v i ∈V, we get two embeddings, namely and Use and Representing the entity embedding matrix in the equity and supply chain knowledge graphs.
[0020] A further technical solution of the present invention is that the hypergraph attention network refers to the equity and supply chain hypergraph attention network ESHAN, which first uses two branch networks, namely the equity hypergraph attention network EHyperGAT and the supply chain hypergraph attention network SHyperGAT, to capture the equity hypergraph G respectively. Q and supply chain hypergraph Gs Then, ESHAN uses the attention mechanism to combine the enterprise representations learned from the above two hypergraphs to learn a comprehensive representation with rich semantics. Specifically, in each branch, the hypergraph G Q or G s It is sent to the hypergraph attention network to perform hypergraph convolution; the hypergraph attention network learns enterprise representation through two modules: the attention node aggregation module and the attention hyperedge aggregation module.
[0021] A further technical solution of the present invention: The specific principle of the attribute hypergraph attention network is as follows:
[0022] First, a new definition of the weight matrix of the hyperedge in the attribute hypergraph is given. In the example, a new diagonal weight matrix W is used A , with different diagonal submatrices Distinguish the weight of each hyperedge type; where represents the number of hyperedge types in the attribute hypergraph, m t represents the number of hyperedges belonging to the tth type; the diagonal submatrix W t The elements on the diagonal can be recorded as vector λ t , all its elements are equal, that is, λ t , represents the weight of the t-th type of hyperedge; in this way, different types of hyperedges are initialized with different weights, while the same type of hyperedges are initialized with the same weight; using all submatrices W t Construct a new hyperedge weight matrix in
[0023] After obtaining the hyperedge weight matrix, a new hyperedge type-dependent attention network (HTDAN) is used to learn enterprise representations. Specifically, HTDAN performs hypergraph convolution through two modules: attention node aggregation and type-dependent hyperedge aggregation.
[0024] Attention node aggregation module: This module learns e by aggregating the information of nodes within the hyperedge j The representation of is therefore equivalent to the equity (or supply chain) hypergraph attention network, which is expressed as follows:
[0025]
[0026] where h k is the node embedding computed by the equity and supply chain hypergraph attention network; similarly, coev jk It can be calculated as follows:
[0027]
[0028] Type-dependent hyperedge aggregation module: In this module, the new type-dependent hyperedge weight matrix W is used A , we learn the representation of enterprise nodes through a type-dependent hyperedge aggregation strategy, and clearly distinguish the influence of different types of hyperedges. Unlike the equity hypergraph attention network, in the process of aggregating hyperedge information connected to nodes to generate node embeddings, we not only use the attention mechanism but also need to combine the weights of various types of hyperedges on the nodes. The specific implementation details are as follows:
[0029]
[0030] in Indicates connection to node v i The weight of the j-th hyperedge of the t-th category; coee ij It is a superedge e j In v i The attention coefficient on is calculated in the same way as before:
[0031]
[0032] For hyperedges of the same type, their weights at a node should be the same; therefore, we calculate the weights of all nodes v i The weight matrix of the t-th hyperedge on ∈V This can be done in the following ways:
[0033]
[0034] Among them, W t is the weight matrix of the redefined t-th hyperedge; is an incidence matrix representing the relationship between t-type hyperedges and G A The relationship between all nodes in the Indicates connection to node v i The set of t-th type hyperedges, Indicates its quantity; It plays a normalizing role here;
[0035] Through the attribute hypergraph attention network, the final enterprise node embedding matrix can be obtained
[0036] A further technical solution of the present invention: the use of enterprise representation for risk classification is specifically: the enterprise representation is input into the softmax layer for node classification.
[0037] The beneficial effects of the present invention are:
[0038] The present invention provides a corporate credit risk assessment method based on a cascaded hypergraph neural network, and proposes a new cascaded graph neural network framework MS-CGNN (Multi-Structure Cascaded Graph Neural Network framework, MS-CGNN) for corporate credit risk assessment. The framework learns multi-level graph convolution features that capture rich pairwise and high-order topological structure information, which is conducive to obtaining richer corporate representations; the present invention also proposes a new hyperedge type-related attention mechanism for heterogeneous hypergraph convolution, which adaptively assigns different weights to each type of corporate hyperedge, which enables MS-CGNN to clearly distinguish the influence of different types of hyperedges and emphasize information from hyperedge types that are important to the corporate credit risk assessment task; a large number of experiments on real-world corporate credit risk assessment datasets show that MS-CGNN can achieve state-of-the-art performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.
[0040] Figure 1 Schematic diagram of the MS-CGNN model of the present invention.
[0041] Figure 2 This is the credit risk assessment indicator system involved in the example of the present invention.
[0042] Figure 3 Schematic diagram of the hypergraph attention network in the method described in an example of the present invention.
[0043] Figure 4 is the hyperedge weight matrix W of the attribute hypergraph involved in the method introduction in the example of the present invention A Schematic diagram of .
[0044] Figure 5 Schematic diagram of the hyperedge type dependent attention network of the method described in an example of the present invention. DETAILED DESCRIPTION
[0045] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0046] This paper proposes a new cascaded graph neural network framework MS-CGNN (Multi-Structure Cascaded Graph Neural Network framework, MS-CGNN) for corporate credit risk assessment. Figure 1 As shown, it contains three modules: enterprise data acquisition and data preprocessing module, cascaded hypergraph neural network module, and enterprise credit risk assessment module. The details are as follows:
[0047] 1. Enterprise data acquisition and data preprocessing module
[0048] 1.1 Enterprise Data Acquisition
[0049] We collected credit-related financial data from the Datayes and CSMAR databases, while our credit risk label data came from the Tianyancha platform. We also collected equity and supply relationships between companies from the CSMAR database and the Tianyancha platform, and company attribute information from the CSMAR database.
[0050] 1.2 Data Preprocessing Module
[0051] Regarding the financial data related to corporate credit, 18 indicators were selected from four aspects: corporate solvency, operating ability, profitability and growth ability to construct a credit risk assessment indicator system. The specific situation is as follows: Figure 2 shown.
[0052] Regarding the relational data of enterprises, the present invention constructs an equity knowledge graph, a supply chain knowledge graph, an equity hypergraph, a supply chain hypergraph, and an attribute hypergraph.
[0053] The specific construction method of the enterprise's equity knowledge graph and supply chain knowledge graph is as follows: For equity relationships, a triple is constructed with the shareholder enterprise as the head entity, the shareholding as the relationship, and the investee enterprise as the tail entity. For supply relationships, including suppliers and customers, a triple is constructed with the enterprise as the head or tail entity, and the supplier or customer as the corresponding relationship.
[0054] The specific construction method of the enterprise's equity hypergraph is: first, each node v i ∈V are set as the central node, find the node v that invests in the enterprise i All enterprises in the equity knowledge graph are mapped to the equity knowledge graph, that is, find the enterprises with v iFor all triples of the tail entity, add a hyperedge containing all these head entities and the center node. Repeat this process until all n nodes are set as center nodes and n hyperedges are added to the hypergraph. Use w(e) to represent the weight of hyperedge e and initialize w(e) to 1. Therefore, the weight matrix is initialized to the identity matrix. Finally, we get the equity hypergraph G Q ={V Q ,E Q ,W Q}, where V Q is the set of nodes, E Q is the set of hyperedges.
[0055] The specific construction method of the enterprise's supply chain hypergraph is: first, each node v i ∈V are set as central nodes, and the enterprises involved in the transaction in their supply chain are found. Then a hyperedge is added that includes all these enterprise nodes and the central node. This process is repeated until all n nodes are set as central nodes. Similarly, the weight matrix is initialized to the identity matrix, and the supply chain hypergraph G is finally obtained. S ={V S ,E S ,W S}, where V S is the set of nodes, E S is the set of hyperedges.
[0056] The specific construction method of the attribute hypergraph of an enterprise is as follows: in addition to the interaction relationship between enterprises, enterprise nodes also contain rich attributes, and an attribute hypergraph can be constructed based on the attribute information of the nodes. Specifically, enterprise nodes with the same industry are divided into the same hyperedge; in addition, enterprise nodes with the same tax status are divided into the same hyperedge. In this way, an attribute hypergraph can be obtained. In the attribute hypergraph, since the sum of the node type and the hyperedge type is greater than 2, the attribute hypergraph is a heterogeneous hypergraph. The attribute hypergraph can be denoted as Where W A is a diagonal matrix representing the hyperedge weights, which is also initialized to the identity matrix.
[0057] 2. Cascaded Hypergraph Neural Network
[0058] 2.1 Knowledge Graph Embedding Network
[0059] The knowledge graph embedding network used in this paper is the TransH model, which aims to learn the relationship-specific enterprise identifiers based on the equity knowledge graph and the supply chain knowledge graph respectively. For each enterprise’s entity node v i ∈V, we get two embeddings, namely and Use and Represents the entity (node) embedding matrix in the equity and supply chain knowledge graph. It should be noted that the initial characteristics of each enterprise used in the knowledge graph embedding network are the 18 indicators included in the risk assessment indicator system constructed in Module 1.2.
[0060] 2.2 Equity and Supply Chain Hypergraph Attention Network
[0061] Equity and Supply Chain Hypergraph Attention Network (ESHAN) first uses two branch networks, namely Equity Hypergraph Attention Network (EHyperGAT) and Supply Chain Hypergraph Attention Network (SHyperGAT), to capture the equity hypergraph (G Q ) and supply chain hypergraph (G s ) to obtain a preliminary enterprise representation. Then, ESHAN uses the attention mechanism to combine the enterprise representations learned from the above two hypergraphs to learn a comprehensive representation with rich semantics. Specifically, in each branch, the hypergraph (G Q or G s ) is sent to the hypergraph attention network to perform hypergraph convolution. The hypergraph attention network learns enterprise representation through two modules: the attention node aggregation module and the attention hyperedge aggregation module, such as Figure 3 The details of each part of the network are as follows:
[0062] 2.2.1EHyperGAT:
[0063] Attention node aggregation module. Given a specific hyperedge e j ∈E, this module aggregates the nodes v within the hyperedge k ∈e j Come learn e j Since different enterprise nodes have different equity hyperedges e j The contributions of the present invention are different. Therefore, the present invention calculates the hyperedge embedding representation f j l The attention mechanism is introduced in the process. Specifically, the attention node aggregation module can be expressed as:
[0064]
[0065] in, σ is a nonlinear function, such as LeakyReLu; P1 is a trainable linear transformation weight matrix; coev jk Represents the hyperedge e j Internal node v k The attention coefficient is used to quantify v k For e j The importance of can be calculated as follows:
[0066]
[0067] Among them, a v is the shared attention mechanism weight vector, and ‖ is the connection operation.
[0068] Attention hyperedge aggregation module. Given a specific node v i ∈V, the module is connected to the hyperedge e via the aggregation node j ∈ε i Come learn v i where the hyperedge representation is obtained by the attention node aggregation module, i.e. It is similar to the attention node aggregation module, where we apply the edge-level attention mechanism. Specifically, the attention hyperedge aggregation module can be formally expressed as:
[0069]
[0070] Among them, P2 is a trainable linear transformation weight matrix, coee ij Represents the hyperedge e j At node v i The attention coefficient on , is calculated as:
[0071]
[0072] Among them, a e is another shared attention mechanism weight vector.
[0073] 2.2.2 SHyperGAT:
[0074] The supply chain hypergraph attention network has a similar workflow to the equity hypergraph attention network. Through the equity and supply chain hypergraph attention networks, the equity embedding matrix can be obtained and supply chain embeddedness matrix
[0075] 2.2.3 Attention Combination of Equity Embedding and Supply Chain Embedding:
[0076] In this part, combined with H q and H s Learn comprehensive corporate representation H c , fusion structure hypergraph G Q and G s different semantic information.
[0077] H c =α q H q +α s H s(5)
[0078] Among them, α q and α q Respectively represent from G Q and G s The attention coefficient of the enterprise representation learned in can be calculated by the following formula:
[0079]
[0080]
[0081] Among them, a c is the attention vector, and W1 represents the trainable linear transformation weight matrix.
[0082] 2.3 Attribute Hypergraph Attention Network
[0083] First, a new definition of the weight matrix of the hyperedge in the attribute hypergraph is given. In the example, a new diagonal weight matrix W is used A , with different diagonal submatrices Distinguish the weight of each hyperedge type. represents the number of hyperedge types in the attribute hypergraph, m t Represents the number of hyperedges belonging to the tth class. The diagonal submatrix W t The elements on the diagonal can be recorded as vector λ t , all its elements are equal, that is, λ t , represents the weight of the t-th hyperedge. In this way, different types of hyperedges are initialized with different weights, while the same type of hyperedges are initialized with the same weight. Using all submatrices W t Construct a new hyperedge weight matrix in Specifically, W A Ke Ru Figure 4 shown.
[0084] After obtaining the hyperedge weight matrix, a new hyperedge type dependent attention network (HTDAN) is used to learn enterprise representation. Figure 5 As shown in Figure 2, HTDAN performs hypergraph convolution through two modules: attention node aggregation and type-dependent hyperedge aggregation.
[0085] Attention node aggregation module. This module learns e by aggregating the information of nodes within the hyperedge j The representation of is therefore equivalent to the equity (or supply chain) hypergraph attention network, which is expressed as follows:
[0086]
[0087] Among them, h k is the node embedding computed by the equity and supply chain hypergraph attention network. Similarly, coev jk It can be calculated as follows:
[0088]
[0089] Type-dependent hyperedge aggregation module. In this module, the new type-dependent hyperedge weight matrix W is used A , we learn the representation of enterprise nodes through a type-dependent hyperedge aggregation strategy, clearly distinguishing the influence of different types of hyperedges. Unlike the equity hypergraph attention network, when aggregating hyperedge information connected to a node to generate node embeddings, we not only use the attention mechanism but also need to combine the weights of various types of hyperedges on the node. The specific implementation details are as follows:
[0090]
[0091] in, Indicates connection to node v i The weight of the j-th hyperedge of the t-th category. ij It is a superedge e j In v i The attention coefficient on is calculated in the same way as before:
[0092]
[0093] For hyperedges of the same type, their weights on a node should be the same. Therefore, calculate the weights of all nodes v i The weight matrix of the t-th hyperedge on ∈V This can be done in the following ways:
[0094]
[0095] Among them, W t is the weight matrix of the redefined t-th hyperedge. is an incidence matrix representing the relationship between t-type hyperedges and G A The relationship between all nodes in . Indicates connection to node v i The set of t-th type hyperedges, Indicates its quantity. It plays a normalization role here.
[0096] Through the attribute hypergraph attention network, the final enterprise node embedding matrix can be obtained
[0097] 3. Enterprise credit risk assessment
[0098] This paper formalizes the enterprise credit risk assessment task into a node classification problem. a After that, H a Input to the softmax layer for node classification, which can be formalized as:
[0099]
[0100] Among them, W c and b c are the weight matrices and biases. Represents the predicted classification label. The goal of the model is to minimize the cross entropy loss between the true label and the predicted label, as shown below:
[0101]
[0102] Among them, D Train ={(u,y)} is the training set, and y represents the true label.
[0103] The present invention provides an enterprise credit risk assessment method based on a cascaded hypergraph neural network, which includes the following three parts: 1) obtaining enterprise credit-related financial data and enterprise interaction relationship data, and constructing a credit risk assessment indicator system, an enterprise knowledge graph, and an enterprise hypergraph; 2) designing a cascaded hypergraph neural network model architecture: ① using a knowledge graph embedding network to obtain knowledge graph embedding; ② using a hypergraph attention network, combined with knowledge graph embedding to learn a comprehensive representation of the enterprise; ③ designing an attention mechanism with hyperedge type correlation, based on the obtained comprehensive enterprise representation to enhance the learning of the enterprise representation, and obtain the final enterprise representation; 3) enterprise credit risk assessment: sending the obtained final enterprise representation to the softmax layer for enterprise credit risk classification, and using the cross-entropy loss function for training.
[0104] The specific steps of the method of the present invention are as follows:
[0105] Step 1: Collect the company's credit-related financial data from the database, build a credit risk assessment indicator system, and collect company relationship data to build the company's equity knowledge graph, supply chain knowledge graph, equity hypergraph, supply chain hypergraph, and attribute hypergraph;
[0106] Step 2: Design a cascaded hypergraph neural network model architecture to obtain enterprise representation;
[0107] Step 2.1: First, use the knowledge graph embedding network to obtain the enterprise's knowledge graph embedding to supplement the missing financial indicator information;
[0108] Step 2.2: In order to capture interactions beyond the pairwise relationships of enterprises and enrich the learned basic enterprise representation, a hypergraph attention network is used to learn a comprehensive representation of enterprises.
[0109] Step 2.3: In order to clearly distinguish the influence of different types of hyperedges and emphasize the information from hyperedge types that are important for the enterprise credit risk assessment task, the attribute hypergraph attention network with hyperedge type-related attention mechanism based on the heterogeneous high-order structure of the enterprise attribute hypergraph is used to enhance the learning of enterprise representation learning and obtain the final enterprise representation.
[0110] Step 3: Use the enterprise representation obtained in step 2 to classify the enterprise to achieve the purpose of assessing the enterprise's credit risk.
[0111] The databases in step 1 refer to the Datayes and CSMAR databases and the Tianyancha platform.
[0112] We collected credit-related financial data from the Datayes and CSMAR databases. We then constructed a credit risk assessment system using 18 indicators across four key areas: solvency, operating capacity, profitability, and growth. Additionally, we sourced credit risk label data from the Tianyancha platform.
[0113] Regarding the collection of enterprise relationship data, the equity relationship and supply relationship between enterprises come from the CSMAR database and Tianyancha platform, and the enterprise attribute information comes from the CSMAR database.
[0114] The specific construction method of the enterprise's equity knowledge graph and supply chain knowledge graph is as follows: For equity relationships, a triple is constructed with the shareholder enterprise as the head entity, the shareholding as the relationship, and the investee enterprise as the tail entity. For supply relationships, including suppliers and customers, a triple is constructed with the enterprise as the head or tail entity, and the supplier or customer as the corresponding relationship.
[0115] The specific construction method of the enterprise's equity hypergraph is: first, each node v i ∈V are set as the central node, find the node v that invests in the enterprise i All enterprises in the equity knowledge graph are mapped to the equity knowledge graph, that is, find the enterprises with v i For all triples of the tail entity, add a hyperedge containing all these head entities and the center node. Repeat this process until all n nodes are set as center nodes and n hyperedges are added to the hypergraph. Use w(e) to represent the weight of hyperedge e and initialize w(e) to 1. Therefore, the weight matrix is initialized to the identity matrix. Finally, we get the equity hypergraph G Q ={V Q ,EQ ,W Q}, where V Q is a set of nodes, E Q is the set of hyperedges, W Q is the hyperedge weight matrix.
[0116] The specific construction method of the enterprise's supply chain hypergraph is: first, each node v i ∈V are set as central nodes, find the enterprises involved in the transaction in their supply chain, and then add a hyperedge containing all these enterprise nodes and the central node. Repeat this process until all n nodes are set as central nodes. Similarly, the weight matrix W S Initialized as the identity matrix, the supply chain hypergraph G is finally obtained S ={V S ,E S ,W S}, where V S is a set of nodes, E S is the set of hyperedges.
[0117] The specific construction method of the attribute hypergraph of an enterprise is as follows: in addition to the interaction relationship between enterprises, enterprise nodes also contain rich attributes, and an attribute hypergraph can be constructed based on the attribute information of the nodes. Specifically, enterprise nodes with the same industry are divided into the same hyperedge; in addition, enterprise nodes with the same tax status are divided into the same hyperedge. In this way, an attribute hypergraph can be obtained. In the attribute hypergraph, since the sum of the node type and the hyperedge type is greater than 2, the attribute hypergraph is a heterogeneous hypergraph. The attribute hypergraph can be denoted as Where W A is a diagonal matrix representing the hyperedge weights, also initialized to the identity matrix, V A is a set of nodes, E A is a set of hyperedges, Indicates the number of hyperedge types.
[0118] The knowledge graph embedding network used in step 2.1 is the TransH model, which aims to learn the relationship-specific enterprise identity based on the equity knowledge graph and the supply chain knowledge graph, respectively. For each enterprise entity node v i ∈V, we get two embeddings, namely and Use and Represents the entity (node) embedding matrix in the equity and supply chain knowledge graph. It should be noted that the initial features of each enterprise used in the knowledge graph embedding network are the 18 indicators included in the risk assessment indicator system constructed in step 1.
[0119] The hypergraph attention network used in step 2.2 refers to the Equity and Supply Chain Hypergraph Attention Network (ESHAN), which first uses two branch networks, namely the Equity Hypergraph Attention Network (EHyperGAT) and the Supply Chain Hypergraph Attention Network (SHyperGAT), to capture the equity hypergraph (G Q ) and supply chain hypergraph (G S ) to obtain a preliminary enterprise representation. Then, ESHAN uses the attention mechanism to combine the enterprise representations learned from the above two hypergraphs to learn a comprehensive representation with rich semantics. Specifically, in each branch, the hypergraph (G Q or G s ) is fed into the Hypergraph Attention Network to perform hypergraph convolution. The Hypergraph Attention Network learns enterprise representations through two modules: the Attention Node Aggregator and the Attention Hyperedge Aggregator. The details of each part of the network are as follows:
[0120] ①EHyperGAT:
[0121] Attention node aggregation module. Given a specific hyperedge e j ∈E, this module aggregates the nodes v within the hyperedge k ∈e j Come learn e j Since different enterprise nodes have different equity hyperedges e j The contribution of the present invention is different, therefore, the hyperedge embedding representation is calculated The attention mechanism is introduced in the process. Specifically, the attention node aggregation module can be expressed as:
[0122]
[0123] in σ is a nonlinear function, such as LeakyReLu; P1 is a trainable linear transformation weight matrix; coev jk Represents the hyperedge e j Internal node v k The attention coefficient is used to quantify v k For e j The importance of can be calculated as follows:
[0124]
[0125] where a v is the shared attention mechanism weight vector, and ‖ is the connection operation.
[0126] Attention hyperedge aggregation module. Given a specific node v i ∈V, the module is connected to the hyperedge e via the aggregation node j∈ε i Come learn v i where the hyperedge representation is obtained by the attention node aggregation module, i.e. It is similar to the attention node aggregation module, where we apply the edge-level attention mechanism. Specifically, the attention hyperedge aggregation module can be formally expressed as:
[0127]
[0128] Where P2 is a trainable linear transformation weight matrix, coee ij Represents the hyperedge e j At node v i The attention coefficient on , is calculated as:
[0129]
[0130] where a e is another shared attention mechanism weight vector.
[0131] ②SHyperGAT:
[0132] The supply chain hypergraph attention network has a similar workflow to the equity hypergraph attention network. Through the equity and supply chain hypergraph attention networks, the equity embedding matrix can be obtained and supply chain embeddedness matrix
[0133] ③Combining attention between equity embedding and supply chain embedding:
[0134] In this part, combined with H q and H s Learn comprehensive corporate representation H c , fusion structure hypergraph G Q and G S different semantic information.
[0135] H c =α q H q +α s H s (5)
[0136] where α q and α q Respectively represent from G Q and G S The attention coefficient of the enterprise representation learned in can be calculated by the following formula:
[0137]
[0138]
[0139] where a c is the attention vector, and W1 represents the trainable linear transformation weight matrix.
[0140] The specific principle of the attribute hypergraph attention network used in step 2.3 is as follows:
[0141] ① First, a new definition of the weight matrix of the hyperedge in the attribute hypergraph is given. In the example, a new diagonal weight matrix W is used A , with different diagonal submatrices Distinguish the weight of each hyperedge type. represents the number of hyperedge types in the attribute hypergraph, m t Represents the number of hyperedges belonging to the tth class. The diagonal submatrix W t The elements on the diagonal can be recorded as vector λ t , all its elements are equal, that is, λ t , represents the weight of the t-th hyperedge. In this way, different types of hyperedges are initialized with different weights, while the same type of hyperedges are initialized with the same weight. Using all submatrices W t Construct a new hyperedge weight matrix in
[0142] ② After obtaining the hyperedge weight matrix, a novel Hyperedge Type-Dependent Attention Network (HTDAN) is used to learn enterprise representations. Specifically, HTDAN performs hypergraph convolution through two modules: attention node aggregation and type-dependent hyperedge aggregation.
[0143] Attention node aggregation module. This module learns e by aggregating the information of nodes within the hyperedge j The representation of is therefore equivalent to the equity (or supply chain) hypergraph attention network, which is expressed as follows:
[0144]
[0145] where h k is the node embedding computed by the equity and supply chain hypergraph attention network. Similarly, coev jk It can be calculated as follows:
[0146]
[0147] Type-dependent hyperedge aggregation module. In this module, the new type-dependent hyperedge weight matrix W is used A, we learn the representation of enterprise nodes through a type-dependent hyperedge aggregation strategy, clearly distinguishing the influence of different types of hyperedges. Unlike the equity hypergraph attention network, when aggregating hyperedge information connected to a node to generate node embeddings, we not only use the attention mechanism but also need to combine the weights of various types of hyperedges on the node. The specific implementation details are as follows:
[0148]
[0149] in Indicates connection to node v i The weight of the j-th hyperedge of the t-th category. ij It is a superedge e j In v i The attention coefficient on is calculated in the same way as before:
[0150]
[0151] For hyperedges of the same type, their weights on a node should be the same. Therefore, calculate the weights of all nodes v i The weight matrix of the t-th hyperedge on ∈V This can be done in the following ways:
[0152]
[0153] Where W t is the weight matrix of the redefined t-th hyperedge. is an incidence matrix representing the relationship between t-type hyperedges and G A The relationship between all nodes in . Indicates connection to node v i The set of t-th type hyperedges, Indicates its quantity. It plays a normalization role here.
[0154] Through the attribute hypergraph attention network, the final enterprise node embedding matrix can be obtained
[0155] The specific methods for assessing corporate credit risk in step 3 are:
[0156] The enterprise credit risk assessment task is formalized as a node classification problem. After obtaining the final node embedding H a After that, H a Input to the softmax layer for node classification, which can be formalized as:
[0157]
[0158] Where Wc and b c are the weight matrices and biases. Represents the predicted classification label. The goal of the model is to minimize the cross entropy loss between the true label and the predicted label, as shown below:
[0159]
[0160] Among them D Train ={(u,y)} is the training set, y represents the true label, and u represents the initial input feature.
[0161] 4. Experiments and Results
[0162] 4.1 Classification Effect
[0163] As shown in Table 1, XGBoost achieves better performance than Logistic Regression because it can handle nonlinear relationships between enterprise features. Since graph-based methods effectively utilize pairwise interactions between enterprises, graph-based methods are significantly superior to traditional statistical machine learning methods. This means that interactions between enterprises carry useful information, which can supplement the credit-related characteristics of individual enterprises and alleviate the problem of insufficient data to a certain extent. With the help of high-order relationships, both HyperGAT and the proposed MS-CGNN outperform GAT. This shows that hypergraphs contain more information than pairwise simple graphs. The HyperGAT model is executed on a hypergraph constructed using all the hyperedges in the three hypergraphs, so it is impossible to distinguish which hypergraph is more valuable for enterprise credit risk assessment. The MS-CGNN proposed in the present invention performs best in all three evaluation indicators, which proves the superiority of the proposed MS-CGNN.
[0164] Table 1 Performance of credit risk assessment tasks on the test set
[0165]
[0166] 4.2 Ablation Experiment
[0167] In order to study the role of each module in the model, an ablation analysis was performed and the results are shown in Table 2.
[0168] Table 2 Ablation experiment analysis
[0169]
[0170] Specifically, MS-CGNN-Without-HTDA is a variant of MS-CGNN, which removes HTDA from the HTDAN module in the complete model. It no longer uses hyperedge-type-dependent attention (HTDA) in the hyperedge-type-dependent attention network module, that is, it uses the original identity matrix as the weight matrix. This observation can verify the effectiveness of our redefined weight matrix and HTDA for learning heterogeneous attribute hypergraphs. In addition, in order to further verify the performance of HTDA, experiments were designed separately on attribute hypergraphs. Among them, HTDAN is the complete hyperedge-type-dependent attention network module, and HTDAN-Without-HTDA refers to removing HTDA from the HTDAN module, that is, directly executing HyperGAT on a single attribute hypergraph. The performance gap between HTDAN-Without-HTDA and HTDAN further illustrates the effectiveness of HTDA.
[0171] By removing the equity hypergraph and supply chain hypergraph one by one, we obtain the MS-CGNN variants SHyperGAT+HTDAN and EHyperGAT+HTDAN. Experimental results show that removing the equity hypergraph or supply chain hypergraph degrades model performance. Comparing the results of SHyperGAT+HTDAN and EHyperGAT+HTDAN reveals that supply chain information is more important. However, different hypergraph information can complement each other and have a positive effect on improving model performance.
[0172] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.
Claims
1. A method for assessing corporate credit risk based on a cascaded hypergraph neural network, characterized in that: include: Obtain corporate credit-related financial data and build a credit risk assessment indicator system; Obtain enterprise interaction relationship data and build the enterprise's equity knowledge graph, supply chain knowledge graph, equity hypergraph, supply chain hypergraph, and attribute hypergraph; Design a cascaded hypergraph neural network model to obtain enterprise representation; The cascaded hypergraph neural network model includes: using a knowledge graph embedding network to obtain equity knowledge graph and supply chain knowledge graph embedding; using equity and supply chain hypergraph attention networks in combination with knowledge graph embedding to learn comprehensive enterprise representations; using an attribute hypergraph attention network with a hyperedge type-related attention mechanism to enhance the learning of enterprise representations based on the obtained comprehensive enterprise representations to obtain the final enterprise representations; the initial features of each enterprise used in the knowledge graph embedding network are indicators included in the credit risk assessment indicator system; Use enterprise representation to classify risks and assess enterprise credit risk.
2. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: Specifically, the equity knowledge graph is as follows: for equity relations, a triple is constructed with the shareholder enterprise as the head entity, the shareholding as the relationship, and the invested enterprise as the tail entity.
3. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: The supply chain knowledge graph specifically includes suppliers and customers, and triples are constructed with the enterprise as the head entity or tail entity and the supplier or customer as the corresponding relationship.
4. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: The equity hypergraph is constructed as follows: first, each node v i ∈V are set as the central node, find the node v that invests in the enterprise i All enterprises in the equity knowledge graph are mapped to the equity knowledge graph, that is, find the enterprises with v i The head entity in all triples of the tail entity; then add a hyperedge containing all these head entities and the central node; repeat this process until all n nodes are set as central nodes and n hyperedges are added to the hypergraph; Use w(e) to represent the weight of hyperedge e, and initialize w(e) to 1; finally, we get the equity hypergraph G Q ={V Q ,E Q ,W Q }, where V Q is a set of nodes, E Q is the set of hyperedges.
5. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: The supply chain hypergraph is constructed as follows: first, each node v i ∈V are set as central nodes, and the enterprises involved in the transaction in their supply chain are found; then a hyperedge is added that includes all these enterprise nodes and the central node; this process is repeated until all n nodes are set as central nodes; finally, the supply chain hypergraph G is obtained S ={V S ,E S ,W S }, where V S is a set of nodes, E S is the set of hyperedges.
6. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: The attribute hypergraph is constructed by: dividing enterprise nodes with the same industry into the same hyperedge; dividing enterprise nodes with the same tax status into the same hyperedge; Let the attribute hypergraph be Where W A is a diagonal matrix representing the hyperedge weights.
7. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1 is characterized in that: The knowledge graph embedding network is a TransH model, which aims to learn relationship-specific enterprise identities based on the equity knowledge graph and the supply chain knowledge graph, respectively. For each enterprise entity node v i ∈V, we get two embeddings, namely and Use and Representing the entity embedding matrix in the equity and supply chain knowledge graphs.
8. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 7 is characterized in that: The hypergraph attention network refers to the equity and supply chain hypergraph attention network ESHAN, which first uses two branch networks, namely the equity hypergraph attention network EHyperGAT and the supply chain hypergraph attention network SHyperGAT, to capture the equity hypergraph G respectively. Q and supply chain hypergraph G S Then, ESHAN uses the attention mechanism to combine the enterprise representations learned from the above two hypergraphs to learn a comprehensive representation with rich semantics. Specifically, in each branch, the hypergraph G Q or G S It is sent to the hypergraph attention network to perform hypergraph convolution; the hypergraph attention network learns enterprise representation through two modules: the attention node aggregation module and the attention hyperedge aggregation module.
9. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 8, characterized in that: The specific principle of the attribute hypergraph attention network is as follows: First, a new definition of the weight matrix of the hyperedge in the attribute hypergraph is given. In the example, a new diagonal weight matrix W is used A , with different diagonal submatrices Distinguish the weight of each hyperedge type; where represents the number of hyperedge types in the attribute hypergraph, m t represents the number of hyperedges belonging to the tth type; the diagonal submatrix W t The elements on the diagonal can be recorded as vector λ t , all its elements are equal, that is, λ t , represents the weight of the t-th type of hyperedge; in this way, different types of hyperedges are initialized with different weights, while the same type of hyperedges are initialized with the same weight; using all submatrices W t Construct a new hyperedge weight matrix in After obtaining the hyperedge weight matrix, a new hyperedge type-dependent attention network (HTDAN) is used to learn enterprise representations. Specifically, HTDAN performs hypergraph convolution through two modules: attention node aggregation and type-dependent hyperedge aggregation. Attention node aggregation module: This module learns e by aggregating the information of nodes within the hyperedge j The representation of is therefore equivalent to the equity or supply chain hypergraph attention network, which is expressed as follows: where h k is the node embedding computed through the equity and supply chain hypergraph attention network; coev jk It can be calculated as follows: Type-dependent hyperedge aggregation module: In this module, the new type-dependent hyperedge weight matrix W is used A , we learn the representation of enterprise nodes through a type-dependent hyperedge aggregation strategy, and clearly distinguish the influence of different types of hyperedges. Unlike the equity hypergraph attention network, in the process of aggregating hyperedge information connected to nodes to generate node embeddings, we not only use the attention mechanism but also need to combine the weights of various types of hyperedges on the nodes. The specific implementation details are as follows: in Indicates connection to node v i The weight of the jth hyperedge of the tth category; coee ij It is a superedge e j In v i The attention coefficient on is calculated in the same way as before: For hyperedges of the same type, their weights at a node should be the same; therefore, we calculate the weights of all nodes v i The weight matrix of the t-th hyperedge on ∈V This can be done in the following ways: Among them, W t is the weight matrix of the redefined t-th hyperedge; is an incidence matrix representing the relationship between t-type hyperedges and G A The relationship between all nodes in the Indicates connection to node v i The set of t-th type hyperedges, Indicates its quantity; It plays a normalizing role here; Through the attribute hypergraph attention network, the final enterprise node embedding matrix can be obtained 10. The enterprise credit risk assessment method based on cascaded hypergraph neural network according to claim 1, characterized in that: The risk classification using enterprise representation is specifically as follows: the enterprise representation is input into the softmax layer for node classification.