Enterprise default prediction method and system based on meta-path denoising and capsule network modeling

Abstract

The invention discloses an enterprise default prediction method and system based on meta-path denoising and capsule network modeling, and belongs to the field of machine learning. The method comprises the following steps: acquiring internal index information and external associated information of an enterprise, and constructing a heterogeneous information network and a meta network of the enterprise according to the external associated information of the enterprise; an enterprise default evaluation model capable of balancing mobility and noise interference, effectively separating enterprise default risk and contract-keeping risk-free information and improving default identification accuracy is constructed; in the constructed enterprise default evaluation model, training various kinds of collected enterprise internal operation information and the constructed enterprise heterogeneous information network and element network to obtain a trained model weight; and acquiring enterprise internal operation and external association information in real business, inputting the enterprise internal operation and external association information into the constructed enterprise default evaluation model, and evaluating the possibility of enterprise default in the real business by using the trained model weight. According to the invention, more accurate enterprise default risk identification can be realized.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular, to an enterprise default prediction method and system based on meta-path denoising and capsule network modeling. Background Art

[0002] Small and medium-sized enterprises play a crucial role in promoting the growth of the national economy. Therefore, analyzing the enterprise default risk is not only of great significance for maintaining economic stability, but also can provide key references for investment decisions, and at the same time help enterprises optimize management and layout. For this purpose, some enterprise default prediction models introduce internal index data of enterprises and external associated default risks to evaluate the default risk of each enterprise.

[0003] Many traditional default assessment models rely on statistical analysis and traditional machine learning methods to predict enterprise default risks, especially for mining internal default risks based on index data. For example, a method for calculating the default probability PD of small and micro enterprise loans (CN111754341B) calculates the default probability by statistically analyzing the historical data of small and micro enterprise loans, using the transition matrix and Markov chain, and makes a forward-looking correction to the result by combining the adjustment parameters provided by business personnel, and finally obtains the accurate default rate in different scenarios; a method for establishing an enterprise default risk model based on xgboost (CN109508864B) sorts out enterprise data, cuts the data by using a sliding window, and establishes a prediction model with the help of the XGBoost algorithm, and the optimized model is used for enterprise default risk prediction. However, small and medium-sized enterprises generally face the problems of irregular or imperfect financial reports, and the practical application of these methods faces many challenges. In addition, these methods ignore the propagation effect of external default risks among enterprises. In recent years, with the advantages of graph models in representing complex relationships gradually emerging, some studies have begun to try to apply them to default prediction, especially in simulating the propagation of external default risks. Such as methods like HAT, ComRisk, "An Enterprise Default Prediction Method, Device, Medium and Electronic Device" (CN112990946B), etc. try to model external risks through heterogeneous graph models, but these methods generally face the problem of insufficient interpretability.

[0004] To address the above challenges, researchers have proposed a default prediction model based on metapaths. This method generates external association paths for enterprises to learn the explicitly propagated default risks, improving the interpretability of the model. However, current default assessment models still face challenges such as limited transferability and susceptibility to noise interference. Specifically, this type of method can generally be divided into two categories: The first category is models that rely on manually preset metapaths. Due to over-reliance on the original data, this limits the transferability of the module. The second category is models that utilize all metapaths not exceeding a fixed length. This approach introduces a large amount of irrelevant noise, reducing the accuracy of information aggregation. In addition, existing default assessment models usually represent the default risk information and non-default risk-free information of enterprises together in a compressed manner. This practice easily confuses the information, resulting in poor enterprise assessment results.

[0005] In summary, existing metapath-based default assessment methods have limitations in terms of transferability and reducing noise interference, and it is difficult to achieve an effective balance. At the same time, existing default assessment methods compress the risk and risk-free information of enterprises into a single representation, leading to information confusion and thus poor assessment results. These two aspects of problems urgently need to be improved. Summary of the Invention

[0006] To overcome the limitations of the prior art, the present invention proposes an enterprise default prediction method and system based on metapath denoising and capsule network modeling. This method can better balance transferability and noise interference, and by separating the enterprise default risk information and non-default risk-free information, achieve more accurate identification of enterprise default risks.

[0007] To achieve the above object, the technical solution of the present invention includes the following content.

[0008] An enterprise default prediction method based on metapath denoising and capsule network modeling, the method comprising:

[0009] Collect the internal business information and external association information of the target enterprise, and construct the heterogeneous information network and metanetwork of the enterprise based on the external association information of the enterprise; wherein, the nodes in the heterogeneous information network include: enterprise nodes, industry nodes, regional nodes, and stakeholder nodes, and the metanetwork is abstracted and generated based on the heterogeneous information network;

[0010] Input the internal index information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise;

[0011] Input the heterogeneous information network and metanetwork of the target enterprise into the explicit external default risk encoder based on metapaths to obtain the explicit external default risk embedding of the target enterprise;

[0012] Input the heterogeneous information network of the target enterprise into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the target enterprise;

[0013] Fuse the internal default risk embedding, explicit external default risk embedding and implicit external default risk embedding, and combine with the capsule network to obtain the default prediction result of the target enterprise.

[0014] Further, the inputting the internal index information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise includes:

[0015] Extract the numerical features and / or discrete features of the internal index information, perform standardization and normalization processing on the numerical features, and convert the discrete features into embedding vectors; wherein, the numerical features include: registered capital, paid-in capital, registration time and litigation duration, and the discrete features include: litigation reason, court level and litigation result;

[0016] Concatenate the embedding vectors of the discrete features with the processed numerical features and perform mapping through a linear layer to obtain the internal default risk embedding of the target enterprise.

[0017] Further, the inputting the heterogeneous information network and meta-network of the enterprise into the explicit external default risk encoder based on the meta-path to obtain the explicit external default risk embedding of the target enterprise includes:

[0018] Based on the meta-network, walk in the heterogeneous information network to generate path instances of all meta-data types not exceeding a fixed length for the target enterprise;

[0019] Calculate the effective value S of each meta-path type P P ;

[0020] According to the effective value S P Sort the meta-path types and select the top K meta-path types with higher rankings;

[0021] Among the top K meta-path types with higher rankings, perform weighted aggregation on the path instances of the meta-path type P to obtain the semantic embedding Z under the meta-path type P P ;

[0022] Use the attention mechanism to fuse the semantic embedding Z P , to obtain the explicit external default risk embedding Z of the target.

[0023] Further, the effective value of the meta-path wherein, T P represents the total number of all path instances of the meta-path type P, Denote the number of path instances where both end nodes in the meta-path type P are defaulting enterprises. Denote the number of path instances where both end nodes in the meta-path type P are compliant enterprises. Denote the number of path instances where the target node in the meta-path type P is a compliant enterprise and the source node is a defaulting enterprise. Denote the number of path instances where the target node in the meta-path type P is a defaulting enterprise and the source node is a compliant enterprise.

[0024] Furthermore, inputting the heterogeneous information network of the target enterprise into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the target enterprise includes:

[0025] Initialize the node embeddings of all nodes in the heterogeneous information network.

[0026] Use the l-th layer of the graph neural network to calculate the embeddings of each node aggregating neighbor information under different external association relationships π. And based on the node embeddings and the embeddings of aggregating neighbor information under all external association relationships π obtain the risk pattern embedding node embedding of this node. Among them, the graph neural network has a total of L layers.

[0027] Concatenate the node embeddings to obtain the default risk pattern embedding matrix P of all nodes. (L) ;

[0028] Extract the risk pattern embedding matrix of all enterprise nodes from the default risk pattern embedding matrix P of all nodes. (L) Extract the risk pattern embedding matrix of all enterprise nodes from the default risk pattern embedding matrix P of all nodes.

[0029] Perform global attention calculation on the risk pattern embedding matrix to obtain the enterprise embedding matrix of each enterprise node.

[0030] Based on the enterprise embedding matrix of the target enterprise and the risk pattern embedding matrix obtain the implicit external default risk embedding of the target enterprise.

[0031] Furthermore, fuse the internal default risk embedding, explicit external default risk embedding, and implicit external default risk embedding, and combine with the capsule network to obtain the default prediction result of the target enterprise, including:

[0032] Fuse the internal default risk embedding, explicit external default risk embedding, and implicit external default risk embedding to obtain the fused embedding.

[0033] Generate the embedding o of the target enterprise in state s based on the fusion embedding and using the dynamic routing mechanism of the capsule network s ; where the state s = {r, nr}, r represents the default risk state, and nr represents the non-default risk state

[0034] Calculate the default risk score ‖o r ‖ and the non-default risk score ‖o nr ‖; where ‖·‖ represents the modulus operation; based on the default risk score ‖o r ‖ and the non-default risk score ‖o r ‖, calculate the default probability of the target enterprise

[0035] Compare the default probability score with a set threshold to obtain the default prediction result of the target enterprise

[0036] Furthermore, generate the embedding o of the target enterprise in state s based on the fusion embedding and using the dynamic routing mechanism of the capsule network s , including

[0037] Extract the feature embedding u of the k-th feature from the fusion embedding k ;

[0038] Based on the feature embedding, obtain the feature embedding t of the k-th feature extracted by the state s capsule from the fusion embedding s,k ;

[0039] Based on the feature embedding t s,k , use the dynamic routing technology to fuse different feature embeddings to obtain the embedding t of the target enterprise in state s s ;

[0040] Based on the embedding t s , generate the embedding o of the target enterprise in state s s .

[0041] An enterprise default prediction system based on meta-path denoising and capsule network modeling, the system includes

[0042] A data collection module for collecting the internal operation information and external association information of the target enterprise, and constructing a heterogeneous information network and a meta-network of the enterprise based on the external association information of the enterprise; where the nodes in the heterogeneous information network include: enterprise nodes, industry nodes, regional nodes, and stakeholder nodes, and the meta-network is abstracted and generated based on the heterogeneous information network

[0043] The first encoding module is used to input the internal index information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise.

[0044] The second encoding module is used to input the heterogeneous information network and meta-network of the target enterprise into the explicit external default risk encoder based on the meta-path to obtain the explicit external default risk embedding of the target enterprise.

[0045] The third encoding module is used to input the heterogeneous information network of the target enterprise into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the target enterprise.

[0046] The default prediction module is used to fuse the internal default risk embedding, the explicit external default risk embedding and the implicit external default risk embedding, and combine with the capsule network to obtain the default prediction result of the target enterprise.

[0047] An electronic device, the electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the enterprise default prediction method based on meta-path denoising and capsule network modeling described in any one of the above is implemented.

[0048] A computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the enterprise default prediction method based on meta-path denoising and capsule network modeling described in any one of the above is implemented.

[0049] Compared with the prior art, the present invention has at least the following beneficial effects.

[0050] 1) The explicit association risk encoder based on the meta-path of the present invention focuses on generating and filtering invalid paths, thereby retaining the key association information in default prediction, effectively balancing the model's mobility and noise interference, and at the same time hierarchically aggregating external risk information to nodes, enhancing the interpretability of risk propagation.

[0051] 2) The implicit association risk modeling component implemented by the heterogeneous graph of the present invention is committed to capturing global implicit association information to make up for the information loss caused by the limited length of the meta-path.

[0052] 3) Based on the internal default risk embedding, the explicit external default risk embedding and the implicit external default risk embedding, the present invention better fuses and utilizes different risk information based on the dynamic routing technology of the capsule network to achieve accurate default prediction. Description of the Drawings

[0053] Figure 1 It is a flowchart of an enterprise default prediction method based on meta-path denoising and capsule network modeling.

[0054] Figure 2 This is the flow chart of the model of the present invention in the training stage.

[0055] Figure 3 This is the flow chart of the model of the present invention in the prediction stage.

[0056] Figure 4 This is the result graph of the performance of the present invention and existing models on common publicly available datasets. Detailed implementation manners

[0057] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0058] The enterprise default prediction method based on meta-path denoising and capsule network modeling of the present invention, as Figure 1 shown, includes the following steps.

[0059] Step 1: Collect the internal index information and external association information of the enterprise, and construct the heterogeneous information network and meta-network of the enterprise based on the external association information of the enterprise.

[0060] The present invention aims to improve the accuracy and interpretability of default prediction through the deep fusion and accurate modeling of multi-dimensional information. First, the method collects the internal information of the enterprise, including numerical features and discrete features, for comprehensive data collection and processing. Specifically, the numerical features include the registered capital, paid-in capital, registration time, litigation duration, etc. of the enterprise; the discrete features cover information such as the litigation reason, court level, litigation result, etc. of the enterprise.

[0061] In addition, the method also collects various external association information of the enterprise with other enterprises, industries, regions, and stakeholders, so as to comprehensively reflect the external risk factors of the enterprise. According to these different types of association information, construct the heterogeneous information network G = {V, E} of the enterprise, where V is the node set of the enterprise and related entities, and E is the set of association relationships between nodes. Based on this heterogeneous information network, further abstract the meta-network T G = {A, R}; where T G is the meta-template of G, A represents the set of abstracted node types, R represents the set of abstracted edge types, satisfying the object mapping and the relationship mapping ψ: E → R.

[0062] Step 2: Construct an enterprise default assessment model that can balance migration and noise interference, effectively separate the enterprise default risk and non-default risk-free information, and improve the accuracy of default identification.

[0063] The present invention proposes an enterprise default prediction method based on meta-path denoising and capsule network modeling, which is essentially a binary classification model, that is, to judge whether an enterprise defaults. The training flow chart of this model is as shown in Figure 2 Figure 1, and includes: inputting the internal indicators of an enterprise into an internal default risk encoder to obtain the internal default risk embedding of the enterprise; inputting the heterogeneous information network and meta-network of the enterprise into an explicit external default risk encoder based on meta-paths to obtain the explicit external default risk embedding of the enterprise; inputting the heterogeneous information network of the enterprise into an implicit external default risk encoder based on a heterogeneous graph to obtain the implicit external default risk embedding of the enterprise; fusing the internal default risk, explicit external default risk, and implicit external default risk of the enterprise, and obtaining the default risk probability score and non-default risk probability score of the enterprise based on the dynamic routing technology of the capsule network; calculating the enterprise default probability based on these two probability scores. Among them, the detailed processes of internal default risk, explicit external default risk, implicit external default risk, risk fusion, and evaluation in this model are shown in the following steps.

[0064] Step 2.1: Input the internal indicator information of the enterprise into the internal default risk encoder to obtain the internal default risk embedding of the enterprise.

[0065] For the internal indicator data of the enterprise, it is divided into numerical type and discrete type according to the feature type. For numerical features, standardization and normalization methods (such as Min-Max normalization and Z-Score standardization) are used for processing to eliminate the influence of dimension and prevent some features from dominating in the model. For discrete features, the Embedding embedding technology is used to convert them into embedding vectors. Subsequently, the embedding vectors of discrete features are concatenated with numerical features and mapped through a linear layer, so as to obtain the internal default risk embedding h i of enterprise c i .

[0066] Step 2.2: Input the heterogeneous information network and meta-network of the enterprise into an explicit external default risk encoder based on meta-paths to obtain the explicit external default risk embedding of the enterprise.

[0067] Given the heterogeneous information network G and meta-network T G , a large number of meta-paths can be generated. However, most of them are irrelevant to the given prediction task. To balance transferability and reduce the interference of irrelevant information, an automatic meta-path selection method is designed. Specifically, based on considering the computational cost and the given prediction task, only meta-paths with both end nodes being enterprises and a length not exceeding a fixed length are considered. An effective meta-path should mainly provide default risk information or non-default risk information, focusing on one of the types. In addition, if the proportion of different information is too similar, the function of this meta-path may become unclear. Given the meta-path P, calculate its effectiveness Among them represents the total number of path instances of the meta-path P, represents the number of instances where both end nodes are defaulting enterprises among all path instances of the meta-path type P, represents the number of instances where both end nodes are compliant enterprises among all path instances of the meta-path type P, represents the number of instances where the target node is a compliant enterprise and the source node is a defaulting enterprise among all path instances of the meta-path type P, represents the number of instances where the target node is a defaulting enterprise and the source node is a compliant enterprise among all path instances of the meta-path type P. Sort all types of meta-paths according to the valid values of the meta-paths, and select the top K meta-paths with higher rankings for capturing the explicit default risk embedding of enterprises.

[0068] To accurately model the external default risk characteristics of enterprises, a hierarchical attention structure is designed. This structure contains two levels, namely the instance level and the semantic level. The instance level aggregates the information along the meta-path instances and calculates the semantic embedding for each node on different meta-paths. The semantic level further considers the semantic differences between different meta-paths and integrates their embeddings to generate the external default risk embedding along the explicit path.

[0069] First, the instance-level fusion. For the target enterprise node c i and the given meta-path type P, there are many different path instances. The importance of these path instances to the target enterprise is different, which depends on the source node c k . In addition, a dynamic attention layer is designed to perform weighted aggregation on the path instances of the meta-path type P of the target node c i to obtain the semantic embedding under the meta-path type P Among them represents the current embedding of enterprise c i and W P represents the embedding transformation matrix of the meta-path type P, represents the set of neighbor enterprises associated with enterprise c i under the meta-path type P, represents the enterprise c i as the target node and the enterprise c k as the source node of the meta-path instance p ik of the embedding, || represents the concatenation operation. represents the fusion attention score of the meta-path instance p ik where represents the learnable computational weight matrix under the meta-path type P, σ represents the activation function, and exp represents the exponential function.

[0070] Secondly, semantic-level fusion. To integrate the semantic information of different meta-paths of each enterprise node c i together, an attention mechanism is proposed to fuse the semantic representations where T i represents the set of meta-paths related to enterprise c i . denotes the fusion weight of meta-path type P for enterprise c i , where W b represents a learnable computational weight matrix.

[0071] Step 2.3: Input the heterogeneous information network of the enterprise into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the enterprise.

[0072] Due to the limitation of the meta-path length, the associations between some enterprises may not be fully utilized. To address this issue, an implicit association default risk encoder based on the heterogeneous graph is proposed, which includes a relation-aware heterogeneous graph convolutional network and a global attention-based implicit association modeling module. The former aims to capture the default risk patterns of enterprises, while the latter explores the implicit associations between enterprises with similar patterns from a global perspective to enhance the effect of default prediction.

[0073] First, the relation-aware heterogeneous graph convolutional network. In the heterogeneous information network G, calculate the embedding of the enterprise aggregating neighbor information under different relations where represents the set of neighbor nodes of node v i under relation π, represents the k-th neighbor node, represents the current embedding of node v k . represents the contribution degree weight of node v k to node v i , where W π represents the matrix of learnable computational weights under relation π, represents the embedding of the edge relation between node v i and node v k . Calculate the risk pattern embedding of the enterprise where θ represents a learnable weight parameter for balancing neighbor information and self-information, σ represents the activation function, represents the current embedding of node v i , W G represents a globally learnable parameter matrix for mapping the embeddings under different relations to the same space. represents the embedding under relation π The fusion weights, where and both represent the embedded vectors after mapping, both are learnable mapping parameters. Subsequently, the results of all enterprises after L-layer aggregation are concatenated row by row to obtain the default risk pattern embedding matrix P of all enterprises (L) , where the default risk pattern embedding matrix P (L) refers to the concatenation of the embeddings of all enterprises row by row.

[0074] Secondly, using the implicit association modeling module based on global attention, calculate the external implicit default risk embedding representation of each enterprise where FFN represents the feed-forward network to enhance the feature representation through multiple non-linear mappings, γ represents a learnable weight parameter for balancing local and global information, represents the risk pattern embedding matrix for all enterprises extracted from P (L) ; represents the implicit risk embedding of enterprise c i . represents the enterprise embedding matrix obtained through global attention, where represents the attention learnable parameter matrix, represents the attention weight matrix, both represent the attention learnable parameter matrix.

[0075] Step 2.4: Fuse the internal default risk, explicit external default risk, and implicit external default risk of the enterprise, and obtain the default risk probability score and non-default risk probability score of the enterprise based on the dynamic routing technology of the capsule network. Calculate the default probability of the enterprise based on the probability scores of the default risk and non-default risk states.

[0076] The present invention uses the risk aggregation layer to fuse the internal default risk embedding h of the enterprise i , the explicit external default risk embedding Z i and the implicit external default risk embedding to obtain the fused embedding where T r ={int, exp, imp} represents the set of risk types, including internal risk int, explicit external risk exp, and implicit external risk imp; represents the fusion weight corresponding to the t-th type of risk. represents the embedded vector after mapping corresponding to the t-th type of risk, including r i int =W int h i , r i exp =Wexp Z i , where W int is the learnable embedding transformation matrix for internal risks, and W exp is the learnable embedding transformation matrix for explicit external risks, and W imp is the learnable embedding transformation matrix for implicit external risks.

[0077] Use the dynamic routing mechanism of the capsule network to generate embeddings for different states s ∈ S of enterprise c i where ‖·‖ obtains the norm of the vector, and S = {r, nr} represents the set of enterprise states, r represents the default risk state, and nr represents the non-default risk-free state; denotes the embedding of enterprise c in state s under the action of the dynamic routing mechanism; where i denotes the fusion weight for the k-th feature calculated by the capsule in state s through the dynamic routing mechanism; denotes the k-th feature embedding extracted from the fusion embedding r by the capsule in state s, where W i denotes the learnable embedding transformation matrix of the capsule in state s for the k-th feature embedding; s,k denotes the k-th feature embedding extracted from the fusion embedding r by the capsule in state s, where W i denotes the learnable embedding transformation matrix for the k-th feature; k denotes the learnable embedding transformation matrix for the k-th feature;

[0078] Calculate the default risk probability score i and the non-default risk-free probability score of enterprise c

[0079] The above-mentioned calculation of the default probability score of the enterprise exp represents the exponential function.

[0080] Step 3: In the constructed enterprise default assessment model, train the various internal business information of the collected enterprises, the constructed enterprise heterogeneous information network and the meta-network to obtain the trained model weights.

[0081] Use the internal information and external association information of the collected training enterprises to construct the enterprise heterogeneous information network and the meta-network, and input them into the constructed enterprise default assessment model to obtain the default probability i of enterprise c

[0082] In each training cycle, the loss function for training C train represents the set of enterprises used for training, and yi Denote enterprise c i 's true label, that is, whether the enterprise is in a default state, Denote enterprise c i 's predicted default risk probability. By using Adam to optimize the objective function, the learning rate can be dynamically adjusted during model training. To improve the robustness of the trained model, techniques such as weight decay and regularization are applied.

[0083] Step 4: Collect the internal operation and external association information of enterprises in real business, and input it into the constructed enterprise default assessment model. Use the trained model weights to evaluate the default possibility of enterprises in real business.

[0084] The present invention uses the data flow in real business as Figure 3 shown;

[0085] Collect the internal information and external association information of enterprises in real business, and construct the heterogeneous information network and meta-network of enterprises;

[0086] Load the weights of the trained default model;

[0087] Input the collected data to obtain the default probability of each test enterprise sample c i ;

[0088] When exceeds 0.5, it means that the model predicts that the enterprise defaults, otherwise it means that the model predicts that the enterprise abides by the contract.

[0089] Figure 4 is the experimental effect comparison chart of the method proposed by the present invention and different existing types of methods, where MDCN is the model of the present invention. The dataset selected by the present invention is a real-world dataset for enterprise default prediction.

[0090] In summary, the present invention includes two novel modules, namely the external risk modeling module and the risk assessment module. Specifically, to better model external risks, the present invention designs an explicit association risk encoder based on meta-paths and an implicit association risk modeling component implemented through a heterogeneous graph. The former focuses on generating and filtering invalid paths, thereby retaining the key association information in default prediction, effectively balancing the model's transferability and noise interference, and at the same time hierarchically aggregating external risk information onto nodes, enhancing the interpretability of risk propagation. The latter is committed to capturing global implicit association information to make up for the information loss caused by the limited length of meta-paths. In addition, to better fuse and utilize different risk information to achieve accurate default prediction, the present invention designs a risk assessment method based on a capsule network.

[0091] Finally, the method of the present invention is only a preferred embodiment and is not used to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting corporate default based on meta-path denoising and capsule network modeling, characterized in that: The method comprises: Collect the internal business information and external related information of the target enterprise, and build the enterprise's heterogeneous information network and meta-network based on the enterprise's external related information; wherein the nodes in the heterogeneous information network include: enterprise nodes, industry nodes, regional nodes and stakeholder nodes, and the meta-network is generated based on the abstraction of the heterogeneous information network; Input the internal indicator information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise; Input the target enterprise's heterogeneous information network and meta-network into the meta-path-based explicit external default risk encoder to obtain the target enterprise's explicit external default risk embedding; Input the target enterprise's heterogeneous information network into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the target enterprise; Internal default risk embedding, explicit external default risk embedding and implicit external default risk embedding are integrated with capsule network to obtain the default prediction results of the target enterprise.

2. The method according to claim 1, characterized in that The step of inputting the internal indicator information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise includes: Extracting numerical features and / or discrete features of internal indicator information, standardizing and normalizing the numerical features, and converting the discrete features into embedded vectors; wherein the numerical features include: registered capital, paid-in capital, registration time, and litigation duration; and the discrete features include: litigation cause, court level, and litigation result; The embedding vector of the discrete feature is concatenated with the processed numerical feature and mapped through a linear layer to obtain the internal default risk embedding of the target enterprise.

3. The method according to claim 1, characterized in that The step of inputting the heterogeneous information network and meta-network of the enterprise into the meta-path-based explicit external default risk encoder to obtain the explicit external default risk embedding of the target enterprise includes: Based on the meta-network, walking in the heterogeneous information network, a path instance of all metadata types not exceeding a fixed length is generated for the target enterprise; Calculate the effective value S for each path type P P ; According to the effective value S P Sort the meta-path types and select the top K meta-path types; Among the top K ranked meta-path types, weighted aggregation is performed on the path instances of meta-path type P to obtain the semantic embedding Z under meta-path type P. P ; Use attention mechanism to fuse semantic embedding Z P , and obtain the explicit external default risk embedding Z of the target.

4. The method according to claim 3, characterized in that: Valid values ​​for the meta path Among them, T P represents the total number of all path instances of meta-path type P, It represents the number of path instances in which both end nodes of the meta-path type P are defaulting enterprises. Indicates the number of path instances in which both end nodes of meta-path type P are contract-abiding enterprises. represents the number of path instances in meta-path type P whose target node is abiding enterprise and whose source node is breaching enterprise, Represents the number of path instances in the meta-path type P whose target node is the defaulting enterprise and whose source node is the abiding enterprise.

5. The method according to claim 1, characterized in that: The step of inputting the heterogeneous information network of the target enterprise into an implicit external default risk encoder based on a heterogeneous graph to obtain an implicit external default risk embedding of the target enterprise includes: Initialize node embeddings of all nodes in a heterogeneous information network Using the lth layer of the graph neural network, we calculate the embedding of aggregated neighbor information of each node under different external association relations π And based on node embedding and the embedding of aggregated neighbor information under all external association relations π Get the risk model embedding node embedding of the node Wherein, the graph neural network has a total of L layers; Splicing node embedding Get the default risk pattern embedding matrix P of all nodes (L) ; Embedding matrix P from default risk patterns of all nodes (L) Extract the risk pattern embedding matrix of all enterprise nodes Embedding matrix for risk model Perform global attention calculation to obtain the enterprise embedding matrix of each enterprise node Based on the enterprise embeddedness matrix of the target enterprise and the risk pattern embedding matrix Get the implicit external default risk embedding of the target enterprise 6. The method according to claim 1, characterized in that By integrating internal default risk embedding, explicit external default risk embedding and implicit external default risk embedding, and combining with capsule network, the default prediction results of the target enterprise are obtained, including: The internal default risk embedding, the explicit external default risk embedding and the implicit external default risk embedding are integrated to obtain the integrated embedding; Based on fusion embedding, the dynamic routing mechanism of capsule network is used to generate embedding o for the target enterprise in state s. s ; Wherein, state s = {r, nr}, r represents the default risk state, and nr represents the compliance risk-free state; Calculate the default risk score of the target enterprise r ‖ and compliance risk-free score‖o nr ‖; where ‖·‖ represents the modulo length operation; Based on the default risk score r ‖ and compliance risk-free score‖o r ‖, calculate the default probability of the target enterprise The probability of default score By comparing with a set threshold, the default prediction result of the target enterprise is obtained.

7. The method according to claim 6, characterized in that Based on fusion embedding, the dynamic routing mechanism of capsule network is used to generate embedding o for the target enterprise in state s. s ,include: Extract the feature embedding u of the kth feature from the fused embedding k ; Based on the feature embedding, obtain the feature embedding t of the kth feature extracted by the state s capsule from the fusion embedding s,k ; Based on the feature embedding t s,k , using dynamic routing technology to fuse different feature embeddings to obtain the embedding t of the target enterprise in state s s ; Based on embedding s , generate the embedding o under the target enterprise state s s .

8. An enterprise default prediction system based on meta-path denoising and capsule network modeling, characterized in that: The system comprises: The data collection module is used to collect the internal business information and external related information of the target enterprise, and to construct the heterogeneous information network and meta-network of the enterprise based on the external related information of the enterprise; wherein the nodes in the heterogeneous information network include: enterprise nodes, industry nodes, regional nodes and stakeholder nodes, and the meta-network is generated based on the abstraction of the heterogeneous information network; The first encoding module is used to input the internal indicator information of the target enterprise into the internal default risk encoder to obtain the internal default risk embedding of the target enterprise; The second encoding module is used to input the heterogeneous information network and meta-network of the target enterprise into the explicit external default risk encoder based on the meta-path to obtain the explicit external default risk embedding of the target enterprise; The third encoding module is used to input the heterogeneous information network of the target enterprise into the implicit external default risk encoder based on the heterogeneous graph to obtain the implicit external default risk embedding of the target enterprise; The default prediction module is used to integrate internal default risk embedding, explicit external default risk embedding and implicit external default risk embedding, and combine with capsule network to obtain the default prediction result of the target enterprise.

9. An electronic device, characterized in that: The electronic device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the enterprise default prediction method based on meta-path denoising and capsule network modeling as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by the processor, the enterprise default prediction method based on meta-path denoising and capsule network modeling as described in any one of claims 1 to 7 is implemented.

Images