A method and device for identifying a risk enterprise and a storage medium

Abstract

Embodiments of the present application disclose a risk enterprise identification method and device and a storage medium, and relate to the field of information security. The method comprises: obtaining a target data set of each enterprise, the target data set comprising a plurality of tax information of the enterprise; constructing an enterprise relationship graph according to the plurality of tax information; pre-training the enterprise relationship graph by using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain a pre-embedding representation of an enterprise node; training the pre-embedding representation by using a graph autoencoder (GAE) to obtain a final embedding representation of the enterprise node; and calculating an enterprise similarity of the enterprise by using a cosine similarity algorithm according to the final embedding representation of the enterprise node, the enterprise similarity being used to identify whether the enterprise is a risk enterprise. The present application realizes potential risk enterprise key monitoring, timely early warning and risk response.

Description

Technical Field

[0001] This application relates to the field of information security, and in particular to a method, apparatus and storage medium for identifying risky enterprises. Background Technology

[0002] Currently, identifying risky enterprises is a crucial aspect of building smart tax systems. By integrating tax data from multiple dimensions, deeply mining big data on taxation, and leveraging machine learning to make the data speak for itself, preventing tax risks has become a key focus and urgent issue in the tax field.

[0003] With the rapid development of modern economy and society, complex relationships have formed between enterprises due to transactions, guarantees, shareholdings, and personnel appointments. If an individual enterprise is identified as a "blacklist" enterprise in a major tax violation case determined by the tax authorities, other related enterprises may also face certain risks, requiring close monitoring of these other enterprises and timely risk response.

[0004] Modeling enterprise similarity allows for chain-like processing of tax data from multiple enterprises, transforming seemingly worthless data into valuable data assets through association and analysis. Further model analysis can then generate increasingly clear enterprise profiles and group characteristics.

[0005] However, existing studies often rely on single data sources and simplistic methods for calculating enterprise similarity, tending to overemphasize local data while neglecting the overall picture. These methods analyze only tax data directly related to the target enterprise, ignoring tax data from other related enterprises and taxpayers, and failing to consider deeper, multi-source relationships between enterprises. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method, apparatus and storage medium for identifying risky enterprises, in order to solve the problem that existing enterprise similarity calculation methods rely on a single data source and analyze enterprise tax data in isolation, thus ignoring the deep-seated relationships between enterprises.

[0007] In a first aspect, embodiments of the present invention provide a method for identifying risky enterprises, the method comprising:

[0008] Obtain the target dataset for each enterprise, which includes various tax information of the enterprise;

[0009] Based on the aforementioned various tax information, a corporate relationship graph is constructed;

[0010] The enterprise relationship graph is pre-trained using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes.

[0011] The pre-embedded representation is trained using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node.

[0012] Based on the final embedded representation of the enterprise node, the enterprise similarity is calculated using a cosine similarity algorithm. This enterprise similarity is used to identify whether an enterprise is a risky enterprise.

[0013] Optionally, the method further includes:

[0014] Receive a search instruction input by the user, the search instruction including the first company name of the risky enterprise;

[0015] Based on the first company name, obtain the similarity of the company matching the first company name and the second company name of the suspected risky company associated with the company similarity;

[0016] The second company name is output to the user.

[0017] Optionally, constructing a corporate relationship graph based on the various tax information specifically includes:

[0018] The various types of tax information are cleaned and processed to obtain valid tax information;

[0019] The enterprise is used as a graph node, and the feature information of the graph node is obtained from the valid tax information;

[0020] The valid tax information is analyzed to obtain the relationship types between the graph nodes, and the relationship types are used as the association edges between multiple graph nodes.

[0021] Calculate the weights of the associated edges;

[0022] The enterprise relationship graph is constructed based on the graph nodes, the feature information of the graph nodes, the associated edges, and the weights of the associated edges.

[0023] Optionally, calculating the weight of the associated edge specifically includes:

[0024] Based on the analytic hierarchy process (AHP) algorithm and historical experience, the associated edges are compared pairwise according to their importance to form a judgment matrix.

[0025] A consistency check is performed on the judgment matrix to obtain the initial weights of the associated edges;

[0026] The product of the initial weight and the correlation coefficient is used as the final weight of the associated edge;

[0027] The correlation coefficient is obtained by the following formula: k = m / N; m is the number of transactions between the first enterprise and the designated enterprise; N is the total number of transactions between the first enterprise and other enterprises besides the first enterprise.

[0028] Optionally, the step of pre-training the enterprise relationship graph using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes specifically includes:

[0029] Positive sample instances are obtained by using the first feature matrix composed of the feature information of the graph nodes and the first adjacency matrix composed of the relationships between the graph nodes;

[0030] Using the first adjacency matrix, a second adjacency matrix is ​​obtained, and the first adjacency matrix is ​​the same as the second adjacency matrix;

[0031] Construct an erosion function, and use the erosion function to shuffle and rearrange the first feature matrix to randomly obtain the second feature matrix;

[0032] Negative sample instances are obtained using the second feature matrix and the second adjacency matrix;

[0033] The first feature matrix and the first adjacency matrix of the positive sample instance are input into the GCN, which serves as an encoder, to obtain the first local features of the positive sample instance.

[0034] The second feature matrix and the second adjacency matrix of the negative sample instance are input into the GCN, which serves as the encoder, to obtain the second local features of the negative sample instance.

[0035] The first local feature is input into the readout function to obtain the global feature at the graph level. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs.

[0036] A discriminator is constructed to score the local-global pairs of the positive samples to obtain a first score for the local-global pairs of the positive samples, and to score the local-global pairs of the negative samples to obtain a second score for the local-global pairs of the negative samples.

[0037] The first score is compared with a vector of all 1s, and the difference between the first score and the vector of all 1s is used as the first loss; the second score is compared with a vector of all 0s, and the difference between the second score and the vector of all 0s is used as the second loss. The sum of the first loss and the second loss is the value of the noise contrast objective function.

[0038] Based on gradient descent, the noise contrast objective function is minimized, and the parameters of the encoder and the readout function are updated. The noise contrast objective function includes the discriminator. The noise contrast objective function is used to make the discriminator score the local-global pairs of positive samples and the local-global pairs of negative samples, so that the first score is closer to the vector of all 1s and the second score is closer to the vector of all 0s, thereby widening the gap between the first score and the second score.

[0039] The process involves repeatedly executing steps to obtain positive sample instances using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes. The steps are based on gradient descent to minimize the noise contrast objective function and update all steps between the parameters of the encoder and the readout function until a preset condition is met.

[0040] The model that minimizes the noise contrastive objective function during training is taken as the optimal model.

[0041] Using the optimal model, the first local feature obtained after inputting the positive sample instance into the GCN is used as the pre-embedded representation;

[0042] The preset conditions include a preset number of training sessions and a preset number of times the accuracy is maintained.

[0043] Optionally, the step of training the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node specifically includes:

[0044] The pre-embedded representation is input into the GAE, and the latent representation of the graph node is obtained through the graph convolutional encoder;

[0045] Based on the latent representations, the enterprise relationship graph is reconstructed using a decoder to obtain a reconstructed graph;

[0046] Using cross-entropy as the loss function, and taking the enterprise relationship graph and the reconstructed graph as input, we obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph.

[0047] Based on gradient descent, the loss function is minimized, and the parameters of the encoder are updated.

[0048] The process involves repeatedly inputting the pre-embedded representation into the GAE, obtaining the latent representation of the graph node through a graph convolutional encoder, minimizing the loss function based on gradient descent, and updating all steps between the parameters of the encoder until the preset condition is met.

[0049] The model that minimizes the loss function during training is taken as the optimal model.

[0050] The optimal model is used to input the pre-embedded representation into the GAE to obtain the latent features as the final embedded representation.

[0051] Secondly, embodiments of the present invention provide a risk enterprise identification device, the device comprising:

[0052] The first acquisition unit is used to acquire a target dataset for each enterprise, the target dataset including various tax information of the enterprise;

[0053] A construction unit is used to construct an enterprise relationship graph based on the various tax information.

[0054] The pre-training unit is used to pre-train the enterprise relationship graph using graph convolutional neural network (GCN) and mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes.

[0055] The training unit is used to train the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node.

[0056] The calculation unit is used to calculate the enterprise similarity of the enterprise based on the final embedded representation of the enterprise node using a cosine similarity algorithm. The enterprise similarity is used to identify whether an enterprise is a risky enterprise.

[0057] Optionally, the device further includes:

[0058] A receiving unit is configured to receive a search instruction input by a user, wherein the search instruction includes the first company name of the risky enterprise;

[0059] The second acquisition unit is used to acquire, based on the first enterprise name, the enterprise similarity matching the first enterprise name and the second enterprise name of the suspected risk enterprise associated with the enterprise similarity;

[0060] The output unit is used to output the second company name to the user.

[0061] The output unit is used to output the second company name to the user.

[0062] Optionally, the construction unit is specifically used to perform data cleaning processing on the various types of tax information to obtain valid tax information;

[0063] The enterprise is used as a graph node, and the feature information of the graph node is obtained from the valid tax information;

[0064] The valid tax information is analyzed to obtain the relationship types between the graph nodes, and the relationship types are used as the association edges between multiple graph nodes.

[0065] Calculate the weights of the associated edges;

[0066] The enterprise relationship graph is constructed based on the graph nodes, the feature information of the graph nodes, the associated edges, and the weights of the associated edges.

[0067] Optionally, the building unit is further specifically used to, based on the hierarchical analysis algorithm and according to historical experience, compare the associated edges pairwise according to their importance to form a judgment matrix;

[0068] A consistency check is performed on the judgment matrix to obtain the initial weights of the associated edges;

[0069] The product of the initial weight and the correlation coefficient is used as the final weight of the associated edge;

[0070] The correlation coefficient is obtained by the following formula: k = m / N; m is the number of transactions between the first enterprise and the designated enterprise; N is the total number of transactions between the first enterprise and other enterprises besides the first enterprise.

[0071] Optionally, the pre-training unit is specifically used to obtain positive sample instances by using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes;

[0072] Using the first adjacency matrix, a second adjacency matrix is ​​obtained, and the first adjacency matrix is ​​the same as the second adjacency matrix;

[0073] Construct an erosion function, and use the erosion function to shuffle and rearrange the first feature matrix to randomly obtain the second feature matrix;

[0074] Negative sample instances are obtained using the second feature matrix and the second adjacency matrix;

[0075] The first feature matrix and the first adjacency matrix of the positive sample instance are input into the GCN, which serves as an encoder, to obtain the first local features of the positive sample instance.

[0076] The second feature matrix and the second adjacency matrix of the negative sample instance are input into the GCN, which serves as the encoder, to obtain the second local features of the negative sample instance.

[0077] The first local feature is input into the readout function to obtain the global feature at the graph level. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs.

[0078] A discriminator is constructed to score the local-global pairs of the positive samples to obtain a first score for the local-global pairs of the positive samples, and to score the local-global pairs of the negative samples to obtain a second score for the local-global pairs of the negative samples.

[0079] The first score is compared with a vector of all 1s, and the difference between the first score and the vector of all 1s is used as the first loss; the second score is compared with a vector of all 0s, and the difference between the second score and the vector of all 0s is used as the second loss. The sum of the first loss and the second loss is the value of the noise contrast objective function.

[0080] Based on gradient descent, the noise contrast objective function is minimized, and the parameters of the encoder and the readout function are updated. The noise contrast objective function includes the discriminator. The noise contrast objective function is used to make the discriminator score the local-global pairs of positive samples and the local-global pairs of negative samples, so that the first score is closer to the vector of all 1s and the second score is closer to the vector of all 0s, thereby widening the gap between the first score and the second score.

[0081] The process involves repeatedly executing steps to obtain positive sample instances using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes. The steps are based on gradient descent to minimize the noise contrast objective function and update all steps between the parameters of the encoder and the readout function until a preset condition is met.

[0082] The model that minimizes the noise contrastive objective function during training is taken as the optimal model.

[0083] Using the optimal model, the first local feature obtained after inputting the positive sample instance into the GCN is used as the pre-embedded representation;

[0084] The preset conditions include a preset number of training sessions and a preset number of times the accuracy is maintained.

[0085] Optionally, the training unit is specifically used to input the pre-embedded representation into the GAE and obtain the latent representation of the graph node through a graph convolutional encoder;

[0086] Based on the latent representations, the enterprise relationship graph is reconstructed using a decoder to obtain a reconstructed graph;

[0087] Using cross-entropy as the loss function, and taking the enterprise relationship graph and the reconstructed graph as input, we obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph.

[0088] Based on gradient descent, the loss function is minimized, and the parameters of the encoder are updated.

[0089] The process involves repeatedly inputting the pre-embedded representation into the GAE, obtaining the latent representation of the graph node through a graph convolutional encoder, minimizing the loss function based on gradient descent, and updating all steps between the parameters of the encoder until the preset condition is met.

[0090] The model that minimizes the loss function during training is taken as the optimal model.

[0091] The optimal model is used to input the pre-embedded representation into the GAE to obtain the latent features as the final embedded representation.

[0092] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising: a housing, a processor, a memory, a circuit board, and a power supply circuit, wherein the circuit board is disposed within the space enclosed by the housing, and the processor and the memory are disposed on the circuit board; the power supply circuit is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, for executing the risk enterprise identification method described in the first aspect above.

[0093] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the risk enterprise identification method described in the first aspect above.

[0094] This invention provides a method, apparatus, and storage medium for identifying risky enterprises. After acquiring the target dataset for each enterprise, an enterprise relationship graph is constructed. The enterprise relationship graph is pre-trained using GCN and DGI to obtain pre-embedded representations of enterprise nodes. Then, GAE is used to train the pre-embedded representations to obtain the final embedded representations of the enterprise nodes. Using a cosine similarity algorithm, the enterprise similarity is calculated based on the final embedded representations, and finally, the enterprise similarity is used to identify whether an enterprise is a risky enterprise.

[0095] The aforementioned solution addresses the problems of existing enterprise similarity calculation methods, which rely on a single data source and analyze enterprise tax data in isolation, neglecting deeper relationships between enterprises. It enables focused monitoring, timely early warning, and risk response for enterprises with potential risks. Attached Figure Description

[0096] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0097] Figure 1 A flowchart illustrating a method for identifying risky enterprises provided in an embodiment of the present invention;

[0098] Figure 2 A schematic diagram of the embedded representation learning model for enterprise nodes provided in an embodiment of the present invention;

[0099] Figure 3 This is a schematic diagram of a risk enterprise identification device provided in an embodiment of the present invention;

[0100] Figure 4 This is a schematic diagram of an electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0101] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0102] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0103] The following is in conjunction with the appendix Figure 1 The solutions provided in the embodiments of the present invention will be described in detail below. Figure 1This is a flowchart illustrating a method for identifying risky enterprises according to an embodiment of the present invention. In this embodiment, the implementing entity is an electronic device. This electronic device can be a terminal device, such as a personal computer or desktop computer. It can also be a server. Figure 1 As shown, the method for identifying risky enterprises provided in this embodiment of the invention specifically includes the following steps:

[0104] Step 110: Obtain the target dataset for each enterprise, which includes various tax information of the enterprise.

[0105] In this embodiment of the invention, the electronic device obtains target datasets of multiple enterprises from government departments (e.g., tax departments), and each target dataset includes various tax information of the enterprise.

[0106] For example, tax information includes taxpayer information data, value-added tax invoice data, and enterprise registration data. Among them, taxpayer information data mainly includes taxpayer name, taxpayer identification number, legal representative name, legal person certificate number, financial officer name, financial officer certificate number, tax agent name, tax agent certificate number, production and operation address, business registration date, enterprise size, industry, and other information.

[0107] Value-added tax (VAT) invoice data mainly includes information such as the buyer's taxpayer identification number, the seller's taxpayer identification number, the seller's name, and the buyer's name.

[0108] Enterprise business registration data mainly includes information such as industry, unified social credit code, taxpayer identification number, registered capital, and legal representative.

[0109] Step 120: Construct an enterprise relationship graph based on the various tax information.

[0110] In this embodiment of the invention, after the electronic device acquires various types of tax information, it cleans and mines the various types of tax information to construct an enterprise relationship graph.

[0111] Furthermore, the enterprise relationship graph can be constructed through the following process. First, various types of tax information are cleaned to obtain valid tax information. Then, graph nodes, graph node features, and the associated edges between graph nodes are constructed, and the weights of the associated edges are calculated. Finally, the enterprise relationship graph is constructed using the graph nodes, graph node features, and associated edges. The process of constructing the enterprise relationship graph will be described in detail in subsequent embodiments; here, it is only briefly described.

[0112] Understandably, the enterprise relationship graph is a graph network.

[0113] Step 130: Using Graph Convolutional Neural Network (GCN) and Mutual Information Maximization (DGI) model, the enterprise relationship graph is pre-trained to obtain pre-embedded representations of enterprise nodes.

[0114] In this embodiment of the invention, after the electronic device constructs an enterprise relationship graph, it uses a graph convolutional neural network (GCN) and a deep graph Infomax (DGI) model to pre-train the enterprise relationship graph to obtain a pre-embedded representation of the enterprise nodes.

[0115] Furthermore, the pre-embedded representation of the enterprise node can be obtained through the following process. First, positive sample instances and negative sample instances are obtained through the enterprise relationship graph. Then, the positive sample instances and negative sample instances are respectively input into the GCN to obtain the first local feature and the second local feature. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs. The discriminator scores the positive sample local-global pairs and the negative sample local-global pairs respectively. The difference between the first score and the vector with all 1s is used as the first loss, and the difference between the second score and the vector with all 0s is used as the second loss. The sum of the first loss and the second loss is the value of the noise contrastive objective function. Finally, the above steps are repeated multiple times until a preset condition is met. The model corresponding to the minimum value of the noise contrastive objective function during training is taken as the optimal model. The first local feature obtained after inputting the positive sample instance into the GCN using the optimal model is used as the pre-embedded representation. The process of obtaining the pre-embedded representation of the enterprise node will be described in detail in subsequent embodiments, and only a brief description is given here.

[0116] Step 140: Train the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node.

[0117] In this embodiment of the invention, after the electronic device obtains the pre-embedded representation of the enterprise node, it uses a graph auto-encoder (GAE) to train the pre-embedded representation to obtain the final embedded representation of the enterprise node.

[0118] Furthermore, the final embedded representation of the enterprise node can be obtained through the following process. First, the pre-embedded representation is input into the GAE, and the latent representation of the graph node is obtained through a graph convolutional encoder. Then, based on the representation, the enterprise relationship graph is reconstructed through a decoder to obtain a reconstructed graph. Using cross-entropy as the loss function, the enterprise relationship graph and the reconstructed graph are used as inputs to obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph. Finally, the above steps are repeated multiple times until a preset condition is met. The model corresponding to the minimum value of the loss function during training is taken as the optimal model. The latent features obtained by inputting the pre-embedded representation into the GAE through the optimal model are taken as the final embedded representation. The process of obtaining the final embedded representation of the enterprise node will be described in detail in subsequent embodiments; only a brief overview is provided here.

[0119] Step 150: Based on the final embedded representation of the enterprise node, calculate the enterprise similarity using the cosine similarity algorithm. The enterprise similarity is used to identify whether an enterprise is a risky enterprise.

[0120] In this embodiment of the invention, after the electronic device obtains the final embedded representation of the enterprise node, it calculates the enterprise similarity using a cosine similarity algorithm. This enterprise similarity is used to identify whether an enterprise is a high-risk enterprise.

[0121] It is understandable that the cosine similarity algorithm is a well-known existing algorithm, and the specific calculation process will not be explained in detail here.

[0122] In this embodiment of the invention, after acquiring the target dataset for each enterprise, the electronic device constructs an enterprise relationship graph. The enterprise relationship graph is pre-trained using GCN and DGI to obtain pre-embedded representations of enterprise nodes. Then, GAE is used to train the pre-embedded representations to obtain the final embedded representations of the enterprise nodes. Using a cosine similarity algorithm, the enterprise similarity is calculated based on the final embedded representations, and finally, the enterprise similarity is used to identify whether an enterprise is a high-risk enterprise.

[0123] The aforementioned solution addresses the problems of existing enterprise similarity calculation methods, which rely on a single data source and analyze enterprise tax data in isolation, neglecting deeper relationships between enterprises. It enables focused monitoring, timely early warning, and risk response for enterprises with potential risks.

[0124] Optionally, in this embodiment of the invention, the process of an electronic device receiving a search instruction input by a user and outputting suspected risky enterprises based on the search instruction is also included.

[0125] Specifically, the user pre-obtains the names of risky companies that are associated with or similar to the "blacklisted (risky) companies". The user enters a search command into the electronic device, which includes the first name of the risky company.

[0126] After receiving a search command, the electronic device parses the command and retrieves the first company name. Based on the first company name, the electronic device obtains the similarity of companies matching the first company name, and based on the company similarity, retrieves the second company name of a suspected risky company that is highly associated with the company (e.g., companies with the same company similarity, companies with an associated edge, or companies with a similar relationship type).

[0127] Understandably, there must be at least one second company name.

[0128] The electronic device outputs a second company name to the user (which can be displayed on a monitor) so that the customer can obtain the second company name.

[0129] The process of constructing an enterprise relationship graph based on various tax information in the embodiments of the present invention is described in detail below.

[0130] Specifically, after acquiring various types of tax information, electronic devices clean the information, that is, they check the information and remove "dirty" data such as missing data, abnormal data, and erroneous data to obtain valid tax information.

[0131] Electronic devices construct graph nodes, graph node features, and the edges connecting graph nodes. Typically, enterprises and individuals can be used as graph nodes, and relationships between enterprises, between enterprises and individuals, and between individuals can be used as edges connecting graph nodes. However, considering the sparsity of edges and data redundancy, in this embodiment of the invention, the electronic device uses enterprises as graph nodes and obtains feature information of graph nodes from valid tax information. Simultaneously, the electronic device performs association analysis on the valid tax information to obtain the types of association relationships between graph nodes, and uses these types of association relationships as edges connecting multiple graph nodes.

[0132] For example, electronic devices convert relationships between enterprises and individuals, and between individuals, into association types such as "same legal representative," "same financial officer," "same tax agent," and "legal representative and financial officer are the same person." The electronic devices treat these association types as edges between graph nodes. They also use registered capital, business registration date, taxpayer status code, regional code, and industry code as feature information for graph nodes, encoding this feature information using one-hot codes.

[0133] The electronic device calculates the weights of associated edges. Based on the analytic hierarchy process (AHP) and historical experience (e.g., expert experience), it compares associated edges pairwise according to their importance, forming a judgment matrix. The electronic device performs a consistency check on the judgment matrix to obtain the initial weights w of the associated edges. i .

[0134] Furthermore, considering the issue that different companies may have the same type of relationship but different degrees of connection. For example, Company A may have transaction relationships with both Company B and Company C, but Company A's transactions with Company B are relatively more frequent. Therefore, the degree of connection between Company A and Company B should be higher than the degree of connection between Company A and Company C. Based on this, the electronic device will calculate the product of the initial weight and the degree of connection coefficient k as k×w. i As the final weight of the associated edge.

[0135] For example, if the total number of transactions between company A and other companies (excluding company A) is N, and company A has m transactions with company B (the designated company), then the correlation coefficient between graph node A and graph node B is k = m / N.

[0136] It should be noted that if there are multiple types of relationships between two graph nodes, the one with the largest weight value among the multiple weight values ​​of the related edge will be used as the final weight.

[0137] Once the electronic device obtains the graph nodes, their feature information, associated edges, and their weights, it can complete the construction of the enterprise relationship graph weighted graph.

[0138] The following details the process by which the electronic device in this embodiment of the invention utilizes GCN and DGI to pre-train the enterprise relationship graph, obtaining pre-embedded representations of enterprise nodes, and then uses GAE to train these pre-embedded representations to obtain the final embedded representations of the enterprise nodes. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of the embedded representation learning model for enterprise nodes provided in an embodiment of the present invention.

[0139] exist Figure 2 In the process, electronic devices will transfer the feature information of the map nodes. The first feature matrix is ​​formed by the relationships between the graph nodes in the electronic device. Where N is the number of nodes in the enterprise relationship graph. The feature representation of node i. Let F represent the entire real number field, and let F represent the feature information of length F, that is, the length of the node embedding feature.

[0140] The electronic device obtains a positive sample instance (X,A) using the first feature matrix and the first adjacency matrix. Since the enterprise relationship graph is a weighted graph, the values ​​stored in the first adjacency matrix are no longer just 0 or 1.

[0141] The electronic device keeps the first adjacency matrix unchanged and obtains the second adjacency matrix, that is, the second adjacency matrix is ​​the same as the first adjacency matrix. Electronic device construction erosion function Using an erosion function and a row-wise shuffle method, the electronic device shuffles the first feature matrix and rearranges it row by row to randomly obtain the second feature matrix. Where N and M have the same value, the corrosion function Indicates: Input Then, by using the erosion function, the feature matrix is ​​changed to obtain...

[0142] The electronic device obtains negative sample instances using the second feature matrix and the second adjacency matrix. Use GCN as the encoder The electronic device inputs the first feature matrix X and the first adjacency matrix A into the GCN. The encoder learns features by continuously aggregating the neighbors around the graph nodes, thereby obtaining the first local features of positive sample instances. Here, F′ indicates that the node embedding feature information is still of length F, but the feature information of the graph node has changed.

[0143] Similarly, the electronic device will use the second feature matrix and the second adjacency matrix The input is fed into the GCN. The encoder learns features by continuously aggregating the neighbors of the graph nodes, thereby obtaining the second local features of the negative sample instances.

[0144] The electronic device inputs the first local feature into the readout function. This yields global features at the graph level. At the same time, the electronic device will have the first local feature With global features As a local-global pair of positive samples The second local feature With global features As a negative sample local-global pair

[0145] Electronic device discriminator To each Scoring is performed to obtain the first score for the local-global pair of positive samples and the second score for the local-global pair of negative samples. The discriminator takes the feature representation of the F length of the two graph nodes as input and outputs a probability score.

[0146] The electronic device compares the first score with a vector of all 1s and uses the difference between the first score and the vector of all 1s as the first loss; the electronic device compares the second score with a vector of all 0s and uses the difference between the second score and the vector of all 0s as the second loss. The sum of the first loss and the second loss is the value of the noise contrast objective function.

[0147] Based on gradient descent, the electronic device minimizes the noise contrastive objective function and updates the parameters of the encoder and readout function. The noise contrastive objective function includes a discriminator, which, when scoring positive and negative local-global pairs, aims to make the first score closer to a vector of all 1s and the second score closer to a vector of all 0s, thereby widening the gap between the first and second scores.

[0148] The objective function for a noise-contrast model with standard binary cross-entropy loss is:

[0149]

[0150] in, It represents expectations.

[0151] The electronic device repeatedly executes the aforementioned steps (from obtaining positive sample instances through the first feature matrix composed of the feature information of the graph nodes and the first adjacency matrix composed of the relationships between the graph nodes, to minimizing the noise contrast objective function based on gradient descent, and updating all the parameters included in the encoder and readout function) until the preset conditions are met.

[0152] The electronic device selects the model that minimizes the noise-contrast objective function during training as the optimal model. Using the optimal model, the electronic device uses the first local features obtained after inputting positive sample instances into the GCN as the pre-embedded representation (H,A) of the enterprise node.

[0153] The preset conditions include the preset number of training sessions and the preset number of times the accuracy is maintained.

[0154] In one example, the number of training iterations specifically refers to the number of times the model is trained. For instance, the model is trained 100 times, and training stops after reaching 100 iterations.

[0155] In another example, the number of times accuracy is maintained specifically refers to training the model once with test data after each training iteration to obtain an accuracy rate. If the accuracy rate remains consistent for 10 consecutive iterations, then the model stops training.

[0156] The electronic device takes the pre-embedded representation (H,A) as input to the GAE and obtains the latent representation Z of the graph nodes through a graph convolutional encoder. Based on the latent representation Z, the electronic device reconstructs the enterprise relationship graph through a decoder to obtain the reconstructed graph.

[0157] The electronic device uses cross-entropy as the loss function, taking the enterprise relationship graph and the reconstructed graph as input, to obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph. Based on gradient descent, the loss function is minimized, and the parameters of the encoder are updated.

[0158] The electronic device repeatedly performs the aforementioned steps (inputting the pre-embedded representation into the GAE, obtaining the latent representation of the graph nodes through the graph convolutional encoder, minimizing the loss function based on gradient descent, and updating the parameters of the encoder) until the preset conditions are met.

[0159] The electronic device selects the model that minimizes the loss function during training as the optimal model. Using the optimal model, the electronic device inputs the pre-embedded representation into the latent features obtained after training as the final embedded representation.

[0160] It is understood that the preset conditions are the same as those described above, and will not be repeated here.

[0161] In this embodiment of the invention, considering the deep-seated multi-source relationships between enterprises, tax data is integrated from multiple dimensions to deeply explore the diverse relationships between enterprises, such as upstream and downstream, sales and purchases, common legal persons and common related taxpayers, and to construct an enterprise relationship graph, which can more accurately identify enterprises with potential tax risks.

[0162] Furthermore, the risk enterprise identification method provided in this embodiment of the invention is mainly applied to the scenario of identifying tax risk enterprises, but it can also be used in other business scenarios such as social networking and e-commerce.

[0163] Figure 3 This is a schematic diagram of a risk enterprise identification device provided in an embodiment of the present invention, such as... Figure 3 As shown, the risk enterprise identification device in this embodiment may include: a first acquisition unit 310, a construction unit 320, a pre-training unit 330, a training unit 340, and a calculation unit 350.

[0164] The first acquisition unit 310 is used to acquire a target dataset for each enterprise, the target dataset including various tax information of the enterprise;

[0165] Construction unit 320 is used to construct an enterprise relationship graph based on the various tax information;

[0166] The pre-training unit 330 is used to pre-train the enterprise relationship graph using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain a pre-embedded representation of the enterprise nodes.

[0167] Training unit 340 is used to train the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node.

[0168] The calculation unit 350 is used to calculate the enterprise similarity of the enterprise based on the final embedded representation of the enterprise node using a cosine similarity algorithm. The enterprise similarity is used to identify whether the enterprise is a risky enterprise.

[0169] Optionally, the device further includes: a receiving unit (not shown in the figure) for receiving a search instruction input by a user, the search instruction including the first company name of the risky enterprise;

[0170] The second acquisition unit (not shown in the figure) is used to acquire, based on the first enterprise name, the enterprise similarity matching the first enterprise name and the second enterprise name of the suspected risk enterprise associated with the enterprise similarity;

[0171] An output unit (not shown in the figure) is used to output the second company name to the user.

[0172] Optionally, the construction unit 320 is specifically used to perform data cleaning processing on the various types of tax information to obtain valid tax information;

[0173] The enterprise is used as a graph node, and the feature information of the graph node is obtained from the valid tax information;

[0174] The valid tax information is analyzed to obtain the relationship types between the graph nodes, and the relationship types are used as the association edges between multiple graph nodes.

[0175] Calculate the weights of the associated edges;

[0176] The enterprise relationship graph is constructed based on the graph nodes, the feature information of the graph nodes, the associated edges, and the weights of the associated edges.

[0177] Optionally, the construction unit 320 is further specifically used to, based on the hierarchical analysis algorithm and according to historical experience, compare the associated edges pairwise according to their importance to form a judgment matrix;

[0178] A consistency check is performed on the judgment matrix to obtain the initial weights of the associated edges;

[0179] The product of the initial weight and the correlation coefficient is used as the final weight of the associated edge;

[0180] The correlation coefficient is obtained by the following formula: k = m / N; m is the number of transactions between the first enterprise and the designated enterprise; N is the total number of transactions between the first enterprise and other enterprises besides the first enterprise.

[0181] Optionally, the pre-training unit 330 is specifically used to obtain positive sample instances by using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes;

[0182] Using the first adjacency matrix, a second adjacency matrix is ​​obtained, and the first adjacency matrix is ​​the same as the second adjacency matrix;

[0183] Construct an erosion function, and use the erosion function to shuffle and rearrange the first feature matrix to randomly obtain the second feature matrix;

[0184] Negative sample instances are obtained using the second feature matrix and the second adjacency matrix;

[0185] The first feature matrix and the first adjacency matrix of the positive sample instance are input into the GCN, which serves as an encoder, to obtain the first local features of the positive sample instance.

[0186] The second feature matrix and the second adjacency matrix of the negative sample instance are input into the GCN, which serves as the encoder, to obtain the second local features of the negative sample instance.

[0187] The first local feature is input into the readout function to obtain the global feature at the graph level. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs.

[0188] A discriminator is constructed to score the local-global pairs of the positive samples to obtain a first score for the local-global pairs of the positive samples, and to score the local-global pairs of the negative samples to obtain a second score for the local-global pairs of the negative samples.

[0189] The first score is compared with a vector of all 1s, and the difference between the first score and the vector of all 1s is used as the first loss; the second score is compared with a vector of all 0s, and the difference between the second score and the vector of all 0s is used as the second loss. The sum of the first loss and the second loss is the value of the noise contrast objective function.

[0190] Based on gradient descent, the noise contrast objective function is minimized, and the parameters of the encoder and the readout function are updated. The noise contrast objective function includes the discriminator. The noise contrast objective function is used to make the discriminator score the local-global pairs of positive samples and the local-global pairs of negative samples, so that the first score is closer to the vector of all 1s and the second score is closer to the vector of all 0s, thereby widening the gap between the first score and the second score.

[0191] The process involves repeatedly executing steps to obtain positive sample instances using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes. The steps are based on gradient descent to minimize the noise contrast objective function and update all steps between the parameters of the encoder and the readout function until a preset condition is met.

[0192] The model that minimizes the noise contrastive objective function during training is taken as the optimal model.

[0193] Using the optimal model, the first local feature obtained after inputting the positive sample instance into the GCN is used as the pre-embedded representation;

[0194] The preset conditions include a preset number of training sessions and a preset number of times the accuracy is maintained.

[0195] Optionally, the training unit 340 is specifically used to input the pre-embedded representation into the GAE and obtain the latent representation of the graph node through a graph convolutional encoder;

[0196] Based on the latent representations, the enterprise relationship graph is reconstructed using a decoder to obtain a reconstructed graph;

[0197] Using cross-entropy as the loss function, and taking the enterprise relationship graph and the reconstructed graph as input, we obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph.

[0198] Based on gradient descent, the loss function is minimized, and the parameters of the encoder are updated.

[0199] The process involves repeatedly inputting the pre-embedded representation into the GAE, obtaining the latent representation of the graph node through a graph convolutional encoder, minimizing the loss function based on gradient descent, and updating all steps between the parameters of the encoder until the preset condition is met.

[0200] The model that minimizes the loss function during training is taken as the optimal model.

[0201] The optimal model is used to input the pre-embedded representation into the GAE to obtain the latent features as the final embedded representation.

[0202] The apparatus of this embodiment can be used to perform Figure 1 , Figure 2 The technical solutions of the method embodiments shown are similar in principle and in effect, and will not be described again here.

[0203] Accordingly, the risk enterprise identification device provided in this embodiment of the invention can also be implemented using another structure. Figure 4 This is a schematic diagram of an embodiment of an electronic device provided by the present invention, which can realize the present invention. Figure 1-2 The process of the illustrated embodiment is as follows: Figure 4 As shown, the aforementioned electronic device may include: a housing 41, a processor 42, a memory 43, a circuit board 44, and a power supply circuit 45. The circuit board 44 is disposed within the space enclosed by the housing 41, and the processor 42 and memory 43 are mounted on the circuit board 44. The power supply circuit 45 supplies power to the various circuits or devices of the aforementioned electronic device. The memory 43 stores executable program code. The processor 42 reads the executable program code stored in the memory 43 to run a program corresponding to the executable program code, thereby executing the methods described in the foregoing embodiments.

[0204] For details on the specific execution process of the above steps by processor 42, and the steps further executed by processor 42 through running executable program code, please refer to the present invention. Figure 1-2 The description of the illustrated embodiments will not be repeated here.

[0205] The electronic device is a device that provides computing services. It consists of a processor, hard drive, memory, system bus, etc. The electronic device is similar to the general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0206] Accordingly, embodiments of the present invention provide a computer-readable storage medium storing one or more programs. These programs can be executed by one or more processors to implement the risk enterprise identification method described in the foregoing embodiments.

[0207] It should be noted that, in this document, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0208] The various embodiments in this specification are described in a related manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0209] In particular, the device embodiment is basically similar to the method embodiment, so the description is relatively simple. For relevant details, please refer to the description of the method embodiment.

[0210] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0211] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0212] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0213] In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0214] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0215] For ease of description, the above apparatus is described by dividing it into various functional units / modules. Of course, in implementing this invention, the functions of each unit / module can be implemented in one or more software and / or hardware.

[0216] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

Claims

1. A method for identifying risky enterprises, characterized in that, The method includes: Obtain the target dataset for each enterprise, which includes various tax information of the enterprise; Based on the aforementioned various tax information, a corporate relationship graph is constructed; The enterprise relationship graph is pre-trained using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes. The pre-embedded representation is trained using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node. Based on the final embedded representation of the enterprise node, the enterprise similarity is calculated using a cosine similarity algorithm. The enterprise similarity is used to identify whether an enterprise is a risky enterprise. Specifically, the enterprise relationship graph is pre-trained using a graph convolutional neural network (GCN) and a mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes, including: Positive sample instances are obtained by using a first feature matrix composed of feature information of the enterprise relationship graph nodes and a first adjacency matrix composed of relationships between the graph nodes; a second adjacency matrix is ​​obtained using the first adjacency matrix, the first adjacency matrix being identical to the second adjacency matrix; an erosion function is constructed, and the first feature matrix is ​​shuffled and rearranged using the erosion function to randomly obtain the second feature matrix; negative sample instances are obtained using the second feature matrix and the second adjacency matrix; the first feature matrix and the first adjacency matrix included in the positive sample instances are input into the GCN, which acts as an encoder, to obtain the first local features of the positive sample instances; the second feature matrix and the second adjacency matrix included in the negative sample instances are input into the GCN, which acts as an encoder, to obtain the second local features of the negative sample instances; the first local features are then processed... The local features are input to the readout function to obtain graph-level global features. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs. The autoencoder (GAE) is used as a discriminator to score the positive sample local-global and negative sample local-global pairs respectively. The difference between the first score and a vector with all 1s is used as the first loss, and the difference between the second score and a vector with all 0s is used as the second loss. The sum of the first loss and the second loss is used as the value of the noise contrastive objective function. Finally, the above steps are repeated multiple times until a preset condition is met. The model corresponding to the minimum value of the noise contrastive objective function during training is used as the optimal model. Using the optimal model, the first local feature obtained after inputting positive sample instances into the graph convolutional neural network (GCN) is used as the pre-embedded representation.

2. The method according to claim 1, characterized in that, The method further includes: Receive a search instruction input by the user, the search instruction including the first company name of the risky enterprise; Based on the first company name, obtain the similarity of the company matching the first company name and the second company name of the suspected risky company associated with the company similarity; The second company name is output to the user.

3. The method according to claim 1, characterized in that, The construction of the enterprise relationship graph based on the aforementioned multiple tax information specifically includes: The various types of tax information are cleaned and processed to obtain valid tax information; The enterprise is used as a graph node, and the feature information of the graph node is obtained from the valid tax information; The valid tax information is analyzed to obtain the relationship types between the graph nodes, and the relationship types are used as the association edges between multiple graph nodes. Calculate the weights of the associated edges; The enterprise relationship graph is constructed based on the graph nodes, the feature information of the graph nodes, the associated edges, and the weights of the associated edges.

4. The method according to claim 3, characterized in that, The calculation of the weight of the associated edge specifically includes: Based on the analytic hierarchy process (AHP) algorithm and historical experience, the associated edges are compared pairwise according to their importance to form a judgment matrix. A consistency check is performed on the judgment matrix to obtain the initial weights of the associated edges; The product of the initial weight and the correlation coefficient is used as the final weight of the associated edge; The correlation coefficient is obtained by the following formula: ; N represents the number of transactions between the first company and the designated company; N represents the total number of transactions between the first company and all other companies except the first company.

5. The method according to claim 3, characterized in that, The method involves using the autoencoder (GAE) as a discriminator to score local-global pairs of positive samples and local-global pairs of negative samples. The difference between the first score and a vector consisting entirely of 1s is used as the first loss, and the difference between the second score and a vector consisting entirely of 0s is used as the second loss. The sum of the first and second losses is used as the value of the noise contrastive objective function. Finally, the aforementioned steps are repeated multiple times until a preset condition is met. The model corresponding to the minimum value of the noise contrastive objective function during training is taken as the optimal model. Using the optimal model, the first local features obtained after inputting positive sample instances into the graph convolutional neural network (GCN) are used as the pre-embedded representation, specifically including: A discriminator is constructed to score the local-global pairs of the positive samples to obtain a first score for the local-global pairs of the positive samples, and to score the local-global pairs of the negative samples to obtain a second score for the local-global pairs of the negative samples. The first score is compared with a vector of all 1s, and the difference between the first score and the vector of all 1s is used as the first loss; the second score is compared with a vector of all 0s, and the difference between the second score and the vector of all 0s is used as the second loss. The sum of the first loss and the second loss is the value of the noise contrast objective function. Based on gradient descent, the noise contrast objective function is minimized, and the parameters of the encoder and the readout function are updated. The noise contrast objective function includes the discriminator. The noise contrast objective function is used to make the discriminator score the local-global pairs of positive samples and the local-global pairs of negative samples, so that the first score is closer to the vector of all 1s and the second score is closer to the vector of all 0s, thereby widening the gap between the first score and the second score. The process involves repeatedly executing steps to obtain positive sample instances using a first feature matrix composed of the feature information of the graph nodes and a first adjacency matrix composed of the relationships between the graph nodes. The steps are based on gradient descent to minimize the noise contrast objective function and update all steps between the parameters of the encoder and the readout function until a preset condition is met. The model that minimizes the noise contrastive objective function during training is taken as the optimal model. Using the optimal model, the first local feature obtained after inputting the positive sample instance into the GCN is used as the pre-embedded representation; The preset conditions include a preset number of training sessions and a preset number of times the accuracy is maintained.

6. The method according to claim 5, characterized in that, The step of training the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node specifically includes: The pre-embedded representation is input into the GAE, and the latent representation of the graph node is obtained through the graph convolutional encoder; Based on the latent representations, the enterprise relationship graph is reconstructed using a decoder to obtain a reconstructed graph; Using cross-entropy as the loss function, and taking the enterprise relationship graph and the reconstructed graph as input, we obtain the difference between the first adjacency matrix of the enterprise relationship graph and the third adjacency matrix of the reconstructed graph. Based on gradient descent, the loss function is minimized, and the parameters of the encoder are updated. The process involves repeatedly inputting the pre-embedded representation into the GAE, obtaining the latent representation of the graph node through a graph convolutional encoder, minimizing the loss function based on gradient descent, and updating all steps between the parameters of the encoder until the preset condition is met. The model that minimizes the loss function during training is taken as the optimal model. The optimal model is used to input the pre-embedded representation into the GAE to obtain the latent features as the final embedded representation.

7. A device for identifying risky enterprises, characterized in that, The device includes: The first acquisition unit is used to acquire a target dataset for each enterprise, the target dataset including various tax information of the enterprise; A construction unit is used to construct an enterprise relationship graph based on the various tax information. The pre-training unit is used to pre-train the enterprise relationship graph using graph convolutional neural network (GCN) and mutual information maximization model (DGI) to obtain pre-embedded representations of enterprise nodes. The training unit is used to train the pre-embedded representation using a graph autoencoder (GAE) to obtain the final embedded representation of the enterprise node. The calculation unit is used to calculate the enterprise similarity of the enterprise based on the final embedded representation of the enterprise node using a cosine similarity algorithm. The enterprise similarity is used to identify whether an enterprise is a risky enterprise. The training unit is further configured to: firstly, obtain positive sample instances using a first feature matrix composed of feature information of the enterprise relationship graph nodes and a first adjacency matrix composed of relationships between the graph nodes; secondly, obtain a second adjacency matrix using the first adjacency matrix, wherein the first adjacency matrix is ​​identical to the second adjacency matrix; construct an erosion function to shuffle and rearrange the first feature matrix using the erosion function, and randomly obtain the second feature matrix; obtain negative sample instances using the second feature matrix and the second adjacency matrix; input the first feature matrix and the first adjacency matrix included in the positive sample instances into the GCN acting as an encoder to obtain the first local features of the positive sample instances; and input the second feature matrix and the second adjacency matrix included in the negative sample instances into the GCN acting as an encoder to obtain the second local features of the negative sample instances. The first local feature is input into the readout function to obtain graph-level global features. The first local feature and the global feature are used as positive sample local-global pairs, and the second local feature and the global feature are used as negative sample local-global pairs. The autoencoder (GAE) is used as a discriminator to score the positive sample local-global and negative sample local-global pairs respectively. The difference between the first score and a vector with all 1s is used as the first loss, and the difference between the second score and a vector with all 0s is used as the second loss. The sum of the first loss and the second loss is used as the value of the noise contrastive objective function. Finally, the above steps are repeated multiple times until a preset condition is met. The model corresponding to the minimum value of the noise contrastive objective function during training is used as the optimal model. The first local feature obtained by inputting positive sample instances into the graph convolutional neural network (GCN) through the optimal model is used as the pre-embedded representation.

8. The apparatus according to claim 7, characterized in that, The device further includes: A receiving unit is configured to receive a search instruction input by a user, wherein the search instruction includes the first company name of the risky enterprise; The second acquisition unit is used to acquire, based on the first enterprise name, the enterprise similarity matching the first enterprise name and the second enterprise name of the suspected risk enterprise associated with the enterprise similarity; The output unit is used to output the second company name to the user.

9. An electronic device, characterized in that, The electronic device includes: a housing, a processor, a memory, a circuit board, and a power supply circuit, wherein the circuit board is disposed inside the space enclosed by the housing, and the processor and the memory are disposed on the circuit board; the power supply circuit is used to supply power to various circuits or devices of the electronic device; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, for executing the risk enterprise identification method according to any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the risk enterprise identification method according to any one of claims 1-6.

Images