Enterprise credit risk assessment method and device based on self-encoding neural network
Through the method based on the self-coding neural network, the low-dimensional feature representation of enterprise credit risk data is extracted, which solves the problem of inaccurate enterprise credit risk assessment in the existing technology, and achieves more accurate risk assessment and efficient utilization of regulatory resources.
Patent Information
- Application Number
- CN202411768343.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-04
- Publication Date
- 2025-03-28
AI Technical Summary
There is a lack of effective and accurate corporate credit risk assessment methods in the prior art, making it difficult to accurately judge and manage corporate credit risks.
Using the method based on the autoencoding neural network, the original data is compressed into a low-dimensional but key information-rich feature representation, and then the enterprise credit risk assessment is carried out.
By reducing data dimensions and enhancing data effectiveness, more accurate corporate credit risk assessment is achieved, supporting differentiated supervision and reasonable resource allocation.
Smart Images
Figure CN119722285A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for evaluating corporate credit risk based on an autoencoding neural network. Background Art
[0002] The judgment and management of corporate credit risk plays a vital role in the market economy. It is not only related to the sustainable development of enterprises, but also the cornerstone of ensuring the stability of market order. Especially in the Guangxi Zhuang Autonomous Region, as the market supervision department has taken measures such as policy guidance, strengthening services and improving mechanisms, the market vitality has been significantly improved. However, with the rapid growth of the number of business entities, the importance of credit risk management has become increasingly prominent. The judgment of corporate credit risk is usually based on credit risk information including registration, administrative licensing, administrative penalties, annual report disclosure, etc., as well as dynamic credit information such as complaints, Internet public opinion, and foreign investment. The results of corporate credit risk classification are an important reference for implementing differentiated supervision, aiming to achieve the rational allocation and efficient use of regulatory resources, thereby improving regulatory effectiveness. However, there is no more effective and accurate corporate credit risk assessment method in the existing technical means. Summary of the invention
[0003] The purpose of the present invention is to overcome the shortcomings of the prior art. The present invention provides a method and device for enterprise credit risk assessment based on an autoencoding neural network. Through the powerful feature extraction capability of the autoencoding neural network model, the original data is compressed and converted into a low-dimensional but key information-rich feature representation, which reduces the dimension of the data and enhances the validity of the data; thereby, the enterprise credit risk assessment can be performed more accurately.
[0004] In order to solve the above technical problems, the present invention provides a method for assessing corporate credit risk based on an autoencoder neural network, the method comprising:
[0005] Obtain the corporate credit risk data and corporate activity data corresponding to the target enterprise;
[0006] Preprocessing the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain preprocessed data corresponding to the target enterprise;
[0007] The pre-processed data corresponding to the target enterprise is input into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and the enterprise credit risk assessment result of the target enterprise is output.
[0008] Optionally, obtaining the enterprise credit risk data and enterprise activity data corresponding to the target enterprise includes:
[0009] Based on the crawler algorithm, data crawling and processing are performed on the target website to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise; wherein the enterprise credit risk data refers to the basic attribute information, operating status information, regulatory enforcement information, association relationship information and comprehensive evaluation information in the universal enterprise credit risk classification model indicator system; the enterprise activity data refers to the operating status domain indicator data, operating activity domain indicator data and cross-domain indicator adjustment parameter data.
[0010] Optionally, preprocessing the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain preprocessed data corresponding to the target enterprise includes:
[0011] Merging the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to form merged data;
[0012] The combined data is cleaned and normalized to form pre-processed data corresponding to the target enterprise.
[0013] Optionally, performing data cleaning on the merged data includes:
[0014] Performing outlier identification processing on the merged data based on a statistical method to obtain outliers in the merged data;
[0015] Correct or delete abnormal values in the merged data based on the business logic and data characteristics of the target enterprise to obtain processed merged data;
[0016] Based on filling or filling or deleting the null values in the processed merged data, the cleaned merged data is obtained.
[0017] Optionally, the autoencoding neural network model includes an autoencoder and a decoder; wherein the autoencoder is composed of a connection layer with an input dimension of 784 and two layers of intermediate dimensions of 1000 and an output dimension of n;
[0018] The decoder consists of a connection layer with an input dimension of n and two layers of intermediate dimension 1000 and an output dimension of 784.
[0019] Optionally, the autoencoder adopts a mean square error loss function during training to regress and predict the input parameters of the autoencoder; the autoencoder adopts a cross entropy loss function during training to predict and classify enterprise data.
[0020] Optionally, the step of inputting the pre-processed data corresponding to the target enterprise into a trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing includes:
[0021] After the pre-processed data corresponding to the target enterprise is input into the trained and converged auto-encoding neural network model, the pre-processed data is subjected to enterprise feature representation learning processing by using an automatic encoding and decoding structure in the trained and converged auto-encoding neural network model to obtain an enterprise feature representation learning result;
[0022] The enterprise feature representation learning result is used in a hypersphere separator in a trained and converged autoencoder neural network model to perform enterprise credit risk judgment modeling and analysis processing, and the enterprise credit risk assessment result of the target enterprise is output.
[0023] In addition, an embodiment of the present invention further provides an enterprise credit risk assessment device based on an autoencoder neural network, the device comprising:
[0024] Acquisition module: used to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise;
[0025] Output processing module: used to pre-process the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain pre-processed data corresponding to the target enterprise;
[0026] Risk assessment module: used to input the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and output the enterprise credit risk assessment result of the target enterprise.
[0027] Optionally, the autoencoding neural network model includes an autoencoder and a decoder; wherein the autoencoder is composed of a connection layer with an input dimension of 784 and two layers of intermediate dimensions of 1000 and an output dimension of n;
[0028] The decoder consists of a connection layer with an input dimension of n and two layers of intermediate dimension 1000 and an output dimension of 784.
[0029] Optionally, the risk assessment module is also used to: after inputting the preprocessed data corresponding to the target enterprise into the trained and converged autoencoding neural network model, use an automatic encoding and decoding structure in the trained and converged autoencoding neural network model to perform enterprise feature representation learning processing on the preprocessed data to obtain enterprise feature representation learning results; perform enterprise credit risk judgment modeling and analysis processing on the enterprise feature representation learning results in a hypersphere separator in the trained and converged autoencoding neural network model, and output the enterprise credit risk assessment results of the target enterprise.
[0030] In an embodiment of the present invention, through the powerful feature extraction capability of the autoencoder neural network model, the original data is compressed and converted into a low-dimensional feature representation rich in key information, which reduces the dimension of the data and enhances the validity of the data; thereby enabling a more accurate assessment of corporate credit risk. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0032] Figure 1 is a flow chart of a method for assessing corporate credit risk based on an autoencoder neural network in an embodiment of the present invention;
[0033] Figure 2 It is a schematic diagram of the structural composition of an enterprise credit risk assessment device based on an autoencoding neural network in an embodiment of the present invention. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] For example, see Figure 1 , Figure 1 It is a flowchart of the enterprise credit risk assessment method based on the autoencoding neural network in an embodiment of the present invention.
[0036] like Figure 1 As shown, a method for assessing corporate credit risk based on an autoencoding neural network comprises:
[0037] S11: Obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise;
[0038] In the specific implementation process of the present invention, the enterprise credit risk data and enterprise activity data corresponding to the target enterprise are obtained, including: crawling data in the target website based on a crawler algorithm to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise; wherein the enterprise credit risk data is the basic attribute information, operating status information, supervision and law enforcement information, association relationship information and comprehensive evaluation information in the universal enterprise credit risk classification model indicator system; the enterprise activity data is the operating status domain indicator data, operating activity domain indicator data and cross-domain indicator adjustment parameter data.
[0039] Specifically, firstly, data is crawled and processed in the target website through crawler algorithms to obtain the corresponding enterprise credit risk data and enterprise activity data of the target enterprise; among them, enterprise credit risk data is a key indicator to measure the credit status of enterprises in economic activities, reflecting the reliability of enterprises in fulfilling economic commitments, including the ability to repay debts, pay rents, and pay taxes. The level of corporate credit has a direct impact on its market reputation, financing capabilities, and business expansion. First, credit is the cornerstone of enterprises' participation in the market economy, and good credit helps to establish stable business relationships; secondly, credit rating helps enterprises improve their management level, discover and improve weak links in management; finally, credit rating can also help enterprises reduce financing costs, improve financing efficiency, and gain competitive advantages in activities such as bidding. The indicator system of the general enterprise credit risk classification model (second edition) includes first-level indicators of five basic dimensions: basic attribute information, operating status information, regulatory enforcement information, related relationship information, and comprehensive evaluation information, 22 second-level indicators, and 73 third-level indicators. Basic attribute information: reflects the risk factors based on the characteristics of the enterprise group, including two second-level indicators of enterprise scale and enterprise background, and 9 third-level indicators. Business status information: reflects the risk factors based on behavioral characteristics, including 7 secondary indicators such as access permit matters, registration and filing matters, information disclosure status, business status, tax and social security payment status, intellectual property and certification status, and credit commitment, and 31 tertiary indicators. Supervision and law enforcement information: reflects the risk factors based on historical supervision and law enforcement records, including 6 secondary indicators such as credit repair, abnormal business list, dishonest persons subject to execution and serious dishonest list (blacklist), administrative penalties, administrative inspections, and judicial litigation information, and 16 tertiary indicators. Related relationship information: reflects the risk factors based on corporate related relationships, including 4 secondary indicators such as credit evaluation, illegal dishonesty information of relevant personnel of the enterprise, illegal dishonesty of related enterprises, and credit risk categories of related enterprises, and 12 tertiary indicators. Comprehensive evaluation information: reflects the risk factors based on various evaluation information, including 3 secondary indicators such as complaints and reports, public opinion evaluation, and social attention, and 5 tertiary indicators.
[0040] The enterprise activity data includes the business status domain indicator data, business activity domain indicator data and cross-domain indicator adjustment parameter data; the basic indicators of the business status domain are divided into: national tax payment status, local tax payment status, social security payment status, and provident fund payment status. The basic indicators of the business activity domain are divided into: the number of enterprise change filings, the category of enterprise change filings, the number of branch offices started, the number of foreign investment enterprises, the total amount of foreign investment, the number of investments, the total amount of investments, the number of investment capital change types, the number of investment capital changes, the number of migration applications, the number of branch registrations and revocations, the number of social security arrears months, the number of provident fund arrears months, and the number of information results searched by the full name of the enterprise. The adjustment parameters are set according to the business status of the enterprise as shown in Table 1 below:
[0041] Table 1
[0042]
[0043] The enterprise's operating status indicators and social security and provident fund contributions are not affected by the industry, so in these indicators, the adjustment parameter is 1. For other indicators, the calculation method of the industry adjustment parameter is as follows:
[0044]
[0045] in, is the mean value of the indicator; The industry average of the indicator.
[0046] S12: Preprocessing the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain preprocessed data corresponding to the target enterprise;
[0047] In the specific implementation process of the present invention, the enterprise credit risk data and enterprise activity data corresponding to the target enterprise are pre-processed to obtain the pre-processed data corresponding to the target enterprise, including: merging the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to form merged data; performing data cleaning on the merged data, and normalizing the cleaned merged data to form the pre-processed data corresponding to the target enterprise.
[0048] Furthermore, the data cleaning processing of the merged data includes: performing outlier identification processing on the merged data based on statistical methods to obtain outliers in the merged data; correcting or deleting outliers in the merged data based on the business logic and data characteristics of the target enterprise to obtain processed merged data; filling or deleting empty values in the processed merged data to obtain cleaned merged data.
[0049] Specifically, data processing is an important step in data analysis and the use of models to assess corporate credit risk, and is directly related to the accuracy and reliability of the model; credit risk and activity data are collected from different sources, and they contain information about the company in different aspects. Through a carefully designed data merging strategy, the two data sets are merged into a unified data set. Then, the merged data is thoroughly cleaned; outlier processing is an important part of this stage; a variety of statistical methods are used to identify outliers in the data, and based on business logic and data characteristics, it is decided to correct or delete these outliers. At the same time, the problem of null values in the data is also handled; the presence of null values may affect the training and prediction of the model, so strategies such as filling or deletion are used to solve this problem. Finally, the data is normalized to ensure the consistency of the weights of different features in model prediction.
[0050] S13: Inputting the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and outputting the enterprise credit risk assessment result of the target enterprise.
[0051] In the specific implementation process of the present invention, the autoencoding neural network model includes an autoencoder and a decoder; wherein the autoencoder adopts an input dimension of 784, two layers of intermediate dimension 1000, and an output dimension of n connection layer composition; the decoder adopts an input dimension of n, two layers of intermediate dimension 1000, and an output dimension of 784 connection layer composition.
[0052] Furthermore, the autoencoder adopts a mean square error loss function during training to regress and predict the input parameters of the autoencoder; the autoencoder adopts a cross entropy loss function during training to predict and classify enterprise data.
[0053] Furthermore, the inputting of the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model for enterprise credit risk assessment includes: after inputting the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model, using an automatic encoding and decoding structure in the trained and converged auto-encoding neural network model to perform enterprise feature representation learning processing on the pre-processed data to obtain an enterprise feature representation learning result; performing enterprise credit risk judgment modeling and analysis processing on the enterprise feature representation learning result in a hypersphere separator in the trained and converged auto-encoding neural network model, and outputting the enterprise credit risk assessment result of the target enterprise.
[0054] Specifically, in order to accurately and efficiently utilize multi-dimensional information such as enterprise attributes, business status, and business activities, and to support the correlation analysis between enterprise activity and regional economic activity, a data analysis architecture based on a deep self-encoding and decoding network is proposed. The proposed architecture can deeply utilize and mine the nonlinear mapping characteristics of the encoding network and the complementary characteristics between multi-dimensional information to ensure that while learning a complete feature representation for enterprise activity, it can further resist the influence of abnormal samples, and thereby obtain a unified feature representation of enterprise activity for different regions and different types of industries. At the same time, reverse decoding learning is performed on the activity of the regional economy, and an implicit intrinsic correlation analysis between the two is established in the middle layer of the network structure.
[0055] Train the self-encoding and decoding neural network to judge the enterprise activity and credit risk: the encoder adopts a fully connected layer with an input dimension of 784, two layers of intermediate dimension 1000, and an output dimension of n; the decoder adopts an input dimension of n, two layers of intermediate dimension 1000, and an output dimension of 784. The decoder and decoder training use the mean square error loss function to regress and predict the input parameters of the encoder; the decoder will generate a large amount of prediction data based on the input of the encoder. When the sample size is small, this method can be used to expand the data set to better train the classifier.
[0056] The encoder is composed of a fully connected layer with an input dimension of 784, an intermediate dimension of 1000, and an output dimension of n. The training loss uses the cross entropy loss function to predict and classify enterprise data.
[0057] In the enterprise feature representation learning based on the automatic encoding and decoding structure, given a set of enterprise sample data sets X = {X (1) ,X (2) ,…,X (V)}, where V represents the enterprise information data collected from V perspectives (enterprise attributes, business status, business activities, etc.), is the characteristic matrix at the vth angle, D v is the feature dimension at the vth angle, and N is the number of enterprise samples. The analysis model proposed in the project fully utilizes the consistency and complementarity of information from multiple angles of the enterprise, obtains a more intrinsic and robust enterprise data representation based on the automatic encoding and decoding structure from the perspective of multi-perspective subspace learning.
[0058] Subspace learning assumes that there is a common subspace that can characterize information from different angles. Its purpose is to make full use of information from various perspectives. It assumes that there is a shared common representation H, which can reconstruct the characteristics of each enterprise sample at each angle through a set of mappings, namely X (v) =f v(H), Different from the traditional learning scheme, the proposed scheme intends to use the initialized representation as input to reconstruct features, which reduces the loss in the process of information transmission and can better represent the original intrinsic features. Assuming that under a given common representation, the information at each angle is conditionally independent, the key to multi-view subspace learning is to effectively encode information from different available perspectives:
[0059] p(X|H)=p(X (1) |H)p(X (2) |H)…p(X (V) |H);
[0060] Given a sample X, for the public representation H Perform likelihood modeling:
[0061]
[0062] in, h n Corresponding to the potential representation of the nth enterprise sample, is the parameter of the subspace encoding network, which controls the reconstruction mapping f from the public representation H to the sample feature X v (.), accordingly, is the constructed reconstruction loss. It can be found that maximizing the likelihood function is equivalent to minimizing the reconstruction loss:
[0063]
[0064] Among them U (v) is the output of the subspace encoding network, and its column vector Specifically:
[0065]
[0066] in and They represent the weight and bias of the network at the vth perspective respectively, and σ(.) is the activation function of the network.
[0067] Based on the hypersphere classification model, the enterprise credit risk judgment modeling analysis is carried out. After learning the unified feature representation of enterprise samples, a classification model based on the hypersphere is built to solve the problem of enterprise credit risk judgment based on the different abnormal situations of abnormal enterprises. Assuming that the samples of normal operating enterprises are all located within the hypersphere, they are taken as the target samples as the inner points, while the samples of abnormal operating enterprises are considered as the outer points and are distributed outside the hypersphere. The farther the distance from the center of the sphere, the greater the degree of abnormality of the enterprise sample. From the above, it can be seen that H = [v 1 ,…,h r ,h r+1 ,…,hr+h ], where h 1 ,…,h r It is represented as the new feature vector of the labeled normal operating enterprise sample, h r+1 ,…,h r+h is the new feature vector of the enterprise to be tested for operating status. The objective function is constructed as follows:
[0068]
[0069] Where t is the center of the circle, R is the radius of the hypersphere, ξ i is the relaxation factor and C is the penalty factor.
[0070] When testing the autoencoder neural network model, the evaluation indicators used include MAE and RMSE; MAE, or mean absolute error, is the average of the absolute values of the difference between the predicted value and the true value. It does not consider the positive or negative of the predicted value, but pays more attention to the size of the absolute error, so it can intuitively represent the size of the gap between the model predicted value and the true value. The advantage of MAE is that it gives equal importance to the size of all errors, which makes it better than the mean square error to reflect the degree of deviation of the predicted value. When the predicted value is completely consistent with the true value, the MAE value is 0, indicating that the model prediction result is perfect; the larger the error, the larger the MAE value. RMSE, or root mean square error, is one of the indicators to measure image quality or the accuracy of the prediction model. Its calculation principle is to first calculate the square of the difference between the true value and the predicted value, then sum and average, and finally take the square root. The smaller the RMSE value, the more similar the image is or the more accurate the prediction result of the model is. RMSE is of the same origin as statistical indicators such as standard deviation. It is mainly used to reflect the degree of discretization of data and is used as a criterion for introducing and eliminating variables in regression analysis.
[0071] In an embodiment of the present invention, through the powerful feature extraction capability of the autoencoder neural network model, the original data is compressed and converted into a low-dimensional feature representation rich in key information, which reduces the dimension of the data and enhances the validity of the data; thereby enabling a more accurate assessment of corporate credit risk.
[0072] For example 2, please refer to Figure 2 , Figure 2 It is a schematic diagram of the structural composition of an enterprise credit risk assessment device based on an autoencoding neural network in an embodiment of the present invention.
[0073] like Figure 2 As shown, a corporate credit risk assessment device based on an auto-encoding neural network comprises:
[0074] Acquisition module 21: used to obtain enterprise credit risk data and enterprise activity data corresponding to the target enterprise;
[0075] In the specific implementation process of the present invention, the enterprise credit risk data and enterprise activity data corresponding to the target enterprise are obtained, including: crawling data in the target website based on a crawler algorithm to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise; wherein the enterprise credit risk data is the basic attribute information, operating status information, supervision and law enforcement information, association relationship information and comprehensive evaluation information in the universal enterprise credit risk classification model indicator system; the enterprise activity data is the operating status domain indicator data, operating activity domain indicator data and cross-domain indicator adjustment parameter data.
[0076] Specifically, firstly, data is crawled and processed in the target website through crawler algorithms to obtain the corresponding enterprise credit risk data and enterprise activity data of the target enterprise; among them, enterprise credit risk data is a key indicator to measure the credit status of enterprises in economic activities, reflecting the reliability of enterprises in fulfilling economic commitments, including the ability to repay debts, pay rent, and pay taxes. The level of enterprise credit risk has a direct impact on its market reputation, financing ability, and business expansion. First, credit is the cornerstone of enterprises' participation in the market economy, and good credit helps to establish stable business relationships; secondly, credit rating helps enterprises improve their management level, discover and improve weak links in management; finally, credit rating can also help enterprises reduce financing costs, improve financing efficiency, and gain competitive advantages in activities such as bidding. The indicator system of the general enterprise credit risk classification model (second edition) includes first-level indicators of five basic dimensions: basic attribute information, operating status information, regulatory enforcement information, related relationship information, and comprehensive evaluation information, 22 second-level indicators, and 73 third-level indicators. Basic attribute information: reflects the risk factors based on the characteristics of the enterprise group, including two second-level indicators of enterprise scale and enterprise background, and 9 third-level indicators. Business status information: reflects the risk factors based on behavioral characteristics, including 7 secondary indicators such as access permit matters, registration and filing matters, information disclosure status, business status, tax and social security payment status, intellectual property and certification status, and credit commitment, and 31 tertiary indicators. Supervision and law enforcement information: reflects the risk factors based on historical supervision and law enforcement records, including 6 secondary indicators such as credit repair, abnormal business list, dishonest persons subject to execution and serious dishonest list (blacklist), administrative penalties, administrative inspections, and judicial litigation information, and 16 tertiary indicators. Related relationship information: reflects the risk factors based on corporate related relationships, including 4 secondary indicators such as credit evaluation, illegal dishonesty information of relevant personnel of the enterprise, illegal dishonesty of related enterprises, and credit risk categories of related enterprises, and 12 tertiary indicators. Comprehensive evaluation information: reflects the risk factors based on various evaluation information, including 3 secondary indicators such as complaints and reports, public opinion evaluation, and social attention, and 5 tertiary indicators.
[0077] The enterprise activity data includes the business status domain indicator data, business activity domain indicator data and cross-domain indicator adjustment parameter data; the basic indicators of the business status domain are divided into: national tax payment status, local tax payment status, social security payment status, and provident fund payment status. The basic indicators of the business activity domain are divided into: the number of enterprise change filings, the category of enterprise change filings, the number of branch offices started, the number of foreign investment enterprises, the total amount of foreign investment, the number of investments, the total amount of investments, the number of investment capital change types, the number of investment capital changes, the number of migration applications, the number of branch registrations and revocations, the number of social security arrears months, the number of provident fund arrears months, and the number of information results searched by the full name of the enterprise. The adjustment parameters are set according to the business status of the enterprise as shown in Table 1 below:
[0078] Table 1
[0079]
[0080]
[0081] The enterprise's operating status indicators and social security and provident fund contributions are not affected by the industry, so in these indicators, the adjustment parameter is 1. For other indicators, the calculation method of the industry adjustment parameter is as follows:
[0082]
[0083] in, is the mean value of the indicator; The industry average of the indicator.
[0084] Output processing module 22: used to pre-process the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain pre-processed data corresponding to the target enterprise;
[0085] In the specific implementation process of the present invention, the enterprise credit risk data and enterprise activity data corresponding to the target enterprise are pre-processed to obtain the pre-processed data corresponding to the target enterprise, including: merging the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to form merged data; performing data cleaning on the merged data, and normalizing the cleaned merged data to form the pre-processed data corresponding to the target enterprise.
[0086] Furthermore, the data cleaning processing of the merged data includes: performing outlier identification processing on the merged data based on statistical methods to obtain outliers in the merged data; correcting or deleting outliers in the merged data based on the business logic and data characteristics of the target enterprise to obtain processed merged data; filling or deleting empty values in the processed merged data to obtain cleaned merged data.
[0087] Specifically, data processing is an important step in data analysis and the use of models to assess corporate credit risk, and is directly related to the accuracy and reliability of the model; credit risk and activity data are collected from different sources, and they contain information about the company in different aspects. Through a carefully designed data merging strategy, the two data sets are merged into a unified data set. Then, the merged data is thoroughly cleaned; outlier processing is an important part of this stage; a variety of statistical methods are used to identify outliers in the data, and based on business logic and data characteristics, it is decided to correct or delete these outliers. At the same time, the problem of null values in the data is also handled; the presence of null values may affect the training and prediction of the model, so strategies such as filling or deletion are used to solve this problem. Finally, the data is normalized to ensure the consistency of the weights of different features in model prediction.
[0088] Risk assessment module 23: used for inputting the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and outputting the enterprise credit risk assessment result of the target enterprise.
[0089] In the specific implementation process of the present invention, the autoencoding neural network model includes an autoencoder and a decoder; wherein the autoencoder adopts an input dimension of 784, two layers of intermediate dimension 1000, and an output dimension of n connection layer composition; the decoder adopts an input dimension of n, two layers of intermediate dimension 1000, and an output dimension of 784 connection layer composition.
[0090] Furthermore, the risk assessment module is also used to: after inputting the preprocessed data corresponding to the target enterprise into the trained and converged autoencoding neural network model, use the automatic encoding and decoding structure in the trained and converged autoencoding neural network model to perform enterprise feature representation learning processing on the preprocessed data to obtain the enterprise feature representation learning result; perform enterprise credit risk judgment modeling and analysis processing on the enterprise feature representation learning result in the hypersphere separator in the trained and converged autoencoding neural network model, and output the enterprise credit risk assessment result of the target enterprise.
[0091] Specifically, in order to accurately and efficiently utilize multi-dimensional information such as enterprise attributes, business status, and business activities, and to support the correlation analysis between enterprise activity and regional economic activity, a data analysis architecture based on a deep self-encoding and decoding network is proposed. The proposed architecture can deeply utilize and mine the nonlinear mapping characteristics of the encoding network and the complementary characteristics between multi-dimensional information to ensure that while learning a complete feature representation for enterprise activity, it can further resist the influence of abnormal samples, and thereby obtain a unified feature representation of enterprise activity for different regions and different types of industries. At the same time, reverse decoding learning is performed on the activity of the regional economy, and an implicit intrinsic correlation analysis between the two is established in the middle layer of the network structure.
[0092] Train the self-encoding and decoding neural network to judge the enterprise activity and credit risk: the encoder adopts a fully connected layer with an input dimension of 784, two layers of intermediate dimension 1000, and an output dimension of n; the decoder adopts an input dimension of n, two layers of intermediate dimension 1000, and an output dimension of 784. The decoder and decoder training use the mean square error loss function to regress and predict the input parameters of the encoder; the decoder will generate a large amount of prediction data based on the input of the encoder. When the sample size is small, this method can be used to expand the data set to better train the classifier.
[0093] The encoder is composed of a fully connected layer with an input dimension of 784, an intermediate dimension of 1000, and an output dimension of n. The training loss uses the cross entropy loss function to predict and classify enterprise data.
[0094] In the enterprise feature representation learning based on the automatic encoding and decoding structure, given a set of enterprise sample data sets X = {X (1) ,X (2) ,…,X (V)}, where V represents the enterprise information data collected from V perspectives (enterprise attributes, business status, business activities, etc.), It is in v The characteristic matrix at angles, D v is the feature dimension at the vth angle, and N is the number of enterprise samples. The analysis model proposed in the project fully utilizes the consistency and complementarity of information from multiple angles of the enterprise, obtains a more intrinsic and robust enterprise data representation based on the automatic encoding and decoding structure from the perspective of multi-perspective subspace learning.
[0095] Subspace learning assumes that there is a common subspace that can characterize information from different angles. Its purpose is to make full use of information from various perspectives. It assumes that there is a shared common representation H, which can reconstruct the characteristics of each enterprise sample at each angle through a set of mappings, namely X (v) =f v(H), Different from the traditional learning scheme, the proposed scheme intends to use the initialized representation as input to reconstruct features, which reduces the loss in the process of information transmission and can better represent the original intrinsic features. Assuming that under a given common representation, the information at each angle is conditionally independent, the key to multi-view subspace learning is to effectively encode information from different available perspectives:
[0096] p(X|X)=p(X (1) |H)p(X (2) |H)…p(X (V) |H);
[0097] Given a sample X, for the public representation H Perform likelihood modeling:
[0098]
[0099] in, h n Corresponding to the potential representation of the nth enterprise sample, is the parameter of the subspace encoding network, which controls the reconstruction mapping f from the public representation H to the sample feature X v (.), accordingly, is the constructed reconstruction loss. It can be found that maximizing the likelihood function is equivalent to minimizing the reconstruction loss:
[0100]
[0101] Among them U (v) is the output of the subspace encoding network, and its column vector Specifically:
[0102]
[0103] in and They represent the weight and bias of the network at the vth perspective respectively, and σ(.) is the activation function of the network.
[0104] The enterprise credit risk judgment modeling analysis based on the hypersphere classification model, after learning to obtain a unified feature representation of enterprise samples, based on the consideration of the different abnormal situations of abnormal enterprises, a classification model based on the hypersphere is built to solve the problem of enterprise credit risk judgment. Assuming that the samples of normal operating enterprises are all located within the hypersphere, they are taken as the target samples as the inner points, while the samples of abnormal operating enterprises are considered as the outer points and are distributed outside the hypersphere. The farther the distance from the center of the sphere, the greater the degree of abnormality of the enterprise sample. From the above, it can be seen that H = [h 1 ,…,h r ,h r+1 ,…,hr+h ], where h 1 ,…,h r It is represented as the new feature vector of the labeled normal operating enterprise sample, h r+1 ,…,h r+h is the new feature vector of the enterprise to be tested for operating status. The objective function is constructed as follows:
[0105]
[0106] Where t is the center of the circle, R is the radius of the hypersphere, ξ i is the relaxation factor and C is the penalty factor.
[0107] When testing the autoencoder neural network model, the evaluation indicators used include MAE and RMSE; MAE, or mean absolute error, is the average of the absolute values of the difference between the predicted value and the true value. It does not consider the positive or negative of the predicted value, but pays more attention to the size of the absolute error, so it can intuitively represent the size of the gap between the model predicted value and the true value. The advantage of MAE is that it gives equal importance to the size of all errors, which makes it better than the mean square error to reflect the degree of deviation of the predicted value. When the predicted value is completely consistent with the true value, the MAE value is 0, indicating that the model prediction result is perfect; the larger the error, the larger the MAE value. RMSE, or root mean square error, is one of the indicators to measure image quality or the accuracy of the prediction model. Its calculation principle is to first calculate the square of the difference between the true value and the predicted value, then sum and average, and finally take the square root. The smaller the RMSE value, the more similar the image is or the more accurate the prediction result of the model is. RMSE is of the same origin as statistical indicators such as standard deviation. It is mainly used to reflect the degree of discretization of data and is used as a criterion for introducing and eliminating variables in regression analysis.
[0108] In an embodiment of the present invention, through the powerful feature extraction capability of the autoencoder neural network model, the original data is compressed and converted into a low-dimensional feature representation rich in key information, which reduces the dimension of the data and enhances the validity of the data; thereby enabling a more accurate assessment of corporate credit risk.
[0109] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.
[0110] In addition, the above is a detailed introduction to a method and device for assessing corporate credit risk based on an autoencoding neural network provided by an embodiment of the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A method for assessing corporate credit risk based on an autoencoder neural network, characterized in that: The method comprises: Obtain the corporate credit risk data and corporate activity data corresponding to the target enterprise; Preprocessing the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain preprocessed data corresponding to the target enterprise; The pre-processed data corresponding to the target enterprise is input into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and the enterprise credit risk assessment result of the target enterprise is output.
2. The enterprise credit risk assessment method according to claim 1, characterized in that: The obtaining of the enterprise credit risk data and enterprise activity data corresponding to the target enterprise includes: Based on the crawler algorithm, data crawling and processing are performed on the target website to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise; wherein the enterprise credit risk data refers to the basic attribute information, operating status information, regulatory enforcement information, association relationship information and comprehensive evaluation information in the universal enterprise credit risk classification model indicator system; the enterprise activity data refers to the operating status domain indicator data, operating activity domain indicator data and cross-domain indicator adjustment parameter data.
3. The enterprise credit risk assessment method according to claim 1, characterized in that: The preprocessing of the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain the preprocessed data corresponding to the target enterprise includes: Merging the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to form merged data; The combined data is cleaned and normalized to form pre-processed data corresponding to the target enterprise.
4. The enterprise credit risk assessment method according to claim 3, characterized in that: The performing data cleaning processing on the merged data includes: Performing outlier identification processing on the merged data based on a statistical method to obtain outliers in the merged data; Correct or delete abnormal values in the merged data based on the business logic and data characteristics of the target enterprise to obtain processed merged data; Based on filling or filling or deleting the null values in the processed merged data, the cleaned merged data is obtained.
5. The enterprise credit risk assessment method according to claim 1, characterized in that: The autoencoder neural network model includes an autoencoder and a decoder; wherein the autoencoder is composed of a connection layer with an input dimension of 784 and two layers of intermediate dimensions of 1000 and an output dimension of n; The decoder consists of a connection layer with an input dimension of n and two layers of intermediate dimension 1000 and an output dimension of 784.
6. The enterprise credit risk assessment method according to claim 5, characterized in that: The autoencoder adopts a mean square error loss function during training to regress and predict the input parameters of the autoencoder; the autoencoder adopts a cross entropy loss function during training to predict and classify enterprise data.
7. The enterprise credit risk assessment method according to claim 1, characterized in that: The step of inputting the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing includes: After the pre-processed data corresponding to the target enterprise is input into the trained and converged auto-encoding neural network model, the pre-processed data is subjected to enterprise feature representation learning processing by using an automatic encoding and decoding structure in the trained and converged auto-encoding neural network model to obtain an enterprise feature representation learning result; The enterprise feature representation learning result is used in a hypersphere separator in a trained and converged autoencoder neural network model to perform enterprise credit risk judgment modeling and analysis processing, and the enterprise credit risk assessment result of the target enterprise is output.
8. An enterprise credit risk assessment device based on an autoencoding neural network, characterized in that: The device comprises: Acquisition module: used to obtain the enterprise credit risk data and enterprise activity data corresponding to the target enterprise; Output processing module: used to pre-process the enterprise credit risk data and enterprise activity data corresponding to the target enterprise to obtain pre-processed data corresponding to the target enterprise; Risk assessment module: used to input the pre-processed data corresponding to the target enterprise into the trained and converged auto-encoding neural network model to perform enterprise credit risk assessment processing, and output the enterprise credit risk assessment result of the target enterprise.
9. The enterprise credit risk assessment device according to claim 8, characterized in that: The autoencoder neural network model includes an autoencoder and a decoder; wherein the autoencoder is composed of a connection layer with an input dimension of 784 and two layers of intermediate dimensions of 1000 and an output dimension of n; The decoder consists of a connection layer with an input dimension of n and two layers of intermediate dimension 1000 and an output dimension of 784.
10. The enterprise credit risk assessment device according to claim 1, characterized in that: The risk assessment module is also used to: After the pre-processed data corresponding to the target enterprise is input into the trained and converged auto-encoding neural network model, the pre-processed data is subjected to enterprise feature representation learning processing by using an automatic encoding and decoding structure in the trained and converged auto-encoding neural network model to obtain an enterprise feature representation learning result; The enterprise feature representation learning result is used in a hypersphere separator in a trained and converged autoencoder neural network model to perform enterprise credit risk judgment modeling and analysis processing, and the enterprise credit risk assessment result of the target enterprise is output.