Logistic regression enterprise credit assessment method and system based on additive model

Abstract

The invention discloses a logistic regression enterprise credit assessment method and system based on an additive model, belongs to the technical field of big data processing, and aims to solve the technical problem of how to realize robust enterprise credit assessment based on the additive model. Comprising the following steps: analyzing and fusing multi-source heterogeneous data of a user to construct a multi-source data standard library; selecting index data based on the multi-source data standard library, and constructing an index system for enterprise credit assessment based on the selected index data; preprocessing index data in the index system; constructing an evaluation model and a loss function of the evaluation model based on logistic regression and an additive model, performing model optimization on the evaluation model through loss function minimization, and performing model verification on the optimized evaluation model; and evaluating the to-be-evaluated enterprise through the trained evaluation model to obtain an evaluation result including the predicted label and the probability value, and judging the credit risk level of the enterprise based on the evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and in particular to a logistic regression enterprise credit assessment method and system based on an additive model. Background Art

[0002] The challenges faced by traditional corporate credit assessment methods when processing large amounts of financial data include:

[0003] Abnormal data interference: Traditional corporate credit assessment methods often have difficulty maintaining accuracy when processing financial data containing outliers. Abnormal data may be generated for various reasons, such as data entry errors, fraudulent behavior, or abnormal market fluctuations. These outliers may interfere with the model's training process, causing the model to misjudge or miss judgments in actual applications;

[0004] Insufficient ability to capture nonlinear relationships: Financial data often contain complex nonlinear relationships. Traditional corporate credit assessment methods are usually modeled based on linear assumptions, so it is difficult to effectively capture these nonlinear relationships. This limits the performance of the model in a complex and changing financial environment and may lead to unsatisfactory risk control results.

[0005] The characteristic of the additive model is that it can directly deal with the nonlinear relationship between the response variable and multiple explanatory variables when dealing with nonlinear relationships. This feature makes it excellent in capturing complex patterns in the data. In terms of improving robustness, the application of the additive model can decompose the prediction function into the sum of multiple sub-functions, each sub-function corresponds to a characteristic variable. This structure makes the model more robust to abnormal data. Even if some characteristic variables have abnormal values, it will not have a significant impact on the prediction results of the entire model.

[0006] How to achieve robust corporate credit assessment based on additive models is a technical problem that needs to be solved. Summary of the invention

[0007] The technical task of the present invention is to address the above shortcomings and provide a logistic regression enterprise credit assessment method and system based on an additive model to solve the technical problem of how to achieve robust enterprise credit assessment based on an additive model.

[0008] In a first aspect, the present invention provides a logistic regression enterprise credit assessment method based on an additive model, comprising the following steps:

[0009] Analyze and fuse the user's multi-source heterogeneous data, form a standard data set based on the analyzed and fused multi-dimensional data, and store the standard data set in the constructed multi-source data standard library;

[0010] Select indicator data based on a multi-source data standard library, and build an indicator system for corporate credit assessment based on the selected indicator data;

[0011] Preprocess the indicator data in the indicator system;

[0012] An evaluation model and a loss function of the evaluation model are constructed based on logistic regression and additive models, the evaluation model is optimized by minimizing the loss function, and the optimized evaluation model is verified to obtain a final trained evaluation model, wherein the evaluation model uses performance indicators as input feature variables and is used to perform binary classification evaluation on enterprises based on the feature variables to output predicted labels and probability values;

[0013] The enterprise to be evaluated is evaluated through the trained evaluation model to obtain the evaluation results including predicted labels and probability values, and the credit risk level of the enterprise is judged based on the evaluation results.

[0014] As a preferred embodiment, the user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in a multi-source data standard library, API interface data, structured data, and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes information related to judicial, business, investment, and intellectual property rights; third-party data includes information related to user external guarantees, water and electricity usage, equity pledge, land mortgage and transfer, and pledge of important shareholders; structured data includes basic enterprise information, change information, blacklist, and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements;

[0015] When analyzing and fusing users' multi-source heterogeneous data, for structured data, a unified standard data table is established through data alignment and fusion comparison methods. A standard data table covering multiple ranges is established for information of a certain dimension. Data between different data sources are fused and compared to establish a unified standard data set. Data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with quality that meets requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

[0016] Preferably, when preprocessing the indicator data in the indicator system, invalid value processing, same value statistics and missing value statistics are performed.

[0017] As a preference, an evaluation model is constructed based on logistic regression and additive model, including the following steps:

[0018] Logistic regression is a statistical method for classification problems. The calculation formula is as follows:

[0019]

[0020] Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable;

[0021] Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is:

[0022]

[0023] Among them, z=β0+β1x1+β2x2+…+β n x n ;

[0024] In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as:

[0025]

[0026] Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1;

[0027] By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows:

[0028]

[0029] Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm;

[0030] According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model to transform the prediction function

[0031] f(x)=f1(x1)+f2(x2)+…+f n (x n )

[0032] Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain:

[0033]

[0034] Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as:

[0035]

[0036] Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and;

[0037] For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as:

[0038]

[0039] Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel;

[0040] Correspondingly, the loss function of the evaluation model based on logistic regression and additive model is constructed, including the following steps:

[0041] After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as:

[0042]

[0043] Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function;

[0044]

[0045] Then the total loss function is obtained, which is expressed as:

[0046]

[0047] Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

[0048] As a preference, when optimizing the evaluation model by minimizing the loss function, the center c of the Gaussian kernel is updated by the gradient descent method. k , width parameter σ k And the weight coefficient α k , including the following steps:

[0049] Calculate the gradient: For the weight coefficient α k , the gradient is expressed as:

[0050]

[0051] Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function.

[0052] For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as:

[0053]

[0054] in,

[0055] For the Gaussian kernel width parameter σ k The gradient of is expressed as:

[0056]

[0057] in,

[0058] The value of each parameter is updated through the gradient descent formula, and the calculation formula is as follows:

[0059]

[0060] Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate;

[0061] Repeat the iteration until the loss function converges to a stable value or reaches the preset number of iterations;

[0062] For the optimized evaluation model, cross-validation is performed on the optimized evaluation model based on indicators including AUC and KS.

[0063] In a second aspect, the present invention provides a logistic regression enterprise credit assessment system based on an additive model, which is used to perform credit assessment on an enterprise through a logistic regression enterprise credit assessment method based on an additive model as described in any one of the first aspects, wherein the system includes a standard library construction module, an indicator system construction module, an indicator processing module, a model optimization module, and an enterprise assessment module;

[0064] The standard library construction module is used to perform the following: analyze and fuse the multi-source heterogeneous data of the user, form a standard data set based on the multi-dimensional data after analysis and fusion, and store the standard data set in the constructed multi-source data standard library;

[0065] The indicator system construction module is used to perform the following: selecting indicator data based on a multi-source data standard library, and constructing an indicator system for enterprise credit assessment based on the selected indicator data;

[0066] The indicator processing module is used to perform the following: pre-processing of indicator data in the indicator system;

[0067] The model optimization module is used to perform the following: construct an evaluation model and a loss function of the evaluation model based on logistic regression and additive models, optimize the evaluation model by minimizing the loss function, and verify the optimized evaluation model to obtain the final trained evaluation model, wherein the evaluation model uses the performance index as the input feature variable, and is used to perform a binary evaluation of the enterprise based on the feature variable to output the predicted label and probability value;

[0068] The enterprise assessment module is used to perform the following: evaluate the enterprise to be assessed through the trained assessment model, obtain the assessment results including the predicted labels and probability values, and judge the credit risk level of the enterprise based on the assessment results.

[0069] As a preferred embodiment, the user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in a multi-source data standard library, API interface data, structured data, and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes information related to judicial, business, investment, and intellectual property rights; third-party data includes information related to user external guarantees, water and electricity usage, equity pledge, land mortgage and transfer, and pledge of important shareholders; structured data includes basic enterprise information, change information, blacklist, and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements;

[0070] When analyzing and fusing the user's multi-source heterogeneous data, the standard library construction module is used to perform the following: for structured data, a unified standard data table is established through data alignment and fusion comparison methods, a standard data table covering multiple ranges is established for information of a certain dimension, data between different data sources are fused and compared, a unified standard data set is established, data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with data quality that meets the requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

[0071] Preferably, when preprocessing the indicator data in the indicator system, the indicator preprocessing module is used to perform invalid value processing, same value statistics and missing value statistics.

[0072] Preferably, the model optimization module is used to perform the following evaluation model construction based on logistic regression and additive model:

[0073] Logistic regression is a statistical method for classification problems. The calculation formula is as follows:

[0074]

[0075] Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable;

[0076] Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is:

[0077]

[0078] Among them, z=β0+β1x1+β2x2+…+β n x n ;

[0079] In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as:

[0080]

[0081] Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1;

[0082] By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows:

[0083]

[0084] Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm;

[0085] According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model to transform the prediction function

[0086] f(x)=f1(x1)+f2(x2)+…+f n (x n )

[0087] Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain:

[0088]

[0089] Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n ) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as:

[0090]

[0091] Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and;

[0092] For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as:

[0093]

[0094] Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel;

[0095] Correspondingly, the loss function of the evaluation model based on logistic regression and additive model is constructed, including the following steps:

[0096] After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as:

[0097]

[0098] Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function;

[0099]

[0100] Then the total loss function is obtained, which is expressed as:

[0101]

[0102] Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

[0103] Preferably, when the evaluation model is optimized by minimizing the loss function, the model optimization module is used to update the center c of the Gaussian kernel by the gradient descent method.k , width parameter σ k And the weight coefficient α k , including the following steps:

[0104] Calculate the gradient: For the weight coefficient α k , the gradient is expressed as:

[0105]

[0106] Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function.

[0107] For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as:

[0108]

[0109] in,

[0110] For the Gaussian kernel width parameter σ k The gradient of is expressed as:

[0111]

[0112] in,

[0113] The value of each parameter is updated through the gradient descent formula, and the calculation formula is as follows:

[0114]

[0115] Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate;

[0116] Repeat the iteration until the loss function converges to a stable value or reaches the preset number of iterations;

[0117] For the optimized evaluation model, the model optimization module is used to cross-validate the optimized evaluation model based on indicators including AUC and KS.

[0118] The logistic regression enterprise credit assessment method and system based on additive model of the present invention has the following advantages:

[0119] 1. Improved the robustness of the evaluation model to abnormal data: When constructing the evaluation model, an additive model is introduced on the basis of logistic regression to decompose the prediction function of the enterprise credit scoring method into the sum of multiple sub-functions, each of which corresponds to a characteristic variable. This structure ensures that even if some characteristic variables have abnormal values, it will not have a significant impact on the prediction results of the entire model, thereby improving the robustness of the model;

[0120] 2. Enhanced the ability of the evaluation model to capture nonlinear relationships: Performance indicators related to enterprise evaluation often contain complex nonlinear relationships, and traditional enterprise credit scoring methods are usually modeled based on linear assumptions, which makes it difficult to effectively capture these nonlinear relationships. The kernel function is used to optimize the evaluation model, mapping the original feature space to a high-dimensional feature space, thereby effectively capturing the nonlinear relationship in the data. This not only improves the generalization ability of the model, but also enables the model to more accurately identify and handle the impact of nonlinear factors, further improving the accuracy and reliability of the enterprise credit scoring method;

[0121] 3. Improve the accuracy and reliability of credit scoring: Through the introduction of kernel functions, the evaluation model can automatically identify and extract nonlinear features that have a significant impact on risk control results. The extraction of these nonlinear features further improves the accuracy and reliability of corporate credit scoring, providing a more scientific basis for financial institutions' credit decisions;

[0122] 4. Optimize the enterprise credit scoring method: By integrating the additive model and kernel function optimization strategy, it not only solves the problems faced by traditional risk control models when processing large amounts of financial data, but also provides new ideas and methods for subsequent enterprise credit scoring research. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, 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.

[0124] The present invention is further described below in conjunction with the accompanying drawings.

[0125] Figure 1 This is a flowchart of a logistic regression enterprise credit assessment method based on an additive model in Example 1. DETAILED DESCRIPTION

[0126] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments may be combined with each other.

[0127] The embodiment of the present invention provides a logistic regression enterprise credit assessment method and system based on an additive model, which is used to solve the technical problem of how to achieve robust enterprise credit assessment based on an additive model.

[0128] Embodiment 1:

[0129] The invention discloses a logistic regression enterprise credit assessment method based on an additive model, which comprises five steps: standard library construction, indicator system construction, indicator processing, model optimization and enterprise assessment.

[0130] Step S100: Standard library construction: Analyze and fuse the user's multi-source heterogeneous data, form a standard data set based on the analyzed and fused multi-dimensional data, and store the standard data set in the constructed multi-source data standard library.

[0131] In this embodiment, the user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in the multi-source data standard library, API interface data, structured data and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes judicial, business, investment, and intellectual property related information; third-party data includes information related to users' external guarantees, water and electricity consumption, equity pledge, land mortgage and transfer, and important shareholders' pledge; structured data includes basic enterprise information, change information, blacklist and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements.

[0132] When analyzing and fusing users' multi-source heterogeneous data, for structured data, a unified standard data table is established through data alignment and fusion comparison methods. A standard data table covering multiple ranges is established for information of a certain dimension. Data between different data sources are fused and compared to establish a unified standard data set. Data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with quality that meets requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

[0133] Step S200: Indicator system construction: selecting indicator data based on a multi-source data standard library, and constructing an indicator system for enterprise credit assessment based on the selected indicator data.

[0134] In this embodiment, based on the established multi-source data standard library, an indicator system for credit scoring is constructed. The data for constructing the indicators mainly comes from industrial and commercial, change, administrative punishment, dishonesty, judicial, debt, taxation and other data; in addition, relationship data such as investment, guarantee, branch, legal person, etc. are also used to construct some indicators.

[0135] Step S300: Index processing: pre-processing the index data in the index system.

[0136] When this embodiment performs indicator processing, the preprocessing of indicator data is completed through invalid value processing, same value statistics, missing value statistics and other processes; invalid value processing mainly removes information such as telephone codes and address codes; same value statistics processing removes fields that only contain same values ​​and fields with a same value rate higher than 80%; missing value statistics processing removes all missing fields and fields with a field missing rate higher than 80%.

[0137] Step S400 model optimization: construct an evaluation model and a loss function of the evaluation model based on logistic regression and additive models, optimize the evaluation model by minimizing the loss function, and verify the optimized evaluation model to obtain the final trained evaluation model, wherein the evaluation model uses performance indicators as input feature variables, and is used to perform binary evaluation of enterprises based on the feature variables to output predicted labels and probability values.

[0138] As a specific implementation of model optimization, an evaluation model is constructed based on logistic regression and additive models, including the following steps:

[0139] (1) Traditional logistic regression model construction, loss function and optimization process

[0140] Logistic regression is a statistical method for classification problems. The calculation formula is as follows:

[0141]

[0142] Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable;

[0143] Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is:

[0144]

[0145] Among them, z=β0+β1x1+β2x2+…+β n xn ;

[0146] In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as:

[0147]

[0148] Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1;

[0149] By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows:

[0150]

[0151] Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm;

[0152] (2) Constructing an evaluation model by integrating logistic regression with additive models

[0153] According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model, the prediction function:

[0154] f(x)=f1(x1)+f2(x2)+…+f n (x n )

[0155] Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain:

[0156]

[0157] Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n ) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as:

[0158]

[0159] Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and;

[0160] For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as:

[0161]

[0162] Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel.

[0163] Correspondingly, the loss function of the evaluation model based on logistic regression and additive model is constructed, including the following steps:

[0164] (1) After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as:

[0165]

[0166] Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function;

[0167]

[0168] (2) Apply the above formula to the entire training set, and then obtain the total loss function, which is expressed as:

[0169]

[0170] Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

[0171] Correspondingly, when optimizing the evaluation model by minimizing the loss function, the center c of the Gaussian kernel is updated by the gradient descent method. k , width parameter σ k And the weight coefficient α k , including the following steps:

[0172] (1) Calculate the gradient: For the weight coefficient α k , the gradient is expressed as:

[0173]

[0174] Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function.

[0175] For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as:

[0176]

[0177] in,

[0178] For the Gaussian kernel width parameter σ k The gradient of is expressed as:

[0179]

[0180] in,

[0181] (2) Update the value of each parameter using the gradient descent formula. The calculation formula is as follows:

[0182]

[0183] Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate;

[0184] (3) Repeat the above steps until the loss function converges to a stable value or reaches the preset number of iterations.

[0185] For the optimized evaluation model, cross-validation is performed on the optimized evaluation model based on indicators including AUC and KS.

[0186] Step S500: Enterprise evaluation: the enterprise to be evaluated is evaluated through the trained evaluation model to obtain an evaluation result including a predicted label and a probability value, and the credit risk level of the enterprise is determined based on the evaluation result.

[0187] The method of this embodiment decomposes the prediction function of the risk control model into multiple sub-functions by introducing an additive model to improve the robustness of the model to abnormal data; then, the kernel function is used to optimize the model, mapping the original feature space to a high-dimensional feature space, effectively capturing the nonlinear relationship in the data, and enhancing the generalization ability of the model; finally, through the introduction of the kernel function, nonlinear features that have a significant impact on the risk control results are automatically identified and extracted, further improving the accuracy and reliability of corporate credit scores.

[0188] In specific applications, the application of additive models can decompose the prediction function of the enterprise credit scoring method into the sum of multiple sub-functions, each of which corresponds to a characteristic variable, to improve the robustness of the model to abnormal data. In the optimization of kernel functions, kernel functions are used to map the original feature space to high-dimensional feature space, capture nonlinear relationships in the data, and improve the generalization ability and accuracy of the model. In addition, in nonlinear factor mining, kernel functions are used to automatically identify and extract nonlinear features that have a significant impact on risk control results, further improving the accuracy and reliability of enterprise credit scores.

[0189] The method of this embodiment effectively reduces the interference of abnormal data on the accuracy of the enterprise credit scoring model, better adapts to the complex and changing financial environment, and provides strong support for the optimization of enterprise credit scoring.

[0190] Embodiment 2:

[0191] The invention discloses a logistic regression enterprise credit evaluation system based on an additive model, comprising a standard library construction module, an indicator system construction module, an indicator processing module, a model optimization module and an enterprise evaluation module.

[0192] The standard library construction module is used to perform the following: analyze and fuse the user's multi-source heterogeneous data, form a standard data set based on the analyzed and fused multi-dimensional data, and store the standard data set in the constructed multi-source data standard library.

[0193] In this embodiment, the user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in the multi-source data standard library, API interface data, structured data and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes judicial, business, investment, and intellectual property related information; third-party data includes information related to users' external guarantees, water and electricity consumption, equity pledge, land mortgage and transfer, and important shareholders' pledge; structured data includes basic enterprise information, change information, blacklist and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements.

[0194] When analyzing and fusing the user's multi-source heterogeneous data, the standard library construction module is used to perform the following: For structured data, a unified standard data table is established through data alignment and fusion comparison methods, a standard data table covering multiple ranges is established for information of a certain dimension, data between different data sources are fused and compared, a unified standard data set is established, data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with data quality that meets the requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

[0195] The indicator system construction module is used to perform the following: select indicator data based on a multi-source data standard library, and construct an indicator system for enterprise credit assessment based on the selected indicator data.

[0196] In this embodiment, based on the established multi-source data standard library, an indicator system for credit scoring is constructed. The data for constructing the indicators mainly comes from industrial and commercial, change, administrative punishment, dishonesty, judicial, debt, taxation and other data; in addition, relationship data such as investment, guarantee, branch, legal person, etc. are also used to construct some indicators.

[0197] The indicator processing module is used to perform the following: pre-processing of indicator data in the indicator system.

[0198] When the indicator processing is performed in this embodiment, the indicator processing module is used to perform the following: preprocessing of the indicator data is completed through invalid value processing, same value statistics, missing value statistics and other processes; invalid value processing is mainly to remove telephone codes, address codes and other information; same value statistics processing removes fields that only contain same values ​​and fields with a same value rate higher than 80%; missing value statistics processing removes all missing fields and fields with a field missing rate higher than 80%.

[0199] The model optimization module is used to perform the following: construct an evaluation model and a loss function of the evaluation model based on logistic regression and additive models, optimize the evaluation model by minimizing the loss function, and verify the optimized evaluation model to obtain the final trained evaluation model, wherein the evaluation model uses performance indicators as input feature variables, and is used to perform binary classification evaluation on the enterprise based on the feature variables to output predicted labels and probability values.

[0200] As a specific implementation of model optimization, the model optimization module is used to perform the following operations to build an evaluation model based on logistic regression and additive models:

[0201] (1) Traditional logistic regression model construction, loss function and optimization process

[0202] Logistic regression is a statistical method for classification problems. The calculation formula is as follows:

[0203]

[0204] Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable;

[0205] Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is:

[0206]

[0207] Among them, z=β0+β1x1+β2x2+…+β n x n ;

[0208] In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as:

[0209]

[0210] Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1;

[0211] By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows:

[0212]

[0213] Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm;

[0214] (2) Constructing an evaluation model by integrating logistic regression with additive models

[0215] According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model, the prediction function:

[0216] f(x)=f1(x1)+f2(x2)+…+f n (x n )

[0217] Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain:

[0218]

[0219] Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n ) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as:

[0220]

[0221] Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and;

[0222] For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as:

[0223]

[0224] Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel.

[0225] Correspondingly, the model optimization module is used to execute the following loss function to build an evaluation model based on logistic regression and additive models:

[0226] (1) After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as:

[0227]

[0228] Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function;

[0229]

[0230] (2) Apply the above formula to the entire training set, and then obtain the total loss function, which is expressed as:

[0231]

[0232] Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

[0233] Correspondingly, when optimizing the evaluation model by minimizing the loss function, the model optimization module is used to update the center c of the Gaussian kernel by gradient descent. k , width parameter σ k And the weight coefficient α k , including the following steps:

[0234] (1) Calculate the gradient: For the weight coefficient α k , the gradient is expressed as:

[0235]

[0236] Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function.

[0237] For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as:

[0238]

[0239] in,

[0240] For the Gaussian kernel width parameter σ k The gradient of is expressed as:

[0241]

[0242] in,

[0243] (2) Update the value of each parameter using the gradient descent formula. The calculation formula is as follows:

[0244]

[0245] Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate;

[0246] (3) Repeat the above steps until the loss function converges to a stable value or reaches the preset number of iterations.

[0247] For the optimized evaluation model, the model optimization module is used to cross-validate the optimized evaluation model based on indicators including AUC and KS.

[0248] The enterprise assessment module is used to perform the following: evaluate the enterprise to be assessed through the trained assessment model, obtain the assessment results including the predicted labels and probability values, and judge the credit risk level of the enterprise based on the assessment results.

[0249] The system of this embodiment can execute the method disclosed in Embodiment 1 to conduct credit assessment on an enterprise.

[0250] The present invention has been shown and described in detail above through the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above multiple embodiments, those skilled in the art can know that the means in the above different embodiments can be combined to obtain more embodiments of the present invention, and these embodiments are also within the protection scope of the present invention.

Claims

1. A logistic regression enterprise credit assessment method based on an additive model, characterized in that: The steps include: Analyze and fuse the user's multi-source heterogeneous data, form a standard data set based on the analyzed and fused multi-dimensional data, and store the standard data set in the constructed multi-source data standard library; Select indicator data based on a multi-source data standard library, and build an indicator system for corporate credit assessment based on the selected indicator data; Preprocess the indicator data in the indicator system; An evaluation model and a loss function of the evaluation model are constructed based on logistic regression and additive models, the evaluation model is optimized by minimizing the loss function, and the optimized evaluation model is verified to obtain a final trained evaluation model, wherein the evaluation model uses performance indicators as input feature variables and is used to perform binary classification evaluation on enterprises based on the feature variables to output predicted labels and probability values; The enterprise to be evaluated is evaluated through the trained evaluation model to obtain the evaluation results including predicted labels and probability values, and the credit risk level of the enterprise is judged based on the evaluation results.

2. The logistic regression enterprise credit assessment method based on additive model according to claim 1 is characterized in that: The user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in the multi-source data standard library, API interface data, structured data and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes information related to judicial, business, investment, and intellectual property rights; third-party data includes information related to users' external guarantees, water and electricity consumption, equity pledge, land mortgage and transfer, and pledge of important shareholders; structured data includes basic enterprise information, change information, blacklist and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements; When analyzing and fusing users' multi-source heterogeneous data, for structured data, a unified standard data table is established through data alignment and fusion comparison methods. A standard data table covering multiple ranges is established for information of a certain dimension. Data between different data sources are fused and compared to establish a unified standard data set. Data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with quality that meets requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

3. The logistic regression enterprise credit assessment method based on additive model according to claim 1 is characterized in that: When preprocessing the indicator data in the indicator system, invalid value processing, same value statistics and missing value statistics are performed.

4. The logistic regression enterprise credit assessment method based on additive model according to claim 1, characterized in that: Building an evaluation model based on logistic regression and additive models includes the following steps: Logistic regression is a statistical method for classification problems. The calculation formula is as follows: Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable; Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is: Where, z=β0+β1x1+β2x2+…+β n x n ; In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as: Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1; By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows: Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm; According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model to transform the prediction function f(x)=f1(x1)+f2(x2)+…+f n (x n ) Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain: Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n ) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as: Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and; For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as: f(x)= Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel; Correspondingly, the loss function of the evaluation model based on logistic regression and additive model is constructed, including the following steps: After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as: Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function; Then the total loss function is obtained, which is expressed as: Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

5. The logistic regression enterprise credit assessment method based on additive model according to claim 4 is characterized in that: When optimizing the evaluation model by minimizing the loss function, the center c of the Gaussian kernel is updated by the gradient descent method. k , width parameter σ k And the weight coefficient α k , including the following steps: Calculate the gradient: For the weight coefficient α k , the gradient is expressed as: Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function. For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as: in, For the Gaussian kernel width parameter σ k The gradient of is expressed as: in, The value of each parameter is updated through the gradient descent formula, and the calculation formula is as follows: Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate; Repeat the iteration until the loss function converges to a stable value or reaches the preset number of iterations; For the optimized evaluation model, cross-validation is performed on the optimized evaluation model based on indicators including AUC and KS.

6. A logistic regression enterprise credit assessment system based on an additive model, characterized in that: Used to perform credit assessment on an enterprise by a logistic regression enterprise credit assessment method based on an additive model as described in any one of claims 1 to 5, the system comprising a standard library construction module, an indicator system construction module, an indicator processing module, a model optimization module and an enterprise assessment module; The standard library construction module is used to perform the following: analyze and fuse the multi-source heterogeneous data of the user, form a standard data set based on the multi-dimensional data after analysis and fusion, and store the standard data set in the constructed multi-source data standard library; The indicator system construction module is used to perform the following: selecting indicator data based on a multi-source data standard library, and constructing an indicator system for enterprise credit assessment based on the selected indicator data; The indicator processing module is used to perform the following: pre-processing of indicator data in the indicator system; The model optimization module is used to perform the following: construct an evaluation model and a loss function of the evaluation model based on logistic regression and additive models, optimize the evaluation model by minimizing the loss function, and verify the optimized evaluation model to obtain the final trained evaluation model, wherein the evaluation model uses the performance index as the input feature variable, and is used to perform a binary evaluation of the enterprise based on the feature variable to output the predicted label and probability value; The enterprise assessment module is used to perform the following: evaluate the enterprise to be assessed through the trained assessment model, obtain the assessment results including the predicted labels and probability values, and judge the credit risk level of the enterprise based on the assessment results.

7. The logistic regression enterprise credit assessment system based on additive model according to claim 6, characterized in that: The user's multi-source heterogeneous data includes government data, Internet data, third-party data, stock data stored in the multi-source data standard library, API interface data, structured data and semi-structured data; among them, government data covers information related to industry and commerce, social security, provident fund, administrative penalties, and enterprise certification; Internet data includes information related to judicial, business, investment, and intellectual property rights; third-party data includes information related to users' external guarantees, water and electricity consumption, equity pledge, land mortgage and transfer, and pledge of important shareholders; structured data includes basic enterprise information, change information, blacklist and certification information; semi-structured data includes information related to the enterprise's judicial, judgment documents, and court announcements; When analyzing and fusing the user's multi-source heterogeneous data, the standard library construction module is used to perform the following: for structured data, a unified standard data table is established through data alignment and fusion comparison methods, a standard data table covering multiple ranges is established for information of a certain dimension, data between different data sources are fused and compared, a unified standard data set is established, data with complementary relationships are aligned and fused, data with redundant relationships are deduplicated, data with data quality that meets the requirements is selected, and multi-dimensional data is fused and compared to form a standard data set.

8. The logistic regression enterprise credit assessment system based on additive model according to claim 6, characterized in that: When preprocessing the indicator data in the indicator system, the indicator preprocessing module is used to perform invalid value processing, same value statistics and missing value statistics.

9. The logistic regression enterprise credit assessment system based on additive model according to claim 6, characterized in that: The model optimization module is used to perform the following evaluation model construction based on logistic regression and additive model: Logistic regression is a statistical method for classification problems. The calculation formula is as follows: Among them, p represents the probability of an event occurring, which represents the probability of corporate default or overdue in corporate credit assessment, β0 represents the intercept term, β1, ... β n represents the coefficient of the corresponding characteristic variable, x1,...x n represents the characteristic variable; Logistic regression uses the Sigmoid function to map the results of linear regression to the (0,1) interval to obtain the probability value. The formula of the Sigmoid function is: Where, z=β0+β1x1+β2x2+…+β n x n ; In logistic regression, model parameters are solved by maximum likelihood estimation based on the likelihood function. The likelihood function is expressed as: Where N represents the number of samples, p i represents the predicted probability value of the i-th sample, yi represents the actual label of the i-th sample, and the actual label is 0 or 1; By simplifying the likelihood function by taking the logarithm of the likelihood function, we get the log-likelihood function, which is expressed as follows: Solve the parameter β that maximizes the log-likelihood function by using the gradient ascent method or Newton optimization algorithm; According to the enterprise risk control needs and data characteristics, select the characteristic variable x n , using the additive model to transform the prediction function f(x)=f1(x1)+f2(x2)+…+f n (x n ) Each subfunction f n (x n ) can be any nonlinear function, which can be combined through the logistic regression formula to obtain: Select the Gaussian kernel function to convert the input feature variable x=(x1,...x n ) is mapped into high-dimensional space, and the Gaussian kernel function is expressed as: Among them, c represents the center of the Gaussian kernel, σ represents the width parameter of the Gaussian kernel, and ‖.‖ represents the Euclidean norm between the feature variable and; For each feature variable x i Select one or more Gaussian kernels and set the corresponding center and width parameters. The evaluation model is represented in high-dimensional space as: f(x)= Where K represents the number of Gaussian kernels, α k Represents the weight coefficient of the kth Gaussian kernel; Correspondingly, the loss function of the evaluation model based on logistic regression and additive model is constructed, including the following steps: After mapping the input feature variables into a high-dimensional space, the logarithmic loss function is used to measure the difference between the probability predicted by the evaluation model and the actual label. For a binary classification problem, the logarithmic loss function is expressed as: Where y represents the actual label, p represents the probability predicted by the evaluation model, and the probability predicted by the evaluation model maps the function value f(x) in the high-dimensional space into the interval (0,1) through the Sigmoid function; Then the total loss function is obtained, which is expressed as: Where N represents the number of training samples, x j and j are the feature variable and actual label of the jth sample respectively.

10. The logistic regression enterprise credit assessment system based on additive model according to claim 9, characterized in that: When the evaluation model is optimized by minimizing the loss function, the model optimization module is used to update the center c of the Gaussian kernel by the gradient descent method. k , width parameter σ k And the weight coefficient α k , including the following steps: Calculate the gradient: For the weight coefficient α k , the gradient is expressed as: Among them, σ ′ (z) = σ(z)(1-σ(z)), which is the derivative of the Sigmoid function. For the dth dimension, the Gaussian kernel center c k The gradient of is expressed as: in, For the Gaussian kernel width parameter σ k The gradient of is expressed as: in, The value of each parameter is updated through the gradient descent formula, and the calculation formula is as follows: Among them, θ represents the parameters that need to be updated, which are the center c of the Gaussian kernel k , width parameter σ k And the weight coefficient α k , η represents the learning rate; Repeat the iteration until the loss function converges to a stable value or reaches the preset number of iterations; For the optimized evaluation model, the model optimization module is used to cross-validate the optimized evaluation model based on indicators including AUC and KS.

Images