Small and micro enterprise risk identification method based on privacy calculation and logistic regression model
By combining privacy calculation and logistic regression models, the problem of insufficient privacy protection in traditional credit risk assessment is solved, and more accurate and reliable risk identification is achieved, which is suitable for the inclusive financial field of small and micro enterprises.
Patent Information
- Application Number
- CN202411845240.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-04-25
AI Technical Summary
Traditional credit risk assessment relies on historical data and financial statements, and there is a problem of insufficient privacy protection, making it difficult to effectively identify the risks of small and micro enterprises.
The risk identification method of small and micro enterprises based on privacy computing and logistic regression models is adopted to obtain and process feature data, and privacy interrogation, feature engineering and model training are carried out to ensure data privacy and security, while improving the accuracy and generalization capabilities of the model.
It realizes more accurate risk identification and more reliable data processing processes, enhances data privacy protection, solves the problem that financial institutions have difficulty obtaining comprehensive information from enterprises, and improves the accuracy and stability of risk assessment.
Smart Images
Figure CN119887366A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for identifying risks of small and micro enterprises, and more specifically, to a method for identifying risks of small and micro enterprises based on privacy computing and a logistic regression model, which belongs to the field of banking business-financial risk prevention and control and the field of computer technology. Background Art
[0002] In recent years, the rapid development of machine learning technology has brought unprecedented changes to the field of inclusive finance. By deeply mining and analyzing large amounts of data, machine learning models can more accurately assess the credit status of borrowers, help risk management personnel make better decisions, and thus reduce business costs and risks. At the same time, the banking industry currently has relatively complete post-loan risk management measures for large and medium-sized corporate customers, but lacks effective means of post-loan risk management for small and micro enterprises. There are a large number of inclusive small and micro enterprises, and risks cannot be identified through manual judgment. Therefore, it is necessary to use technical means such as big data and artificial intelligence to automatically and efficiently identify risky customers in batches.
[0003] Traditional credit risk assessment often relies on historical data and financial statements, and usually does not involve complex data processing techniques. These methods may not be perfect in terms of privacy protection, because data may be collected and processed without the full consent of the data owner. And the privacy and security of data often rely on the security of data storage and processing. However, these systems may be vulnerable to data leakage and unauthorized access, especially during data sharing and analysis. Privacy computing technology enables joint analysis of data and model training without leaking the data of all parties. This technology ensures the privacy and security of data during the calculation process, thereby improving the reliability of data sharing.
[0004] Traditional models may be limited in performance due to limited data sources and low model complexity. In addition, traditional methods may have difficulty dealing with data imbalance and are not very accurate because the form is very simple and it is difficult to fit the true distribution of the data. Logistic regression models are trained by maximizing the likelihood function and can provide better prediction performance. Combined with privacy computing technology, richer data sources can be used to improve the accuracy and generalization ability of the model without sacrificing data privacy. Summary of the invention
[0005] In order to solve the above-mentioned problems of the prior art, the present invention provides a small and micro enterprise risk identification method based on privacy computing and logistic regression model, which has technical characteristics that can solve the problems of traditional credit risk assessment relying on historical data and financial statements and insufficient privacy protection.
[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:
[0007] The present invention provides a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model, the method comprising the following steps:
[0008] Step S1: Obtain the feature data required for modeling; obtain the relevant feature data of the two small and micro enterprises through database operations, perform data deduplication processing, and select relevant features from all features;
[0009] Step S2: Privacy intersection: On the premise that the data of both parties are not shared, each participant encrypts and calculates its own data locally to generate a ciphertext result;
[0010] Step S3: Feature engineering: further perform feature engineering based on the data set after the privacy intersection sample alignment is completed in step S2;
[0011] Step S4: Horizontal division: further divide the data set according to the data set processed by feature engineering in step S3;
[0012] Step S5: Logistic regression model training;
[0013] Step S6: Model evaluation: After the model training is completed, the model effect is analyzed using the evaluation indicators commonly used in machine learning;
[0014] Step S7: Model deployment application; save the trained model in the privacy computing platform and deploy it to the production environment for actual prediction tasks.
[0015] Preferably, step 3 specifically includes:
[0016] Step S31: missing value filling: performing missing value filling;
[0017] Step S32: Standard deviation normalization: Calculate the mean μ of each feature and its standard deviation σ, and further normalize each feature data point to have zero mean and unit variance;
[0018] Step S33: Extreme value processing: determine the percentile you want to use, and use this percentile as the replacement value for the extreme value; specifically: after the data set is sorted in ascending order, the value at the x%*nth position will replace the extreme value, where x% is the selected percentile and n is the total amount of data in the data set; for each feature, use a statistical method to calculate the value corresponding to the selected percentile, and replace the identified extreme value with the calculated percentile value;
[0019] Step S34: evidence weight conversion;
[0020] Perform Woe conversion of evidence weight value: first, all features are binned (discretized);
[0021] The binning strategy used is equal frequency binning, which groups continuous variables into a series of intervals and ensures that the number of data points in each box is equal. The specific process is as follows: sort the data in ascending order, determine the number of bins k, divide the data into k intervals (for example, k is 12), calculate the total data volume n and the data capacity of each interval n / k (for example, n is 64572); determine the specific boundary value of each interval according to the number of bins and allocate data to specific intervals. After completing the binning, count the proportion of good customers and bad customers in each bin, that is, and Among themGood i and Bad i are the number of good customers and bad customers in the ith interval, and Good sum and Bad sum Represent the total number of good customers and bad customers in the data set, and then calculate the weight of evidence (woe). The calculation formula is: ： Check whether the woe value in each bin satisfies monotonicity. If not, return to the binning step for re-binning. When there are only good customers or bad customers in a bin, the woe value needs to be repaired. The common method is to add 0.5 to both the numerator and the denominator, that is,
[0022] Preferably, step 5 specifically includes:
[0023] Step S51: configure the model; initialize the logistic regression model, set the initial parameters of the model, the expansion coefficient (representing the encryption accuracy), the error value (the allowable error value between the model prediction result and the actual result), the optimization method (the optimization method used to minimize the loss function of the regression equation), the weights of both parties, the batch size (the number of samples selected for one training), the maximum number of iterations (the maximum number of times the optimizer repeatedly iterates the training data before finding the best model weight), the regularization method (whether to constrain the complexity of the model by adding a penalty term in the model), and the loss function;
[0024] Step S52: Model definition: The goal of the logistic regression model is to output a predicted value, a predicted classification probability value p, which is obtained by converting the Sigmoid function Where z is a linear combination, z = θ0 + θ1x1 + θ2x2 + ... + θ m θ m , where θ0,θ1,θ2,...,θ m are model parameters, and x0,x1,x2,...,x m is the eigenvalue;
[0025] Step S53: Gradient calculation: In order to update the parameters, it is necessary to calculate the gradient of the loss function with respect to each parameter. The gradient is the slope of the loss function at the current parameter value, which indicates the rate of change of the loss function in this direction. For parameter θ j The gradient is calculated as follows: in is the derivative of the Sigmoid function,
[0026] Step S54: Parameter update: According to the gradient calculated in step S53, the parameters are further updated according to the SGDM method. The specific calculation method is: θ t+1 =θ t -ηv t , where θ t refers to the model parameters after the tth iteration, v t =βv t-1 +(1-β)▽J(θ t ) is the momentum of the tth iteration, β is the momentum decay coefficient, ▽J(θ t ) is the parameter θ t The gradient calculated according to step S53 is here, and η is the learning rate;
[0027] Step S55: Iterative training: cyclically execute the above steps S53 and S54 until the set maximum number of iterations is completed and the training process is completed.
[0028] Preferably, the participants in step 2 exchange encrypted calculation results through the established secure network dedicated line, and without exposing their own original data, perform sample alignment based on the matching fields of both parties, retain the enterprises that exist in the data sets of both parties, and both parties generate a matching column match_id for matching the different features of the same enterprise on both parties.
[0029] Preferably, step 31 adopts mean filling, that is, filling the missing values of the feature column by calculating the mean of the non-missing values of the feature column.
[0030] Preferably, in step 4, the matched data set is divided into a training set and a test set in a ratio of 8:2, with the training set accounting for 80% and the test set accounting for 20%.
[0031] Preferably, the analysis focus values in step 6 include the KS value (Kolmogorov-Smirnov) and AUC value (Area Under Cureve) of the focus model to perform model performance analysis.
[0032] Preferably, the minimum value replacement percentile selected in step S33 is 1%, and the maximum value replacement percentile is 99%.
[0033] Preferably, the parameter initialization mode of the model in step S51 is selected as random uniform initialization, and the initial weight of the model is initialized by randomly sampling values from a uniform distribution. The expansion coefficient of the model is selected as 1000, the error value is selected as 0, the optimization method is selected as Stochastic Gradient Descent with Momentum (SGDM), the weights of the two parties are not adjusted, the batch size is selected as 128, the maximum number of iterations is 30, the regularization method is selected as L1 regularization method with Lasso coefficient of 0.1, and the loss function is logarithmic loss (Log Loss). For a single sample, the loss function value is: L(y,p) = -[ylog(p) + (1-y)log(1-p)], and for the entire training set, the loss function is the average of all sample losses.
[0034] Preferably, in step S54, β takes a value of 0.1 and η takes a value of 0.1.
[0035] Beneficial effects: The present invention solves the problem that traditional credit risk assessment relies on historical data and financial statements and has insufficient privacy protection. By introducing the combination of privacy computing technology and logistic regression model, more accurate risk identification and more reliable data processing process are achieved; the application of privacy computing technology ensures the privacy and security of data in the calculation process, so that the data of all parties can be jointly analyzed and model trained without leaking individual data, thereby improving the reliability of data sharing, not only enhancing the protection of data privacy, but also solving the problem that financial institutions such as banks are difficult to obtain comprehensive information of enterprises. Through privacy computing technology, government data and bank data can be combined for analysis under the premise of protecting privacy to improve the accuracy of risk assessment; the use of logistic regression model provides better prediction performance by maximizing the likelihood function for training, thereby effectively coping with data imbalance and improving the accuracy of the model. Combined with privacy computing technology, this method can use richer data sources to make the model have stronger generalization ability, so as to more accurately identify the risks of small and micro enterprises. The application of this technology can improve the accuracy and stability of risk identification without sacrificing data privacy, bringing more possibilities and benefits to the field of inclusive finance. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a model training flow chart of the present invention.
[0037] Figure 2 This is the main architecture diagram of a privacy computing platform of the present invention.
[0038] Figure 3 It is a comparison chart of the training set results of the present invention relative to the bank's unilateral data modeling.
[0039] Figure 4 It is a comparison chart of the test set results of the present invention relative to the bank's unilateral data modeling.
[0040] Figure 5 It is a box diagram of the prediction results of the present invention. DETAILED DESCRIPTION
[0041] The present invention will be further described below in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.
[0042] like Figure 1-5 The figure shows a specific embodiment of a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model. The embodiment is a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model. The method includes the following steps:
[0043] Step S1: Acquire feature data required for modeling.
[0044] Through the operation of bank and government databases, we obtain relevant characteristic data of small and micro enterprises on both sides, deduplicate the data, and screen out relatively relevant features from all the features based on expert experience.
[0045] In this embodiment, the specific features used after screening by the bank are as follows:
[0046] Transaction amount of encrypted social credit code companies’ accounts in the past 9 months
[0047] Business days Amount of account payment in the last 12 months
[0048] Annual average daily balance of corporate loans Number of account loan transactions in the past three months
[0049] Corporate loan balance Number of account loan transactions in the past 6 months
[0050] Number of over-the-counter transactions in the past 12 months Number of account loan transactions in the past 9 months
[0051] Number of POS transactions in the past 12 months Number of account loan transactions in the past 12 months
[0052] Number of account transactions in the past 12 months Amount of account loan transactions in the past 3 months
[0053] Total financial balance Account loan transaction amount in the past 6 months
[0054] Head office effective customer mark account loan transaction amount in the last 9 months
[0055] Number of products covered in the past 12 months Account loan transaction amount in the past 12 months
[0056] Average daily balance of corporate deposits in the past 12 months Number of account loan transactions in the past 3 months
[0057] Average daily balance of corporate loans in the past 12 months
[0058] Average daily balance of total financial volume in the past 12 months Number of account loan transactions in the past 6 months
[0059] Number of corporate account transactions in the past three months Number of account loan transactions in the past nine months
[0060] Number of corporate account transactions in the past 6 months Number of account loan transactions in the past 12 months
[0061] Number of corporate account transactions in the past 9 months Amount of account loan transactions in the past 3 months
[0062] Number of corporate account transactions in the past 12 months Amount of account loan transactions in the past 6 months
[0063] Account transaction amount in the last 3 months Account loan transaction amount in the last 9 months
[0064] The company's account transaction amount in the last 6 months and the account loan transaction amount in the last 12 months
[0065] Specific features on the government side include:
[0066] Encrypted social credit code
[0067] Total value-added tax paid by enterprises in the past three months
[0068] Total corporate income tax paid by the company in the past 24 months
[0069] Total VAT paid by enterprises in the past six months
[0070] Average monthly VAT payment of enterprises in the past 12 months
[0071] Total value-added tax paid by enterprises in the past 12 months
[0072] Total amount of other taxes paid by enterprises in the past six months
[0073] Total VAT paid by enterprises in the past 24 months
[0074] The average number of employees who paid pension contributions in the enterprise in the past year
[0075] Tax types paid by enterprises in the past 12 months
[0076] Total tax paid by enterprises in the past 12 months
[0077] Total amount of other taxes paid by the enterprise in the past 24 months
[0078] Tax types paid by enterprises in the past 24 months
[0079] Total tax paid by the enterprise in the past 24 months
[0080] The last time the enterprise's tax credit rating was assessed was this year.
[0081] Number of employees in the company's latest annual report
[0082] The duration of the enterprise's participation in social insurance
[0083] Company Type
[0084] The number of years since the company started operations
[0085] Total corporate income tax paid by enterprises in the past 12 months
[0086] The enterprise entity has all types of vehicles in its current stock
[0087] Step S2: Privacy intersection. Under the premise that the data of both parties are not shared, each participant encrypts and calculates its own data locally to generate a ciphertext result. Figure 2 As shown, the participants exchange encrypted calculation results through the established secure network dedicated line, and align samples based on the matching fields of both parties (the "encrypted social credit code" field in the embodiment) without exposing their own original data. That is, the companies that exist in the data sets of both parties are retained, and both parties generate a matching column (match_id) for matching the different features of the same company on both sides.
[0088] Step S3: Feature engineering. Based on the data set after the privacy intersection sample alignment in step S2, further feature engineering is performed.
[0089] Step S31: Missing value filling: First, missing value filling is performed. In this embodiment, the method used is mean filling, that is, the missing values of the feature column are filled by calculating the mean of the non-missing values of the feature column.
[0090] Step S32: Standard deviation normalization. Calculate the mean μ of each feature, and calculate its standard deviation σ to further normalize each feature data point so that it has zero mean and unit variance.
[0091] Step S33: Extreme value processing. Determine the percentile you want to use, which will be used as the replacement value for the extreme value. That is, after the data set is sorted in ascending order, the value at the x%*nth position will replace the extreme value, where x% is the selected percentile and n is the total amount of data in the data set. (In this embodiment, the replacement percentile for the minimum value is 1%, and the replacement percentile for the maximum value is 99%). For each feature, use a statistical method to calculate the value corresponding to the selected percentile. Replace the identified extreme value with the calculated percentile value.
[0092] Step S34: Weight of Evidence conversion (woe); further perform woe conversion. First, all features are binned, that is, discretized. The binning strategy adopted in this embodiment is equal-frequency binning, which groups continuous variables into a series of intervals and ensures that the number of data points in each box is roughly equal. The specific process is: data sorting, and ascending sorting is adopted in this embodiment. Determine the number of bins k, that is, divide the data into k intervals (k is 12 in this specific embodiment), calculate the total data volume n and the data capacity n / k of each interval (n is 64572 in this specific embodiment). Further determine the specific boundary value of each interval and allocate data to specific intervals according to the number of bins. After the binning is completed, the proportion of good customers and bad customers in each bin is counted, that is and Among themGood i and Bad i are the number of good customers and bad customers in the ith interval, and Good sum and Bad sum Represent the total number of good customers and bad customers in the data set respectively. The weight of evidence (woe) is calculated later. The calculation formula is: Check whether the woe value in each bin satisfies monotonicity. If not, return to the binning step and re-bin. When there are only good customers or bad customers in a bin, the woe value needs to be repaired. The common method is to add 0.5 to both the numerator and the denominator, that is,
[0093] Step S4: Horizontal division. According to the dataset processed by feature engineering in step S3, the dataset is further divided. The matched dataset is divided into training set and test set in a ratio of 8:2. The training set accounts for 80% and the test set accounts for 20%.
[0094] Step S5: Logistic regression model training, specifically including:
[0095] Step S51: configure the model.
[0096] Initialize the logistic regression model. Set the model's initial parameters, expansion coefficient (representing encryption accuracy), error value (the allowable error value between the model prediction result and the actual result), optimization method (the optimization method used to minimize the loss function of the regression equation), weights of both parties involved, batch size (the number of samples selected for one training), maximum number of iterations (the maximum number of times the optimizer repeatedly iterates the training data before finding the best model weight), regularization method (whether to constrain the complexity of the model by adding penalty terms in the model), loss function, etc. In this specific embodiment, the model parameter initialization method is selected as random uniform initialization, and the initial weights of the model are initialized by randomly sampling values from a uniform distribution. The expansion coefficient of the model is selected as 1000, the error value is selected as 0, the optimization method is selected as Stochastic Gradient Descent with Momentum (SGDM), the weights of the participating parties are not adjusted, the batch size is selected as 128, the maximum number of iterations is 30, the regularization method is the L1 regularization method with a Lasso coefficient of 0.1, the loss function is logarithmic loss (Log Loss), for a single sample, the loss function value is: L(y,p) = -[ylog(p) + (1-y)log(1-p)], for the entire training set, the loss function is the average of all sample losses.
[0097] Step S52: Model definition. The goal of the logistic regression model is to output a predicted value, a predicted classification probability value p, which is converted by the Sigmoid function. Where z is a linear combination, z = θ0 + θ1x1 + θ2x2 + ... + θ m θ m , where θ0,θ1,θ2,...,θ m are model parameters, and x0,x1,x2,...,x m is the eigenvalue.
[0098] Step S53: Gradient calculation. In order to update the parameters, it is necessary to calculate the gradient of the loss function with respect to each parameter. The gradient is the slope of the loss function at the current parameter value, indicating the rate of change of the loss function in this direction. j The gradient is calculated as follows: in is the derivative of the Sigmoid function,
[0099] Step S54: Parameter update. According to the gradient calculated in step S53, the parameters are further updated according to the SGDM method. The specific calculation method is: θ t+1 =θ t -ηv t , where θt refers to the model parameters after the tth iteration, v t =βv t-1 +(1-β)▽J(θ t ) is the momentum of the tth iteration, β is the momentum decay coefficient, ▽J(θ t ) is the parameter θ t is the gradient calculated according to step S53, and η is the learning rate. In this specific embodiment, β takes a value of 0.1 and η takes a value of 0.1.
[0100] Step S55: Iterative training. The above steps S53 and S54 are executed cyclically until the set maximum number of iterations is completed and the training process is completed.
[0101] Step S6: Model evaluation. After the model training is completed, the model effect is analyzed through a series of evaluation indicators commonly used in machine learning. The method of the present invention trains a risk control score card model, so the model's KS value (Kolmogorov-Smirnov) and AUC value (Area Under Cureve) are mainly focused on for model performance analysis.
[0102] In this specific embodiment, the specific indicators of the model obtained by training are as follows:
[0103]
[0104] Furthermore, in a specific embodiment, in order to prove the necessity of introducing privacy computing technology, the model using data from both the government and the bank is compared with the performance of the model using only the bank's data. In addition to using only the data provided by the bank, all model settings and training processes are consistent with the privacy computing model. The final performance indicator comparison is shown in Figure 3 and Figure 4 It can be seen that the introduction of privacy computing and the use of bilateral data improves the performance of the model while ensuring security, especially its performance on the test set, proving that bilateral data modeling improves the generalization ability of the model and has practical application value.
[0105] Further observation of the proportion of bad customers in each interval after binning according to the predicted values, and observation of the risk control prediction ability of the model, the result graph is in Figure 5 middle.
[0106] Step S7: Model deployment and application. The trained model is saved in the privacy computing platform and deployed to the production environment for actual prediction tasks.
[0107] Finally, it should be noted that the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.
Claims
1. A method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model, characterized by: The method comprises the following steps: Step S1: Obtain the feature data required for modeling; obtain the relevant feature data of the two small and micro enterprises through database operations, perform data deduplication processing, and select relevant features from all features; Step S2: Privacy intersection: On the premise that the data of both parties are not shared, each participant encrypts and calculates its own data locally to generate a ciphertext result; Step S3: feature engineering; Further feature engineering is performed based on the dataset after the privacy intersection sample alignment is completed in step S2; Step S4: Horizontal division: further divide the data set according to the data set processed by feature engineering in step S3; Step S5: Logistic regression model training; Step S6: Model evaluation: After the model training is completed, the model effect is analyzed through evaluation indicators; Step S7: Model deployment application; save the trained model in the privacy computing platform and deploy it to the production environment for actual prediction tasks.
2. According to claim 1, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: Step 3 specifically includes: Step S31: Missing value filling: performing missing value filling; Step S32: Standard deviation normalization: Calculate the mean μ of each feature and its standard deviation σ, and further normalize each feature data point to have zero mean and unit variance; Step S33: Extreme value processing: determine the percentile you want to use, and use this percentile as the replacement value for the extreme value; specifically: after the data set is sorted in ascending order, the value at the x%*nth position will replace the extreme value, where x% is the selected percentile and n is the total amount of data in the data set; for each feature, use a statistical method to calculate the value corresponding to the selected percentile, and replace the identified extreme value with the calculated percentile value; Step S34: evidence weight conversion; Perform Woe conversion of evidence weight value: first, all features are binned (discretized); The binning strategy used is equal frequency binning, which groups continuous variables into a series of intervals and ensures that the number of data points in each box is equal. The specific process is as follows: sort the data in ascending order, determine the number of bins k, divide the data into k intervals, calculate the total data volume n and the data capacity n / k of each interval; determine the specific boundary value of each interval according to the number of bins and allocate data to specific intervals. After completing the binning, count the proportion of good customers and bad customers in each bin, that is, and Among themGood i and Bad i are the number of good customers and bad customers in the ith interval, and Good sum and Bad sum Respectively represent the total number of good customers and bad customers in the data set, and then calculate the evidence weight value. The calculation formula is: Check whether the woe value in each bin satisfies monotonicity. If not, return to the binning step for re-binning. When there are only good customers or bad customers in a bin, the woe value needs to be repaired. The common method is to add 0.5 to both the numerator and the denominator, that is, 3. A method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model according to claim 1 or 2, characterized in that: Step 5 specifically includes: Step S51: configure the model; initialize the logistic regression model, set the model's initial parameters, expansion coefficient, error value, optimization method, weights of both parties involved, batch size, maximum number of iterations, regularization method, and loss function; Step S52: Model definition: The goal of the logistic regression model is to output a predicted value, a predicted classification probability value p, which is obtained by converting the Sigmoid function Where z is a linear combination, z = θ0 + θ1x1 + θ2x2 + ... + θ m θ m , where θ0,θ1,θ2,...,θ m are model parameters, and x0,x1,x2,...,x m is the eigenvalue; Step S53: Gradient calculation: In order to update the parameters, it is necessary to calculate the gradient of the loss function with respect to each parameter. The gradient is the slope of the loss function at the current parameter value, which indicates the rate of change of the loss function in this direction. For parameter θ j The gradient is calculated as follows: in is the derivative of the Sigmoid function, Step S54: Parameter update: According to the gradient calculated in step S53, the parameters are further updated according to the SGDM method. The specific calculation method is: θ t+1 =θ t -ηv t , where θ t refers to the model parameters after the tth iteration, v t =βv t-1 +(1-β)▽J(θ t ) is the momentum of the tth iteration, β is the momentum decay coefficient, ▽J(θ t ) is the parameter θ t The gradient calculated according to step S53 is here, and η is the learning rate; Step S55: Iterative training: cyclically execute the above steps S53 and S54 until the set maximum number of iterations is completed and the training process is completed.
4. According to claim 1, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: In step 2, the participants exchange encrypted calculation results through the secure network dedicated line. Without exposing their own original data, they align samples based on the matching fields of both parties, retaining the companies that exist in both parties' data sets. At the same time, both parties generate a matching column match_id to match the different features of the same company on both sides.
5. According to claim 2, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: The method adopted in step 31 is mean filling, that is, the missing values of the feature column are filled by calculating the mean of the non-missing values of the feature column.
6. According to claim 1, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: In step 4, the matched data set is divided into a training set and a test set in a ratio of 8:2, with the training set accounting for 80% and the test set accounting for 20%.
7. According to claim 1, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: In step 6, the analysis focus values include the KS value and AUC value of the focus model to perform model performance analysis.
8. According to claim 2, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: The minimum value replacement percentile selected in step S33 is 1%, and the maximum value replacement percentile is 99%.
9. According to claim 3, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: In step S51, the parameter initialization method of the model is selected as random uniform initialization, and the initial weight of the model is initialized by randomly sampling values from a uniform distribution. The expansion coefficient of the model is selected as 1000, the error value is selected as 0, the optimization method is selected as random gradient descent with momentum, the weights of the two parties involved are not adjusted, the batch size is selected as 128, the maximum number of iterations is 30, the regularization method is selected as the L1 regularization method with a Lasso coefficient of 0.1, and the loss function is logarithmic loss. For a single sample, the loss function value is: L(y,p)=-[ylog(p)+(1-y)log(1-p)], and for the entire training set, the loss function is the average of all sample losses.
10. According to claim 3, a method for identifying risks of small and micro enterprises based on privacy computing and logistic regression model is characterized by: In step S54, β takes a value of 0.1 and η takes a value of 0.1.