A retail financial big data credit risk control method and system

Abstract

The application provides a retail financial big data credit risk control method and system, and the method comprises the following steps: step 1, collecting modeling sample data sets in a retail financial big data risk control scene, and constructing a credit risk control model according to the modeling sample data sets; step 2, iteratively training the credit risk control model based on a target method, solving model parameters, and automatically optimizing the credit risk control model according to the model parameters to obtain an optimized model; step 3, comprehensively optimizing the optimized model to obtain a target credit risk control model; and step 4, collecting user information in real time, transmitting the real-time collected user information to the target credit risk control model for analysis, and outputting a credit risk control evaluation result of the user; the stability of the obtained target credit risk control model is ensured, and convenience and accuracy are provided for subsequent analysis of user information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a retail financial big data credit risk control method and system. BACKGROUND

[0002] At present, in the field of retail finance, the application of big data risk control and machine learning methods is increasingly popular, and traditional risk control and credit evaluation methods have been difficult to cope with the massive and complex data demand; by introducing advanced machine learning algorithms (such as logistic regression, decision tree, support vector machine, etc.), financial institutions can establish more accurate risk assessment models, thereby improving the accuracy of credit decision-making;

[0003] However, complex machine learning models, although having strong prediction ability, often lack transparency, and the interpretability of the model is crucial in financial risk control; in addition, in the case where data distribution and quality may change over time, the stability of the model is a key factor to ensure its continuous role in practical application;

[0004] Therefore, in order to overcome the above technical problems, the present application provides a retail financial big data credit risk control method and system SUMMARY

[0005] The present application provides a retail financial big data credit risk control method and system, which can effectively construct a credit risk control model by collecting modeling sample data sets in a retail financial big data risk control scenario, thereby performing model iterative training on the credit risk control model through a target method, which can effectively solve the model parameters, and further facilitate the automatic optimization of the credit risk control model, thereby obtaining an optimized model for comprehensive optimization, thereby ensuring the stability of the target credit risk control model, and providing convenience and accuracy for subsequent analysis of user information.

[0006] A retail financial big data credit risk control method, comprising:

[0007] Step 1: Collecting modeling sample data sets in a retail financial big data risk control scenario, and constructing a credit risk control model according to the modeling sample data sets;

[0008] Step 2: performing model iterative training on the credit risk control model based on a target method, solving the model parameters, and automatically optimizing the credit risk control model according to the model parameters, and obtaining an optimized model;

[0009] Step 3: Comprehensive optimization of the optimized model to obtain a target credit risk control model;

[0010] Step 4: Real-time collection of user information, and transmission of the real-time collected user information to the target credit risk control model for analysis, and output of the credit risk control evaluation result of the user.

[0011] Preferably, in a retail financial big data credit risk control method, step 1 involves collecting a modeling sample dataset for a retail financial big data risk control scenario, including:

[0012] After collecting the modeling sample dataset for retail financial big data risk control scenarios, the target variables and predictor variables of the modeling sample dataset are determined. The specific process is as follows:

[0013] Read the machine learning tasks in the financial big data risk control scenario in the retail sector, and determine the target variables of the modeling sample dataset based on the machine learning tasks;

[0014] Obtain the remaining variables from the modeling sample dataset, excluding the target variable;

[0015] Obtain the variable attributes of the target variable, analyze the variable attributes, identify the variables related to the target variable, and use the variables related to the target variable as the variables to be predicted;

[0016] The variables to be predicted are filtered to obtain the predictor variables.

[0017] Preferably, in a retail financial big data credit risk control method, step 1 involves constructing a credit risk control model based on a modeling sample dataset, including:

[0018] The modeling sample dataset is divided into a training set and a test set according to a preset ratio;

[0019] Obtain the initial model;

[0020] The initial model is trained based on the training set to obtain the credit risk control model. The model parameters of the credit risk control model are extracted, and the first evaluation index value of the credit risk control model is determined based on the model parameters.

[0021] The test set is input into the credit risk control model, and the second evaluation index value of the credit risk control model is determined based on the output results.

[0022] The first evaluation index value is compared with the second evaluation index value, and the optimization strategy for the credit risk control model is determined based on the comparison results.

[0023] The credit risk control model is optimized according to the optimization strategy to obtain the final credit risk control model.

[0024] Preferably, in a retail financial big data credit risk control method, step 2 involves iteratively training the credit risk control model based on the target method, solving for the model parameters, and automatically optimizing the credit risk control model based on the model parameters to obtain an optimized model, including:

[0025] Read the predictor variables from the modeling sample dataset, bin the predictor variables to obtain variable bins, and perform one-hot encoding on the predictor variables of the variable bins to obtain a one-hot encoded wide table of predictor variables.

[0026] The coefficient vector in the one-hot encoded wide table is solved according to the objective method, and the score of each variable bin is determined according to the coefficient vector, where each coefficient in the coefficient vector corresponds to the score of a variable bin.

[0027] Set optimization objectives based on the scores of variable binning;

[0028] Based on the needs of retail financial big data risk control business scenarios, constraints are determined, and the credit risk control model is iteratively trained according to the objective method based on the optimization objective. In the process of model iterative training, the objective function of the constraints is solved by convex optimization to obtain the model parameters.

[0029] The credit risk control model is automatically optimized based on the model parameters. At the same time, the maximum number of iterations for iterative training is obtained, and the iterative training stops when the maximum number of iterations is reached, thus obtaining the optimized model.

[0030] Preferably, a retail financial big data credit risk control method involves solving the coefficient vector in a one-hot encoded wide table according to the objective method, including:

[0031] Obtain positive sample data for users who repay on time, and at the same time, obtain negative sample data for users who are overdue.

[0032] Calculate the first mean of the positive sample data after linear transformation in the sample space and the second mean of the negative sample data after linear transformation in the sample space. At the same time, calculate the square of the difference between the first mean and the second mean to obtain the first calculated value.

[0033] Calculate the first covariance of the positive sample data in the sample space after linear transformation, and at the same time, calculate the second covariance of the negative sample data in the sample space after linear transformation. Sum the first covariance and the second covariance to obtain the second calculated value.

[0034] Obtain the target ratio between the first calculated value and the second calculated value;

[0035] The objective method maximizes the target ratio and solves the coefficient vector in the one-hot encoded wide table based on the maximization result.

[0036] Preferably, in a retail financial big data credit risk control method, step 3 involves comprehensively optimizing the model to obtain the target credit risk control model, including:

[0037] The optimization model is first fine-tuned based on the optimization evaluation index.

[0038] A second optimization is performed based on the contribution of each input variable in the optimization model to the optimization objective;

[0039] The target credit risk control model is obtained based on the first and second optimization results.

[0040] Preferably, a retail financial big data credit risk control method involves first-stage optimization of the optimization model based on optimization evaluation indicators, including:

[0041] Obtain the optimization evaluation index for the optimized model, and calculate the optimization evaluation index value of the credit score output by the optimized model on the training set and the optimization evaluation index value of the credit score output by the optimized model on the test set.

[0042] Obtain the difference between the optimized evaluation index value of the credit score output by the optimized model on the training set and the optimized evaluation index value of the credit score output by the optimized model on the test set, and determine the difference threshold.

[0043] The difference is compared with the difference threshold to determine whether the optimization model is stable;

[0044] If the difference is greater than or equal to the difference threshold, the optimized model is determined to be unstable. At the same time, the risk trends of IV and WoE of the unmodeled variables on the training and test sets are read, and the unmodeled variables with the closest risk trends are extracted for variable modeling and the optimized model is first tuned until the optimized model is stable.

[0045] Otherwise, the optimization model is considered stable.

[0046] Preferably, a retail financial big data credit risk control method includes a second optimization based on the contribution of each input variable in the optimization model to the optimization objective, comprising:

[0047] Calculate the first contribution of the input variables to the optimization objective on the training set;

[0048] Calculate the second contribution of the input variables to the optimization objective on the validation set;

[0049] Obtain the contribution difference between the first contribution and the second contribution, and obtain the contribution difference threshold;

[0050] The contribution difference is compared with the contribution difference threshold to determine whether the input variable is stable.

[0051] If the contribution difference is greater than the contribution difference threshold, the input variable is determined to be unstable. At the same time, the variable binning corresponding to the input variable is determined and adjusted to complete the second tuning of the optimization model.

[0052] Preferably, in a retail financial big data credit risk control method, step 4 involves real-time collection of user information, transmission of this information to the target credit risk control model for analysis, and output of a credit risk control assessment result for the user, including:

[0053] Collect user information, obtain the corresponding information text, and at the same time, obtain the credit identifier, locate the information text based on the credit identifier, and determine the key information text corresponding to the credit identifier.

[0054] Obtain the execution objective of the target credit risk control model, and locate the data in the key information text according to the execution objective to determine the key information data corresponding to the execution objective;

[0055] Key information data is input into the target credit risk control model for analysis, and the credit risk control assessment results for users are output based on the analysis results.

[0056] An assessment report is generated based on the credit risk control assessment results. At the same time, the assessment report is segmented into sub-assessment report segments.

[0057] The sub-evaluation report segment is read to determine the sub-report data sequence. A first preset encryption character is added to each sub-report data sequence according to a preset field interval to obtain the first encrypted ciphertext of each sub-evaluation report segment.

[0058] Add a second preset encryption character between the first encrypted ciphertext to obtain the second encrypted ciphertext;

[0059] Package the second encrypted ciphertext to obtain an encrypted data packet, and at the same time, obtain report read permissions;

[0060] The authorized terminal is determined based on the report read permissions, and the encrypted data packet is transmitted to the authorized terminal.

[0061] A retail financial big data credit risk control system includes:

[0062] The model building module is used to collect modeling sample datasets in retail financial big data risk control scenarios and build credit risk control models based on the modeling sample datasets;

[0063] The iterative training and optimization module is used to iteratively train the credit risk control model based on the objective method, solve the model parameters, and automatically optimize the credit risk control model based on the model parameters to obtain an optimized model.

[0064] The tuning module is used to comprehensively tune the optimization model to obtain the target credit risk control model;

[0065] The analysis module is used to collect user information in real time and transmit the collected user information to the target credit risk control model for analysis, and output the credit risk control assessment results for the user.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] By collecting modeling sample datasets from retail financial big data risk control scenarios, credit risk control models can be effectively constructed. Through iterative training of these models using a target method, the model parameters can be effectively solved, facilitating automatic optimization. This allows for comprehensive tuning of the optimized model, ensuring its stability and providing convenience and accuracy for subsequent analysis of user information.

[0068] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0069] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a flowchart of a retail financial big data credit risk control method according to an embodiment of the present invention;

[0072] Figure 2 This is a flowchart of step 1 in a retail financial big data credit risk control method according to an embodiment of the present invention;

[0073] Figure 3 This is a structural diagram of a retail financial big data credit risk control system according to an embodiment of the present invention. Detailed Implementation

[0074] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0075] In one embodiment, this embodiment provides a retail financial big data credit risk control method, such as... Figure 1 As shown, it includes:

[0076] Step 1: Collect a modeling sample dataset for retail financial big data risk control scenarios, and build a credit risk control model based on the modeling sample dataset;

[0077] Step 2: Iteratively train the credit risk control model based on the objective method, solve for the model parameters, and automatically optimize the credit risk control model based on the model parameters to obtain the optimized model;

[0078] Step 3: Perform comprehensive optimization of the model to obtain the target credit risk control model;

[0079] Step 4: Collect user information in real time and transmit the collected user information to the target credit risk control model for analysis, and output the credit risk control assessment results for the user.

[0080] In this embodiment, the target method can be the diversity method.

[0081] In this embodiment, comprehensive optimization of the model can be achieved by optimizing the model based on the overall model results and by calculating the contribution of each input variable in the model to the model optimization objective diversity value.

[0082] In this embodiment, user information may include user account, income level, credit record, and consumption behavior data.

[0083] In this embodiment, the optimized model is the model after automatic optimization of the credit risk control model, and the target credit risk control model can be the model obtained after comprehensive tuning of the optimized model.

[0084] The working principle and beneficial effects of the above technical solution are as follows: By collecting modeling sample datasets in retail financial big data risk control scenarios, a credit risk control model can be effectively constructed. Then, by using the target method to iteratively train the credit risk control model, the model parameters can be effectively solved, which is conducive to the automatic optimization of the credit risk control model. The optimized model can then be comprehensively tuned to ensure the stability of the obtained target credit risk control model, providing convenience and accuracy for subsequent analysis of user information.

[0085] In one embodiment, this embodiment provides a retail financial big data credit risk control method. Step 1 involves collecting a modeling sample dataset for a retail financial big data risk control scenario, including:

[0086] After collecting the modeling sample dataset for retail financial big data risk control scenarios, the target variables and predictor variables of the modeling sample dataset are determined. The specific process is as follows:

[0087] Read the machine learning tasks in the financial big data risk control scenario in the retail sector, and determine the target variables of the modeling sample dataset based on the machine learning tasks;

[0088] Obtain the remaining variables from the modeling sample dataset, excluding the target variable;

[0089] Obtain the variable attributes of the target variable, analyze the variable attributes, identify the variables related to the target variable, and use the variables related to the target variable as the variables to be predicted;

[0090] The variables to be predicted are filtered to obtain the predictor variables.

[0091] In this embodiment, the target variable may include at least values ​​representing normal repayment behavior and values ​​representing overdue behavior.

[0092] In this embodiment, the variable attributes of the target variable can be the degree of correlation between the target variable and other variables, the range of values ​​of the target variable, etc.

[0093] In this embodiment, the analysis of variable attributes can be performed by calculating the Pearson correlation coefficient to determine the degree of correlation through linear analysis, or by using regression analysis to determine which variables have a significant impact on the target variable.

[0094] In this embodiment, the screening of the variable to be predicted can be done by removing missing values, all negative numbers, zero values, and missing value samples from the variable to obtain the predictor.

[0095] In this embodiment, a target variable is determined, which contains at least two possible values. Machine learning tasks in retail financial big data credit risk control scenarios are generally binary classification tasks, primarily aimed at identifying high-quality customers with no delinquency risk and high-risk customers with a high probability of delinquency. Therefore, the target variable in the dataset must contain at least values ​​representing normal repayment behavior and delinquency behavior. That is, the dataset must contain at least two types of samples: normal repayment and delinquency. Predictor variables are then determined, which are used to distinguish between normal repayment behavior and delinquency behavior. These variables generally have a certain correlation with the target variable, commonly including transaction behavior-related variables and multi-risk behavior-related variables.

[0096] To ensure model stability, predictor variables are typically chosen based on their relatively stable mean over time. For interpretability, predictor variables are usually selected based on their monotonic trend in credit risk; if non-monotonic variables are chosen, the reasons for the non-monotonic change must be explainable from a business perspective. In retail financial big data credit risk control scenarios, a severe imbalance between positive and negative samples naturally exists because there are fewer users with overdue payments compared to those with regular repayments. When credit products have been on the market for a long time and have accumulated a large user base, modeling with the full sample size would consume significant computing resources. Therefore, lending institutions often retain all overdue samples (a relatively small sample size) and randomly sample users with regular repayments. Furthermore, the sampling ratio may differ for different types of credit products. This results in sample weights.

[0097] When sample weights are specified in the wide table modeling data, and when these sample weights are to be used during model training, the values ​​in the sample weight column must all be positive and there must be no missing values. All negative, zero, and missing value samples will be removed from the modeling samples.

[0098] The beneficial effects of the above technical solution are: it ensures the stability and interpretability of model training, thereby guaranteeing the quality of the credit risk control model.

[0099] In one embodiment, this embodiment provides a retail financial big data credit risk control method, such as... Figure 2 As shown, in step 1, a credit risk control model is constructed based on the modeling sample dataset, including:

[0100] S101: Divide the modeling sample dataset into a training set and a test set according to a preset ratio;

[0101] S102: Obtain the initial model;

[0102] S103: Train the initial model based on the training set to obtain the credit risk control model, extract the model parameters of the credit risk control model, and determine the first evaluation index value of the credit risk control model based on the model parameters.

[0103] S104: Input the test set into the credit risk control model and determine the second evaluation index value of the credit risk control model based on the output results;

[0104] S105: Compare the values ​​of the first evaluation index with the values ​​of the second evaluation index, and determine the optimization strategy for the credit risk control model based on the comparison results;

[0105] S106: Optimize the credit risk control model according to the optimization strategy to obtain the final credit risk control model.

[0106] In this embodiment, the preset ratio can be determined by the modeler based on business experience and the sample size of the modeling dataset. Common ratios include 6:4, 7:3, and 8:2.

[0107] In this embodiment, the first evaluation index value can be an index value determined by the model parameters extracted after training the initial model based on the training set, used to evaluate the credit risk control model.

[0108] In this embodiment, the second evaluation index value can be an index value for evaluating the credit risk control model based on the test set input into the credit risk control model and the output result. The first evaluation index and the second evaluation index are usually composed of Divergence, KS, AUC, etc.

[0109] In this embodiment, optimization strategies may include: If the evaluation metrics on the training set are good, but the metrics on the test set are significantly poor, this usually indicates an overfitting problem; the following optimization methods can be considered: Increase the amount of data: collect more relevant data to enhance the model's generalization ability. Enhance regularization: such as increasing the regularization coefficient. Reduce model complexity: for example, reduce the number of network layers or neurons; Data augmentation: use more diverse data augmentation techniques; Adjust the learning rate: try to lower the learning rate to make the model learn more smoothly; If the metrics on both the training and test sets are not ideal, it may be necessary to try different model architectures: the current model may not be suitable for this task, further optimization of hyperparameters may be needed, such as finer adjustment of parameters such as the learning rate; Check data quality and preprocessing: ensure that the data is accurate and properly preprocessed; Increase the number of training epochs: perhaps the model has not yet learned sufficiently. If the metrics on the training set are poor, but the metrics on the test set are slightly better, the model's learning ability may be insufficient, and the following can be considered: increase model complexity, check if there are any problems with the data.

[0110] In this embodiment, the modeling dataset is divided into training and test sets. Modelers determine the partitioning ratio based on business experience and the sample size of the dataset; common ratios include 6:4, 7:3, and 8:2. Model parameters are solved on the training set, and evaluation metrics are calculated. The main evaluation metrics include Divergence, KS, and AUC. The model obtained from the training set is then applied to the test set, and evaluation metrics are calculated again. The difference between the evaluation metrics (i.e., the first and second evaluation metric values) on the training and test sets is used to determine how to optimize the model.

[0111] The working principle and beneficial effects of the above technical solution are as follows: By dividing the training set and the test set, the performance of the model on different data can be better evaluated, improving the reliability and generalization ability of the model; obtaining the initial model and training it helps to build a model adapted to specific credit risk control tasks; by extracting model parameters and determining the first evaluation index value, the performance of the model can be quantitatively evaluated; by comparing the first evaluation index value with the second evaluation index value, the differences between the model on the training set and the test set can be analyzed more comprehensively, thereby formulating targeted optimization strategies; optimization based on the optimization strategy can continuously improve the accuracy and effectiveness of the credit risk control model, better cope with actual credit risk scenarios, and the final credit risk control model can more accurately identify risks, ensure the safe and stable operation of credit business, and reduce potential risk losses.

[0112] In one embodiment, this embodiment provides a retail financial big data credit risk control method. In step 2, the credit risk control model is iteratively trained based on the target method, and the model parameters are solved. The credit risk control model is then automatically optimized based on the model parameters to obtain an optimized model, including:

[0113] Read the predictor variables from the modeling sample dataset, bin the predictor variables to obtain variable bins, and perform one-hot encoding on the predictor variables of the variable bins to obtain a one-hot encoded wide table of predictor variables.

[0114] The coefficient vector in the one-hot encoded wide table is solved according to the objective method, and the score of each variable bin is determined according to the coefficient vector, where each coefficient in the coefficient vector corresponds to the score of a variable bin.

[0115] Set optimization objectives based on the scores of variable binning;

[0116] Based on the needs of retail financial big data risk control business scenarios, constraints are determined, and the credit risk control model is iteratively trained according to the objective method based on the optimization objective. In the process of model iterative training, the objective function of the constraints is solved by convex optimization to obtain the model parameters.

[0117] The credit risk control model is automatically optimized based on the model parameters. At the same time, the maximum number of iterations for iterative training is obtained, and the iterative training stops when the maximum number of iterations is reached, thus obtaining the optimized model.

[0118] In this embodiment, the optimization objective can be determined by setting a target value for optimizing the credit risk control model based on the scores of the variable binning.

[0119] The working principle of the above technical solution is as follows: Since the predictive variables used in the retail financial big data risk control scenario are often non-normally distributed, and in order to enhance the stability of the model and reduce the complexity of the model, it is necessary to bin the predictive variables and then perform one-hot encoding on the binned data to obtain a wide table with expanded dimensions; the parameter to be solved in the Divergence method is a vector that performs a linear transformation on the wide table after the one-hot encoding in the sample space, and maps the wide table to 1 dimension. Each element in this vector is the coefficient of each column of the wide table; when optimizing the Divergence objective, the Divergence method is also subject to constraints based on the needs of the retail financial big data risk control business scenario, including: (1) Symmetry constraint, that is, the parameters to be solved by the model, within each variable, satisfy the symmetry of the good and bad samples (weighted) equal distribution about 0, so that the scores within each variable can be positive and negative, and the credit score finally added at the user or sample level is also within a range, which is more in line with business needs; (2) (2) Binning monotonicity constraint: In order to make the model more interpretable and each input variable more interpretable, a monotonicity constraint will be applied to the bins of each variable at the variable level, so that the score of each variable level will show a trend that conforms to the business logic; (3) Correlation constraint: In the wide table, there may be two or more bins with highly similar or completely consistent values. For such bins, an equality constraint will be applied, so that such bins will have the same score, which conforms to the business logic; The iterative solution refers to the convex optimization solution of the objective function with constraints described above. Specifically, based on the diversity value obtained from the previous model training solution, for each non-input variable, its possible contribution value to the optimization objective diversity is calculated. The calculation method is: the non-input variable and the variable used in the previous model training are used to solve and calculate a diversity value. The difference between the non-input variable and the diversity value obtained from the previous model training is the possible contribution value of the non-input variable to the optimization objective. In the first iteration, the initial diversity value is 0. When the maximum number of iterations is reached, the iterative training stops, and the final model result is obtained.

[0120] The beneficial effects of the above technical solution are as follows: the diversity method, due to the inherent characteristics of the loss function, is naturally insensitive to sample imbalance, making it suitable for business scenarios such as retail financial big data credit risk control where positive and negative samples are extremely imbalanced; compared to nonlinear methods, the model has stronger interpretability at the business level and can more intuitively describe the relationship between the final model output (i.e., credit score) and variables; and the addition of constraints based on business needs logic makes the model solution more in line with the requirements of credit scoring in retail financial scenarios.

[0121] In one embodiment, this embodiment provides a retail financial big data credit risk control method, which solves for the coefficient vector in the one-hot encoded wide table according to the target method, including:

[0122] Obtain positive sample data for users who repay on time, and at the same time, obtain negative sample data for users who are overdue.

[0123] Calculate the first mean of the positive sample data after linear transformation in the sample space and the second mean of the negative sample data after linear transformation in the sample space. At the same time, calculate the square of the difference between the first mean and the second mean to obtain the first calculated value.

[0124] Calculate the first covariance of the positive sample data in the sample space after linear transformation, and at the same time, calculate the second covariance of the negative sample data in the sample space after linear transformation. Sum the first covariance and the second covariance to obtain the second calculated value.

[0125] Obtain the target ratio between the first calculated value and the second calculated value;

[0126] The objective method maximizes the target ratio and solves the coefficient vector in the one-hot encoded wide table based on the maximization result.

[0127] In this embodiment, the Divergence method refers to using the Divergence metric as the optimization objective to solve the binary classification optimization problem in big data retail financial credit risk control scenarios. The Divergence metric is a fraction, calculated as follows: the numerator is the heterogeneity between users with normal repayment (positive samples) and users with overdue payments (negative samples), which is the square of the difference between the means of the two types of samples after linear transformation in the sample space; the denominator is the homogeneity within each of the users with normal repayment (positive samples) and users with overdue payments (negative samples), which is calculated by calculating the covariance of each category of samples after linear transformation in the sample space and then summing them; the optimization objective is to maximize this Divergence metric, that is, to solve for a set of coefficient vectors so that the Divergence value calculated after the wide table data is linearly transformed by this set of coefficient vectors is maximized. The coefficients obtained are the scores of each variable bin.

[0128] The working principle and beneficial effects of the above technical solution are as follows: By acquiring positive sample data corresponding to users with normal repayment and obtaining negative sample data corresponding to overdue users, it is easier to calculate the first calculated value and the second calculated value. By determining the target ratio of the first calculated value and the second calculated value, it is easier to maximize the target ratio. Then, based on the maximization result, the coefficient vector in the one-hot encoded wide table is solved. By comprehensively and deeply analyzing the data characteristics, the ability to predict and analyze user repayment behavior is improved, providing a more reliable basis for financial decision-making, which helps to optimize risk management, improve business efficiency, and ensure the sound operation of financial business.

[0129] In one embodiment, this embodiment provides a retail financial big data credit risk control method. In step 3, the optimization model is comprehensively tuned to obtain the target credit risk control model, including:

[0130] The optimization model is first fine-tuned based on the optimization evaluation index.

[0131] A second optimization is performed based on the contribution of each input variable in the optimization model to the optimization objective;

[0132] The target credit risk control model is obtained based on the first and second optimization results.

[0133] In this embodiment, the optimization evaluation index can be a representation of the direction in which the optimization model needs to be tuned and the specific adjustment parameters corresponding to the adjustment.

[0134] In this embodiment, the input variables refer to the variables that are included in the model for analysis and prediction. They are selected, processed and analyzed, and then used as inputs for model construction and operation to help the model perform operations such as decision-making, classification, and regression, thereby enabling the optimization of the model.

[0135] In this embodiment, the degree of contribution is used to characterize the degree of optimization of the input variables to the optimization objective, so as to facilitate the corresponding optimization operation on the optimization model when the degree of contribution is insufficient.

[0136] The working principle and beneficial effects of the above technical solution are as follows: by optimizing the evaluation indicators and the contribution of each input variable to the optimization objective, the optimization model is first tuned and then tuned, respectively, ensuring the performance of the optimization model and thus providing a reliable guarantee for credit risk control assessment.

[0137] In one embodiment, this embodiment provides a retail financial big data credit risk control method, which performs a first optimization of the optimization model based on the optimization evaluation index of the optimization model, including:

[0138] The optimization model is first tuned based on the optimization evaluation index of the optimization model, including: obtaining the optimization evaluation index of the optimization model, and calculating the optimization evaluation index value of the credit score output by the optimization model on the training set and the optimization evaluation index value of the credit score output by the optimization model on the test set.

[0139] Obtain the difference between the optimized evaluation index value of the credit score output by the optimized model on the training set and the optimized evaluation index value of the credit score output by the optimized model on the test set, and determine the difference threshold.

[0140] The difference is compared with the difference threshold to determine whether the optimization model is stable;

[0141] If the difference is greater than or equal to the difference threshold, the optimized model is determined to be unstable. At the same time, the risk trends of IV and WoE of the unmodeled variables on the training and test sets are read, and the unmodeled variables with the closest risk trends are extracted for variable modeling and the optimized model is first tuned until the optimized model is stable.

[0142] Otherwise, the optimization model is considered stable.

[0143] In this embodiment, the model is tuned based on the overall model results. Specifically, this involves observing the KS or diversity of the credit score output by the model on the training and test sets. If the KS or diversity differs too much, it indicates that the model lacks stability and the input variables need to be adjusted. Variables with the largest potential contribution to the model and whose IV and WoE risk trends are closer on the training and test sets should be selected for input. The binning of the input variables can also be adjusted to make the WoE risk trends of the input variables closer on the training and test sets.

[0144] In this embodiment, the optimized evaluation index values ​​of the credit score output by the optimized model on the training set and the optimized evaluation index values ​​of the credit score output by the optimized model on the test set refer to the results obtained by the optimized model analyzing and processing the training set and the test set respectively, and the result value of the optimized evaluation index obtained by quantifying the results obtained by the analysis and processing through the optimized evaluation index.

[0145] In this embodiment, the difference threshold is set in advance and is used to compare and analyze the difference between the optimized evaluation index value of the credit score output by the optimized model on the training set and the optimized evaluation index value on the test set, thereby determining the stability of the optimized model.

[0146] The working principle and beneficial effects of the above technical solution are as follows: By determining the optimized evaluation index value of the credit score output by the optimized model on the training set and the optimized evaluation index value of the credit score output by the optimized model on the test set, and determining the difference between the two, the stability of the optimized model can be accurately and effectively determined based on the relationship between the difference and the difference threshold. At the same time, when the model is unstable, an effective and reliable first adjustment is made to ensure the stable operation of the optimized model and improve the reliability of credit risk control assessment.

[0147] In one embodiment, this embodiment provides a retail financial big data credit risk control method, which performs a second optimization based on the contribution of each input variable in the optimization model to the optimization objective, including:

[0148] Calculate the first contribution of the input variables to the optimization objective on the training set;

[0149] Calculate the second contribution of the input variables to the optimization objective on the validation set;

[0150] Obtain the contribution difference between the first contribution and the second contribution, and obtain the contribution difference threshold;

[0151] The contribution difference is compared with the contribution difference threshold to determine whether the input variable is stable.

[0152] If the contribution difference is greater than the contribution difference threshold, the input variable is determined to be unstable. At the same time, the variable binning corresponding to the input variable is determined and adjusted to complete the second tuning of the optimization model.

[0153] In this embodiment, the contribution of each input variable in the model to the objective diversity value of the model optimization is calculated. Specifically, a certain input variable is removed from the input variables, and the diversity value is calculated using the remaining input variables. The difference between this difference and the model's diversity value is the contribution of that input variable to the objective diversity value. The contribution of the input variables on the training and test sets is calculated separately. If the contribution differs significantly, it indicates that the input variable is not stable enough in the model, and the binning of the input variable needs to be adjusted to make its WoE risk trend more similar on the training and test sets. If it is impossible to make the WoE risk trend of the input variable more similar on the training and test sets, then the variable needs to be replaced.

[0154] In this embodiment, the first contribution can be the degree of contribution of the input variables on the training set to the optimization objective.

[0155] In this embodiment, the second contribution can be the degree of contribution of the input variables to the optimization objective on the validation set.

[0156] In this embodiment, the contribution difference threshold is known in advance and is the minimum standard used to measure whether the contribution difference between the first contribution and the second contribution meets the difference threshold requirement. It can be adjusted.

[0157] The beneficial effects of the above technical solution are as follows: By adjusting the model based on the model results, the potential contribution of unincorporated variables to the model is utilized during the model iteration training process. The binning of incorporated variables is adjusted, thereby making the risk level classification of the population represented by the variable binning more stable, making the model more stable. The contribution of incorporated variables to the model is calculated, and this contribution is linearly additive among incorporated variables, which more intuitively shows the correlation between incorporated variables and model performance. At the same time, by comparing the model horizontally on the training set and the test set, the stability of incorporated variables in the model is directly measured, further enhancing the interpretability of the model at the variable level.

[0158] In one embodiment, this embodiment provides a retail financial big data credit risk control method. In step 4, user information is collected in real time and transmitted to the target credit risk control model for analysis, outputting the credit risk control assessment result for the user, including:

[0159] Collect user information, obtain the corresponding information text, and at the same time, obtain the credit identifier, locate the information text based on the credit identifier, and determine the key information text corresponding to the credit identifier.

[0160] Obtain the execution objective of the target credit risk control model, and locate the data in the key information text according to the execution objective to determine the key information data corresponding to the execution objective;

[0161] Key information data is input into the target credit risk control model for analysis, and the credit risk control assessment results for users are output based on the analysis results.

[0162] An assessment report is generated based on the credit risk control assessment results. At the same time, the assessment report is segmented into sub-assessment report segments.

[0163] The sub-evaluation report segment is read to determine the sub-report data sequence. A first preset encryption character is added to each sub-report data sequence according to a preset field interval to obtain the first encrypted ciphertext of each sub-evaluation report segment.

[0164] Add a second preset encryption character between the first encrypted ciphertext to obtain the second encrypted ciphertext;

[0165] Package the second encrypted ciphertext to obtain an encrypted data packet, and at the same time, obtain report read permissions;

[0166] The authorized terminal is determined based on the report read permissions, and the encrypted data packet is transmitted to the authorized terminal.

[0167] In this embodiment, the information text can be the specific text content corresponding to the user information.

[0168] In this embodiment, the credit identifier can be applied to credit-related symbols and other markings, thereby facilitating the identification of key credit information in the information text.

[0169] In this embodiment, the key information text can be a data fragment in the information text that can represent the specific credit situation.

[0170] In this embodiment, the execution target can be the analysis object corresponding to the target credit risk control model, that is, the data object that can be processed.

[0171] In this embodiment, the key information data can be the result of extracting key fields from the key information text according to the execution target.

[0172] In this embodiment, segmentation processing can be to split the evaluation report into multiple different data segments, that is, to split the evaluation report into sub-evaluation report segments.

[0173] In this embodiment, the sub-report data sequence can be the data sequence corresponding to different sub-evaluation report segments obtained by reading the contents of different sub-evaluation report segments.

[0174] In this embodiment, the preset field interval is set in advance to limit the interval at which the first preset encrypted character is added to the sub-report data sequence. The first preset encrypted character is known in advance and is used to encrypt the sub-report data sequence, that is, to add the first preset encrypted character to the sub-report data sequence. The result of adding the first preset encrypted character is the first encrypted ciphertext, which is the result obtained after encrypting the sub-report data sequence.

[0175] In this embodiment, the second preset encryption character can be the data required to encrypt the first encrypted ciphertext again, that is, the second preset encryption character is added to the first encrypted ciphertext, and the result is the second encrypted ciphertext.

[0176] In this embodiment, report reading permission can be a device terminal that represents the ability to view and read credit risk control assessment results.

[0177] The working principle and beneficial effects of the above technical solution are as follows: By processing the collected user information, key information data related to credit is extracted from the corresponding text information of the user information, providing reliable data support for credit risk control assessment. Secondly, the obtained key information data is input into the target credit risk control model for analysis, so as to accurately and effectively obtain the credit risk control assessment results. Finally, the credit risk control assessment results are used to generate an assessment report, and the assessment report is encrypted to ensure the security and reliability of the credit risk assessment results. Furthermore, the encrypted data packet is transmitted to the corresponding authorized terminal according to the report reading permissions, which improves the reliability of credit risk control assessment and ensures the security of assessment results.

[0178] In one embodiment, this embodiment provides a retail financial big data credit risk control system, such as... Figure 3 As shown, it includes:

[0179] The model building module is used to collect modeling sample datasets in retail financial big data risk control scenarios and build credit risk control models based on the modeling sample datasets;

[0180] The iterative training and optimization module is used to iteratively train the credit risk control model based on the objective method, solve the model parameters, and automatically optimize the credit risk control model based on the model parameters to obtain an optimized model.

[0181] The tuning module is used to comprehensively tune the optimization model to obtain the target credit risk control model;

[0182] The analysis module is used to collect user information in real time and transmit the collected user information to the target credit risk control model for analysis, and output the credit risk control assessment results for the user.

[0183] The working principle and beneficial effects of the above technical solution are as follows: By collecting modeling sample datasets in retail financial big data risk control scenarios, a credit risk control model can be effectively constructed. Then, by using the target method to iteratively train the credit risk control model, the model parameters can be effectively solved, which is conducive to the automatic optimization of the credit risk control model. The optimized model can then be comprehensively tuned to ensure the stability of the obtained target credit risk control model, providing convenience and accuracy for subsequent analysis of user information.

[0184] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A retail financial big data credit risk control method, characterized in that, The method comprises the following steps: Step 1: Collecting modeling sample data set in the retail financial big data risk control scene, and constructing a credit risk control model according to the modeling sample data set; Step 2: Iterative training of the credit risk control model based on the target method, solving the model parameters, and automatically optimizing the credit risk control model according to the model parameters to obtain an optimized model; Step 3: Comprehensive tuning of the optimized model to obtain a target credit risk control model; Step 4: Real-time collection of user information, and transmission of the real-time collected user information to the target credit risk control model for analysis to output a credit risk control evaluation result of the user; In step 2, the credit risk control model is iteratively trained based on the target method, and the model parameters are solved, and the credit risk control model is automatically optimized according to the model parameters to obtain an optimized model, comprising: Reading the prediction variables in the modeling sample data set, and performing binning processing on the prediction variables to obtain variable bins, and simultaneously performing one-hot encoding on the variable bins of the prediction variables to obtain a one-hot encoding wide table of the prediction variables; Solving the coefficient vector in the one-hot encoding wide table according to the target method, and determining the score of each variable bin according to the coefficient vector, wherein one coefficient in the coefficient vector corresponds to the score of one variable bin; Setting an optimization target according to the score of the variable bin; Determining the constraint condition based on the demand of the retail financial big data risk control business scene, and iteratively training the credit risk control model based on the optimization target according to the target method, wherein the objective function of the constraint condition is solved by convex optimization in the process of model iterative training to obtain the model parameters; According to the model parameters, the credit risk control model is automatically optimized, and the maximum number of iterations of the iterative training is obtained, and when the maximum number of iterations is reached, the iterative training stops, and an optimized model is obtained.

2. The retail financial big data credit risk control method of claim 1, wherein, In step 1, the modeling sample data set in the retail financial big data risk control scene is collected, comprising: After collecting the modeling sample data set in the retail financial big data risk control scene, the target variable and the prediction variable of the modeling sample data set are determined, and the specific process is as follows: Reading the machine learning task in the retail field financial big data risk control scene, and determining the target variable of the modeling sample data set according to the machine learning task; Obtaining the remaining variables in the modeling sample data set except the target variable; Obtaining the variable attribute of the target variable, and analyzing the variable attribute to determine the variables related to the target variable, and taking the variables related to the target variable as the predicted variables; Screening the predicted variables to obtain the prediction variables.

3. The retail financial big data credit risk control method of claim 1, wherein, In step 1, the credit risk control model is constructed according to the modeling sample data set, comprising: Dividing the modeling sample data set into a training set and a test set according to a preset proportion; Obtaining an initial model; Training the initial model according to the training set to obtain a credit risk control model, extracting the model parameters of the credit risk control model, and determining the first evaluation index value of the credit risk control model according to the model parameters; Inputting the test set into the credit risk control model, and determining the second evaluation index value of the credit risk control model according to the output result; Comparing the first evaluation index value with the second evaluation index value, and determining the tuning strategy for optimizing the credit risk control model according to the comparison result; According to the tuning strategy, the credit risk control model is tuned to obtain a final credit risk control model.

4. The retail financial big data credit risk control method of claim 1, wherein, According to the target method, the coefficient vector in the one-hot encoding wide table is solved, including: Obtain the positive sample data corresponding to the normal repayment user, and obtain the negative sample data corresponding to the overdue user at the same time; Calculate the first mean value of the positive sample data after linear transformation in the sample space and the second mean value of the negative sample data after linear transformation in the sample space, and at the same time, calculate the square of the difference between the first mean value and the second mean value to obtain the first calculation value; Calculate the first covariance of the positive sample data after linear transformation in the sample space, and at the same time, calculate the second covariance of the negative sample data after linear transformation in the sample space, and sum the first covariance and the second covariance to obtain the second calculation value; Obtain the target ratio between the first calculation value and the second calculation value; Based on the target method, the target ratio is maximized, and the coefficient vector in the one-hot encoding wide table is solved based on the maximization result.

5. The retail financial big data credit risk control method of claim 1, wherein, In step 3, the optimization model is comprehensively tuned to obtain a target credit risk control model, including: According to the optimization evaluation index of the optimization model, the optimization model is first tuned; According to the contribution degree of each input variable in the optimization model to the optimization target, the optimization model is secondly tuned; Based on the first tuning result and the second tuning result, the target credit risk control model is obtained.

6. The retail financial big data credit risk control method of claim 5, wherein, According to the optimization evaluation index of the optimization model, the optimization model is first tuned, including: Obtain the optimization evaluation index of the optimization model, and calculate the optimization evaluation index value of the credit score output by the optimization model on the training set and the optimization evaluation index value of the credit score output by the optimization model on the test set; Obtain the difference between the optimization evaluation index value of the credit score output by the optimization model on the training set and the optimization evaluation index value of the credit score output by the optimization model on the test set, and determine the difference threshold value; Compare the difference with the difference threshold value to determine whether the optimization model is stable; If the difference is greater than or equal to the difference threshold value, it is determined that the optimization model is unstable, and at the same time, the risk trend of the IV and WoE of the non-input variable on the training set and the test set is read, and the non-input variable with the closest risk trend is selected for variable input and the optimization model is first tuned until the optimization model is stable; Otherwise, it is determined that the optimization model is stable.

7. The retail financial big data credit risk control method of claim 5, wherein, According to the contribution degree of each input variable in the optimization model to the optimization target, the optimization model is secondly tuned, including: Calculate the first contribution degree of the input variable to the optimization target on the training set; Calculate the second contribution degree of the input variable to the optimization target on the validation set; Obtain the contribution difference between the first contribution degree and the second contribution degree, and obtain the contribution difference threshold value; Compare the contribution difference with the contribution difference threshold value to determine whether the input variable is stable; If the contribution difference is greater than the contribution difference threshold value, it is determined that the input variable is unstable, and at the same time, the variable bin corresponding to the input variable is determined, and the variable bin corresponding to the input variable is adjusted to complete the second tuning of the optimization model.

8. The retail financial big data credit risk control method of claim 1, wherein, In step 4, real-time user information is collected, and the real-time collected user information is transmitted to the target credit risk control model for analysis to output the credit risk control evaluation result of the user, including: Collect user information, obtain information text corresponding to the user information, and simultaneously obtain a credit identifier. The credit identifier is positioned in the information text to determine key information text corresponding to the credit identifier; An execution target of the target credit risk control model is obtained, and data positioning is performed in the key information text according to the execution target to determine key information data corresponding to the execution target; The key information data is input into the target credit risk control model for analysis, and a credit risk control evaluation result of the user is output according to the analysis result; An evaluation report is generated according to the credit risk control evaluation result, and the evaluation report is segmented into sub-evaluation report segments; The sub-evaluation report segments are read to determine sub-report data sequences, a first preset encryption character is added in each sub-report data sequence according to a preset field interval, and a first encrypted ciphertext of each sub-evaluation report segment is obtained; A second preset encryption character is added between the first encrypted ciphertexts to obtain a second encrypted ciphertext; The second encrypted ciphertext is packaged to obtain an encrypted data packet, and a report reading permission is obtained; An authorized terminal is determined according to the report reading permission, and the encrypted data packet is transmitted to the authorized terminal.

9. A retail financial big data credit risk control system, characterized in that, Comprise: The model construction module is used for collecting modeling sample data set in the retail financial big data risk control scene, and constructing the credit risk control model according to the modeling sample data set; The iterative training optimization module is used for model iterative training of the credit risk control model based on the target method, solving the model parameters, and automatically optimizing the credit risk control model according to the model parameters to obtain the optimized model; The tuning module is used for comprehensive tuning of the optimized model to obtain the target credit risk control model; The analysis module is used for real-time collection of user information, and the real-time collected user information is transmitted to the target credit risk control model for analysis to output the credit risk control evaluation result of the user; In the iterative training optimization module, the credit risk control model is iteratively trained based on the target method, the model parameters are solved, and the credit risk control model is automatically optimized according to the model parameters to obtain the optimized model, comprising: The prediction variables in the modeling sample data set are read and the prediction variables are binned to obtain variable bins, and the prediction variables of the variable bins are one-hot encoded to obtain a one-hot encoding wide table of the prediction variables; The coefficient vector in the one-hot encoding wide table is solved according to the target method, and the score of each variable bin is determined according to the coefficient vector, wherein one coefficient in the coefficient vector corresponds to the score of one variable bin; The optimization target is set according to the score of the variable bin; The constraint condition is determined based on the demand of the retail financial big data risk control business scene, and the credit risk control model is iteratively trained based on the optimization target according to the target method, wherein the objective function of the constraint condition is convexly optimized and solved in the model iterative training process to obtain the model parameters; The credit risk control model is automatically optimized according to the model parameters, the maximum number of iterations of the iterative training is obtained, and the iterative training is stopped when the maximum number of iterations is reached to obtain the optimized model.

Images