Multi-dimensional scoring credit data risk prediction control method and system

Abstract

The invention provides a multi-dimensional scoring credit data risk prediction control method and system, and the method comprises the steps: building a KGB model and an AR model on a known label sample set, and obtaining a KGB model score and an AR model score; expanding each tag sample needing to be deduced to be good or bad in the tag sample set needing to be deduced to be good or bad to be fuzzy expanded good samples and fuzzy expanded bad samples; setting the sum of the fuzzy expanded well sample weight and the fuzzy expanded bad sample weight to be equal to the label sample weight needing to be deduced; carrying out weight operation distribution by using the score of the KGB model and the score of the AR model; and establishing an AGB model on a total sample of the tag sample set needing to deduce the quality, starting iterative operation, judging whether the AGB model has similar prediction performance on different types of customer groups, and performing iterative operation until the convergence of an inference result is achieved, thereby obtaining a final risk prediction inference result.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing financial risk prediction and control, and more specifically, to a multi-dimensional scoring credit data risk prediction and control method and system. Background Art

[0002] In the entire credit chain, the risk control system is like a sieve that filters out relatively high-quality customers layer by layer, and finally decides whether to lend. For 100% of the full application orders, the actual loan orders may only account for 10%, or even lower, which actually depends on the risk control strategy of each lending institution. The subsequent risk control model may be established based on a loan sample that only accounts for 10% of the full application user data. This leads to the problem of survivor bias, ignoring the comprehensiveness of the sample, replacing the overall sample with a local sample, and causing a deviation in the description of the population, resulting in inaccurate estimation of model parameters, thereby affecting the judgment of real risks; in order to solve this problem, rejection inference came into being; the fundamental idea of ​​rejection inference is to assume that if the rejected order is loaned, then the post-loan performance of this order is good or bad, and then the rejected inference good and bad samples (Inferred Good / Bad) are added to the modeling samples to enrich the diversity of the modeling samples, thereby narrowing the difference between the training sample and the population, and reducing the possibility of underestimating the risk of the full sample;

[0003] At present, rejection inference usually uses KGB model scores to infer labels. How to combine AR model scores based on the commonly used KGB model scores, how to establish AR models based on the information of rejected borrowers, how to fully understand the risk characteristics of customers, including potential high-risk customers who have failed to successfully apply for loans, to build a more accurate inference prediction model to optimize customer screening strategies, how to include rejected customers who actually have good repayment ability and low default risk into the service scope to expand the customer base and improve business diversity, etc., have yet to be solved; therefore, it is necessary to propose a multi-dimensional scoring credit data risk prediction control method and system to at least partially solve the problems existing in the prior art. Summary of the invention

[0004] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0005] In order to at least partially solve the above problems, the present invention provides a multi-dimensional scoring credit data risk prediction control method, comprising: S1, establishing a KGB model and an AR model on a known label sample set, and obtaining a KGB model scoring score and an AR model scoring score respectively;

[0006] S2, obtaining a set of samples of good and bad labels to be inferred, and using the fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample;

[0007] S3, set the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establish the cross-correlation relationship between the inference good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; use the KGB model score and the AR model score to perform weight calculation and allocation;

[0008] S4, build the AGB model on all samples of the set of good and bad label samples to be inferred, start iterative calculations, determine whether the AGB model has similar prediction performance on different types of customer groups, and iterate until the inference results converge to obtain the final inference result of risk prediction.

[0009] Preferably, S1 comprises:

[0010] S101, establish a KGB model on samples with known risk performance labels of good and bad, and obtain a KGB model score;

[0011] S102, establishing an AR model on all samples with known rejection labels, and obtaining an AR model scoring score.

[0012] Preferably, S2 comprises:

[0013] S201, obtaining a set of samples of good and bad labels to be inferred; using a fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1;

[0014] S202, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establishing a cross-correlation relationship between the inferred good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and using the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0015] The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

[0016] Preferably, S3 includes:

[0017] S301, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; S302, establishing the cross-correlation relationship between the inferred good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; using the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0018] The cross-correlation relationship between the good and bad probabilities is established based on the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively, including: on samples with good and bad risk performance, 0 and 1 are used as good sample labels and bad sample labels respectively, and the logistic regression model is trained with the KGB model score and the AR model score to obtain the relationship between the overdue probability and the KGB model score and the AR model score, and the Logistic Regression Overdue Probability Prediction Relationship; the Logistic Regression Overdue Probability Prediction Relationship is applied to the rejected samples and the unwithdrawn samples, and the probability of each credit data being inferred to be bad is obtained accordingly; the weight of the sample that needs to be inferred is multiplied by the inferred bad probability to obtain the extended bad sample weight, and similarly, the weight of the sample that needs to be inferred is multiplied by the predicted good probability to obtain the extended good sample weight.

[0019] Preferably, S4 includes:

[0020] S401, establishing an AGB model on all samples of the set of good and bad label samples to be inferred, and performing iterative operations; the iterative operations include: setting a maximum number of iterations, and stopping the iterative operations when the maximum number of iterations is reached or the iteration results converge;

[0021] S403, determine whether the AGB model has similar prediction performance on different types of customer groups; when the AGB model has similar prediction performance on different types of customer groups, the inference results converge and the final inference result of risk prediction is obtained; when the AGB model does not have similar prediction performance on different types of customer groups, re-weight allocation and iterative calculation are performed until the inference results converge and the final inference result of risk prediction is obtained.

[0022] The present invention provides a multi-dimensional scoring credit data risk prediction control system, comprising:

[0023] The sample model multi-dimensional scoring system builds the KGB model and AR model on the known label sample set, and obtains the KGB model scoring score and AR model scoring score respectively;

[0024] The inference sample fuzzy expansion subsystem obtains a set of samples with good and bad labels to be inferred, and uses the fuzzy expansion method to expand each sample with good and bad labels to be inferred in the set of samples with good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample;

[0025] The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0026] The weighted distribution multi-model inference subsystem establishes an AGB model on the full set of samples of the good and bad label sample sets to be inferred, and starts iterative operations to determine whether the AGB model has similar prediction performance on different types of customer groups. It iterates until the inference results converge to obtain the final inference result of risk prediction.

[0027] Preferably, the sample model multidimensional scoring system includes:

[0028] The KGB model scoring subsystem builds a KGB model on samples with known risk performance labels of good and bad, and obtains the KGB model scoring score;

[0029] The AR model scoring subsystem builds an AR model on all samples with known rejection labels to obtain an AR model scoring score. Preferably, the inference sample fuzzy expansion subsystem includes:

[0030] The fuzzy expansion sample expansion subsystem obtains a set of samples of good and bad labels to be inferred; uses the fuzzy expansion method to expand each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1;

[0031] The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0032] The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

[0033] Preferably, the label sample weight setting subsystem includes:

[0034] Expand sample weight setting subsystem, set the sum of fuzzy expansion good sample weight and fuzzy expansion bad sample weight to be equal to the weight of the sample with good and bad labels to be inferred;

[0035] The inferred good or bad probability operation subsystem establishes a correlation relationship with the inferred good or bad probability based on the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively; uses the KGB model score and the AR model score to perform weight operation allocation; establishes a correlation relationship with the inferred good or bad probability based on the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively, including: on samples with good and bad risk performance, use 0 and 1 as good sample labels and bad sample labels respectively, train the logistic Sty regression model with the KGB model score and the AR model score, obtain the relationship between the overdue probability and the KGB model score and the AR model score, and obtain the Sty regression overdue probability prediction relationship; apply the Sty regression overdue probability prediction relationship to the rejected samples and the unwithdrawn samples, and accordingly obtain the probability that each credit data is inferred to be bad; multiply the weight of the sample that needs to be inferred by the inferred bad probability to obtain the extended bad sample weight, and similarly, multiply the weight of the sample that needs to be inferred by the predicted good probability to obtain the extended good sample weight.

[0036] Preferably, the weight allocation multi-model inference subsystem includes:

[0037] The AGB model iteration subsystem establishes an AGB model on all samples of the set of good and bad label samples to be inferred, and performs iterative operations; the iterative operations include: setting a maximum number of iterations, and stopping the iterative operation when the maximum number of iterations is reached or the iteration results converge;

[0038] The risk prediction model inference subsystem determines whether the AGB model has similar prediction performance on different types of customer groups. When the AGB model has similar prediction performance on different types of customer groups, the inference results converge and the final inference results of risk prediction are obtained. When the AGB model does not have similar prediction performance on different types of customer groups, the weight allocation and iterative calculation are redistributed until the inference results converge and the final inference results of risk prediction are obtained.

[0039] Compared with the prior art, the present invention has at least the following beneficial effects:

[0040] The present invention provides a multi-dimensional scoring credit data risk prediction control method and system, which establishes a KGB model and an AR model on a known label sample set, and obtains a KGB model scoring score and an AR model scoring score respectively; obtains a set of good or bad label samples to be inferred, and uses a fuzzy expansion method to expand each good or bad label sample to be inferred in the set of good or bad label samples to be inferred into a fuzzy expansion good sample and a fuzzy expansion bad sample; sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the good or bad label sample to be inferred; establishes a cross-correlation relationship between the good or bad probability of inference according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; establishes an AGB model on all samples of the set of good or bad label samples to be inferred, and starts Iterate operations are performed to determine whether the AGB model has similar prediction performance on different types of customer groups. Iterate until the inference results converge to obtain the final inference results of risk prediction. KGB scoring and AR scoring play an important role in multi-dimensional enhanced model evaluation. The risk characteristics of potential high-risk customers, including those who have failed to successfully apply for loans, can be more comprehensively understood. The information mining capabilities of this part of user data are improved to avoid survivor bias and build a more accurate inference prediction model. From a practical application perspective, this technical solution can effectively optimize the customer screening strategy in risk control, and include rejected customers who actually have better repayment capabilities and lower default risks into the service scope. It monitors and warns of information that has been passed with potential risks and higher default risks, improves system security, reduces bad debt rates, and increases profits while controlling risks.

[0041] The multi-dimensional scoring credit data risk prediction control method and system described in the present invention, other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by technical personnel in the field through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 This is a diagram of an embodiment of a multi-dimensional scoring credit data risk prediction control system described in the present invention.

[0044] Figure 2 This is a diagram of an embodiment of a multi-dimensional scoring credit data risk prediction control method described in the present invention.

[0045] Figure 3 This is another embodiment diagram of the multi-dimensional scoring credit data risk prediction control method described in the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments so that those skilled in the art can implement the invention according to the instructions. Figure 1-Figure 3 As shown, the present invention provides a multi-dimensional scoring credit data risk prediction control method, comprising:

[0047] S1, establish a KGB model and an AR model on a set of known label samples, and obtain the KGB model score and the AR model score respectively;

[0048] S2, obtaining a set of samples of good and bad labels to be inferred, and using the fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample;

[0049] S3, set the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establish the cross-correlation relationship between the inference good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; use the KGB model score and the AR model score to perform weight calculation and allocation;

[0050] S4, build the AGB model on all samples of the set of good and bad label samples to be inferred, start iterative calculations, determine whether the AGB model has similar prediction performance on different types of customer groups, and iterate until the inference results converge to obtain the final inference result of risk prediction.

[0051] The principle and effect of the above technical solution are as follows: the present invention provides a multi-dimensional scoring credit data risk prediction control method, including: establishing a KGB model and an AR model on a known label sample set, obtaining a KGB model scoring score and an AR model scoring score respectively; obtaining a set of good or bad label samples to be inferred, using a fuzzy expansion method to expand each good or bad label sample to be inferred in the set of good or bad label samples to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample; setting the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight to be equal to the weight of the good or bad label sample to be inferred; establishing a cross-correlation relationship between the good or bad probability of inference according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively; using the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; establishing an AGB model on all samples of the set of good or bad label samples to be inferred, starting iterative calculations, judging whether the AGB model has similar prediction performance on different types of customer groups, iterative calculations until the inference results converge, and obtaining the final inference results of risk prediction; KGB scoring and AR scoring multi-dimensional enhanced model evaluation play an important role; KGB model, a known good or bad prediction model, defines the prediction results based on whether there is a default ; 1 is bad, 0 is good, and it is constructed by good and bad samples for prediction; AR model, acceptance rejection model, defines the prediction result by whether it passes; 1=accept, 0=reject; constructed with full application samples for prediction; AGB model of the present invention, full good and bad inference prediction model, defines the prediction output result by whether it defaults; and combined with full inference, defines the prediction result by whether it defaults; 1=bad, 0=good; jointly constructed with good and bad passing samples and rejected samples for prediction; combined with known passing samples and inferred rejection samples for prediction; and combined with full inference for prediction; for potential high-risk customers who fail to successfully apply for loans, the risk characteristics can be more comprehensively understood; the information mining ability of this part of user data is improved, survivor bias is avoided, and a more accurate inference prediction model is constructed; from the perspective of practical application, the technical solution can effectively optimize the customer screening strategy in risk control, and include rejected customers who actually have better repayment ability and lower default risk into the service scope; monitor and warn the passed information with potential risks and higher default risks, improve system security, reduce bad debt rate, increase profits while controlling risks.

[0052] In one embodiment, S1 includes:

[0053] S101, establish a KGB model on samples with known risk performance labels of good and bad, and obtain a KGB model score;

[0054] S102, establishing an AR model on all samples with known rejection labels, and obtaining an AR model scoring score;

[0055] A KGB model is established on samples with known good or bad risk performance labels, and the KGB model score is obtained, including: on samples with known good or bad risk performance labels, the KGB model modeling sample range is limited to user data that has passed loan approval, and a binary target variable is defined based on whether the credit user data has overdue repayment behavior; denoted as = 1 for a bad label, indicating that the user data has defaulted; denoted as = 0 for a good label, indicating that the user data has not defaulted; the KGB model modeling range is limited to samples with labels 0 and 1; a target variable is defined for post-loan default performance for subsequent use distinction, the target variable of user data whose loan application is rejected is denoted as 2, and the target variable of user data with no cash flow performance after loan approval is denoted as -1; the behavior information of credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; a KGB model is constructed on approved samples with known good or bad, which is used to predict the post-loan default probability of user data, and the KGB model score is obtained;

[0056] An AR model is established on the full amount of samples with known rejection labels to obtain an AR model score, including: on the full amount of samples with known rejection labels, the AR model modeling sample range is limited to the data of all users who apply for credit, and a binary target variable 1 is defined based on whether the user data is approved, where 1=1 is rejected, the user data is not approved, and the loan application is rejected; 1=0 is passed, indicating that the user data is approved and the credit application is accepted; the behavioral information of the credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; an AR model is constructed with the full amount of application samples to predict the probability of user data being rejected in the credit application approval, and an AR model score is obtained.

[0057] The principle and effect of the above technical solution are as follows: a KGB model is established on samples with known risk performance labels of good or bad, and a KGB model score is obtained; an AR model is established on all samples with known rejection labels, and an AR model score is obtained;

[0058] A KGB model is established on samples with known good or bad risk performance labels, and the KGB model score is obtained, including: on samples with known good or bad risk performance labels, the KGB model modeling sample range is limited to user data that has passed loan approval, and a binary target variable is defined based on whether the credit user data has overdue repayment behavior; denoted as = 1 for a bad label, indicating that the user data has defaulted; denoted as = 0 for a good label, indicating that the user data has not defaulted; the KGB model modeling range is limited to samples with labels 0 and 1; a target variable is defined for post-loan default performance for subsequent use distinction, the target variable of user data whose loan application is rejected is denoted as 2, and the target variable of user data with no cash flow performance after loan approval is denoted as -1; the behavior information of credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; a KGB model is constructed on approved samples with known good or bad, which is used to predict the post-loan default probability of user data, and the KGB model score is obtained;

[0059] An AR model is established on the full amount of samples with known rejection labels to obtain an AR model score, including: on the full amount of samples with known rejection labels, the AR model modeling sample range is limited to the data of all users who apply for credit, and a binary target variable 1 is defined based on whether the user data is approved, where 1=1 is rejected, the user data is not approved, and the loan application is rejected; 1=0 is passed, indicating that the user data is approved and the credit application is accepted; the behavioral information of the credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; an AR model is constructed with the full amount of application samples to predict the probability of user data being rejected in the credit application approval, and an AR model score is obtained; the risk characteristics of potential high-risk customers, including those who have failed to successfully apply for loans, can be more comprehensively understood.

[0060] In one embodiment, S2 includes:

[0061] S201, obtaining a set of samples of good and bad labels to be inferred; using a fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1;

[0062] S202, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establishing a cross-correlation relationship between the inferred good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and using the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0063] The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

[0064] The principle and effect of the above technical solution are as follows: obtain a set of samples of good or bad labels to be inferred; use the fuzzy expansion method to expand each sample of good or bad labels to be inferred in the set of samples of good or bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are fuzzy expansion good samples and fuzzy expansion bad samples; the target variable of fuzzy expansion good samples is recorded as 0; the target variable of fuzzy expansion bad samples is recorded as 1; set the sum of the weight of fuzzy expansion good samples and the weight of fuzzy expansion bad samples to be equal to the weight of the samples of good or bad labels to be inferred; establish the correlation relationship between the good and bad probability of inference according to the weight of fuzzy expansion good samples and the weight of fuzzy expansion bad samples respectively; use the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; set the fuzzy expansion The sum of the good sample weight and the fuzzy expanded bad sample weight is equal to the weight of the sample with good or bad label to be inferred; the correlation relationship between the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight and the inferred good or bad probability is established respectively, including: the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expanded good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expanded bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample; the information mining ability of this part of user data is improved, survivor bias is avoided, and a more accurate inference prediction model is constructed.

[0065] In one embodiment, S3 includes:

[0066] S301, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the good and bad label samples to be inferred;

[0067] S302, establishing a correlation relationship between the inferred good and bad probabilities according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; and performing weight calculation and allocation using the KGB model scoring score and the AR model scoring score;

[0068] Establishing the correlation between the good and bad probabilities according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively includes: on samples with good and bad risk performance, using 0 and 1 as good sample labels and bad sample labels respectively, training the logistic regression model with the KGB model score and the AR model score, obtaining the relationship between the overdue probability and the KGB model score and the AR model score, and obtaining the Logistic regression overdue probability prediction relationship; applying the Logistic regression overdue probability prediction relationship to the rejected samples and the unwithdrawn samples, and correspondingly obtaining the probability that each credit data is inferred to be bad; multiplying the weight of the sample that needs to be inferred by the inferred bad probability to obtain the extended bad sample weight, and similarly, multiplying the weight of the sample that needs to be inferred by the predicted good probability to obtain the extended good sample weight;

[0069] Assume that different customer groups have similar risk performance under the condition of similar AGB scores; convert similar risk performance into similar default prediction probability; then the relationship between AGB score and prediction probability between different customer groups is similar; since the probability of prediction as good is obtained by the logistic regression model, then there is a linear relationship between it and the AGB score, where is. More intuitively, two variables with a linear relationship can be drawn into a straight line in a two-dimensional rectangular coordinate system. Because two points determine a straight line, this scheme selects the 5% quantile and 95% quantile of each customer group score as judgment points, and compares their corresponding prediction probabilities as the basis for judging whether there are similar risk performances.

[0070] The principle and effect of the above technical solution are: setting the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establishing the correlation between the inferred good or bad probability according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively; using the KGB model score and the AR model score to perform weight calculation and allocation; establishing the correlation between the inferred good or bad probability according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively includes: on samples with good and bad risk performance, using 0 and 1 as good sample labels and bad sample labels respectively, training the logistic Stygian regression model with the KGB model score and the AR model score, obtaining the relationship between the overdue probability and the KGB model score and the AR model score, and obtaining the Stygian regression overdue probability prediction relationship; applying the Stygian regression overdue probability prediction relationship to the rejected samples and the unwithdrawn samples, and correspondingly obtaining the probability of each credit data being inferred to be bad; multiplying the weight of the sample to be inferred by The extended bad sample weight is obtained by taking the probability of inferring the performance as bad, and similarly, the extended good sample weight is obtained by multiplying the weight of the sample to be inferred by the probability of being predicted as good; assuming that different customer groups have similar risk performance under the condition of similar AGB scores; converting similar risk performance into similar default prediction probability; then the relationship between AGB score and prediction probability is similar between different customer groups; since the probability of being predicted as good is obtained by the logistic regression model, then there is a linear relationship between it and the AGB score, where is. More intuitively, two variables with a linear relationship can be drawn into a straight line in a two-dimensional rectangular coordinate system, because two points determine a straight line, so this scheme selects the 5% quantile and the 95% quantile of each customer group score as the judgment point, and compares the corresponding prediction probabilities as the basis for judging whether there are similar risk performances; from the perspective of practical application, this technical scheme can effectively optimize the customer screening strategy in risk control, and include rejected customers who actually have better repayment ability and lower default risk into the service scope.

[0071] In one embodiment, S4 includes:

[0072] S401, establishing an AGB model on all samples of the set of good and bad label samples to be inferred, and performing iterative operations; the iterative operations include: setting a maximum number of iterations, and stopping the iterative operations when the maximum number of iterations is reached or the iteration results converge;

[0073] S403, judging whether the AGB model has similar prediction performance on different types of customer groups; when the AGB model has similar prediction performance on different types of customer groups, the inference results converge, and the final inference results of risk prediction are obtained; when the AGB model does not have similar prediction performance on different types of customer groups, re-weight allocation and iterative calculation are performed until the inference results converge, and the final inference results of risk prediction are obtained;

[0074] To achieve convergence of inference results and obtain the final inference results of risk prediction, the following are included: establishing an AGB model based on the inferred AGB sample; assuming that different customer groups have similar AGB scores and similar risk performance.

[0075] The principle and effect of the above technical solution are: using the KGB model scoring score and the AR model scoring score to perform weight calculation allocation; establishing an AGB model on the full sample of the good and bad label sample set to be inferred, and performing iterative calculation; performing iterative calculation includes: setting a maximum number of iterations, and stopping the iterative calculation when the maximum number of iterations is reached or the iterative result converges; judging whether the AGB model has similar prediction performance on different types of customer groups; when the AGB model has similar prediction performance on different types of customer groups, the inference result converges and the final inference result of risk prediction is obtained; when the AGB model does not have similar prediction performance on different types of customer groups, re-allocating weights and performing iterative calculations until the inference result converges and the final inference result of risk prediction is obtained;

[0076] To achieve convergence of inference results and obtain the final inference results of risk prediction, the following steps are involved: establishing an AGB model based on the inferred AGB sample; assuming that different customer groups have similar AGB scores and similar risk performance;

[0077] (1) After the AGB model is established, the user data is divided into three customer groups: the user data with known overdue performance is recorded as the KGB customer group, the user data with rejected approval is recorded as the Reject customer group, and the user data with approved approval but no cash flow performance is recorded as the Uncash customer group; for the three customer groups, a logistic regression model is established with the AGB score as the predictor variable to obtain the mapping relationship between the model score and the predicted probability;

[0078] (2) Take the 5% quantile and 95% quantile of the scores of the Reject customer group and the KGB customer group respectively; take the maximum value of the 5% quantile of the two as the measurement point A, and take the minimum value of the 95% quantile of the two as the measurement point B; (3) According to the mapping relationship in (1), calculate the predicted probabilities of measurement points A and B under the Reject customer group and the KGB customer group. If the risk probabilities of the Reject customer group and the KGB customer group corresponding to measurement points A and B differ by less than 5%, it is considered that the Reject customer group and the KGB customer group have similar score risk relationships;

[0079] (4) If there is an Uncash customer group in the sample, repeat the calculation steps in (3) for the Uncash customer group and the KGB customer group to determine whether there is a similar score risk relationship;

[0080] (5) If the conditions of (3) and (4) are met, the entire model is considered to have converged, and the final inference result of risk prediction is obtained; if any of the conditions of (3) and (4) is not met, it is necessary to repeat the calculation of the score risk relationship on the KGB customer group using the AGB model score and the AR model score, and reallocate weights for the Reject customer group and the Uncash customer group; if the model does not converge, the iterative calculation is restarted, the predicted probability of the samples that need to be inferred labels is recalculated, and the weights are further reallocated for the expanded samples; the weight allocation method here follows the fuzzy expansion method and the AGB score statistical operation steps are consistent with the embodiment; the AR model score remains unchanged, the modeling scope remains unchanged, and the relationship between the predicted probability and the AGB score and the AR score is established for samples with good and bad risk performance, and then it is applied to the samples that need to be inferred labels to allocate weights for the expanded good and bad samples; the AGB model is reestablished until the convergence conditions are met to obtain the final inference result of risk prediction; the information that has passed with potential risks and high default risks is monitored and warned to improve system security, reduce bad debt rate, and increase profits while controlling risks.

[0081] The present invention provides a multi-dimensional scoring credit data risk prediction control system, comprising:

[0082] The sample model multi-dimensional scoring system builds the KGB model and AR model on the known label sample set, and obtains the KGB model scoring score and AR model scoring score respectively;

[0083] The inference sample fuzzy expansion subsystem obtains a set of samples with good and bad labels to be inferred, and uses the fuzzy expansion method to expand each sample with good and bad labels to be inferred in the set of samples with good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample;

[0084] The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0085] The weighted distribution multi-model inference subsystem establishes an AGB model on the full set of samples of the good and bad label sample sets to be inferred, and starts iterative operations to determine whether the AGB model has similar prediction performance on different types of customer groups. It iterates until the inference results converge to obtain the final inference result of risk prediction.

[0086] The principle and effect of the above technical solution are as follows: the present invention provides a multi-dimensional scoring credit data risk prediction control system, including: a sample model multi-dimensional scoring subsystem, which establishes a KGB model and an AR model on a known label sample set, and obtains a KGB model scoring score and an AR model scoring score respectively; an inference sample fuzzy expansion subsystem, which obtains a set of good or bad label samples to be inferred, and uses a fuzzy expansion method to expand each good or bad label sample to be inferred in the set of good or bad label samples to be inferred into a fuzzy expansion good sample and a fuzzy expansion bad sample; a label sample weight setting subsystem, which sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the good or bad label sample to be inferred; establishes a cross-correlation relationship between the good or bad probability of inference according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; weight allocation multi-model inference score The system establishes an AGB model on the full set of samples of the good and bad label sample sets to be inferred, and starts iterative operations to determine whether the AGB model has similar prediction performance on different types of customer groups. The iterative operation is carried out until the inference result converges to obtain the final inference result of risk prediction; KGB scoring and AR scoring play an important role in multi-dimensional enhanced model evaluation; the risk characteristics of potential high-risk customers, including those who have failed to successfully apply for loans, can be more comprehensively understood; the information mining ability of this part of user data is improved, survivor bias is avoided, and a more accurate inference prediction model is constructed; from the perspective of practical application, this technical solution can effectively optimize the customer screening strategy in risk control, and include rejected customers who actually have better repayment ability and lower default risk into the service scope; monitor and warn the information that has been passed with potential risks and higher default risks, improve system security, reduce bad debt rate, increase profits while controlling risks.

[0087] In one embodiment, the sample model multi-dimensional scoring system includes:

[0088] The KGB model scoring subsystem builds a KGB model on samples with known risk performance labels of good and bad, and obtains the KGB model scoring score;

[0089] The AR model scoring subsystem builds an AR model on all samples with known rejection labels to obtain the AR model scoring score;

[0090] A KGB model is established on samples with known good or bad risk performance labels, and the KGB model score is obtained, including: on samples with known good or bad risk performance labels, the KGB model modeling sample range is limited to user data that has passed loan approval, and a binary target variable is defined based on whether the credit user data has overdue repayment behavior; denoted as = 1 for a bad label, indicating that the user data has defaulted; denoted as = 0 for a good label, indicating that the user data has not defaulted; the KGB model modeling range is limited to samples with labels 0 and 1; a target variable is defined for post-loan default performance for subsequent use distinction, the target variable of user data whose loan application is rejected is denoted as 2, and the target variable of user data with no cash flow performance after loan approval is denoted as -1; the behavior information of credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; a KGB model is constructed on approved samples with known good or bad, which is used to predict the post-loan default probability of user data, and the KGB model score is obtained;

[0091] An AR model is established on the full amount of samples with known rejection labels to obtain an AR model score, including: on the full amount of samples with known rejection labels, the AR model modeling sample range is limited to the data of all users who apply for credit, and a binary target variable 1 is defined based on whether the user data is approved, where 1=1 is rejected, the user data is not approved, and the loan application is rejected; 1=0 is passed, indicating that the user data is approved and the credit application is accepted; the behavioral information of the credit user data is calculated or discretized to obtain a prediction variable, and the prediction variable value types include: numeric, character, and Boolean; the data samples are stratified and sampled according to the proportion of each type of sample in the data, and the sampled data obtains a weight variable; an AR model is constructed with the full amount of application samples to predict the probability of user data being rejected in the credit application approval, and an AR model score is obtained.

[0092] The principle and effect of the above technical solution are as follows: the sample model multi-dimensional scoring subsystem includes: a KGB model scoring subsystem, which establishes a KGB model on samples with known risk performance labels to obtain a KGB model scoring score; an AR model scoring subsystem, which establishes an AR model on all samples with known rejection labels to obtain an AR model scoring score; a KGB model is established on samples with known risk performance labels to obtain a KGB model scoring score, including: on samples with known risk performance labels, the KGB model modeling sample range is limited to user data that has passed loan approval, based on whether the credit user data is issued Define a binary target variable for overdue repayment behavior; record = 1 as a bad label, indicating that the user data has default behavior; record = 0 as a good label, indicating that the user data does not have default behavior; the modeling scope of the KGB model is limited to samples with labels 0 and 1; define a target variable for post-loan default performance to distinguish subsequent use, the target variable of user data whose loan application is rejected is recorded as 2, and the target variable of user data with no cash flow performance after loan approval is recorded as -1; calculate or discretize the behavioral information of credit user data to obtain prediction variables, and the value types of prediction variables include: numeric, character, and Boolean ; According to the proportion of each type of sample in the data, the data samples are stratified and sampled, and the sampled data obtains the weight variable; the KGB model is built on the approved samples with known quality to predict the post-loan default probability of user data and obtain the KGB model score; the AR model is built on the full number of samples with known rejection labels to obtain the AR model score, including: on the full number of samples with known rejection labels, the AR model modeling sample range is limited to all user data applying for credit, and the binary target variable 1 is defined based on whether the user data is approved, and 1=1 is rejected. The user data is not approved and the loan application is rejected. to rejection; 1=0 is passed, indicating that the user data has passed the approval and the credit application has been accepted; the behavioral information of the credit user data is calculated or discretized to obtain the prediction variable, and the value types of the prediction variable include: numeric, character and Boolean; according to the proportion of each type of sample in the data, the data samples are stratified and sampled, and the sampled data obtains the weight variable; the AR model is constructed with the full application sample to predict the probability of user data being rejected in the credit application approval, and the AR model score is obtained; the risk characteristics of potential high-risk customers, including those who have failed to successfully apply for loans, can be more comprehensively understood.

[0093] In one embodiment, the inferred sample fuzzy expansion subsystem includes:

[0094] The fuzzy expansion sample expansion subsystem obtains a set of samples of good and bad labels to be inferred; uses the fuzzy expansion method to expand each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1;

[0095] The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation;

[0096] The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

[0097] The principle and effect of the above technical solution are as follows: the inference sample fuzzy expansion subsystem includes: a fuzzy expansion sample expansion subsystem, which obtains a set of good and bad label samples to be inferred; uses the fuzzy expansion method to expand each good and bad label sample to be inferred in the set of good and bad label samples to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the fuzzy expansion good sample target variable is recorded as 0; the fuzzy expansion bad sample target variable is recorded as 1; the label sample weight setting subsystem, sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the good and bad label sample to be inferred; establishes a cross-correlation relationship with the inference good and bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; uses the KGB model scoring score and the AR model The weighted score is allocated according to the type score; the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight is set to be equal to the weight of the sample with good or bad labels to be inferred; the cross-correlation relationship between the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight and the inferred good or bad probability is established respectively according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight, including: the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expanded good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad labels to be inferred as a good sample; the ratio of the fuzzy expanded bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad labels to be inferred as a bad sample; the information mining ability of this part of user data is improved, survivor bias is avoided, and a more accurate inference prediction model is constructed.

[0098] In one embodiment, the label sample weight setting subsystem includes:

[0099] Expand sample weight setting subsystem, set the sum of fuzzy expansion good sample weight and fuzzy expansion bad sample weight to be equal to the weight of the sample with good and bad labels to be inferred;

[0100] The inference good or bad probability operation subsystem establishes the correlation between the inference good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and uses the KGB model scoring score and the AR model scoring score to perform weight operation allocation;

[0101] Establishing the correlation between the good and bad probabilities according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively includes: on samples with good and bad risk performance, using 0 and 1 as good sample labels and bad sample labels respectively, training the logistic regression model with the KGB model score and the AR model score, obtaining the relationship between the overdue probability and the KGB model score and the AR model score, and obtaining the Logistic regression overdue probability prediction relationship; applying the Logistic regression overdue probability prediction relationship to the rejected samples and the unwithdrawn samples, and correspondingly obtaining the probability that each credit data is inferred to be bad; multiplying the weight of the sample that needs to be inferred by the inferred bad probability to obtain the extended bad sample weight, and similarly, multiplying the weight of the sample that needs to be inferred by the predicted good probability to obtain the extended good sample weight;

[0102] Assume that different customer groups have similar risk performance under the condition of similar AGB scores; convert similar risk performance into similar default prediction probability; then the relationship between AGB score and prediction probability between different customer groups is similar; since the probability of prediction as good is obtained by the logistic regression model, then there is a linear relationship between it and the AGB score, where is. More intuitively, two variables with a linear relationship can be drawn into a straight line in a two-dimensional rectangular coordinate system. Because two points determine a straight line, this scheme selects the 5% quantile and 95% quantile of each customer group score as judgment points, and compares their corresponding prediction probabilities as the basis for judging whether there are similar risk performances.

[0103] The principle and effect of the above technical solution are as follows: the label sample weight setting subsystem includes: the expanded sample weight setting subsystem, which sets the sum of the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight to be equal to the weight of the good or bad label sample to be inferred; the inferred good or bad probability operation subsystem, which establishes the cross-correlation relationship between the inferred good or bad probability according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively; uses the KGB model scoring score and the AR model scoring score to perform weight operation allocation; establishes the cross-correlation relationship between the inferred good or bad probability according to the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively, including: on samples with good and bad risk performance, use 0 and 1 as good sample labels and bad sample labels respectively, and use the KGB model scoring score and the AR model scoring score to train the logistic Stygian regression model to obtain the relationship between the overdue probability and the KGB model scoring score and the AR model scoring score, and obtain the Stygian regression overdue probability prediction relationship; apply the Stygian regression overdue probability prediction relationship to the rejected samples and the unwithdrawn samples, and obtain each credit data accordingly. The probability of inferring performance is bad; multiply the weight of the sample that needs to be inferred by the probability of inferring performance being bad to obtain the extended bad sample weight. Similarly, multiply the weight of the sample that needs to be inferred by the probability of prediction being good to obtain the extended good sample weight; assume that different customer groups have similar risk performance under the condition of similar AGB scores; convert similar risk performance into similar default prediction probability; then the relationship between AGB score and prediction probability between different customer groups is similar; since the probability of prediction being good is obtained by the logistic regression model, then there is a linear relationship between it and the AGB score, where is. More intuitively, two variables with a linear relationship can be drawn into a straight line in a two-dimensional rectangular coordinate system. Because two points determine a straight line, this solution selects the 5% quantile and the 95% quantile of each customer group score as the judgment point, and compares the corresponding prediction probabilities as the basis for judging whether there is a similar risk performance; from the perspective of practical application, this technical solution can effectively optimize the customer screening strategy in risk control, and include rejected customers who actually have better repayment ability and lower default risk into the service scope.

[0104] In one embodiment, the weight allocation multi-model inference subsystem includes:

[0105] The AGB model iteration subsystem establishes an AGB model on all samples of the set of good and bad label samples to be inferred, and performs iterative operations; the iterative operations include: setting a maximum number of iterations, and stopping the iterative operation when the maximum number of iterations is reached or the iteration results converge;

[0106] The risk prediction model inference subsystem determines whether the AGB model has similar prediction performance on different types of customer groups. When the AGB model has similar prediction performance on different types of customer groups, the inference results converge and the final inference results of risk prediction are obtained. When the AGB model does not have similar prediction performance on different types of customer groups, the weight allocation and iterative calculation are re-performed until the inference results converge and the final inference results of risk prediction are obtained.

[0107] To achieve convergence of inference results and obtain the final inference results of risk prediction, the following are included: establishing an AGB model based on the inferred AGB sample; assuming that different customer groups have similar AGB scores and similar risk performance.

[0108] The principle and effect of the above technical solution are as follows: the weight allocation multi-model inference subsystem includes: an AGB model iteration subsystem, which establishes an AGB model on the full number of samples of the good and bad label sample set to be inferred, and performs iterative operations; the iterative operations include: setting a maximum number of iterations, and when the maximum number of iterations is reached or the iteration results converge, the iterative operation stops; the risk prediction model inference subsystem determines whether the AGB model has similar prediction performance on different types of customer groups; when the AGB model has similar prediction performance on different types of customer groups, the inference results converge, and the final inference result of risk prediction is obtained; when the AGB model does not have similar prediction performance on different types of customer groups, the weight allocation and iterative operations are re-performed until the inference results converge, and the final inference result of risk prediction is obtained;

[0109] To achieve convergence of inference results and obtain the final inference results of risk prediction, the following steps are involved: establishing an AGB model based on the inferred AGB sample; assuming that different customer groups have similar AGB scores and similar risk performance;

[0110] (1) After the AGB model is established, the user data is divided into three customer groups: the user data with known overdue performance is recorded as the KGB customer group, the user data with rejected approval is recorded as the Reject customer group, and the user data with approved approval but no cash flow performance is recorded as the Uncash customer group; for the three customer groups, a logistic regression model is established with the AGB score as the predictor variable to obtain the mapping relationship between the model score and the predicted probability;

[0111] (2) Take the 5% quantile and 95% quantile of the scores of the Reject customer group and the KGB customer group respectively; take the maximum value of the 5% quantile of the two as the measurement point A, and take the minimum value of the 95% quantile of the two as the measurement point B; (3) According to the mapping relationship in (1), calculate the predicted probabilities of measurement points A and B under the Reject customer group and the KGB customer group. If the risk probabilities of the Reject customer group and the KGB customer group corresponding to measurement points A and B differ by less than 5%, it is considered that the Reject customer group and the KGB customer group have similar score risk relationships;

[0112] (4) If there is an Uncash customer group in the sample, repeat the calculation steps in (3) for the Uncash customer group and the KGB customer group to determine whether there is a similar score risk relationship;

[0113] (5) If the conditions of (3) and (4) are met, the entire model is considered to have converged, and the final inference result of risk prediction is obtained; if any of the conditions of (3) and (4) is not met, it is necessary to repeat the calculation of the score risk relationship on the KGB customer group using the AGB model score and the AR model score, and reallocate weights for the Reject customer group and the Uncash customer group; if the model does not converge, the iterative calculation is restarted, the predicted probability of the samples that need to be inferred labels is recalculated, and the weights are further reallocated for the expanded samples; the weight allocation method here follows the fuzzy expansion method and the AGB score statistical operation steps are consistent with the embodiment; the AR model score remains unchanged, the modeling scope remains unchanged, and the relationship between the predicted probability and the AGB score and the AR score is established for samples with good and bad risk performance, and then it is applied to the samples that need to be inferred labels to allocate weights for the expanded good and bad samples; the AGB model is reestablished until the convergence conditions are met to obtain the final inference result of risk prediction; the information that has passed with potential risks and high default risks is monitored and warned to improve system security, reduce bad debt rate, and increase profits while controlling risks.

[0114] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A multi-dimensional scoring credit data risk prediction control method, characterized in that: include: S1, establish a KGB model and an AR model on a set of known label samples, and obtain the KGB model score and the AR model score respectively; S2, obtaining a set of samples of good and bad labels to be inferred, and using the fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample; S3, set the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establish the cross-correlation relationship between the inference good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; use the KGB model score and the AR model score to perform weight calculation and allocation; S4, build the AGB model on all samples of the set of good and bad label samples to be inferred, start iterative calculations, determine whether the AGB model has similar prediction performance on different types of customer groups, and iterate until the inference results converge to obtain the final inference result of risk prediction.

2. A multi-dimensional scoring credit data risk prediction control method according to claim 1, characterized in that: S100 includes: S101, establish a KGB model on samples with known risk performance labels of good and bad, and obtain a KGB model score; S102, establishing an AR model on all samples with known rejection labels, and obtaining an AR model scoring score.

3. A multi-dimensional scoring credit data risk prediction control method according to claim 1, characterized in that: S200 includes: S201, obtaining a set of samples of good and bad labels to be inferred; using a fuzzy expansion method, expanding each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1; S202, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the sample with good or bad label to be inferred; establishing a cross-correlation relationship between the inferred good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and using the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

4. A multi-dimensional scoring credit data risk prediction control method according to claim 1, characterized in that: S300 includes: S301, setting the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the good and bad label samples to be inferred; S302, establishing a correlation relationship between the inferred good and bad probabilities according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; and performing weight calculation and allocation using the KGB model scoring score and the AR model scoring score; The cross-correlation relationship between the good and bad probabilities is established based on the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively, including: on samples with good and bad risk performance, 0 and 1 are used as good sample labels and bad sample labels respectively, and the logistic regression model is trained with the KGB model score and the AR model score to obtain the relationship between the overdue probability and the KGB model score and the AR model score, and the Logistic Regression Overdue Probability Prediction Relationship; the Logistic Regression Overdue Probability Prediction Relationship is applied to the rejected samples and the unwithdrawn samples, and the probability of each credit data being inferred to be bad is obtained accordingly; the weight of the sample that needs to be inferred is multiplied by the inferred bad probability to obtain the extended bad sample weight, and similarly, the weight of the sample that needs to be inferred is multiplied by the predicted good probability to obtain the extended good sample weight.

5. The multi-dimensional scoring credit data risk prediction control method according to claim 1, characterized in that: S400 includes: S401, establishing an AGB model on all samples of the set of good and bad label samples to be inferred, and performing iterative operations; performing iterative operations includes: setting a maximum number of iterations, and stopping the iterative operation when the maximum number of iterations is reached or the iteration results converge; the AGB model marks the known risk performance good and bad label samples and the fuzzy expansion samples with AGB scores, and compares the first linear relationship between the AGB score on the known risk performance good and bad label samples and the log (Odds) of the known risk performance good and bad label samples, and the second linear relationship between the AGB score on the fuzzy expansion samples and the log (Odds) of the fuzzy expansion samples, if the proximity between the first linear relationship and the second linear relationship reaches the proximity threshold condition, then stopping the iterative operation; otherwise, updating the fuzzy expansion sample weight according to the AGB score, and continuing to create the AGB score, and performing iterative operations; S403, determine whether the AGB model has similar prediction performance on different types of customer groups; when the AGB model has similar prediction performance on different types of customer groups, the inference results converge and the final inference result of risk prediction is obtained; when the AGB model does not have similar prediction performance on different types of customer groups, re-weight allocation and iterative calculation are performed until the inference results converge and the final inference result of risk prediction is obtained.

6. A multi-dimensional scoring credit data risk prediction control system, characterized in that: include: The sample model multi-dimensional scoring system builds the KGB model and AR model on the known label sample set, and obtains the KGB model scoring score and AR model scoring score respectively; The inference sample fuzzy expansion subsystem obtains a set of samples with good and bad labels to be inferred, and uses the fuzzy expansion method to expand each sample with good and bad labels to be inferred in the set of samples with good and bad labels to be inferred into a fuzzy expanded good sample and a fuzzy expanded bad sample; The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight respectively; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; The weighted distribution multi-model inference subsystem establishes an AGB model on the full set of samples of the good and bad label sample sets to be inferred, and starts iterative operations to determine whether the AGB model has similar prediction performance on different types of customer groups. It iterates until the inference results converge to obtain the final inference result of risk prediction.

7. A multi-dimensional scoring credit data risk prediction control system according to claim 6, characterized in that: The sample model multidimensional scoring system includes: The KGB model scoring subsystem builds a KGB model on samples with known risk performance labels of good and bad, and obtains the KGB model scoring score; The AR model scoring subsystem builds an AR model on all samples with known rejection labels to obtain an AR model scoring score.

8. A multi-dimensional scoring credit data risk prediction control system according to claim 6, characterized in that: The inference sample fuzzy expansion subsystem includes: The fuzzy expansion sample expansion subsystem obtains a set of samples of good and bad labels to be inferred; uses the fuzzy expansion method to expand each sample of good and bad labels to be inferred in the set of samples of good and bad labels to be inferred into two fuzzy expansion samples; the two fuzzy expansion samples are a fuzzy expansion good sample and a fuzzy expansion bad sample; the target variable of the fuzzy expansion good sample is recorded as 0; the target variable of the fuzzy expansion bad sample is recorded as 1; The label sample weight setting subsystem sets the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight to be equal to the weight of the label sample to be inferred; establishes the cross-correlation relationship between the inference good and bad probabilities based on the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and uses the KGB model scoring score and the AR model scoring score to perform weight calculation and allocation; The sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set to be equal to the weight of the sample with good or bad label to be inferred; the cross-correlation relationship between the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight and the inference good or bad probability is established respectively, including: the sum of the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight is set as the sample weight of the good or bad label to be inferred, the ratio of the fuzzy expansion good sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a good sample; the ratio of the fuzzy expansion bad sample weight to the sample weight of the good or bad label to be inferred is the probability of inferring the sample with good or bad label to be inferred as a bad sample.

9. A multi-dimensional scoring credit data risk prediction control system according to claim 6, characterized in that: The label sample weight setting subsystem includes: Expand sample weight setting subsystem, set the sum of fuzzy expansion good sample weight and fuzzy expansion bad sample weight to be equal to the weight of the sample with good and bad labels to be inferred; The inference good or bad probability operation subsystem establishes the correlation between the inference good or bad probability according to the fuzzy expansion good sample weight and the fuzzy expansion bad sample weight; and uses the KGB model scoring score and the AR model scoring score to perform weight operation allocation; The cross-correlation relationship between the good and bad probabilities is established based on the fuzzy expanded good sample weight and the fuzzy expanded bad sample weight respectively, including: on samples with good and bad risk performance, 0 and 1 are used as good sample labels and bad sample labels respectively, and the logistic regression model is trained with the KGB model score and the AR model score to obtain the relationship between the overdue probability and the KGB model score and the AR model score, and the Logistic Regression Overdue Probability Prediction Relationship; the Logistic Regression Overdue Probability Prediction Relationship is applied to the rejected samples and the unwithdrawn samples, and the probability of each credit data being inferred to be bad is obtained accordingly; the weight of the sample that needs to be inferred is multiplied by the inferred bad probability to obtain the extended bad sample weight, and similarly, the weight of the sample that needs to be inferred is multiplied by the predicted good probability to obtain the extended good sample weight.

10. A multi-dimensional scoring credit data risk prediction control system according to claim 6, characterized in that: The weight distribution multi-model inference subsystem includes: The AGB model iteration subsystem establishes an AGB model on all samples of the set of good and bad label samples to be inferred, and performs iterative operations; the iterative operations include: setting a maximum number of iterations, and stopping the iterative operation when the maximum number of iterations is reached or the iteration results converge; The risk prediction model inference subsystem determines whether the AGB model has similar prediction performance on different types of customer groups. When the AGB model has similar prediction performance on different types of customer groups, the inference results converge and the final inference results of risk prediction are obtained. When the AGB model does not have similar prediction performance on different types of customer groups, the weight allocation and iterative calculation are redistributed until the inference results converge and the final inference results of risk prediction are obtained.

Images