Risk control rule updating method and device, and electronic equipment

Abstract

This application discloses a method, apparatus, and electronic device for updating risk control rules. The method includes: acquiring historical and incremental datasets related to credit business, and identifying rules in the risk control rule base whose performance indicators are below a preset threshold for optimization; randomly selecting multiple credit applications from a first credit application to obtain a second credit application, and generating a bias-corrected historical dataset; training an incremental learning model based on the incremental dataset, the rules to be optimized, and the bias-corrected historical dataset to obtain convergence parameters; generating a candidate rule set based on the convergence parameters, determining the model evaluation index of the candidate rule set on the historical dataset, identifying the candidate rule with the highest model evaluation index as the optimized rule, and updating the risk control rule base based on the optimized rule. This application solves the technical problem of risk control rules gradually becoming overly conservative due to the amplification of closed-loop bias during the iterative optimization process of the risk control rule base.

Description

Technical Field

[0001] This application relates to the field of financial technology, and more specifically, to a method, apparatus, and electronic device for updating risk control rules. Background Technology

[0002] In fintech, particularly in the area of ​​SME lending risk control, iterative optimization of the rule base is crucial for ensuring loan security and business sustainability. Traditional methods primarily rely on previously approved loan application data when updating risk control rules, forming the basis of historical datasets. However, this update mechanism suffers from a fundamental flaw: the so-called "closed-loop bias amplification" phenomenon, which manifests itself in the following ways:

[0003] 1. Data Selection Bias: The loan samples in historical datasets only reflect cases that successfully passed risk control rule reviews, failing to cover all applications rejected by the rule base. This is because rejected applications do not enter subsequent loan cycles, thus their actual repayment performance cannot be obtained, resulting in a "selection bias" in the dataset. This bias causes the model to over-rely on approved, high-quality customer data during training, ignoring the more complex and variable aspects of the risk distribution, especially those cases close to the rejection threshold.

[0004] 2. Overly Conservative Trend: As risk control rules iterate, models may tend to adopt stricter and more conservative rules to reduce error rates on historical datasets and thus minimize defaults. While this strategy may seem to improve model accuracy in the short term, it could actually lead to increasingly stringent credit thresholds, excluding small and micro enterprises with good credit. In the long run, this trend will result in a decline in market share and harm the long-term interests of financial institutions.

[0005] 3. Self-reinforcing negative feedback mechanism: Since model updates are based on historical data generated by its own decisions, over time, the updated rules will be more inclined to make judgments that conform to its previous decision patterns, even if these decisions are not necessarily optimal. This mechanism forms a self-reinforcing cycle, that is, the more optimized the model is, the more likely it is to reject new applications, which further reduces the positive samples that can be used for future training, exacerbating the aforementioned tendency towards excessive conservatism.

[0006] 4. Imperceptible Systemic Risks: The consequences of amplified closed-loop bias may not be readily apparent in the short term, as the model's performance on existing data may appear to be improving. However, in the long run, this overly conservative strategy will significantly reduce the amount of loans issued, limit the depth and breadth of financial services, and damage the market competitiveness and reputation of financial institutions.

[0007] In conclusion, the amplification of closed-loop bias is the fundamental reason why the risk control rule base gradually becomes overly conservative during iterative optimization. This phenomenon not only limits the system's ability to grasp the true risk distribution but may also hinder the development and expansion of financial institutions in the micro and small enterprise lending market.

[0008] There is currently no effective solution to the above problems. Summary of the Invention

[0009] This application provides a method, apparatus, and electronic device for updating risk control rules, in order to at least solve the technical problem that risk control rules gradually become overly conservative due to the amplification of closed-loop deviations in the iterative optimization process of the risk control rule base.

[0010] According to one aspect of this application, a method for updating risk control rules is provided, comprising: acquiring a historical dataset and an incremental dataset related to credit business, and identifying rules to be optimized in a risk control rule base whose performance indicators are less than a preset threshold, wherein the incremental dataset is a time-series data batch used to trigger an incremental learning process; acquiring a first credit application, wherein the first credit application is a credit application rejected by the risk control rule base; randomly selecting multiple credit applications from the first credit application to obtain a second credit application, and generating a biased historical dataset based on the historical dataset and the second credit application; training an incremental learning model based on the incremental dataset, the rules to be optimized, and the biased historical dataset, and updating the model parameters during the training process until a convergence condition is reached to obtain convergence parameters, wherein the incremental learning model is used to optimize the rules to be optimized; generating a candidate rule set based on the convergence parameters, determining the model evaluation index of the candidate rule set on the historical dataset, identifying the candidate rule with the highest model evaluation index as the optimized rule, and updating the risk control rule base based on the optimized rule.

[0011] Optionally, multiple credit applications are randomly selected from the first credit application to obtain a second credit application. Based on the historical dataset and the second credit application, a bias-free historical dataset is generated, including: for each first credit application, generating a random number within a preset interval; comparing the random number with a preset exploration rate parameter; if the random number is less than the exploration rate parameter, changing the decision status of the first credit application from rejection to approval, and identifying the first credit application as the second credit application; tracking the second credit application in subsequent credit cycles to obtain its credit performance; once the credit performance is determined, generating an exploration sample set based on the sample pair consisting of the target feature vector corresponding to the second credit application and the credit label corresponding to the credit performance; merging the exploration sample set with the historical dataset to obtain the bias-free historical dataset.

[0012] Optionally, an incremental learning model is trained based on the incremental dataset, the rules to be optimized, and the biased historical dataset. During training, the model parameters are updated until convergence is achieved. This includes: parsing the rules to be optimized into a set of optimizable parameters; determining the credit risk prediction result corresponding to the input sample based on the parameter set and the feature vector of the input sample; summing the prediction loss of the credit risk prediction results of all samples in the incremental dataset to obtain a first loss term; determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base; assigning an importance weight to each exploration sample based on the probability estimate, where the importance weight is negatively correlated with the probability estimate; determining the weighted prediction loss of the credit risk prediction result of each sample in the exploration sample set based on the importance weight, and summing the weighted prediction loss of all samples in the exploration sample set to obtain a second loss term; determining the counterfactual risk minimization objective function for model training based on the first and second loss terms; and iteratively optimizing the counterfactual risk minimization objective function using a gradient descent algorithm to update the model parameters until convergence is achieved.

[0013] Optionally, determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base includes: constructing a probability estimation model based on the historical decision logs of the risk control rule base, wherein each sample in the historical decision log includes a target feature vector and the actual decision result corresponding to the target feature vector, and the probability estimation model is used to output the probability estimate of rejection by the risk control rule base based on the input feature vector; constructing a likelihood function based on the historical decision logs, wherein the likelihood function is the product of the conditional probabilities of all samples; for samples that were rejected in the past, the conditional probability is the rejection probability predicted by the probability estimation model; for samples that were allowed in the past, the conditional probability is the difference between 1 and the rejection probability; optimizing the solution by maximizing the likelihood function to obtain the parameters of the probability estimation model, thus obtaining a fully trained probability estimation model; and using the fully trained probability estimation model to determine the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base.

[0014] Optionally, based on the probability estimate, an importance weight is assigned to each exploration sample, including: calculating the initial weight value corresponding to the exploration sample based on the probability estimate of each exploration sample being rejected by the risk control rule base; and dividing the initial weight value corresponding to the exploration sample by the sum of the initial weight values ​​of all exploration samples in the exploration sample set to obtain the importance weight of the exploration sample.

[0015] Optionally, the gradient descent algorithm is used to iteratively optimize the objective function for minimizing counterfactual risk to update the model parameters until the convergence condition is met. This includes: constructing a training data pool that includes a biased historical dataset and an incremental dataset; and performing the following operations during each iteration: randomly sampling from the training data pool to obtain a batch sample set; calculating the loss value and gradient of the objective function for minimizing counterfactual risk based on the batch sample set; updating the model parameters using an adaptive learning rate method according to the gradient; and repeating the iteration process until the preset iteration termination condition is met.

[0016] Optionally, based on the first loss term and the second loss term, the objective function for minimizing counterfactual risk used for model training is determined, including: selecting multiple candidate values ​​within a preset numerical range, and for each candidate value, performing the following operations: determining the objective function for minimizing counterfactual risk based on the candidate value, and training the incremental learning model based on the objective function for minimizing counterfactual risk to obtain the rule parameter solution; evaluating the performance of the rule parameter solution on the validation set to obtain the model's discriminative ability; among all candidate values, determining the candidate value whose corresponding rule parameter solution exhibits the best model discriminative ability on the validation set as the balancing hyperparameter; and determining the objective function for minimizing counterfactual risk based on the product of the first loss term, the second loss term, and the balancing hyperparameter.

[0017] Optionally, the model evaluation metrics for the candidate rule set are determined on a historical dataset, including: performing batch predictions on each candidate rule in the candidate rule set on the historical dataset to obtain prediction results; calculating the ratio of the total number of correctly predicted samples to the total number of samples based on the prediction results and the actual results to obtain the accuracy; calculating the ratio of the number of correctly predicted performance samples to the actual number of performance samples based on the prediction results and the actual results to obtain the recall; calculating the first cumulative distribution of performance samples on the prediction results and the second cumulative distribution of default samples on the prediction results, and calculating the maximum difference between the first and second cumulative distributions to obtain the model discrimination; calculating the estimated revenue from correctly releasing performance samples and the estimated loss from incorrectly releasing default samples based on the prediction results, and calculating the net value between the estimated revenue and the estimated loss to obtain the expected business revenue; and weighting and summing the accuracy, recall, model discrimination, and expected business revenue based on preset weight coefficients to obtain the model evaluation metrics for each candidate rule.

[0018] According to another aspect of this application, a risk control rule updating device is also provided, comprising: a first acquisition module, used to acquire historical datasets and incremental datasets related to credit business, and determine the rules to be optimized whose performance indicators are less than a preset threshold in the risk control rule base, wherein the incremental dataset is a time-series data batch used to trigger the incremental learning process; a second acquisition module, used to acquire a first credit application, wherein the first credit application is a credit application rejected by the risk control rule base; randomly select multiple credit applications from the first credit application to obtain a second credit application, and generate a biased historical dataset based on the historical dataset and the second credit application; a first processing module, used to train an incremental learning model based on the incremental dataset, the rules to be optimized, and the biased historical dataset, and update the model parameters during the training process until the convergence condition is reached to obtain convergence parameters, wherein the incremental learning model is used to optimize the rules to be optimized; and a second processing module, used to generate a candidate rule set based on the convergence parameters, determine the model evaluation index of the candidate rule set on the historical dataset, determine the candidate rule with the highest model evaluation index as the optimized rule, and update the risk control rule base based on the optimized rule.

[0019] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned method for updating risk control rules.

[0020] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program executes the above-described method for updating risk control rules during runtime.

[0021] According to another aspect of this application, a computer program is also provided, wherein when the computer program is executed by a processor, it implements the above-mentioned method for updating risk control rules.

[0022] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for updating risk control rules.

[0023] In this application, the following methods are employed: acquiring historical and incremental datasets related to credit business; identifying rules to be optimized whose performance indicators are below a preset threshold in the risk control rule base; wherein the incremental dataset is a time-series data batch used to trigger the incremental learning process; acquiring a first credit application, wherein the first credit application is a credit application rejected by the risk control rule base; randomly selecting multiple credit applications from the first credit application to obtain a second credit application; and generating a bias-corrected historical dataset based on the historical dataset and the second credit application; training an incremental learning model based on the incremental dataset, the rules to be optimized, and the bias-corrected historical dataset; and during the training process... The model parameters are updated until the convergence condition is met, and convergence parameters are obtained. The incremental learning model is used to optimize the rules to be optimized. A candidate rule set is generated based on the convergence parameters. The model evaluation index of the candidate rule set is determined on the historical dataset. The candidate rule with the highest model evaluation index is determined as the optimized rule. The risk control rule base is updated based on the optimized rule. This method achieves the purpose of correcting data selection bias and optimizing rule performance, thereby achieving the technical effect of maintaining the robustness of risk control rules. It also solves the technical problem that risk control rules gradually become overly conservative due to the amplification of closed-loop bias in the iterative optimization process of the risk control rule base. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 This is a flowchart of a risk control rule update method according to an embodiment of this application;

[0026] Figure 2 This is a structural diagram of a risk control rule updating device according to an embodiment of this application;

[0027] Figure 3 This is a hardware structure block diagram of a computer terminal for a risk control rule update method according to an embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] According to an embodiment of this application, a method embodiment for updating risk control rules is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0031] Figure 1 This is a flowchart of a risk control rule update method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0032] Step S102: Obtain historical datasets and incremental datasets related to credit business, and identify rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series data batch used to trigger the incremental learning process.

[0033] Historical datasets refer to a comprehensive collection of records accumulated over a long period of time in the lending process of financial institutions, detailing loan applications, approval results, and applicant credit performance for micro and small enterprises. Historical datasets include feature vectors of past applicants (such as corporate financial status, credit score, industry type, and years of operation) and corresponding credit records (such as on-time repayment, overdue payments, and defaults), serving as a crucial basis for training and optimizing risk control rules.

[0034] Incremental datasets refer to batches of credit business data that are newly generated after a specific time point and arranged in a time series. Incremental datasets include the feature profiles and preliminary approval results of the latest applications, but do not yet fully record the true credit performance of each application. The introduction of incremental datasets aims to enable risk control rules to quickly adapt to environmental changes, triggering dynamic updates and optimizations of rules through real-time data feedback, thereby maintaining the timeliness and flexibility of decision-making.

[0035] A risk control rule base is a collection of decision-making rules established by financial institutions to assess the risk of credit applications. Each rule in the rule base defines a mapping relationship between a specific combination of features and a risk level, used to quickly determine whether a credit application meets the lending criteria.

[0036] Performance metrics are indicators used to quantitatively evaluate the effectiveness of risk control rules in processing credit applications, such as accuracy, recall, F1 score, and area under the AUC-ROC curve.

[0037] Preferably, the incremental dataset can be obtained by the following method.

[0038] The latest credit business data and external environment data of micro and small enterprises are obtained from multiple heterogeneous data sources. The data is then cleaned, denoised, and standardized to generate a time-series incremental dataset with a unified format. Where t represents the current data batch, and the data is represented by a feature vector. This is to facilitate subsequent risk control rule processing and incremental learning.

[0039] Step S104: Obtain the first credit application, which is a credit application rejected by the risk control rule base; randomly select multiple credit applications from the first credit application to obtain the second credit application, and generate a biased historical dataset based on the historical dataset and the second credit application.

[0040] Step S104 can be achieved by the following method: Incremental dataset... Enter the currently deployed risk control rule base The trial run is underway, and each rule is calculated based on a preset set of performance indicators. exist Performance values When the presence is detected Rules to be optimized When this happens, the rule optimization process is automatically triggered, in which... A preset threshold is set. For current and subsequent inflows of credit applications, a portion of applications that should have been rejected by the rule base R are randomly allowed with probability ε, forming an exploration sample set E. The feature vector of the samples in the exploration sample set is x, and their final credit performance is y. After the sample results are fully recorded, they are injected into the historical dataset H to form a biased historical dataset.

[0041] Step S106: Train an incremental learning model based on the incremental dataset, the rule to be optimized, and the biased historical dataset. During the training process, update the model parameters until the convergence condition is met to obtain the convergence parameters. The incremental learning model is used to optimize the rule to be optimized.

[0042] In step S106, the incremental learning model (incremental learning engine) loads the rule to be optimized, the bias-reduced historical dataset, and the incremental dataset as input, and parses the rule into a differentiable parameterized form. And adopt the objective function of minimizing counterfactual risk. The trainer uses a mini-batch stochastic gradient descent algorithm to update parameters until the maximum number of iterations is reached or the gradient norm falls below a threshold, ultimately outputting convergent parameters. .

[0043] Step S108: Generate a candidate rule set based on the convergence parameters, determine the model evaluation index of the candidate rule set on the historical dataset, determine the candidate rule with the highest model evaluation index as the optimized rule, and update the risk control rule base based on the optimized rule.

[0044] In step S108, batch prediction of the candidate rule set is performed on the full historical dataset. Accuracy, Recall, KS statistic, and Profit (expected business benefit indicator) are calculated, and a weighted summation is used to calculate the overall score (Score). The weighting coefficients are determined using the analytic hierarchy process (AHP). Further, the candidate rule with the highest overall score is selected as the final optimized rule. After approval, the risk control rule base is seamlessly updated via hot deployment. Replace with This enables rapid, unbiased, iterative updates of risk control rules.

[0045] The above steps first acquire correction data through controlled exploration, then use this data to optimize the rules to remove biases, and finally verify and deploy the rules in a simulation environment based on global data distribution. These steps form a complete, self-examining, and self-correcting automated risk control rule evolution system. This not only enables rapid rule iteration but also ensures the correctness of the iteration direction, ultimately improving risk control accuracy while safeguarding the healthy development of the business and addressing the core risks faced by risk control systems in long-term autonomous operation.

[0046] The following are Figure 1 The steps shown are illustrated and explained by way of example.

[0047] According to some optional embodiments of this application, multiple credit applications are randomly selected from the first credit application to obtain a second credit application. A debiased historical dataset is generated based on the historical dataset and the second credit application. This can be achieved through the following method: For each first credit application, a random number within a preset interval is generated; the random number is compared with a preset exploration rate parameter; if the random number is less than the exploration rate parameter, the decision status of the first credit application is changed from rejection to approval, and the first credit application is identified as the second credit application; in subsequent credit cycles, the second credit application is tracked to obtain its credit performance; once the credit performance is determined, an exploration sample set is generated based on the sample pair consisting of the target feature vector corresponding to the second credit application and the credit label corresponding to the credit performance; the exploration sample set is merged with the historical dataset to obtain the debiased historical dataset.

[0048] In this embodiment, for each instance marked as a first credit application by the risk control rule base, i.e., those credit requests that have been initially rejected based on existing rules, a random number p is generated, which follows a uniform distribution. The principle is to randomly select a number within the preset interval [0,1]. The generated random number p is compared with a pre-set exploration rate parameter ε. If the value of p is less than or equal to the exploration rate ε, the decision status of the first credit application will undergo a proactive change: from the original rejection status to the approval status. Applications approved in this special state are called second credit applications, signifying their selection to participate in the controlled exploration process in order to collect more comprehensive credit behavior data.

[0049] Once a second credit application is approved, its entire process is tracked until its final credit performance (e.g., on-time repayment or default). The credit performance of the second credit application is recorded from approval to credit closure. Where y=1 indicates that the applicant fulfills its obligations on schedule, and y=0 indicates that a default has occurred. Once the corresponding y is fully observed, a sample pair (x,y) consisting of a feature vector and credit performance is generated, where x represents the target feature vector corresponding to the second credit application, i.e., the business and environmental data at the time of application, and y represents the final credit performance of the second credit application.

[0050] Based on the sample pair (x, y) of the second credit application, an exploration sample set E is constructed and expanded. The exploration sample set E aims to accumulate real credit behavior data of credit applications that would normally be rejected but were granted approval through a controlled exploration mechanism, thus compensating for selection bias in the historical dataset. When the number of samples in the exploration sample set E is sufficient and the credit performance of each sample is determined, the exploration sample set E is merged with the historical dataset H to form a bias-free historical dataset. Among them, the debiased historical dataset includes a more comprehensive and balanced credit risk performance, providing a more reliable data foundation for subsequent incremental learning and rule optimization.

[0051] The above steps involve an exploration with a low probability. A portion of credit applications that would otherwise be rejected by the existing risk control rule base are proactively and randomly approved. These approved applications and their subsequent actual credit performance are fully recorded and constitute an exploration sample set E. This sample set is then fused with the historical dataset H to generate a biased historical dataset. The steps described above proactively create feedback information about "data blind spots" (i.e. rejected applications), providing crucial counterfactual data sources for subsequent incremental learning and optimization. This breaks the self-fulfilling data loop and lays the data foundation for generating unbiased estimates.

[0052] According to some alternative embodiments of this application, an incremental learning model is trained based on an incremental dataset, the rule to be optimized, and a biased historical dataset. During training, the model parameters are updated until convergence is achieved. This can be achieved through the following method: The rule to be optimized is parsed into a set of optimizable parameters; based on the parameter set and the feature vector of the input sample, the credit risk prediction result corresponding to the input sample is determined; the prediction loss of the credit risk prediction results of all samples in the incremental dataset is summed to obtain a first loss term; the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base is determined; based on the probability estimate, an importance weight is assigned to each exploration sample, wherein the importance weight is negatively correlated with the probability estimate; based on the importance weight, the weighted prediction loss of the credit risk prediction result of each sample in the exploration sample set is determined, and the weighted prediction loss of all samples in the exploration sample set is summed to obtain a second loss term; based on the first and second loss terms, a counterfactual risk minimization objective function is determined for model training; the gradient descent algorithm is used to iteratively optimize the counterfactual risk minimization objective function to update the model parameters until convergence is achieved.

[0053] In this embodiment, the rule to be optimized is... Bias-free historical datasets and incremental datasets The inputs are fused to construct the training inputs for parameter updates, where Represents the time-series incremental data of the current batch. It is formed by merging the historical dataset and the exploration sample set E. The incremental learning model aims to minimize counterfactual risk during optimization, and its loss function is defined as:

[0054]

[0055] in, As the first loss item, This is the second loss item. The set of parameters for the rule to be optimized. Representation rules in input samples The prediction results above The loss function is used to label the true credit performance of this sample. Cross-entropy is defined as follows:

[0056]

[0057] loss function middle To balance hyperparameters, the relative importance of incremental and exploratory data in the overall loss is controlled. An importance weight w(x) is introduced for the exploratory samples when calculating the loss, defined as:

[0058]

[0059] in, This provides a probability estimate for rejection decisions on input x for the original rule base. This weight enhances the influence of samples in high rejection probability regions of the original rule base, thereby mitigating distribution distortion caused by rule selection bias. The loss function is optimized using stochastic gradient descent, and the parameter set... The solution is updated gradually during the iteration process until it converges to the optimal solution. Finally, the updated rules are output. This enables incremental iterative improvement of the rules to be optimized.

[0060] The above steps introduce an objective function based on a counterfactual risk minimization framework. This replaces the traditional empirical risk minimization objective. The function incorporates the value of the controlled exploration sample set into the optimization process through importance weighting, thereby correcting data selection bias caused by historical rule rejections. These steps enable the incremental learning model to learn unbiased, globally superior risk control rules directly from biased observation data using counterfactual reasoning without service interruption. This fundamentally solves the problem of closed-loop bias amplification and avoids the rule iteration becoming conservative.

[0061] Specifically, determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base can be achieved through the following method: Based on the historical decision logs of the risk control rule base, a probability estimation model is constructed. Each sample in the historical decision log includes a target feature vector and the corresponding actual decision result. The probability estimation model is used to output the probability estimate of rejection by the risk control rule base based on the input feature vector. A likelihood function is constructed based on the historical decision logs. The likelihood function is the product of the conditional probabilities of all samples. For samples historically rejected, the conditional probability is the rejection probability predicted by the probability estimation model. For samples historically approved, the conditional probability is the difference between 1 and the rejection probability. By maximizing the likelihood function, the parameters of the probability estimation model are optimized and solved to obtain the fully trained probability estimation model. Using the fully trained probability estimation model, the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base is determined.

[0062] In this embodiment, the probability estimation model is constructed in the form of logistic regression, and its input is the feature vector of the credit application. The output is an estimate of the probability that the sample will be rejected by the original rule base. The probability estimate is defined as:

[0063]

[0064] Where g(x) is a linear discriminant function, of the form: In this function, Let be the weight vector, representing the linear contribution of each feature dimension to the rejection decision. The bias term adjusts the overall shift of the classification boundary. Model parameters w and b are trained by maximizing the likelihood function of historical decision logs, which contain the correspondence between feature vectors and actual rejection results in the rule base. The objective during training is to maximize the likelihood function.

[0065]

[0066] Where N represents the number of log samples. This indicates that the rule base is for samples The actual decision-making result, if This indicates that the sample was rejected in the past. This indicates that passage has been permitted. Converged parameter estimates can be obtained through log-likelihood transformation and gradient descent optimization. and The trained logistic regression model provides an estimate of the rejection probability for any input feature vector x, and this output serves as the basis for the aforementioned importance weight calculation formula. The denominator of the loss function is used to reflect the relative importance of different samples in the loss function of the incremental learning model, ensuring that the exploration samples receive weight compensation that matches their distribution scarcity during the optimization process.

[0067] Specifically, based on the probability estimate, each exploration sample is assigned an importance weight, which can be achieved as follows: based on the probability estimate of each exploration sample being rejected by the risk control rule base, calculate the initial weight value corresponding to the exploration sample; divide the initial weight value corresponding to the exploration sample by the sum of the initial weight values ​​of all exploration samples in the exploration sample set to obtain the importance weight of the exploration sample.

[0068] For example, a logistic regression function can be used to determine an estimate of the probability that an exploratory sample will be rejected by the original rule base:

[0069]

[0070] in, The parameters w and b have been obtained from historical decision logs. Based on this probability estimate, the unnormalized importance weights are first calculated. This weight is inversely proportional to the probability of a sample being rejected, giving greater weight to exploration samples that have historically been more likely to be rejected. To avoid excessive weight differences between different samples affecting training stability, the unnormalized weights of all exploration samples are normalized to obtain the final importance weights:

[0071]

[0072] Where M represents the size of the exploration sample set E, This represents all samples in the set. The normalized weight vector can be expressed as: The normalized weight vector is passed as input parameters to the counterfactual risk minimization objective function to adjust the contribution of the exploration samples in the loss function. This method enhances the learning ability for rare samples during the optimization process, compensates for the distribution bias caused by rule solidification, and improves the final generated rules. It more closely reflects the actual risk distribution.

[0073] The above steps quantify the rejection risk of the exploration samples through a probability estimation model, and then calculate the initial weight values ​​through inverse proportional calculation and normalization to obtain the importance weights. This not only reflects the recognition of the unique value of the exploration samples, but also ensures the mathematical rigor and engineering feasibility of the training process, providing a solid data foundation and weight guarantee for the iterative optimization of rules based on counterfactual risk minimization in the incremental learning model.

[0074] Specifically, the gradient descent algorithm is used to iteratively optimize the objective function for minimizing counterfactual risk to update the model parameters until the convergence condition is met. This can be achieved as follows: Construct a training data pool that includes a biased historical dataset and an incremental dataset; In each iteration, perform the following operations: Randomly sample from the training data pool to obtain a batch sample set; Based on the batch sample set, calculate the loss value and gradient of the objective function for minimizing counterfactual risk; Update the model parameters using an adaptive learning rate method according to the gradient; Repeat the iterative process until the preset iteration termination condition is met.

[0075] In this embodiment, the incremental learning model consists of a rule parser, a data buffer, a trainer, and a rule generator. The rule parser receives the rules to be optimized. and the rules to be optimized Convert to differentiable parameterized form ,in The parameter set represents the rule, ensuring that the rule can be embedded into the gradient-based optimization process. A data buffer is used to store the historical dataset for bias correction. And the incremental dataset of the current batch This ensures that the training process can utilize both long-accumulated data and the latest incoming data in continuous iterations, thus balancing stability and timeliness. The trainer takes the data provided by the buffer as input and optimizes the objective function of minimizing counterfactual risk. The trainer uses a mini-batch stochastic gradient descent algorithm to update parameters. In each iteration, based on the gradient The parameters are adjusted so that the loss function decreases continuously during training and eventually converges to the optimal solution. The rule generator receives the updated parameters, converts them back from parameterized representation into executable rule expressions, and generates a candidate rule set. These candidate rules retain the results of parameter optimization and serve as input for subsequent simulation verification and deployment steps, enabling rapid, unbiased iterative updates of the rule base.

[0076] When performing parameter optimization, the trainer first randomly samples a mini-batch of data from the data buffer. This dataset contains feature vectors x and corresponding ground truth labels y. At the current iteration step t, the trainer is based on the parameter set... Calculate the loss function:

[0077]

[0078] in, Indicates in the parameter The rule prediction results are then obtained. The trainer then calculates the gradient of the loss function to obtain... Parameter updates employ an adaptive learning rate method, following the formula in each iteration. Update, among which Let be the learning rate for the t-th iteration, dynamically adjusted according to the training progress to ensure convergence stability. The iterative process continues until the preset maximum number of iterations is reached, or when the gradient norm is... The process terminates when the value falls below a set threshold. The final output parameters... To obtain a converged solution, the rule generator then converts it into an executable candidate rule form, which is then used for subsequent simulation verification and deployment.

[0079] The above steps concretize the incremental learning model into a collaborative system comprising four components: a rule parser, a data buffer, a trainer, and a rule generator. The rule parser is responsible for converting non-differentiable business rules into differentiable parameterized forms. The data buffer is responsible for managing the historical data for bias correction. With incremental data The trainer specifically executes a mini-batch stochastic gradient descent optimization algorithm based on the CRM objective function; the rule generator is responsible for optimizing the parameters. The process involves reverse engineering to recompile interpretable and executable business rules. These steps provide an efficient, stable, and interpretable infrastructure for automated rule optimization, ensuring the reliability and engineering feasibility of the closed-loop transformation process from rules to models and back to rules.

[0080] Specifically, the objective function for minimizing counterfactual risk used for model training is determined based on the first loss term and the second loss term. This can be achieved as follows: Select multiple candidate values ​​within a preset numerical range, and for each candidate value, perform the following operations: Determine the objective function for minimizing counterfactual risk based on the candidate value, and train the incremental learning model based on the objective function for minimizing counterfactual risk to obtain the rule parameter solution; Evaluate the performance of the rule parameter solution on the validation set to obtain the model's discriminative ability; Among all candidate values, determine the candidate value whose corresponding rule parameter solution exhibits the best model discriminative ability on the validation set as the balancing hyperparameter; Determine the objective function for minimizing counterfactual risk based on the product of the first loss term, the second loss term, and the balancing hyperparameter.

[0081] In this embodiment, the balancing hyperparameters The determination process can employ a grid search method, with the search range set to [0.1, 10.0]. In each search, a candidate value is fixed. And substitute it into the objective function of the incremental learning engine:

[0082]

[0083] in, Let E be the incremental dataset, E be the exploration sample set, and w(x) be the importance weight. This is the cross-entropy loss function. The trainer is at a fixed... Under the given conditions, the objective function is optimized to obtain the corresponding rule parameter solution. Then, the rules were applied to the validation set. The performance is evaluated using a validation set consisting of a randomly sampled holdout of historical data, ensuring independence from the training data. The evaluation metric can be the Kolmogorov–Smirnov statistic, defined as:

[0084]

[0085] in, This represents the cumulative distribution function of good samples at a scoring threshold z. The cumulative distribution function represents the bad samples, and the KS statistic measures the model's ability to distinguish between defaulted and fulfilled samples. Among all candidate... Calculate the corresponding KS value and select the one that maximizes the KS statistic. As the final balancing hyperparameter, it ensures that the weighting of incremental data and exploration samples reaches the optimal balance during the optimization process, thereby improving the rule's ability to distinguish real business distributions.

[0086] In some optional embodiments, determining the model evaluation metrics for the candidate rule set on a historical dataset can be achieved as follows: On the historical dataset, perform batch predictions for each candidate rule in the candidate rule set to obtain prediction results; based on the prediction results and the actual results, calculate the ratio of the total number of correctly predicted samples to the total number of samples in the overall dataset to obtain the accuracy; based on the prediction results and the actual results, calculate the ratio of the number of correctly predicted fulfillment samples to the actual number of fulfillment samples to obtain the recall; calculate the first cumulative distribution of fulfillment samples on the prediction results and the second cumulative distribution of default samples on the prediction results, and calculate the maximum difference between the first and second cumulative distributions to obtain the model discrimination; based on the prediction results, calculate the estimated revenue from correctly releasing fulfillment samples and the estimated loss from incorrectly releasing default samples, and calculate the net value between the estimated revenue and the estimated loss to obtain the expected business revenue; based on preset weighting coefficients, perform a weighted sum of accuracy, recall, model discrimination, and expected business revenue to obtain the model evaluation metrics for each candidate rule.

[0087] In this embodiment, the simulation verification environment uses the full historical dataset. As a foundational test set, it ensures that candidate rules are tested on the most complete data distribution. The above candidate rule set Perform batch predictions and calculate multi-dimensional performance metrics for each candidate rule.

[0088] Accuracy is defined as:

[0089]

[0090] Wherein, TP represents the number of samples correctly identified as fulfilling the contract, TN represents the number of samples correctly identified as defaulting, FP represents the number of samples that actually defaulted but were wrongly allowed to proceed, and FN represents the number of samples that actually fulfilled the contract but were wrongly rejected.

[0091] Recall rate is defined as: The recall rate is used to represent the rule's ability to cover the fulfillment samples.

[0092] The ability to differentiate was assessed using the Kolmogorov–Smirnov statistic, the formula of which is: ,in Let be the cumulative distribution function of the performance samples. Let be the cumulative distribution function of the defaulted samples.

[0093] The expected business revenue metric is obtained by weighted summation of the forecast results. Where G represents the average revenue from a single loan repayment, and C represents the average loss from a single default. For true credit results, The prediction results are for the candidate rules.

[0094] The above multiple indicators are weighted and aggregated into a comprehensive score. ,in , , , These are pre-defined weighting coefficients used to balance accuracy, coverage, discriminative power, and business value, to meet [the following criteria]. + + + =1 and , , , >0. Finally, the rule with the largest comprehensive score S is selected as the final optimization rule from all candidate rules. This information is then submitted to subsequent review and hot deployment processes to enable rapid updates to the rule base.

[0095] Figure 2 This is a structural diagram of a risk control rule updating device according to an embodiment of this application, such as... Figure 2 As shown, the device includes:

[0096] The first acquisition module 22 is used to acquire historical datasets and incremental datasets related to credit business, and to determine the rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series data batch used to trigger the incremental learning process.

[0097] The second acquisition module 24 is used to acquire the first credit application, which is a credit application rejected by the risk control rule base; randomly select multiple credit applications from the first credit application to obtain the second credit application, and generate a biased historical dataset based on the historical dataset and the second credit application.

[0098] The first processing module 26 is used to train an incremental learning model based on the incremental dataset, the rules to be optimized, and the biased historical dataset. During the training process, the model parameters are updated until the convergence condition is met, and the convergence parameters are obtained. The incremental learning model is used to optimize the rules to be optimized.

[0099] The second processing module 28 is used to generate a candidate rule set based on the convergence parameters, determine the model evaluation index of the candidate rule set on the historical dataset, determine the candidate rule with the highest model evaluation index as the optimized rule, and update the risk control rule base based on the optimized rule.

[0100] Optionally, multiple credit applications are randomly selected from the first credit application to obtain a second credit application. A debiased historical dataset is then generated based on the historical dataset and the second credit application. Specifically, this involves the following steps: For each first credit application, a random number within a preset range is generated; the random number is compared with a preset exploration rate parameter; if the random number is less than the exploration rate parameter, the decision status of the first credit application is changed from rejection to approval, and the first credit application is identified as the second credit application; in subsequent credit cycles, the second credit application is tracked to obtain its credit performance; once the credit performance is determined, an exploration sample set is generated based on the sample pairs consisting of the target feature vector corresponding to the second credit application and the credit label corresponding to the credit performance; the exploration sample set is merged with the historical dataset to obtain the debiased historical dataset.

[0101] Optionally, an incremental learning model is trained based on the incremental dataset, the rules to be optimized, and the biased historical dataset. During training, the model parameters are updated until convergence is achieved. Specifically, this includes the following steps: parsing the rules to be optimized into a set of optimizable parameters; determining the credit risk prediction result corresponding to the input sample based on the parameter set and the feature vector of the input sample; summing the prediction losses of the credit risk prediction results for all samples in the incremental dataset to obtain a first loss term; determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base; assigning an importance weight to each exploration sample based on the probability estimate, where the importance weight is negatively correlated with the probability estimate; determining the weighted prediction loss of the credit risk prediction result for each sample in the exploration sample set based on the importance weight, and summing the weighted prediction losses of all samples in the exploration sample set to obtain a second loss term; determining the counterfactual risk minimization objective function for model training based on the first and second loss terms; and iteratively optimizing the counterfactual risk minimization objective function using a gradient descent algorithm to update the model parameters until convergence is achieved.

[0102] Optionally, determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base includes the following steps: Constructing a probability estimation model based on historical decision logs of the risk control rule base, wherein each sample in the historical decision log includes a target feature vector and the corresponding true decision result, and the probability estimation model is used to output the probability estimate of rejection by the risk control rule base based on the input feature vector; Constructing a likelihood function based on the historical decision logs, wherein the likelihood function is the product of the conditional probabilities of all samples; For samples historically rejected, the conditional probability is the rejection probability predicted by the probability estimation model; For samples historically approved, the conditional probability is the difference between 1 and the rejection probability; Optimizing the solution by maximizing the likelihood function to obtain the parameters of the probability estimation model, resulting in a fully trained probability estimation model; Using the fully trained probability estimation model, determining the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base.

[0103] Optionally, based on the probability estimate, an importance weight is assigned to each exploration sample, specifically including the following steps: based on the probability estimate of each exploration sample being rejected by the risk control rule base, the initial weight value corresponding to the exploration sample is calculated; the initial weight value corresponding to the exploration sample is divided by the sum of the initial weight values ​​of all exploration samples in the exploration sample set to obtain the importance weight of the exploration sample.

[0104] Optionally, the gradient descent algorithm is used to iteratively optimize the objective function for minimizing counterfactual risk to update the model parameters until the convergence condition is met. Specifically, this includes the following steps: constructing a training data pool that includes a biased historical dataset and an incremental dataset; in each iteration, performing the following operations: randomly sampling from the training data pool to obtain a batch sample set; calculating the loss value and gradient of the objective function for minimizing counterfactual risk based on the batch sample set; updating the model parameters using an adaptive learning rate method according to the gradient; and repeating the iteration process until the preset iteration termination condition is met.

[0105] Optionally, based on the first loss term and the second loss term, a counterfactual risk minimization objective function for model training is determined, specifically including the following steps: selecting multiple candidate values ​​within a preset numerical range, and for each candidate value, performing the following operations: determining the counterfactual risk minimization objective function based on the candidate value, and training the incremental learning model based on the counterfactual risk minimization objective function to obtain the rule parameter solution; evaluating the performance of the rule parameter solution on the validation set to obtain the model's discriminative ability; among all candidate values, determining the candidate value whose corresponding rule parameter solution exhibits the optimal model discriminative ability on the validation set as the balancing hyperparameter; and determining the counterfactual risk minimization objective function based on the product of the first loss term, the second loss term, and the balancing hyperparameter.

[0106] Optionally, the model evaluation metrics for the candidate rule set are determined on the historical dataset, specifically including the following steps: On the historical dataset, batch prediction is performed on each candidate rule in the candidate rule set to obtain prediction results; based on the prediction results and the actual results, the ratio of the total number of correctly predicted samples to the total number of samples is calculated to obtain the accuracy; based on the prediction results and the actual results, the ratio of the number of correctly predicted performance samples to the actual number of performance samples is calculated to obtain the recall; the first cumulative distribution of performance samples on the prediction results and the second cumulative distribution of default samples on the prediction results are calculated, and the maximum difference between the first and second cumulative distributions is calculated to obtain the model discrimination; based on the prediction results, the estimated revenue from correctly releasing performance samples and the estimated loss from incorrectly releasing default samples are calculated, and the net value between the estimated revenue and the estimated loss is calculated to obtain the expected business revenue; based on preset weight coefficients, the accuracy, recall, model discrimination, and expected business revenue are weighted and summed to obtain the model evaluation metrics for each candidate rule.

[0107] It should be noted that the above Figure 2The modules in can be program modules (e.g., a set of program instructions that implements a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0108] It should be noted that, Figure 2 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0109] Figure 3 A hardware block diagram of a computer terminal for implementing a method for updating risk control rules is shown. Figure 3 As shown, the computer terminal 30 may include one or more processors 302 (shown as 302a, 302b, ..., 302n in the figure) 302 (processor 302 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 304 for storing data, and a transmission module 306 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 3 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 30 may also include... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown.

[0110] It should be noted that the aforementioned one or more processors 302 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 30. As involved in the embodiments of this application, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0111] The memory 304 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the risk control rule update method in this embodiment. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, thereby realizing the aforementioned risk control rule update method. The memory 304 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 304 may further include memory remotely located relative to the processor 302, and these remote memories can be connected to the computer terminal 30 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0112] The transmission module 306 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 30. In one example, the transmission module 306 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 306 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0113] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 30.

[0114] It should be noted here that, in some optional embodiments, the above... Figure 3 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 3 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0115] It should be noted that, Figure 3 The computer terminal shown is used to execute Figure 1 The method for updating the risk control rules shown above also applies to this electronic device, and will not be repeated here.

[0116] This application also provides a non-volatile storage medium, which includes a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned method for updating risk control rules.

[0117] A non-volatile storage medium performs the following functions: It acquires historical and incremental datasets related to credit business, and identifies rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series batch of data used to trigger the incremental learning process. It acquires a first credit application, which is a credit application rejected by the risk control rule base. It randomly selects multiple credit applications from the first credit application to obtain a second credit application, and generates a biased historical dataset based on the historical dataset and the second credit application. It trains an incremental learning model based on the incremental dataset, the rules to be optimized, and the biased historical dataset, updating the model parameters during training until convergence is achieved, obtaining convergence parameters. The incremental learning model is used to optimize the rules to be optimized. It generates a candidate rule set based on the convergence parameters, determines the model evaluation index of the candidate rule set on the historical dataset, identifies the candidate rule with the highest model evaluation index as the optimized rule, and updates the risk control rule base based on the optimized rule.

[0118] This application also provides an electronic device, including: a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-mentioned risk control rule update method during runtime.

[0119] The processor runs a program that performs the following functions: acquiring historical and incremental datasets related to credit business, and identifying rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series batch of data used to trigger the incremental learning process. It acquires a first credit application, which is a credit application rejected by the risk control rule base. It randomly selects multiple credit applications from the first credit application to obtain a second credit application, and generates a biased historical dataset based on the historical dataset and the second credit application. It trains an incremental learning model based on the incremental dataset, the rules to be optimized, and the biased historical dataset, updating the model parameters during training until convergence is achieved, obtaining convergence parameters. The incremental learning model is used to optimize the rules to be optimized. It generates a candidate rule set based on the convergence parameters, determines the model evaluation index of the candidate rule set on the historical dataset, identifies the candidate rule with the highest model evaluation index as the optimized rule, and updates the risk control rule base based on the optimized rule.

[0120] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0121] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0122] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0127] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for updating risk control rules, characterized in that, include: Acquire historical and incremental datasets related to credit business, and identify rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series data batch used to trigger the incremental learning process. Obtain a first credit application, wherein the first credit application is a credit application rejected by the risk control rule base; randomly select multiple credit applications from the first credit application to obtain a second credit application, and generate a bias-free historical dataset based on the historical dataset and the second credit application; An incremental learning model is trained based on the incremental dataset, the rule to be optimized, and the biased historical dataset. During the training process, the model parameters are updated until the convergence condition is met to obtain convergence parameters. The incremental learning model is used to optimize the rule to be optimized. A candidate rule set is generated based on the convergence parameters. A model evaluation index for the candidate rule set is determined on the historical dataset. The candidate rule with the highest model evaluation index is determined as the optimized rule. The risk control rule base is updated based on the optimized rule.

2. The method according to claim 1, characterized in that, Multiple credit applications are randomly selected from the first credit application to obtain a second credit application. Based on the historical dataset and the second credit application, a bias-corrected historical dataset is generated, including: For each of the first credit applications, generate a random number within a preset range; The random number is compared with a preset exploration rate parameter; If the random number is less than the exploration rate parameter, the decision status of the first credit application is changed from rejection to approval, and the first credit application is identified as the second credit application. In subsequent credit cycles, the second credit application is tracked to obtain its credit performance; Once the credit performance is determined, an exploration sample set is generated based on the sample pairs consisting of the target feature vector corresponding to the second credit application and the credit label corresponding to the credit performance. The exploratory sample set is merged with the historical dataset to obtain the bias-free historical dataset.

3. The method according to claim 2, characterized in that, An incremental learning model is trained based on the incremental dataset, the rule to be optimized, and the bias-reduced historical dataset. During training, the model parameters are updated until the convergence condition is met, including: The rules to be optimized are parsed into a set of optimizable parameters; Based on the parameter set and the feature vector of the input sample, the credit risk prediction result corresponding to the input sample is determined; The predicted loss of the credit risk prediction results for all samples in the incremental dataset is summed to obtain the first loss term; Determine the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base; Based on the probability estimate, each of the exploration samples is assigned an importance weight, wherein the importance weight is negatively correlated with the probability estimate; Based on the importance weights, the weighted prediction loss of the credit risk prediction result for each sample in the exploration sample set is determined, and the weighted prediction loss of all samples in the exploration sample set is summed to obtain the second loss term. Based on the first loss term and the second loss term, determine the counterfactual risk minimization objective function for model training; The gradient descent algorithm is used to iteratively optimize the objective function for minimizing counterfactual risk, thereby updating the model parameters until the convergence condition is met.

4. The method according to claim 3, characterized in that, Determining the probability estimate of each exploratory sample in the exploratory sample set being rejected by the risk control rule base includes: Based on the historical decision logs of the risk control rule base, a probability estimation model is constructed, wherein each sample in the historical decision log includes the target feature vector and the actual decision result corresponding to the target feature vector, and the probability estimation model is used to output the probability estimate of rejection by the risk control rule base according to the input feature vector; Based on the historical decision logs, a likelihood function is constructed, wherein the likelihood function is the product of the conditional probabilities of all samples; for samples that were rejected in the past, the conditional probability is the rejection probability predicted by the probability estimation model; for samples that were allowed in the past, the conditional probability is the difference between 1 and the rejection probability. By maximizing the likelihood function, the parameters of the probability estimation model are obtained through optimization, resulting in a fully trained probability estimation model. Using the fully trained probability estimation model, determine the probability estimate of each exploration sample in the exploration sample set being rejected by the risk control rule base.

5. The method according to claim 3, characterized in that, Based on the probability estimate, each of the exploration samples is assigned an importance weight, including: Based on the probability estimate of each exploration sample being rejected by the risk control rule base, calculate the initial weight value corresponding to the exploration sample; The importance weight of the exploration sample is obtained by dividing the initial weight value corresponding to the exploration sample by the sum of the initial weight values ​​of all exploration samples in the exploration sample set.

6. The method according to claim 3, characterized in that, The gradient descent algorithm is used to iteratively optimize the objective function for minimizing counterfactual risk, updating the model parameters until the convergence condition is met, including: Construct a training data pool that includes the biased historical dataset and the incremental dataset; In each iteration, the following operations are performed: random sampling is performed from the training data pool to obtain a batch sample set; based on the batch sample set, the loss value and gradient of the counterfactual risk minimization objective function are calculated; and the model parameters are updated using an adaptive learning rate method according to the gradient. Repeat the iterative process until the preset iteration termination condition is met.

7. The method according to claim 3, characterized in that, Based on the first loss term and the second loss term, a counterfactual risk minimization objective function for model training is determined, including: Multiple candidate values ​​are selected within a preset numerical range, and for each candidate value, the following operations are performed: based on the candidate value, a counterfactual risk minimization objective function is determined, and the incremental learning model is trained based on the counterfactual risk minimization objective function to obtain the rule parameter solution; the performance of the rule parameter solution is evaluated on a validation set to obtain the model's discriminative ability. Among all the candidate values, the candidate value whose corresponding rule parameter solution exhibits the optimal model discrimination ability on the validation set is determined as the balanced hyperparameter; The objective function for minimizing counterfactual risk is determined based on the product of the first loss term, the second loss term, and the balance hyperparameter.

8. The method according to claim 1, characterized in that, Determine the model evaluation metrics for the candidate rule set on the historical dataset, including: On the historical dataset, batch prediction is performed on each candidate rule in the candidate rule set to obtain prediction results; Based on the predicted results and the actual results, the ratio of the total number of correctly predicted samples to the total number of samples in the whole is calculated to obtain the accuracy rate; Based on the predicted results and the actual results, the recall rate is obtained by calculating the proportion of the number of correctly predicted fulfillment samples to the actual number of all fulfillment samples. Calculate the first cumulative distribution of the performance samples and the second cumulative distribution of the default samples on the prediction results, and calculate the maximum difference between the first cumulative distribution and the second cumulative distribution to obtain the model discrimination. Based on the prediction results, calculate the estimated revenue from correctly releasing fulfilling samples and the estimated loss from incorrectly releasing defaulting samples, and calculate the net value between the estimated revenue and the estimated loss to obtain the expected business revenue. Based on preset weight coefficients, the accuracy, recall, model discrimination, and expected business revenue are weighted and summed to obtain the model evaluation index for each candidate rule.

9. A risk control rule updating device, characterized in that, include: The first acquisition module is used to acquire historical datasets and incremental datasets related to credit business, and to determine the rules to be optimized in the risk control rule base whose performance indicators are less than a preset threshold. The incremental dataset is a time-series data batch used to trigger the incremental learning process. The second acquisition module is used to acquire a first credit application, wherein the first credit application is a credit application rejected by the risk control rule base; randomly select multiple credit applications from the first credit application to obtain a second credit application, and generate a bias-free historical dataset based on the historical dataset and the second credit application; The first processing module is used to train an incremental learning model based on the incremental dataset, the rule to be optimized, and the biased historical dataset, and update the model parameters during the training process until the convergence condition is met to obtain convergence parameters. The incremental learning model is used to optimize the rule to be optimized. The second processing module is used to generate a candidate rule set based on the convergence parameters, determine the model evaluation index of the candidate rule set on the historical dataset, determine the candidate rule with the highest model evaluation index as the optimized rule, and update the risk control rule base based on the optimized rule.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device where the non-volatile storage medium is located to execute the risk control rule update method according to any one of claims 1 to 7.

11. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the method for updating risk control rules as described in any one of claims 1 to 7.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for updating risk control rules as described in any one of claims 1 to 7.

Images