A multi-model integrated small sample credit score offset adaptive correction system

Abstract

The application discloses a multi-model integrated small sample credit score offset adaptive correction system and relates to the technical field of data processing. The system comprises a divergence quantization routing decision module, which acquires target small samples and source domain samples, calculates the JS divergence of each feature dimension, and generates a decoupling mask vector to separate the original feature space into anchor features and drift features; a controlled alignment dynamic demarcation module, which performs an antagonistic feature alignment on the drift features, extracts a deformation vector, and dynamically demarcates the boundary of a decision fuzzy area based on the module length; a partial derivative projection directional correction module, which intercepts samples according to the relative position of the fluctuation boundary and the fuzzy area boundary predicted by a multi-base model, and implements nonlinear directional correction on the intercepted samples based on the dot product of the deformation vector and the partial derivative gradient; and a noise reduction verification closed-loop evaluation module, which generates a final credit score and outputs a verification index of the correction effect. The application realizes adaptive offset correction of credit scores in a small sample scenario and improves the scoring accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a multi-model integrated small-sample credit score offset adaptive correction system. Background Technology

[0002] In the field of credit risk assessment, financial institutions often face a cold start dilemma when expanding into new geographic markets, acquiring new customer groups, or facing new economic cycles: the target sample size available for modeling or scoring is extremely limited, while historically accumulated source domain samples, although sufficient and fully labeled, exhibit significant distributional shifts from the target samples due to changes in customer composition, behavioral patterns, or the macroeconomic environment. This dual challenge of small sample size and distributional shift leads to systematic scoring bias when directly applying historical models, increasing credit risk.

[0003] Existing technologies have the following main shortcomings: First, traditional domain adaptation methods mostly focus on the alignment of global features, ignoring the heterogeneity of feature shifts in credit assessment scenarios—some features, such as basic demographics, have relatively stable distributions, while others, such as recent behavior, have drastic temporal shifts, and uniform alignment can lead to distortion of stable features; Second, although adversarial domain alignment can reduce distribution differences, the alignment process itself introduces sample-level deformation. The impact of this deformation on the orientation of subsequent multi-model ensemble predictions is not quantified and compensated, resulting in false consensus or variance tearing in the model output near the decision boundary; Third, most methods lack a self-verification mechanism for the correction effect, making it difficult for the business side to judge the reliability of the scoring results and the actual effectiveness of the correction operation;

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-model integrated small-sample credit score offset adaptive correction system to address the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A multi-model integrated small sample credit scoring offset adaptive correction system includes the following modules: a divergence quantification routing decision module is used to obtain the target small sample set and the source domain sample set, calculate the JS divergence of the target small sample set and the source domain sample set on each feature dimension, generate a decoupling mask vector based on the JS divergence, and separate the original feature space into anchor features and drift features based on the decoupling mask vector;

[0008] The controlled alignment dynamic delimitation module is used to perform feature alignment on drift features to obtain aligned drift features, restore and concatenate the anchor features and aligned drift features to obtain merged aligned features, input the merged aligned features into the base model set to obtain the initial prediction probability set; extract the vector difference between the drift features before alignment and the aligned drift features to obtain the deformation vector, and calculate the lower boundary and upper boundary of the decision fuzzy region based on the L2 norm of the deformation vector;

[0009] The partial derivative projection orientation correction module is used to calculate the mean and standard deviation of the initial prediction probability set to construct the prediction distribution fluctuation boundary, determine the relative positional relationship between the prediction distribution fluctuation boundary and the decision ambiguity zone boundary to obtain the interception state; calculate the partial derivative gradient vector of each base model output probability with respect to the drift feature dimension for the intercepted sample, perform a dot product operation on the partial derivative gradient vector and the deformation vector to obtain the orientation deviation magnitude, and perform nonlinear inverse correction based on the orientation deviation magnitude to obtain the corrected prediction probability set;

[0010] The noise reduction verification closed-loop evaluation module is used to aggregate the set of corrected prediction probabilities to obtain the integrated default probability, and then map the integrated default probability to the scoring interval to obtain the final offset adaptive corrected credit score.

[0011] In a preferred embodiment, the JS divergence between the target small sample set and the source domain sample set on each feature dimension is calculated. Specifically, for each feature dimension, the kernel density of the target small sample set and the source domain sample set is estimated for the values ​​of that feature dimension to obtain the target domain probability density function and the source domain probability density function.

[0012] The estimated probability density function is truncated by a lower bound and then renormalized.

[0013] Calculate the arithmetic mean of the KL divergences in two directions between the two probability density functions, and use it as the JS divergence for that feature dimension.

[0014] In a preferred embodiment, the decoupling mask vector is generated based on JS divergence, specifically as follows:

[0015] Preset feature dimension divergence determination threshold;

[0016] If the JS divergence of a feature dimension is less than the feature dimension divergence judgment threshold, then the mask component corresponding to that feature dimension takes a value of 0 and is judged as an anchored feature.

[0017] If the JS divergence of a feature dimension is greater than or equal to the feature dimension divergence determination threshold, then the mask component corresponding to that feature dimension is set to 1, and it is determined to be a drift feature.

[0018] In a preferred embodiment, the divergence-based quantitative routing decision module is also used to perform system startup determination:

[0019] The overall mean offset is obtained by calculating the arithmetic mean of the JS divergence of all feature dimensions;

[0020] When the overall offset mean is greater than or equal to the routing start threshold, the controlled alignment dynamic delimitation module, the partial derivative projection orientation correction module, and the noise reduction verification closed-loop evaluation module are activated.

[0021] When the overall mean offset is less than the routing start threshold, the target small sample set is directly input into the base model set to obtain the score.

[0022] In a preferred embodiment, feature alignment of the drift features is performed using an adversarial learning framework, specifically:

[0023] Deploy a feature alignment network and a domain discrimination network. The feature alignment network is used to perform transformations on the target domain-shifted features, and the domain discrimination network is used to determine whether the input features come from the source domain or the target domain.

[0024] During training, the feature alignment network and the domain discriminator network are optimized alternately. The optimization goal of the feature alignment network is to reduce the domain classification accuracy of the domain discriminator network to below the domain discriminator convergence threshold.

[0025] In a preferred embodiment, the lower boundary and upper boundary of the decision fuzzy region are calculated based on the L2 norm of the deformation vector, specifically as follows:

[0026] Centered on the standard decision boundary point, the lower boundary and upper boundary of the decision fuzzy region are obtained by symmetrically expanding the sum of the weighted quantities after the logarithmic saturation mapping of the basic safety radius and the deformation vector.

[0027] Among them, the logarithmic saturation mapping uses the natural logarithm function to ensure that the expansion increases monotonically with the increase of deformation amplitude, but the growth rate gradually saturates.

[0028] In a preferred embodiment, determining the relative positional relationship between the predicted distribution fluctuation boundary and the decision ambiguity region boundary to obtain the interception state specifically includes:

[0029] When the mean of the initial predicted probability set falls between the lower boundary and the upper boundary of the decision fuzzy region, it is determined to be a mutually exclusive state B.

[0030] When the mean of the initial predicted probability set is outside the decision fuzzy region, but the boundary of the predicted distribution fluctuation overlaps with the boundary of the decision fuzzy region, it is determined to be a mutually exclusive state C.

[0031] When the predicted distribution fluctuation boundary and the decision fuzzy area boundary do not intersect, it is determined to be a mutually exclusive state A;

[0032] For samples in mutually exclusive states B and C, a partial derivative orientation correction process is triggered; for samples in mutually exclusive state A, the initial prediction probability set is directly passed through.

[0033] In a preferred embodiment, nonlinear inverse correction is performed based on the orientation deviation magnitude, specifically as follows:

[0034] When the orientation bias is positive, it indicates that there is an artificially high component in the output of the base model, and a negative subtraction correction is performed.

[0035] When the orientation bias amplitude is negative, it indicates that there is a false low component in the base model output, and positive compensation correction is performed.

[0036] The nonlinear inverse correction uses a correction formula with an exponential saturation mechanism and restricts the correction result to a legal probability range through numerical clipping operations.

[0037] In a preferred embodiment, the noise reduction verification closed-loop evaluation module is also used to perform closed-loop verification of the correction effect:

[0038] The difference between the variance of the initial predicted probability set and the variance of the corrected predicted probability set is calculated to obtain the noise reduction gain index.

[0039] When the noise reduction gain index is greater than the noise reduction gain significance threshold, the correction effect is considered significant.

[0040] In a preferred embodiment, the noise reduction verification closed-loop evaluation module is further configured to:

[0041] Business confidence labels are generated based on noise reduction gain indicators and interception status indicators. Business confidence labels include high confidence consensus is stable, high confidence has been cleared, and medium confidence is pending verification.

[0042] Generate batch-level statistical reports, which include the percentage of intercepted samples within the batch, the average expansion rate of the dynamic blurry region boundary, and the average noise reduction gain within the batch.

[0043] The technical effects and advantages of the multi-model integrated small sample credit score offset adaptive correction system of the present invention are as follows: fine-grained quantification of feature distribution offset and structural decoupling are achieved through the divergent quantification routing decision module. The dimensional JS divergence calculation and adaptive threshold determination separate the original feature space into anchored features and drift features, preserving the original information of stable dimensions while providing targeted input for accurate alignment, fundamentally avoiding the pollution of stable features by global alignment. Through a controlled alignment dynamic delimitation module, an innovative transformation of alignment costs into early warning signals is achieved. Based on adversarial feature alignment, sample-level deformation vectors are extracted, and the decision fuzzy region boundary is dynamically delimited based on their magnitude using nonlinear saturation mapping, transforming the invisible alignment cost into a visualized risk boundary, providing a preliminary basis for subsequent interception and correction. Through a partial derivative projection orientation correction module, accurate interception and orientation correction of samples within the fuzzy region are achieved. Based on the overlap determination of the mean-variance fluctuation boundary predicted by the multi-base model and the dynamic fuzzy region boundary, high-risk samples are accurately identified. Furthermore, the dot product of the partial derivative gradient and the deformation vector is calculated to obtain the orientation deviation amplitude, and a nonlinear inverse correction with an exponential saturation mechanism is executed to eliminate the probability of falsely high or low values, significantly improving the accuracy of integrated scoring in offset scenarios. A closed-loop evaluation module for noise reduction verification is constructed to establish a closed-loop verification system for the correction effect. Calculate the variance denoising gain before and after correction, generate multi-level business confidence labels, and provide batch-level statistical reports to provide quantitative support for model monitoring and business decision-making, achieving seamless integration of technical effects and business implementation. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a multi-model integrated small-sample credit score offset adaptive correction system according to the present invention.

[0045] Figure 2 This is a schematic diagram of the divergent quantitative routing decision module of a multi-model integrated small-sample credit score offset adaptive correction system of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example

[0048] Please see Figure 1 As shown, this invention discloses a multi-model integrated small-sample credit score offset adaptive correction system, comprising the following modules:

[0049] Please see Figure 2 As shown, the divergence quantification routing decision module is used to obtain the target small sample set and the source domain sample set, calculate the JS divergence of the target small sample set and the source domain sample set on each feature dimension, generate a decoupling mask vector based on the JS divergence, and separate the original feature space into anchored features and drift features based on the decoupling mask vector.

[0050] In practical applications of credit assessment, assessment agencies often face the dilemma of extremely limited sample size available for credit assessment when expanding into new geographic markets or new customer channels. The discrete quantitative routing decision module of this invention serves as the data entry point and decision guidance end of the entire system, specifically:

[0051] First, the target small sample set to be evaluated and the historical source domain sample set for reference are obtained from the business system; wherein, the target small sample set is defined as... This indicates that it comes from new channels or new markets. The multidimensional feature vector of each evaluation object to be scored; the target small sample refers to the credit sample to be scored that is in a new business scenario, a new customer group, or a new time period, with insufficient sample size and a data distribution that differs significantly from historical training data. Include There are 3 credit assessment feature dimensions; where i represents the i-th target small sample, i=1,2,... D represents the total number of credit assessment feature dimensions for a single credit sample; the source domain sample set is defined as... This indicates the accumulation of historical credit during the credit modeling phase. Features of a sample with known credit performance, where The source domain refers to a set of historical credit samples that are sufficient in number, have a stable distribution, have been labeled, and can be used for model training and distribution reference. The core task of this module is to quantify the degree of difference in feature distribution between two sample sets, and based on the spatial distribution characteristics of this difference, to separate the original feature space into a stable anchored feature part and a drifted feature part that has shifted, thereby providing structured input for subsequent accurate alignment and orientation correction.

[0052] It should be noted that the original feature space of this invention refers to a high-dimensional feature space composed of the user's original feature dimensions after standardization and normalization in credit risk assessment business. The original features include, but are not limited to: user basic attribute features, performance history features, asset and liability features, behavioral time series features, equipment and environment features, etc. Each dimension in the original feature space corresponds to an independent business feature, and all dimensions together constitute a complete feature expression system for credit risk assessment.

[0053] Furthermore, perform dimension-wise feature divergence calculation for each feature dimension in the original feature space. and Extract the target small sample set respectively With source domain sample set The marginal distribution along this feature dimension; to quantify the degree of difference between two marginal distributions, this invention uses the Kullback-Leibler divergence, i.e., the symmetric form of KL divergence, JS divergence, as a metric.

[0054] Specifically, regarding the first Each feature dimension, firstly for the target small sample set The kernel density is estimated based on the value of this feature dimension, and the bandwidth is adaptively selected using the silver window rule. ,in The standard deviation of the sample is 1. Interquartile range, Given the sample size, the probability density function is obtained. Furthermore, to avoid logarithmic singularity in JS divergence calculation caused by the probability density function being zero, the system applies a lower bound truncation to the estimated probability density function, removing all values ​​less than the lower bound truncation value. Replace the density value with , The initial value is set to Then, the truncated density function is renormalized to ensure its integral is 1; similarly, for the source domain sample set... The feature dimension values ​​are used to perform kernel density estimation using the same Gaussian kernel function and bandwidth selection method to obtain the probability density function. Based on this, the first JS divergence across feature dimensions Defined as the arithmetic mean of the KL divergences in two directions, its calculation expression is:

[0055] In the formula, Represents probability distribution Relative to probability distribution The KL divergence is used to quantify the degree of difference between two probability distributions;

[0056] Traverse all Given each feature dimension, we obtain the dimension divergence vector. After completing the dimension-by-dimensional divergence calculation, decoupling mask generation is performed; the system presets the feature dimension divergence judgment threshold. This threshold is used to distinguish whether a significant shift has occurred in the feature dimension, and its initial value is set to 0.15; the threshold is determined based on the feature dimension divergence. Binarization is performed on each feature dimension to generate a decoupling mask vector. The first decoupling mask The generation rule for the components corresponding to each feature dimension is as follows: if the divergence value of the feature dimension is less than the feature dimension divergence determination threshold, then the component is set to 0, indicating that the feature dimension is determined to be an anchored feature; if the divergence value of the feature dimension is greater than or equal to the feature dimension divergence determination threshold, then the component is set to 1, indicating that the feature dimension is determined to be a drift feature; the formal expression of this determination rule is:

[0057] In the formula, This is an indicator function. It returns 0 when the condition in parentheses is true, indicating that the feature dimension is an anchored feature; otherwise, it returns 1, indicating that the feature dimension is a drifting feature.

[0058] Based on decoupling mask vector Perform a dimension-level separation operation on the original feature space: for any target small sample Anchoring features It is composed of the components corresponding to the dimensions with a mask value of 0 in the original feature vector. Its dimension number is equal to the number of components with a value of 0 in the mask vector, denoted as . Drift characteristics It is composed of the components corresponding to the dimensions with a mask value of 1 in the original feature vector. Its dimension number is equal to the number of components with a value of 1 in the mask vector, denoted as . ,in After alignment is completed in subsequent modules, the system will restore and concatenate the identity-protected anchor features and the aligned drift features according to the dimensional order in the original feature space to form a D-dimensional merged and aligned feature vector. ;

[0059] Based on decoupling mask vector Similarly, for the source domain sample set Perform the same dimensionality separation operation to extract the drift feature set of source domain samples. , which serves as the source domain reference input for the domain discriminant network during adversarial training; source domain drift features are directly input into the domain discriminant network in their original form without undergoing feature alignment network transformation;

[0060] After completing the feature separation operation, this module performs a system startup check; the system calculates the arithmetic mean of the divergence of all feature dimensions to obtain the overall offset mean. Its calculation expression is:

[0061] System preset routing startup threshold Used to determine the target small sample set With source domain sample set Whether the overall offset between them reaches a level that requires initiating a correction link, the initial value of this threshold is set to 0.10;

[0062] When the overall deviation is the mean Greater than or equal to the routing startup threshold When the system determines that the distribution shift is significant, it automatically activates subsequent modules; when the overall shift mean... Less than the routing start threshold At that time, the system determined that the distribution shift was slight and directly set the target small sample set. The input base model set is used to obtain the score; this routing decision mechanism ensures the reasonable allocation of system computing resources and avoids performing unnecessary correction operations on samples with stable distributions.

[0063] It should be noted that the base model set of this invention refers to a collection of heterogeneous prediction models with different structures and training mechanisms. Base models include, but are not limited to, logistic regression models, gradient boosting tree models, random forest models, support vector machine models, and shallow neural network models. The base model set is used to perform parallel predictions on the same input feature, forming a multi-model output distribution, providing a basis for subsequent integrated scoring and bias correction.

[0064] The controlled alignment dynamic delimitation module receives anchor features and drift features, performs feature alignment on the drift features to obtain aligned drift features, restores and concatenates the anchor features and aligned drift features to obtain merged aligned features, inputs the merged aligned features into the base model set to obtain the initial prediction probability set; extracts the vector difference between the drift features before alignment and the aligned drift features to obtain the deformation vector, and calculates the lower boundary and upper boundary of the decision fuzzy region based on the 2 norm of the deformation vector;

[0065] It receives the target domain anchoring feature set, target domain drift feature set, source domain drift feature set, and decoupling mask vector transmitted from the upstream divergence quantization routing decision module, performs controlled partial feature alignment operation, and simultaneously extracts deformation vectors and pre-defines the decision ambiguity region boundary for each sample during the alignment process; specifically:

[0066] This module first performs partial adversarial alignment operations; targeting drift features. System deployment feature alignment network The network employs a fully connected structure, comprising an input layer, hidden layers, and an output layer. The system pre-sets the hidden layer dimension coefficient k, which is the ratio of the number of hidden layer neurons to the dimension of the drift feature. The initial value is set to 2, therefore the number of hidden layer neurons is... , The dimension of the drift feature is represented; the optimization objective of the feature alignment network is to minimize the distribution distance between the transformed drift feature and the corresponding feature in the source domain in the high-dimensional feature mapping space, thereby achieving feature distribution alignment between domains; for any target small sample Its drift characteristics are processed by a feature alignment network. Alignment drift features are obtained after transformation. Meanwhile, anchoring features Perform identity protection, do not participate in any transformation operations, and ensure that the stable feature dimensions shared by the source domain and the target domain are not disturbed;

[0067] Subsequently, based on the decoupling mask vector The anchored features and the aligned drift features are restored and concatenated according to the dimensional order of the original feature space to obtain the merged aligned features. Specifically, for the first feature in the original feature space... Each feature dimension, if Then the feature dimension takes the anchored feature. The corresponding component value, if Then this feature dimension takes the alignment drift feature. The corresponding component values; this restoration and splicing operation ensures the merging and alignment of features. The arrangement of the feature dimensions is completely consistent with the arrangement of the original feature space dimensions during the training of the base model, thus ensuring that the base model can directly receive the feature for prediction.

[0068] Then, perform initial predictions using a multi-base model; the system is pre-deployed. An ensemble model pool consisting of heterogeneous base models is denoted as . Each base model employs a differentiated algorithm architecture, including but not limited to gradient boosting trees, deep neural networks, and support vector machines; for each target small sample, merge alignment features are used. Enter them one by one The first base model is used to obtain the initial predicted confidence probability value output by each base model; the second base model is used to obtain the predicted confidence probability value. The base model for the first The initial prediction for each target small sample is denoted as . ,in Let represent the probability that the i-th target small sample is predicted as a failure; by pooling the outputs of all base models, we obtain the initial set of predicted probabilities for that sample. ;

[0069] It should be noted that the feature alignment network The training employs an adversarial learning framework, with the system simultaneously deploying a domain discriminant network. The network receives aligned drift features and determines whether they originate from the source or target domain. The domain discrimination network employs a fully connected structure, containing one hidden layer and one sigmoid output layer. During training, the feature alignment network and the domain discrimination network are optimized alternately. The loss function of the domain discrimination network is the binary cross-entropy loss.

[0070] Where u is the global index subscript of the training samples of the domain discriminant network and ;

[0071] During training, the input to the domain discrimination network is a mixture of the source domain drift feature set and the target domain drift feature set after alignment. Calculate the classification loss for the full mixed sample, driving the domain discriminant network to learn the distribution differences of features between the two domains; the input features of the domain discriminant network are denoted as... When the sample comes from the source domain, The original values ​​of the source domain drift features, and the domain labels. When the sample comes from the target domain, The target domain drift features after alignment network transformation, domain labels ;

[0072] The loss function of the feature alignment network is the inverse of the domain discrimination loss, i.e. The alignment network is optimized to prevent the domain discriminant network from distinguishing between samples from the two domains. Training uses the Adam optimizer with an initial learning rate of 0.001. In each training epoch, the domain discriminant network is updated one step at a time, followed by the alignment network, for a total of 100 epochs. The training convergence condition is that the domain classification accuracy of the domain discriminant network on the mixed samples of the full source domain drift features and target domain alignment drift features in the current training batch drops to the domain discriminant convergence threshold. the following, The initial value is set to 0.55, which is close to the level of random guessing. At this point, it is determined that the distribution of the drift features between the two domains is sufficiently aligned. If the training epochs reach the preset maximum training epoch limit and the domain classification accuracy still does not drop to the limit, the value is determined to be 0.55. The alignment network parameters corresponding to the maximum training epoch are then used as the final parameter output, at which point it is determined that the two domain distributions are fully aligned.

[0073] The core technical features of this module lie in the extraction of partial feature deformation vectors and the pre-delineation of the boundary of the decision fuzzy region; for each target small sample Extract the high-dimensional vector difference of its drift features before and after alignment; where This refers to the original drift characteristics obtained after separating the sample. The original drift feature is transformed by a feature alignment network to obtain the aligned drift feature; the difference between the two is defined as the deformation vector. Its calculation expression is: Deformation vector This indicates the direction of deviation of the drift feature before alignment relative to the drift feature after alignment, that is, the residual direction and amplitude of the original drift feature deviating from the alignment target position;

[0074] The deformation vector carries information about the direction and magnitude of the alignment operation applied to the sample's drift feature; the L2 norm of the deformation vector reflects the severity of the alignment operation and is defined as the deformation amplitude. Its calculation is ,in Represents the set of drift feature dimension indices. Depends on the deformation vector The component in the d-th drift feature dimension;

[0075] Based on deformation amplitude This module pre-defines the decision fuzzy region boundary for each target small sample; for the first... For a small sample of targets, the dynamic expansion of its decision fuzzy region is calculated, defined as the sum of the weighted amount of the basic safety radius and the deformation amplitude after a logarithmic saturation mapping. This logarithmic mapping ensures that the expansion monotonically increases with the deformation amplitude, but the growth rate gradually saturates. The lower and upper boundaries of the decision fuzzy region are defined by the standard decision boundary point. The calculation expression is as follows: (The expression is incomplete and requires further context.)

[0076] Lower boundary of decision fuzzy region ,

[0077] Upper boundary of decision fuzzy region ;

[0078] In the formula, Using the natural logarithm function, the nonlinear saturation characteristics of the expansion of the lower and upper boundaries of the decision fuzzy region are realized. This is the standard decision-making threshold, with an initial value of 0.5, representing the equal probability judgment benchmark for the probability of default in the credit scoring system; To preset the basic safety radius, its initial value is set to 0.05. This parameter defines the width of the basic tolerance interval on both sides of the standard decision boundary point under zero deformation conditions. Preset deformation sensitivity coefficient Its initial value is set to 0.08. This parameter adjusts the intensity of the influence of deformation amplitude on the width of the blurred area.

[0079] For example, suppose that after alignment, the L2 norm of the deformation vector of a small target sample is 1.5; substituting this into the above formula, the deformation contribution term is calculated as follows: Then the lower boundary of the fuzzy region of this small sample is The upper boundary is This boundary is different from the basic boundary. Significant widening means that subsequent modules need to handle the falling into The predicted probabilities for the intervals will be subject to more rigorous review and correction;

[0080] The data structure passed through by this module to the downstream partial derivative projection orientation correction module includes: the initial predicted probability set for each sample. Deformation vector and the pre-defined boundaries of the decision-making ambiguity zone. This data structure provides downstream modules with all the input elements needed to perform precise interception and targeted correction.

[0081] The partial derivative projection orientation correction module receives the initial predicted probability set, deformation vector, and the lower and upper boundaries of the decision ambiguity region. It calculates the mean and standard deviation of the initial predicted probability set to construct the predicted distribution fluctuation boundary, determines the relative positional relationship between the predicted distribution fluctuation boundary and the decision ambiguity region boundary to obtain the interception state, calculates the partial derivative gradient vector of each base model output probability with respect to the drift feature dimension for the intercepted samples, performs a dot product operation between the partial derivative gradient vector and the deformation vector to obtain the orientation deviation magnitude, and performs nonlinear inverse correction based on the orientation deviation magnitude to obtain the corrected predicted probability set. Specifically:

[0082] This module first performs multi-dimensional fuzzy region interception and judgment; and then performs an initial prediction probability set for each target small sample. Calculate its statistical characteristics; the mean of the initial predicted probability set is defined as , representing the average consensus of all base models in predicting the probability of failure for this small sample; the standard deviation of the initial predicted probability set is defined as This indicates the degree of dispersion among the predictions of each base model; based on the mean and standard deviation of the initial prediction probability set, a fluctuation boundary of the prediction distribution is constructed, including the upper fluctuation boundary of the prediction distribution. and the fluctuation boundary under the predicted distribution ;

[0083] Based on the obtained predicted distribution fluctuation boundary, each target small sample is divided into three mutually exclusive states according to the relative positional relationship between its predicted distribution and the boundary of the decision fuzzy region. The determination rule is as follows:

[0084] First, determine whether the condition for mutually exclusive state B is met, i.e., the mean of the initial predicted probability set. Does it meet the requirements? If the conditions are met, it is determined to be a mutually exclusive state B, i.e., a mean trap state, indicating that the average consensus of the multi-base model has fallen into the decision ambiguity region; if the mutually exclusive state B is not met, it is further determined whether the predicted distribution fluctuation boundary and the decision ambiguity region boundary overlap; the specific determination condition is: the predicted distribution fluctuation boundary Less than the upper boundary of the decision fuzzy region And the predicted distribution fluctuates at the boundary. Greater than the lower boundary of the decision fuzzy region If both conditions are met simultaneously, the system is determined to be in mutually exclusive state C, i.e., variance tearing state, indicating that although the mean is outside the fuzzy region, the fluctuation range of the prediction divergence between models has invaded the fuzzy region; if the above two conditions are not met simultaneously, the system is determined to be in mutually exclusive state A, i.e., absolutely safe state, indicating that the fluctuation range of the prediction distribution has no intersection with the decision fuzzy region, the prediction consensus of all base models is robust and not affected by deformation; the system directly transmits the initial prediction probability set to the target small sample in mutually exclusive state A without performing correction operations; and for the samples in mutually exclusive states B and C, it forcibly triggers the subsequent partial derivative orientation correction process;

[0085] For the intercepted mutual exclusion states B and C target small samples, this module performs partial derivative gradient vector calculation and orientation deviation magnitude quantization; for the first... Each base model is used to calculate its output probability. Regarding merge alignment features The partial derivative gradient vector of the drift feature dimension component Specifically, the system first calculates Regarding the full merge alignment feature all gradient in each dimension Then, based on the decoupling mask vector Extract the components corresponding to the drift dimension to form The partial derivative gradient vector of dimension The gradient vector and the deformation vector The dimensions are the same, all of them are dimension;

[0086] It should be noted that for base models whose output probabilities are differentiable with respect to input features, such as deep feedforward network models and logistic regression models, the system directly calculates the probability through backpropagation. Regarding merge alignment features full-dimensional gradient Then, based on the decoupling mask vector Extract the components corresponding to the drift dimension to obtain The partial derivative gradient vector of dimension ;

[0087] For a basis model where the output probability is not differentiable with respect to the input features, the system uses the finite difference method to approximate the partial derivative gradient vector; specifically, for the drift feature... The system constructs two perturbation feature vectors in one dimension. and Both are based on the merge alignment feature. Based on this, small perturbations are added and reduced only at the original feature space locations corresponding to this drift dimension. , The initial value is set to 0.01; the two perturbation feature vectors are input into the base model respectively, and the difference between the two outputs is divided by twice the perturbation amount as the approximate partial derivative of that dimension: Perform the above calculations sequentially on all drift feature dimensions to obtain the approximate partial derivative gradient vector of the basis model, which is then used as the partial derivative gradient vector of the non-differentiable basis model.

[0088] This will then affect the partial derivative gradient vector. With deformation vector Perform a vector dot product operation to obtain the orientation deviation magnitude. The orientation deviation magnitude describes the directional push effect of the alignment operation on the output probability along the sensitive direction of the base model; the expression for calculating the orientation deviation magnitude is:

[0089] In the formula, Represents the partial derivative gradient vector With deformation vector The angle between them;

[0090] When the directional deviation amplitude When it is positive, it indicates that the deformation vector In the partial derivative gradient vector The projection in the direction is positive, meaning that the original drift feature is located on the side of the drift feature along the direction of increasing probability after alignment. This means that although the alignment operation pulls the feature from its original position to the source domain distribution, the direction of this pull is exactly the direction of decreasing probability. Incomplete convergence during the alignment process may cause the feature to not be pulled away from the original high probability region, so that the base model output still has a probability component higher than the true value. Therefore, negative subtraction needs to be performed to compensate for this residual bias.

[0091] When the directional deviation amplitude When the value is negative, it indicates that the projection of the deformation vector onto the direction of the partial derivative gradient vector is negative. That is, the original drift feature is located on the side of the probability decreasing direction. The pulling direction of the alignment operation is exactly the direction of the probability increasing, which may cause the base model output to be lower than the true value of the probability component. Therefore, positive compensation needs to be performed.

[0092] When the directional deviation amplitude When it approaches zero, it indicates that the deformation vector and the partial derivative gradient vector are approximately orthogonal, and the alignment operation has almost no directional effect on the output probability of the base model;

[0093] Based on orientation deviation amplitude Perform nonlinear reverse orientation correction; the system presets the correction strength coefficient. This is used to adjust the overall strength of the correction amplitude, and its initial value is set to 0.12. To prevent over-correction from causing the probability value to exceed the limit, this invention adopts a correction formula with an exponential saturation mechanism. The expression for calculating the corrected probability is:

[0094] In the formula, This is a numerical clipping operator that restricts the correction result to a valid probability range. , This is a sign function that returns the sign of the input. The term constitutes a saturation mechanism. When the absolute value of the orientation deviation amplitude is small, the correction amount is approximately linear with the orientation deviation amplitude. As the absolute value of the orientation deviation amplitude continues to increase, the correction amount gradually approaches the correction strength coefficient. The upper limit;

[0095] An example illustrating the correction process: Suppose the initial predicted probability of the second base model for a small sample of an intercepted target is... The calculated orientation deviation of the base model is as follows: A positive value indicates an artificially inflated value; substituting this into the correction formula, the correction term is calculated as follows: The corrected probability is This correction operation pulls down the spurious probability that the deformation vector has been pushed up, eliminating the bias introduced by the alignment operation.

[0096] This module performs a full analysis of each intercepted target small sample. Each base model prediction is sequentially subjected to the above correction process, and the resulting set of corrected prediction probabilities is aggregated. And transmit the corrected prediction probability set to downstream modules. The interception status flags are as follows: 0 for mutual exclusion state A, 1 for mutual exclusion state B, and 2 for mutual exclusion state C.

[0097] Furthermore, for a small sample of unintercepted targets in mutually exclusive state A, the corrected prediction probability set is equivalent to the initial prediction probability set.

[0098] The noise reduction verification closed-loop evaluation module is used to aggregate the corrected predicted probability set to obtain the integrated default probability, and then map the integrated default probability to a scoring interval to obtain the final offset adaptive corrected credit score; specifically:

[0099] This module first performs the final credit score generation; for each target small sample The integrated failure probability of the sample is obtained by performing equal-weighted average aggregation on its corrected prediction probability set. Its calculation expression is: ;

[0100] This leads to the conversion of the probability of default into a business-wide credit scoring system, using a linear mapping to divide the probability interval. Mapped to rating range ,in This indicates the lower limit of the score, with an initial value set at 300; This represents the upper limit of the score, initially set at 850; the final offset adaptive correction credit score is calculated using the following expression:

[0101] ;

[0102] The above formula demonstrates the negative correlation between the probability of default and credit score: the lower the probability of default, the higher the credit score;

[0103] Furthermore, closed-loop verification of the correction effect is performed, and the variance of the initial predicted probability set is calculated. variance of the corrected prediction probability set Their calculation expressions are as follows: , In the formula, the mean of the corrected prediction probability set is the integrated failure probability calculated above. ;

[0104] System extracts noise reduction gain index This index reflects the effect of the correction operation on compressing the dispersion of the base model predictions; its extraction formula is: Preset noise reduction gain significance threshold This is used to determine whether the compression effect of the correction operation on the predicted dispersion of the base model reaches a statistically significant level; its initial value is set to 0.005. When the noise reduction gain index... Greater than the noise reduction gain significance threshold At that time, the correction effect was deemed significant;

[0105] Based on the noise reduction gain index and the interception status indicator, a service confidence label is generated; the label generation rules are as follows:

[0106] When the interception status is 0, meaning the target small sample is in mutually exclusive state A, the high-confidence consensus of the generated label is stable, indicating that the multi-model prediction of the target small sample itself has a high degree of consistency and no correction intervention is required.

[0107] When the interception status is marked as 1 or 2, meaning the target small sample is in mutually exclusive state B or mutually exclusive state C, and the noise reduction gain index... Greater than the noise reduction gain significance threshold When the generated label has a high confidence level, it means that the target small sample has successfully eliminated false divergences in the decision ambiguity area after targeted correction.

[0108] When the interception status is marked as 1 or 2, indicating that the target small sample is in the noise reduction gain index Less than or equal to the noise reduction gain significance threshold If the confidence level in the generated labels is pending verification, it indicates that the correction effect has not met expectations, and manual verification is recommended.

[0109] This module ultimately outputs the evaluation result data structure for each target small sample, containing the following fields: final offset adaptively corrected credit score. Integrated probability of dishonesty Interception status indicator, noise reduction gain index This module also generates batch-level statistical reports, including the following summary indicators: and business confidence labels;

[0110] The percentage of intercepted samples within a batch is the ratio of the total number of samples in state B and state C within the batch to the total number of target small samples within the batch.

[0111] The average expansion rate of the dynamic blur region boundary is defined as the arithmetic mean of the ratios of the half-width of the dynamic blur region to the basic safety radius of all target small samples within a batch, expressed as a percentage; where, the... The half-width of the dynamic blur region for a small sample of a target is defined as follows: The basic safety radius is Therefore, the expansion rate of a single sample is... The average batch expansion rate is ;

[0112] The average noise reduction gain within a batch is the sum of the noise reduction gains of all target small samples within the batch, divided by the total number of target small samples to obtain the average value.

[0113] This statistical report provides a macroscopic visualization of the correction effect of this invention for the business side, making it easier for business personnel to assess the severity of small sample bias and the effectiveness of the system's response;

[0114] In summary, this invention achieves structured separation of anchored and drift features through feature divergence calculation and decoupling mask generation in the divergence quantification routing decision module; it transforms alignment costs into early warning signals through deformation vector extraction and dynamic boundary calculation in the controlled alignment dynamic delimitation module; it accurately eliminates false consensus in ambiguous areas through mean-variance two-dimensional interception and dot product projection correction in the partial derivative projection orientation correction module; and it achieves quantitative evaluation and business implementation of correction effects through closed-loop verification and label generation in the noise reduction verification closed-loop evaluation module.

[0115] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0116] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0117] Those skilled in the art will recognize that the modules and algorithm modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0118] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0119] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0120] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-model ensemble small-sample credit scoring bias adaptive correction system, characterized in that, The module includes the following modules: the divergence quantification routing decision module is used to obtain the target small sample set and the source domain sample set, calculate the JS divergence of the target small sample set and the source domain sample set on each feature dimension, generate a decoupling mask vector based on the JS divergence, and separate the original feature space into anchored features and drift features based on the decoupling mask vector; The controlled alignment dynamic delimitation module is used to perform feature alignment on drift features to obtain aligned drift features, restore and concatenate the anchor features and aligned drift features to obtain merged aligned features, input the merged aligned features into the base model set to obtain the initial prediction probability set; extract the vector difference between the drift features before alignment and the aligned drift features to obtain the deformation vector, and calculate the lower boundary and upper boundary of the decision fuzzy region based on the L2 norm of the deformation vector; The partial derivative projection orientation correction module is used to calculate the mean and standard deviation of the initial prediction probability set to construct the prediction distribution fluctuation boundary, determine the relative positional relationship between the prediction distribution fluctuation boundary and the decision ambiguity zone boundary to obtain the interception state; calculate the partial derivative gradient vector of each base model output probability with respect to the drift feature dimension for the intercepted sample, perform a dot product operation on the partial derivative gradient vector and the deformation vector to obtain the orientation deviation magnitude, and perform nonlinear inverse correction based on the orientation deviation magnitude to obtain the corrected prediction probability set; The noise reduction verification closed-loop evaluation module is used to aggregate the set of corrected prediction probabilities to obtain the integrated default probability, and then map the integrated default probability to the scoring interval to obtain the final offset adaptive corrected credit score.

2. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 1, characterized in that, The JS divergence between the target small sample set and the source domain sample set is calculated on each feature dimension. Specifically, for each feature dimension, the kernel density of the target small sample set and the source domain sample set is estimated for the values ​​of that feature dimension to obtain the probability density function of the target domain and the probability density function of the source domain. The estimated probability density function is truncated by a lower bound and then renormalized. Calculate the arithmetic mean of the KL divergences in two directions between the two probability density functions, and use it as the JS divergence for that feature dimension.

3. The multi-model integrated small-sample credit scoring offset adaptive correction system according to claim 2, characterized in that, The decoupling mask vector is generated based on JS divergence, specifically as follows: Preset feature dimension divergence determination threshold; If the JS divergence of a feature dimension is less than the feature dimension divergence judgment threshold, then the mask component corresponding to that feature dimension takes a value of 0 and is judged as an anchored feature. If the JS divergence of a feature dimension is greater than or equal to the feature dimension divergence determination threshold, then the mask component corresponding to that feature dimension is set to 1, and it is determined to be a drift feature.

4. The multi-model integrated small-sample credit scoring bias adaptive correction system according to claim 3, characterized in that, The divergence quantization routing decision module is also used to perform system startup determination: The overall mean offset is obtained by calculating the arithmetic mean of the JS divergence of all feature dimensions; When the overall offset mean is greater than or equal to the routing start threshold, the controlled alignment dynamic delimitation module, the partial derivative projection orientation correction module, and the noise reduction verification closed-loop evaluation module are activated. When the overall mean offset is less than the routing start threshold, the target small sample set is directly input into the base model set to obtain the score.

5. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 1, characterized in that, An adversarial learning framework is used to perform feature alignment on drift features, specifically: Deploy a feature alignment network and a domain discrimination network. The feature alignment network is used to perform transformations on the target domain-shifted features, and the domain discrimination network is used to determine whether the input features come from the source domain or the target domain. During training, the feature alignment network and the domain discriminator network are optimized alternately. The optimization goal of the feature alignment network is to reduce the domain classification accuracy of the domain discriminator network to below the domain discriminator convergence threshold.

6. The multi-model ensemble small-sample credit scoring offset adaptive correction system according to claim 5, characterized in that, The lower and upper boundaries of the decision fuzzy region are calculated based on the L2 norm of the deformation vector, specifically as follows: Centered on the standard decision boundary point, the lower boundary and upper boundary of the decision fuzzy region are obtained by symmetrically expanding the sum of the weighted quantities after the logarithmic saturation mapping of the basic safety radius and the deformation vector. Among them, the logarithmic saturation mapping uses the natural logarithm function to ensure that the expansion increases monotonically with the increase of deformation amplitude, but the growth rate gradually saturates.

7. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 1, characterized in that, Determining the relative positional relationship between the predicted distribution fluctuation boundary and the decision ambiguity region boundary yields the interception status, specifically including: When the mean of the initial predicted probability set falls between the lower boundary and the upper boundary of the decision fuzzy region, it is determined to be a mutually exclusive state B. When the mean of the initial predicted probability set is outside the decision fuzzy region, but the boundary of the predicted distribution fluctuation overlaps with the boundary of the decision fuzzy region, it is determined to be a mutually exclusive state C. When the predicted distribution fluctuation boundary and the decision fuzzy area boundary do not intersect, it is determined to be a mutually exclusive state A; For samples in mutually exclusive states B and C, a partial derivative orientation correction process is triggered; for samples in mutually exclusive state A, the initial prediction probability set is directly passed through.

8. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 7, characterized in that, Nonlinear inverse correction is performed based on the orientation deviation amplitude, specifically as follows: When the orientation bias is positive, it indicates that there is an artificially high component in the output of the base model, and a negative subtraction correction is performed. When the orientation bias amplitude is negative, it indicates that there is a false low component in the base model output, and positive compensation correction is performed. The nonlinear inverse correction uses a correction formula with an exponential saturation mechanism and restricts the correction result to a legal probability range through numerical clipping operations.

9. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 1, characterized in that, The noise reduction verification closed-loop evaluation module is also used to perform closed-loop verification of the correction effect: The difference between the variance of the initial predicted probability set and the variance of the corrected predicted probability set is calculated to obtain the noise reduction gain index. When the noise reduction gain index is greater than the noise reduction gain significance threshold, the correction effect is considered significant.

10. The multi-model ensemble small-sample credit scoring bias adaptive correction system according to claim 9, characterized in that, The noise reduction verification closed-loop evaluation module is also used for: Business confidence labels are generated based on noise reduction gain indicators and interception status indicators. Business confidence labels include high confidence consensus is stable, high confidence has been cleared, and medium confidence is pending verification. Generate batch-level statistical reports, which include the percentage of intercepted samples within the batch, the average expansion rate of the dynamic blurry region boundary, and the average noise reduction gain within the batch.

Images