A personal credit risk assessment method based on sample equalization enhancement and an interpretable gradient boosting tree

Abstract

This invention discloses a personal credit risk assessment method based on sample balance enhancement and interpretable gradient boosting trees, belonging to the fields of financial technology and intelligent risk control technology. The method includes: acquiring original personal credit risk sample data; performing missing value imputation, outlier identification and correction, field consistency checks, variable encoding, and continuous feature scaling on the sample data; constructing target labels based on a predetermined observation period; dividing the samples into training, validation, and test sets, wherein the test set maintains the original class distribution; performing sample balance enhancement processing only on a minority of classes in the training set; establishing a gradient boosting tree risk assessment model based on the sample balance enhancement training set; performing parameter optimization and training control on the model; inputting the sample to be assessed into the trained gradient boosting tree risk assessment model to obtain the default probability output; converting the default probability output into at least one of risk level, approval suggestion, credit limit control information, interest rate stratification information, or warning indicator according to a preset threshold or hierarchical mapping rule; performing feature attribution and interpretation analysis on the gradient boosting tree risk assessment model to obtain the contribution results of input features to the overall model output and individual sample output, and jointly outputting the default probability, risk level, and feature contribution description. This invention can improve the ability to identify default minority samples under class imbalance conditions, enhance the ability to model complex nonlinear risk relationships, and improve the interpretability of model output results. It is applicable to scenarios such as pre-loan approval, credit line assessment, pricing stratification, mid-loan early warning, and post-loan re-scoring.

Description

Technical Field

[0001] This invention relates to the fields of financial technology, intelligent risk control, data mining and machine learning applications, and in particular to a computer implementation method for personal credit risk assessment.

[0002] More specifically, the present invention relates to an assessment method for identifying default risk in personal credit business. The method comprehensively utilizes sample balance enhancement technology, gradient boosting tree modeling technology, and model interpretation technology to predict the probability that a borrower will experience serious delinquency or default within a predetermined period, and provides a quantitative explanation of the reasons for the prediction results.

[0003] This invention is applicable to banks, consumer finance companies, microfinance institutions, online lending matchmaking platforms, guarantee institutions, credit scoring service institutions, and risk management middleware systems. It can be deployed in business scenarios such as pre-loan access, credit approval, credit limit assessment, pricing tiering, mid-loan early warning, and post-loan re-scoring.

[0004] With the development of digital finance and inclusive finance, personal lending is showing a trend of continuous expansion in scale, a continuous expansion of the user base, and a continuous diversification of data sources. Traditional risk identification methods that rely on manual review and a limited number of static indicators are no longer sufficient to meet the business needs of high efficiency, low default rates, and strong compliance explanation requirements. Against this backdrop, how to build a highly discriminative and explainable risk model using borrowers' credit history, income level, debt structure, credit limit usage, and delinquency behavior has become a key technical issue in personal lending risk control.

[0005] This invention is particularly applicable to scenarios with the following distinct data characteristics: a large sample size of borrowers, multiple variable dimensions, and strong nonlinear relationships between features; a significantly larger number of normal performance samples than default samples, and a highly unbalanced category distribution; in addition to requiring the model to have high predictive power, the output results are also required to explain which factors lead to an increase or decrease in risk.

[0006] Against this technological backdrop, traditional statistical models have certain advantages in interpretability, but they have limitations in identifying complex nonlinear risk patterns. While complex tree models can improve identification capabilities, they often fail to adequately learn the minority defaulter samples that truly need to be identified if the imbalance in sample class distribution is not addressed. Therefore, this invention focuses on the core issue of "how to improve the risk identification capability of a single master model under class imbalance conditions while simultaneously outputting interpretable results." Technical Background

[0007] Personal credit risk assessment refers to the technical process of quantitatively estimating the probability of a borrower's future delinquency, default, or other credit risk events based on their past credit history, repayment ability, debt characteristics, and related behavioral variables, and then applying the estimation results to credit granting decisions, risk pricing, and post-loan management. Essentially, it is a typical supervised binary classification or probability estimation problem.

[0008] Existing personal credit risk assessment technologies can be mainly divided into two categories. The first category is traditional statistical modeling methods, with logistic regression being a typical example. This method has the advantages of simple models, clear parameter interpretation, low deployment costs, and regulatory friendliness, and has long been one of the mainstream tools for credit scoring modeling in the financial industry. The second category is machine learning methods, typically including random forests, gradient boosting trees, XGBoost, and LightGBM. These methods can automatically learn complex nonlinear relationships and variable interactions, and in scenarios with high dimensionality, large samples, and complex behavioral variables, they usually have better predictive capabilities than traditional linear models.

[0009] However, in personal lending, both traditional statistical models and complex machine learning models generally face the following practical problems: First, the proportion of defaulting samples is usually significantly lower than the proportion of normal samples, exhibiting a clear class imbalance. If the original data is used directly to train the model, the model is prone to bias towards the majority class, resulting in a seemingly high overall accuracy but insufficient identification of truly defaulting customers. Second, common variables in personal lending data, such as debt ratio, credit limit utilization, income, number of overdue payments, loan amount, and family burden, have complex nonlinear relationships, making it difficult for traditional linear models to fully extract the discriminative information from these relationships. Third, while complex models may offer higher predictive power, in financial scenarios, simply providing the probability of default or a risk label is insufficient. Both the business and regulatory sides usually need to clarify the reasons for the risk assessment, such as whether high risk is driven by factors like an excessively high debt ratio, insufficient income, a recent increase in overdue payments, or excessive credit limit utilization.

[0010] 1.1 Comparison of Existing Authorized Patents

[0011] The first existing comparative patent is CN111652710B, published as "A Personal Credit Risk Assessment Method Based on Ensemble Tree Feature Extraction and Logistic Regression". As seen in the published text, this patent first collects personal historical credit data, preprocesses the data by handling missing values, outliers, and standardization, then extracts and transforms features using ensemble tree models such as GBDT, XGBoost, and LightGBM, and finally establishes a Logistic Regression model after feature fusion to complete the personal credit risk assessment. Its core idea is "ensemble tree feature extraction + Logistic regression modeling".

[0012] While this patent addresses the shortcomings of traditional Logistic Regression in uncovering nonlinear relationships, it still has the following deficiencies: First, the main model remains Logistic Regression, and its nonlinear capabilities primarily rely on pre-conversion feature transformations, without directly utilizing the risk ranking capabilities of the tree model itself. Second, the disclosed approach focuses on feature extraction and fusion, without making minority class sample enhancement a necessary step in the main process. Third, the solution lacks a single-sample attribution explanation mechanism based on Shapley values, making it impossible to directly explain why a particular customer is classified as high-risk.

[0013] The second existing comparative patent is CN110414716B, published as "A Method and System for Predicting the Probability of Corporate Credit Default Based on LightGBM". This patent predicts the probability of corporate credit default by collecting corporate credit behavior footprint information, constructing a training dataset, preprocessing and feature engineering, and then building a LightGBM model. This patent has proven that LightGBM is suitable for credit-related risk identification scenarios.

[0014] However, this patent differs from the present invention in the following ways: First, the assessment object is the corporate defaulter, not the individual loan borrower, and the data field system and risk formation logic are significantly different; Second, although the disclosed content uses LightGBM, it does not explicitly incorporate the class imbalance enhancement on the training set side into the main method structure; Third, the focus is on the output of the corporate default probability, and it does not form an explanatory module around the explanation of the individual risk reasons in the personal loan approval process.

[0015] The third prior art patent is CN108550077A, entitled "A Method and System for Personal Credit Risk Assessment Based on Large-Scale Unbalanced Credit Data". This patent focuses on personal credit risk assessment under conditions of large-scale, high-dimensional, and unbalanced credit data, reflecting the importance attached to the problem of class imbalance.

[0016] However, compared with this invention, the patent still has the following differences: First, the solution focuses more on large-scale credit data processing and the construction of weighted or classifiers, without using a gradient boosting tree single-master model as the core, without making the training set sample equalization enhancement the main processing step, and without constructing a unified method structure with feature attribution explanation as the output module; Second, the focus is on the overall nature of the "unbalanced credit data assessment system", rather than an interpretable single-master model prediction method designed for personal credit default risk scenarios; Third, the disclosed technology does not highlight the explanation of the local feature contribution of individual borrowers.

[0017] 1.2 Problems with existing technology

[0018] Based on the above analysis, existing personal credit risk assessment technologies still have at least the following problems.

[0019] First, there is the problem of insufficient learning from minority class samples. Even with a strong classification model, if the number of defaulting customers in the training sample is significantly small, the model's training objective is easily dominated by the majority class samples. This causes the model to learn more about the distribution characteristics of customers with good repayment records, while failing to adequately characterize the boundaries of defaulting customers. For financial institutions, missing truly high-risk customers often has a higher business cost than misidentifying low-risk customers. Therefore, improving the model's ability to learn from minority class samples is one of the core issues affecting the model's application value.

[0020] Secondly, there is the problem of insufficient representation of complex risk relationships. Personal credit risk is not determined by a single variable, but is often the result of the combined effects of multiple variables. If the model cannot learn the higher-order interactions and nonlinear patterns between variables, it will be difficult to form a more accurate risk ranking.

[0021] Third, there is the issue of insufficient interpretability of the output results. While traditional scorecards offer strong interpretability, their expressive power is limited in complex scenarios; and while complex models can improve accuracy, they may result in black-box outputs. If the model cannot clearly answer questions such as "What is the most important factor leading to the increased risk for this customer?", "Does a certain feature increase or decrease risk?", and "What is the relative magnitude of the variable's contribution?", it will affect the acceptability of the model results at the business and regulatory levels.

[0022] Fourth, the problem of incomplete method chain. Many existing solutions often stop at a certain local technical improvement, such as only emphasizing feature engineering, only emphasizing model accuracy, only emphasizing sample weighting, or only emphasizing classifier selection, without building a complete integrated process of "data preprocessing - training set balancing enhancement - single master model training - probability output - risk stratification - interpretation result output".

[0023] Fifth, there is the issue of insufficient engineering implementation. Many theoretical models perform well in experimental environments, but when implemented in real-time scoring systems, they often encounter problems such as inconsistent feature definitions, missing model interpretations, and difficulties in connecting threshold strategies. Therefore, a truly usable credit risk assessment method not only needs to outperform existing solutions in terms of model accuracy, but also needs to be easy to deploy, have interpretable results, and be compatible with rules.

[0024] Based on the above issues, it is evident that in personal credit risk assessment scenarios, simply relying on raw category distribution samples to directly train a classification model can easily lead to the model developing a stronger fit to the majority class samples during the training phase, while failing to adequately learn the minority class samples that truly need to be identified as defaulters. For financial institutions, overall model accuracy is not the only objective. A more valuable technical direction lies in improving the ability to identify truly high-risk customers while maintaining overall ranking capabilities. Especially given the multi-dimensional nature of personal credit data, the complexity of variable relationships, and the low proportion of defaulting samples, it is difficult to develop a risk assessment method that balances predictive power and business usability if the imbalance in category distribution is not addressed and effective modeling of complex nonlinear relationships is lacking.

[0025] Meanwhile, risk models in financial scenarios are not merely offline research tools, but need to be applied to actual business processes such as approval, pricing, credit limit control, and post-loan early warning. Therefore, the model not only needs to output default probabilities, but also needs to be further converted into risk levels, risk scores, or strategy signals that can be understood and invoked by business rules; furthermore, it should be able to explain the main reasons why the model makes a certain risk judgment. In other words, there is an urgent need in this field for a personal credit risk assessment method that uses training set sample balancing enhancement as a core preliminary step, employs a single gradient boosting tree model as the main risk model, and performs feature attribution interpretation on the model output. This would improve default identification capabilities under imbalanced personal credit data conditions and enhance the interpretability and implementability of the output results. Summary of the Invention

[0026] 1. Purpose of the invention

[0027] The purpose of this invention is to provide a personal credit risk assessment method based on sample balance enhancement and interpretable gradient boosting tree, so as to solve the problems of insufficient default identification ability caused by class imbalance, limited nonlinear representation ability of traditional linear models, and lack of explanation of complex model results in the prior art.

[0028] A further objective of this invention is to provide a risk assessment method suitable for engineering deployment in financial institutions, which, while ensuring prediction accuracy, can output risk level, risk probability, and key influencing factors to support various business operations such as automated approval, manual review, differentiated pricing, and post-loan early warning.

[0029] This invention also aims to provide a technical solution that integrates sample processing, model building, and result interpretation, so that there is a clear connection between the model training process, the model evaluation process, and the business decision-making process, thereby improving the deployability, maintainability, and auditability of the method in real business environments.

[0030] 1.1 Technical Solution

[0031] To achieve the above objectives, the present invention adopts the following technical solution.

[0032] A personal credit risk assessment method based on sample balance enhancement and interpretable gradient boosting trees includes the following steps:

[0033] S1. Obtain original personal credit risk sample data, which includes at least the borrower's basic attribute information, repayment history information, income and debt information, credit account usage information, and default label information.

[0034] S2, preprocessing the original personal credit risk sample data, the preprocessing includes missing value imputation, outlier identification and correction, field consistency check, variable encoding and continuous feature scaling adjustment.

[0035] S3. Construct a target label based on the predetermined observation period. The target label is used to characterize whether the sample has experienced serious overdue payment, default, or preset bad credit event during the observation period.

[0036] S4 divides the samples into a training set, a validation set, and a test set, where the test set retains the original class distribution.

[0037] S5 performs sample equalization enhancement processing only on minority class samples in the training set to increase the representation density of default samples in the training set.

[0038] S6. A gradient boosting tree risk assessment model is established based on the training set after sample equalization and enhancement, with the LightGBM model being the preferred choice.

[0039] S7 optimizes the learning rate, tree depth, number of leaf nodes, minimum number of leaf node samples, subsampling ratio, and regularization parameters of the LightGBM model through cross-validation, grid search, random search, or other hyperparameter optimization methods.

[0040] S8 inputs the sample to be evaluated into the trained LightGBM model to obtain the default probability output.

[0041] S9 converts the default probability output into risk level, approval suggestion, quota control information, interest rate tier information or early warning indicator according to preset thresholds or hierarchical mapping rules.

[0042] S10, Perform feature attribution analysis on the LightGBM model to obtain the contribution results of each input feature to the overall output of the model and the output of a single sample.

[0043] S11 outputs the probability of default, risk level, and characteristic contribution description for use by the risk control system or business approval system.

[0044] In S1, the basic attribute information includes at least one of age, gender, marital status, education level, and occupation category; the income and debt information includes at least one of monthly income, income stability, debt ratio, monthly debt repayment amount, loan balance, and credit card balance; the repayment history information includes at least one of overdue number of days (30 to 59 days), overdue number of days (60 to 89 days), and overdue number of days (over 90 days); and the credit account usage information includes at least one of credit limit usage ratio, number of open credit loans, number of real estate loans, account activity, and number of credit inquiries.

[0045] In S2, missing value imputation preferably uses one or more of the following: median imputation, stratified median imputation, mode imputation, or model estimation imputation. For variables that are susceptible to extreme values, such as monthly income and household size, median imputation is preferred.

[0046] In S2, outlier identification and correction includes at least one of the following: pruning continuous variables such as debt ratio, revenue, and credit limit utilization rate according to a predetermined quantile threshold; truncating count variables such as the number of overdue payments according to a business confidence interval; and removing or logically correcting obviously erroneous records.

[0047] Step S5 is a key step in this invention to improve the learning sufficiency of minority class samples in default. Since the number of normally fulfilling samples in personal credit risk samples is usually far greater than the number of default samples, if a risk model is built directly based on the original training set, the training objective is easily dominated by the majority class samples. This can lead to insufficient identification of real defaulting customers even with a high overall accuracy. This invention improves the model's ability to identify minority class samples by performing sample balancing enhancement processing only within the training set, without changing the true distribution of the validation and test sets.

[0048] In S5, the sample balancing enhancement process preferably adopts the SMOTE algorithm, which includes: for each minority class sample in the training set, calculating its nearest neighbors in the minority class sample space; selecting at least one sample from the nearest neighbors and interpolating it with the current target sample in the feature space; generating a new minority class synthetic sample according to a preset ratio and adding it to the training set; repeating the above process until the minority class samples and majority class samples in the training set reach a preset balance ratio.

[0049] In a preferred embodiment, the generation of SMOTE synthetic samples satisfies the condition for any minority class sample. and its neighboring samples Generate new samples Its expression is:

[0050]

[0051] in, It is a random number between 0 and 1.

[0052] In step S6, the gradient boosting tree risk assessment model is the single master risk model of this invention. Unlike schemes that employ multiple models in parallel collaboration or multi-stage fusion, this invention does not focus on dual-model collaboration. Instead, it combines training set sample balancing enhancement with a single gradient boosting tree master model, enabling the model to simultaneously achieve strong nonlinear expressive power and good minority class recognition capabilities within a unified training framework. The LightGBM model is preferred because it can efficiently model complex feature relationships in structured tabular data and adapt to threshold effects and local interactions commonly found in personal credit risk variables.

[0053] In a preferred embodiment, the present invention performs sample balancing augmentation only on the training set, and not on the validation and test sets. This is because the main purpose of the training set is to provide learning samples for the model; augmenting the default minority class samples improves the model's sufficiency in learning the minority class boundary. The main purpose of the validation and test sets is to evaluate the model's generalization ability under real business data distributions. If the validation and test sets are also artificially augmented, their original class distribution will be altered, causing the evaluation results to deviate from the actual ratio of normal to default samples in real credit scenarios, thereby weakening the objectivity of the model performance evaluation.

[0054] Furthermore, if sample equalization augmentation is also performed on the validation and test sets, the data distribution faced by the model during the evaluation phase will no longer represent the sample distribution in the real online environment. In this case, even if the model exhibits high classification accuracy or recall on the augmented validation and test sets, it cannot fully demonstrate its risk identification capability in the real online application flow. Especially in personal lending, the proportion of defaulted samples is usually low, and the true goal of the model after deployment is to identify a small number of high-risk customers against a backdrop of low default rates. Therefore, this invention, by "augmenting only the training set while keeping the original distribution of the validation and test sets unchanged," enables the model to fully learn the characteristics of the minority defaulted samples during the training phase and to undergo data testing closer to the real business environment during the evaluation phase.

[0055] In a preferred implementation, after sample balancing enhancement of the training set, the model's learning boundary for minority class samples is expanded, while the validation and test sets maintain their original distribution. This facilitates a comprehensive evaluation of model performance from multiple dimensions. For example, the AUC value reflects the model's overall ranking ability, the KS value reflects the model's ability to distinguish between positive and negative samples, the recall rate reflects the model's ability to capture real defaulting customers, and the bad debt stratification result reflects the model's risk stratification effect under the actual business volume. These metrics are only truly meaningful for business reference when the validation and test sets maintain their original class distribution.

[0056] Furthermore, the reason this invention limits sample equalization augmentation to the training set is that the augmented samples themselves are artificially synthesized samples generated through interpolation. Their main purpose is to improve the learning of the local space of minority class samples during the training phase, rather than to replace real evaluation samples. Therefore, limiting the augmentation operation to the training phase helps to achieve a balance between "model learning sufficiency" and "model evaluation authenticity." It is precisely because of this approach that this invention can improve the ability of a single gradient boosting tree master model to identify defaulting minority class samples in class imbalance scenarios, while avoiding overestimation of model performance due to distortion of the evaluation set distribution.

[0057] In S6, the LightGBM model learns the complex nonlinear feature relationships in individual credit risk samples by minimizing the objective loss function through histogram-based feature binning and iteratively constructing a decision tree using a leaf-first growth method. Its objective function can be generally expressed as:

[0058]

[0059] In a preferred embodiment, the approximate objective function for the t-th iteration can be expressed as:

[0060]

[0061] In a preferred embodiment, the optimal leaf node weight for a given leaf node can be expressed as:

[0062]

[0063] In S7, the preferred model optimization metrics include one or more of AUC, KS value, recall, precision, and F1 score, with AUC and KS value serving as the main ranking ability metrics, and recall and F1 score serving as default identification ability metrics.

[0064] In S8, the default probability value can be further mapped to a risk score value. Preferably, the default probability can be converted into a standardized score through a logarithmic probability mapping relationship to facilitate integration with existing credit approval systems and scoring card systems. The relationship between the risk score and the default probability can be expressed as:

[0065]

[0066] In S9, risk level classification can be based on default probability or risk score, with multiple tiered thresholds set as follows:

[0067]

[0068] In S10, the SHAP method is preferably used for feature attribution explanation. The contribution value of any feature can generally be expressed as:

[0069]

[0070] In S10, the SHAP interpretation results include at least the following three categories: First, global interpretation results, which are used to output the feature importance ranking across the entire sample range; second, local interpretation results, which are used to output the direction and value of each feature in a single sample that increases or decreases the default probability; and third, visualization interpretation results, which are used to generate readable outputs such as bar charts, waterfall charts, force charts, and dependency graphs.

[0071] Step S10 converts the output of the single master risk model into an explainable explanation of the causes of risk formation. In financial risk control scenarios, the model output must not only be able to "determine" but also "explain" it clearly. By performing feature attribution analysis on the single gradient boosting tree master model, the ranking of important features within the overall model can be obtained, as well as the effect of each feature in a single loan sample on the risk outcome, thus providing a basis for manual review, strategy optimization, and model governance. Therefore, this invention forms a complete technical chain of "training set sample balancing enhancement—single master model training—risk output—single model explanation".

[0072] 1.2 Beneficial Effects

[0073] Compared with the traditional direct modeling scheme of Logistic regression, this invention uses LightGBM to model the nonlinear and interactive relationships between variables, which can more fully explore the complex discrimination patterns between debt ratio, income, number of overdue payments and credit limit utilization, thereby improving the risk identification capability.

[0074] Compared with tree modeling schemes that do not address class imbalance, this invention improves the ability to identify defaulting customers by introducing sample balancing enhancement processing into the training set, allowing minority class sample information to be learned more fully during the training process.

[0075] Compared to solutions that only output risk probabilities without explaining their causes, this invention further outputs feature contribution values ​​through SHAP, which can explain which variables drive the high-risk judgment, making it more suitable for the audit and manual review needs of financial institutions.

[0076] Compared with existing technologies, this invention forms a complete methodological structure, from raw sample preprocessing, label construction, training set enhancement, single-master model training, parameter optimization, hierarchical mapping to interpretation result output. This integrated design makes this invention not only suitable for offline modeling verification, but also easily embedded into existing risk control platforms, approval systems, and model governance systems, possessing strong engineering feasibility and industrial application value. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. The drawings described below are only a part of the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a schematic diagram of the overall process of the method of the present invention, used to illustrate the main steps of original data input, data preprocessing, training set balancing and enhancement, model training, risk output and interpretation generation.

[0079] Figure 2 This is a schematic diagram of the data preprocessing process in this invention, used to illustrate the processes of missing value imputation, outlier handling, variable normalization, and sample partitioning.

[0080] Figure 3 This is a schematic diagram of the SMOTE sample equalization enhancement process in this invention, used to illustrate the generation relationship between minority class samples, nearest neighbor samples and newly synthesized samples.

[0081] Figure 4 This is a schematic diagram of the risk model training process based on LightGBM in this invention, used to illustrate the parameter input, training iteration, verification, and early stopping control process.

[0082] Figure 5 This is a schematic diagram of the risk output and interpretation result generation process in this invention, used to illustrate the formation relationship between default probability, risk level, and SHAP feature contribution description.

[0083] Figure 6 This is a schematic diagram of the application architecture of the present invention deployed in a personal credit approval system, used to illustrate the calling relationship between the data layer, model service layer, interpretation service layer and business decision layer.

[0084] The annotations in the attached figures are explained as follows:

[0085] 100. Overall process of personal credit risk assessment methods;

[0086] 110. Raw sample data acquisition module;

[0087] 120. Data preprocessing module;

[0088] 121. Missing value processing unit;

[0089] 122. Outlier Correction Unit;

[0090] 123. Field consistency verification unit;

[0091] 124. Variable Encoding and Scale Processing Unit;

[0092] 130. Tag building module;

[0093] 140. Sample set partitioning module;

[0094] 150. Training set sample balancing enhancement module;

[0095] 151. Minority class sample identification unit;

[0096] 152. Nearest Neighbor Sample Search Unit;

[0097] 153. Synthetic Sample Generation Unit;

[0098] 154. Enhanced sample output unit;

[0099] 160. Risk assessment model training module;

[0100] 161. Parameter initialization unit;

[0101] 162. Model Iterative Training Unit;

[0102] 163. Cross-validation unit;

[0103] 164. Early Stop Control Unit;

[0104] 170. Risk Prediction Output Module;

[0105] 171. Default probability output unit;

[0106] 172. Risk Score Mapping Unit;

[0107] 173. Risk Level Classification Unit;

[0108] 180. Feature Attribution Explanation Module;

[0109] 181. Global Feature Importance Analysis Unit;

[0110] 182. Local Sample Contribution Analysis Unit;

[0111] 183. Visual explanation result generation unit;

[0112] 190. Business Decision Output Module;

[0113] 191. Approval suggestion output unit;

[0114] 192. Quota control output unit;

[0115] 193. Interest Rate Adjustment Output Unit;

[0116] 194. Early warning information output unit;

[0117] 600. Application architecture of the personal loan approval system;

[0118] 610. Data Access Layer;

[0119] 620. Feature Processing Layer;

[0120] 630. Model Service Layer;

[0121] 640. Explain the service layer;

[0122] 650. Business decision-making level. Detailed Implementation

[0123] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, reasonable improvements and formal changes made based on the technical concept of the present invention should fall within the scope of protection of the present invention. It should be noted that the technical focus of the present invention is not on using multiple models in parallel to complete risk assessment, but on improving the learning ability of a single gradient boosting tree master model by balancing training set samples under the condition of imbalanced personal credit data, and forming a complete method chain by combining feature attribution explanation output. In other words, the gradient boosting tree model used in the present invention is the single master model for the final risk output, and the sample balancing enhancement step is a key preliminary step to support the single master model in effectively learning default minority class samples. Unlike the technical route of dual-model collaboration, model fusion or branch fallback, the present invention does not rely on multiple models to jointly form the final risk output, but improves the risk identification ability by enhancing training samples and optimizing the training process of the single master model. Therefore, the inventive point of the present invention should be understood as "improvement of single master model risk assessment method under imbalanced conditions", rather than improvement of multi-model collaborative system.

[0124] 1. In this embodiment, historical sample data of personal credit is used as the basis for modeling. The sample data may come from public credit default datasets, anonymized historical credit data of commercial banks, historical loan and repayment records of consumer finance platforms, or a combination of the above multiple sources of data.

[0125] Each sample includes at least one or more of the following characteristics: age, monthly income, debt ratio, unsecured credit line usage rate, number of delinquencies in 30 to 59 days, number of delinquencies in 60 to 89 days, serious delinquency related indicators for over 90 days, number of open loans, number of real estate loans, and number of dependents in the household.

[0126] In this embodiment, the target label is defined as whether the borrower has committed serious delinquency or default during the predetermined observation period. Preferably, whether serious delinquency of more than 90 days occurs within the next 24 months can be used as a binary label.

[0127] This embodiment is implemented according to the following steps:

[0128] S1. Obtain raw personal credit sample data and form a basic sample library.

[0129] S2 uses the median to fill in missing variables such as monthly income and number of family members; it uses quantile or business threshold pruning to remove extreme outliers in debt ratio, number of overdue payments, and credit limit utilization; and it performs further standardization on continuous variables when necessary.

[0130] S3, the preprocessed samples are divided into training set, validation set, and test set. Preferably, the subsequent sample balancing enhancement operation is performed only on the training set, while the validation set and test set maintain their original class distribution.

[0131] S4, Perform SMOTE sample augmentation on the training set. Let S be the minority class sample set in the training set. For any target sample x in set S, find its k nearest neighbors in the minority class space; randomly select a nearest neighbor sample x_i from the k nearest neighbors; generate a new sample according to the aforementioned SMOTE formula and add it to the training set.

[0132] S5, a LightGBM risk assessment model is built based on the training set processed by SMOTE.

[0133] S6 uses the validation set to select parameters for the model, with AUC and KS values ​​as the main evaluation metrics, and recall, precision and F1 score are combined to comprehensively judge the model performance.

[0134] S7. After training is complete, input the test set samples into the model to obtain the default probability P.

[0135] S8 calculates a risk score based on the default probability P and maps it to a risk level.

[0136] S9 performs SHAP interpretation on the trained LightGBM model.

[0137] In a preferred embodiment, the input features used in this invention are not limited to the original fields themselves, but also include derived variables, proportional variables, count variables, and window statistical variables constructed based on the original fields. For personal credit risk assessment, although the original fields can directly reflect the borrower's basic attributes, account status, and historical repayment information, in many cases, the truly discriminative risk information is not only reflected in a single original field, but more often in the proportional relationships, relative relationships, and time window changes among multiple fields. Therefore, without departing from the technical concept of this invention, the original fields can be appropriately derived according to the business scenario to form a feature set more suitable for learning by the gradient boosting tree master model.

[0138] In a preferred embodiment, the features can be organized as follows: The first category is basic identity attribute features, used to describe the borrower's stability and basic demographic characteristics, such as age, marital status, education level, and occupation category; the second category is income and debt repayment ability features, used to describe the borrower's cash flow status and debt pressure, such as monthly income, monthly debt repayment amount, debt ratio, and income stability; the third category is credit account usage features, used to describe the borrower's credit line utilization and account activity, such as credit line utilization ratio, number of open credit accounts, number of real estate loans, and number of active accounts; the fourth category is historical repayment features, used to describe the borrower's past repayment performance, such as the number of overdue payments in different overdue periods, the maximum number of overdue days in history, and historical serious overdue records; the fifth category is external behavior and inquiry features, used to describe recent financing needs and potential liquidity pressure, such as the number of credit inquiries, application frequency, and short-term account changes.

[0139] In a preferred embodiment, the derived variables may include at least one of the following: a ratio variable between income and debt, a coverage variable between monthly income and monthly debt repayment, a occupancy variable between loan balance and credit limit, a cumulative variable of the number of overdue payments within different time windows, an aggregate variable of the number of credit inquiries over different periods, and a variable representing changes in account activity. By constructing the above derived variables, the single gradient boosting tree master model can more easily learn the relative and changing relationships between personal credit risk variables, thereby improving its ability to express complex risk patterns.

[0140] In a preferred embodiment, the window statistical variables can be constructed based on different time windows such as the past month, past three months, past six months, and past twelve months. For example, the number of new overdue payments in the past three months, the number of credit inquiries in the past six months, and the proportion of maximum credit limit usage in the past twelve months can be statistically analyzed. These time window characteristics can reflect the recent risk trends and periodic behavioral changes of borrowers, thereby enhancing the model's ability to perceive short-term risk increases.

[0141] It should be noted that the core of this invention does not lie in limiting a specific variable name or a specific derived variable, but rather in incorporating the original fields, derived variables, and window statistical variables into the training process of a single gradient boosting tree master model through a feature organization method suitable for personal credit risk identification. This improves the ability to identify defaulting minority samples while ensuring balanced enhancement of the training set samples. Therefore, any variable organization method and feature construction method adopted that still serve the above-mentioned technical objectives can be considered a reasonable implementation of this invention.

[0142] 2. In this embodiment, to improve the reproducibility of the technical solution of the present invention, examples are provided for the sample size, number of features and key parameters.

[0143] In a preferred embodiment, the unified sample database contains 50,000 to 500,000 personal credit history samples, with positive samples being those that have experienced serious delinquency or default, and negative samples being those that have not experienced serious delinquency or default; preferably, the proportion of positive samples is between 5% and 15%.

[0144] In a preferred embodiment, the number of input features is twenty to eighty, preferably thirty to fifty. These features include both original variables and ratio features, count features, and window aggregation features constructed based on business experience.

[0145] In a preferred embodiment, the LightGBM model parameters include: a learning rate of 0.01 to 0.1; a maximum depth of 5 to 10; a leaf node number of 16 to 128; a minimum leaf node sample number of 20 to 200; a feature sampling ratio of 0.6 to 1; and a sample sampling ratio of 0.6 to 1.

[0146] In a preferred embodiment, the number of SMOTE nearest neighbors is between three and ten, preferably five. The ratio of positive to negative samples in the enhanced training set is preferably adjusted to approximately one to one or one to two.

[0147] In a preferred embodiment, the risk level thresholds are divided into three, four, or five levels. Taking five levels as an example, they can be divided into low risk, low-to-medium risk, medium risk, medium-to-high risk, and high risk. The thresholds are dynamically adjusted based on historical bad debt rates, approval rates, and manual review resources.

[0148] In a preferred embodiment, the above parameter range is not arbitrarily set, but is determined comprehensively based on the structural characteristics, category distribution features, and model generalization requirements of the personal credit risk samples. Specifically, when the learning rate is set too high, although the model can reduce the loss function quickly in the early stages of training, it is also more likely to form unstable splits near the boundaries of minority class samples, leading to increased fluctuations in the validation set performance. When the learning rate is set too low, the number of training epochs increases significantly, training costs rise, and it is not conducive to rapid convergence with a limited sample size. Therefore, this invention preferably controls the learning rate between 0.01 and 0.1 to balance training stability and convergence efficiency.

[0149] In a preferred embodiment, the maximum depth, the number of leaf nodes, and the minimum number of leaf node samples all affect the complexity of the tree model. If the maximum depth is too shallow, the model will struggle to express the complex nonlinear relationships between individual credit variables; if the maximum depth is too deep, it may overfit to locally synthesized samples on the training set after sample equalization and augmentation. Similarly, too few leaf nodes will limit the model's expressive power, while too many leaf nodes will amplify noise in the training set. Therefore, this invention preferably sets the maximum depth to five to ten, the number of leaf nodes to sixteen to one hundred and twenty-eight, and the minimum number of leaf node samples to twenty to two hundred to balance the model's expressive power and generalization ability.

[0150] In a preferred embodiment, the number of SMOTE nearest neighbors also needs to be determined in conjunction with the distribution of minority class samples. If the number of nearest neighbors is too small, the generated new samples will be overly confined to the local neighborhood, making it difficult to effectively expand the boundaries of minority class samples; if the number of nearest neighbors is too large, minority class samples with large differences in space may be forcibly connected, causing distortion of the synthesized samples. Therefore, the present invention preferably controls the number of nearest neighbors between three and ten, and preferably five, to balance the smoothness of sample boundaries and the rationality of synthesized samples.

[0151] In a preferred embodiment, the ratio of positive to negative samples in the enhanced training set should not be expanded without constraint. Simply increasing the minority class samples to a much higher number than the majority class samples can easily cause the model to become overly biased towards the defaulting class during training, thus affecting its true performance on the test set. Therefore, this invention preferably controls the ratio of positive to negative samples in the enhanced training set to around 1:1 or 1:2, ensuring that the model can fully learn the characteristics of defaulting samples without significantly deviating from the actual business distribution.

[0152] In a preferred embodiment, the present invention further maps the default probability output by the model into a risk score and a risk level, aiming to enhance the connection between the model output and actual business rules. For risk control systems, while the default probability directly reflects the relative likelihood of a risk event occurring in a sample, business operations typically prefer to use standardized risk scores or risk levels to formulate approval, credit limit, pricing, and early warning strategies. Therefore, the present invention converts the default probability into a score through a logarithmic probability mapping relationship, and then further converts it into a risk level based on preset thresholds or stratification rules, enabling the model results to be directly used by the approval engine, rule engine, and early warning system.

[0153] In a preferred embodiment, parameters A and B in the scoring mapping can be set according to the institution's scoring card caliber, score range design, and risk preference. For example, when an institution wants a more dispersed distribution of score ranges, parameter B can be adjusted appropriately to improve score sensitivity; when an institution wants the overall scores to fall within a specific score range, parameter A can be adjusted to achieve overall shifting. Thus, the scoring mapping in this invention is not a mechanical conversion of model probabilities, but an important step in converting model outputs into a business-friendly representation.

[0154] In a preferred embodiment, the risk level threshold should be determined by comprehensively considering the distribution of bad debt rates in historical samples, approval rates, manual review capabilities, product profitability, and stage-specific risk appetite. For business stages with lower risk appetite, the proportion of medium-high and high-risk assessments can be appropriately increased to enhance the tightening effect of risk control; for expansion stages or products targeting high-quality customer groups, the boundaries between low-risk and medium-low-risk can be moderately relaxed to improve the coverage of automatic approvals. In other words, the risk level threshold in this invention is not fixed but can be dynamically adjusted according to changes in business strategies.

[0155] In a preferred embodiment, the dynamic adjustment of the threshold can be implemented in at least one of the following ways: First, recalibrate the threshold based on the recent bad debt stratification results; second, adjust the boundary between medium-risk and medium-high-risk based on the changing trends of manual review pass rate and rejection rate; third, set differentiated thresholds based on the risk differences of different products, channels, or customer groups; fourth, appropriately tighten or relax the risk level classification standards at the global level based on changes in the macro environment or risk policies. By adopting the above methods, the risk level output by this invention can maintain high applicability under different business cycles and different operating strategies.

[0156] Furthermore, in a preferred embodiment, the scoring mapping and threshold adjustment do not alter the single-master model structure of the present invention; rather, they perform business-oriented transformations on the results after the model output. In other words, the core of the present invention remains enhancing the default identification capability of the single gradient boosting tree-master model through training set sample balancing, while the scoring mapping and threshold adjustment further ensure that the model results can stably serve different business scenarios. Through this design, the present invention maintains the consistency of the core risk model while improving the adaptability of the output results in approval, pricing, quota control, and early warning scenarios.

[0157] As can be seen from the above parameter setting logic, this invention does not select a certain model parameter or a certain enhancement ratio in isolation, but rather designs an overall system around the stable training and effective generalization of a single gradient boosting tree master model under the condition of class-imbalanced personal credit data.

[0158] 3. In this embodiment, the present invention is used in a pre-loan automatic approval scenario. The system includes a data access module, a feature processing module, a model training module, a real-time scoring module, and an interpretation and output module.

[0159] The data access module extracts necessary fields from the credit reporting system, credit application system, core accounting system, and historical repayment system to form the original feature vector of the applicant to be evaluated. The feature processing module performs missing value imputation, outlier correction, and variable mapping according to the same standards as the training phase, ensuring consistency between offline training and online scoring. The model training module periodically retrains the model using historical samples and performs SMOTE enhancement only on the training set. Upon receiving a sample from a new applicant, the real-time scoring module calls the trained LightGBM model to output the default probability. The explanation output module simultaneously outputs a risk cause description, which includes at least the main features leading to increased risk, the direction of contribution of each feature to the risk value, and the magnitude of that contribution. Based on the default probability, risk level, and explanation results, the business approval engine generates at least one of the following approval actions: automatic approval, transfer to manual review, rejection, credit limit reduction, or price increase.

[0160] In a preferred embodiment, the feature attribution explanation results output by this invention are not only used for model analysis but can also directly serve the manual review process. In personal loan approval scenarios, when a sample is determined by the model to be of medium risk, medium-high risk, or high risk, the business system, in addition to outputting the default probability and risk level, can also simultaneously output the main features driving the risk outcome and their contribution direction. For example, for a sample awaiting approval, if the model output shows that a high debt ratio, an increase in the number of overdue payments in the recent period, and an excessively high credit limit utilization rate are the main factors driving the increased default probability, then manual reviewers can quickly determine whether the customer's risk stems from short-term debt pressure, historical deterioration in repayment, or excessive credit occupation. In this way, manual review no longer relies solely on the single risk score given by the model but can combine key risk driving factors to make a more targeted judgment on the customer.

[0161] Furthermore, in a preferred embodiment, the interpretation results can be sorted according to the absolute value of their contribution, outputting only the top few most significant positive risk factors and negative mitigation factors. For example, the top three features that increase the probability of default and the top one or two features that decrease the probability of default can be preferentially output. The advantage of this approach is that it avoids the interpretation results being too information-intensive for human reading, while enabling business personnel to quickly grasp the core basis for the model's judgment. For the auditing system, the above interpretation results can also be displayed in the form of text descriptions, visual bar charts, or structured fields, thus forming result descriptions that can be directly used for manual review, audit review, and complaint handling.

[0162] In a preferred embodiment, if a customer is deemed high-risk, but manual review reveals that the primary risk stems from a temporary factor, such as a short-term increase in credit inquiries, unverified income data, or anomalies in a variable due to data collection delays, the manual reviewer can further verify the relevant fields based on the interpretation results. If the verification reveals that the input data needs correction, the features can be recalculated and the model re-scored. Therefore, the interpretation results of this invention not only statically explain the model's reasons but also serve as an important auxiliary basis for manual review, data verification, and risk control debriefing.

[0163] Furthermore, in a preferred embodiment, the SHAP interpretation results can also serve to optimize risk control strategies. When business personnel discover that certain features consistently act as major positive risk factors in a large number of high-risk samples, they can adjust admission rules, tighten credit limits, or supplement pre-verification conditions accordingly. When certain features consistently act as risk mitigation factors, they can be appropriately referenced in credit limit allocation or pricing strategies. Through the above methods, this invention achieves a connection between "risk prediction results—risk formation causes—manual review and strategy adjustment," thereby enhancing the business value of the model interpretation output.

[0164] 4. In this embodiment, the present invention is used in an online real-time approval scenario. After receiving a customer's application request, the system completes feature generation, model invocation, risk stratification, interpretation generation, and result return in the real-time scoring process.

[0165] The online inference process includes:

[0166] S1, Receive an application request, the application request including at least the applicant's identity information, the application amount information, and the application channel information;

[0167] S2 calls the data access layer to obtain characteristic data related to the applicant from the credit reporting system, accounting system, behavior system and historical record system;

[0168] S3 calls the feature processing layer to perform missing data completion, anomaly correction, field mapping, and variable generation on the original data;

[0169] S4, call the LightGBM model service to obtain the default probability value and risk score;

[0170] S5, call the risk stratification module to map the default probability value and risk score to risk level and business action;

[0171] S6, calls the interpretation service layer to output the global importance summary and the contribution results of key features of a single sample;

[0172] S7 returns the risk level, business action, and explanation results to the approval system.

[0173] In a preferred embodiment, the total online inference time is controlled to be less than one second; for high-concurrency scenarios, the response time can be further reduced through model service caching, feature result caching, or asynchronous interpretation and generation mechanisms.

[0174] 5. In this embodiment, the present invention is not limited to pre-loan approval, but can also be used for post-loan customer risk warning.

[0175] The system periodically collects data on existing customers' recent repayment behavior, changes in credit limit usage, debt changes, and new overdue information, forming a rolling update sample. This updated sample is then input into a pre-trained LightGBM model to obtain a new default probability output. When a customer's default probability significantly increases compared to the previous period, and the increase exceeds a preset threshold, the system classifies them as an increased-risk customer. Simultaneously, the system uses SHAP (Shape-Based Analysis) to interpret the results and identify the primary causes of the increased risk, such as a surge in credit limit usage, a decrease in income, an increase in recent overdue payments, or an increase in new loan burdens. These causes are then pushed to loan management personnel to support collection prioritization and the development of proactive intervention strategies.

[0176] In a preferred embodiment, the change in risk can be expressed as:

[0177]

[0178] when When the value exceeds the preset warning threshold, the system triggers a warning.

[0179] In a preferred embodiment, the present invention further performs hierarchical mapping on the default probability or risk score output by the model to form a risk level structure suitable for business calls. The risk levels can be divided into three, four, or five levels according to business needs. Taking a five-level division as an example, the samples can be divided into low risk, low-to-medium risk, medium risk, medium-to-high risk, and high risk.

[0180] When a sample is classified as low-risk, it indicates a low probability of default and stable main risk characteristics, and it can proceed to the automatic approval process or the standard credit limit allocation process. When a sample is classified as low-to-medium risk, it indicates that its overall risk is controllable but there are individual factors of concern, and it can proceed to the simplified manual review process or the prudent credit limit allocation process. When a sample is classified as medium-risk, it indicates that its probability of default has increased significantly, and it usually requires further judgment based on the application amount, product type, and channel characteristics, and it can proceed to the regular manual review process. When a sample is classified as medium-to-high risk, it indicates that its risk characteristics are relatively obvious, and it is preferable to implement a strategy of reducing the credit limit, increasing the price, or restricting access. When a sample is classified as high-risk, it indicates that its likelihood of default has increased significantly, and it is preferable to proceed to the process of rejection, key observation, or key post-loan warning.

[0181] In a preferred embodiment, the risk level stratification thresholds are not fixed but can be dynamically adjusted based on historical bad debt rates, approval rates, manual review capabilities, product profitability, and risk appetite. For example, for high-value products, the criteria for determining medium-to-high and high risk can be appropriately increased; for high-frequency, small-amount products, the proportion of automatic approval can be moderately increased, while strengthening post-loan warnings for customers with medium to high risk; for business phases with lower risk appetite, the threshold settings can be tightened overall, resulting in a corresponding increase in the proportion of medium-to-high and high-risk customers.

[0182] In a preferred embodiment, risk stratification is used not only for pre-loan approval but also for post-loan early warning and credit limit management. For example, in a post-loan scenario, if an existing customer's current credit score is still at the low-to-medium risk level, but their risk level continues to rise, the system can trigger an early warning. In a credit limit management scenario, differentiated credit limit adjustment strategies can be implemented for customers based on their risk level. In a pricing scenario, interest rate ranges can also be mapped accordingly based on risk level. Therefore, the risk level output by this invention is not merely a display label but can be directly transformed into a decision-making basis for different business actions.

[0183] As can be seen from the above risk stratification example, this invention further transforms the default probability output by the single gradient boosting tree master model into a risk level structure that is easy for business to understand and rule to invoke, thereby enhancing the connection between the model output and business processes such as approval, early warning, quota and pricing.

[0184] 6. In this embodiment, to verify the improvement effect of the method of the present invention compared with the traditional linear benchmark model, a Logistic regression model can be trained simultaneously as a control benchmark model. The control model uses the same original samples, the same label definitions, and the same training set, validation set, and test set partitioning method as the present invention, but uses Logistic regression instead of LightGBM at the main model level.

[0185] The comparative training process includes: performing the same preprocessing and training / testing division process on the same batch of samples; training the Logistic regression model with the preprocessed samples; calculating the AUC, KS value, recall, precision, and F1 value of the Logistic regression model and the method of this invention on the test set; and comparing the differences between the two models in terms of ranking ability, default identification ability, and output interpretation ability.

[0186] In the implementation using a standalone Logistic regression model, the system has advantages in standardized score output and parameter interpretation, but its ability to identify complex nonlinear risk patterns is relatively insufficient.

[0187] In the implementation of the single-master model LightGBM scheme of this invention, the system has advantages in complex risk pattern recognition and minority class sample capture, while the SHAP interpretation mechanism preserves the transparency of the results.

[0188] As can be seen from the above comparison, the present invention does not merely replace the model type, but forms a complete method chain between sample augmentation, single-master model training and interpretation output. Therefore, it is superior to traditional linear benchmark schemes in terms of recognition capability and method integrity.

[0189] 7. In this embodiment, to improve the stability of the invention during long-term operation, a model update and monitoring mechanism can be further set up. Since the personal credit market environment, customer characteristics, channel structure, and risk preferences may all change over time, if the model is not updated for a long period, problems such as feature distribution drift, threshold failure, and decreased recognition ability may occur.

[0190] The model update and monitoring process can be performed as follows:

[0191] S1, statistically analyze the distribution characteristics, scoring results, and actual performance results of online scoring samples according to a predetermined period;

[0192] S2 calculates the distribution stability index of core features, the recent AUC of the model, KS value, stratified bad debt rate, and the index of changes in the interpretation results;

[0193] S3: When any indicator exceeds the preset threshold or falls below the preset standard, the model update evaluation process is triggered.

[0194] S4, reconstruct the recent sample set, and perform missing data handling, anomaly correction, label generation, and training set sample balancing enhancement;

[0195] S5, train the new version of the LightGBM model and compare its performance with the old model on the same validation and test sets;

[0196] S6: Once the new model meets the deployment criteria, release the new version of the model and retain the old version of the model for rollback.

[0197] S7 will continue to monitor the input distribution, output distribution, and business hit rate of the new model during the observation period after its launch.

[0198] Through the above mechanism, this invention can work continuously and stably as a long-term single-master model risk assessment method.

[0199] In a preferred embodiment, the method of the present invention can also perform a pre-deployment verification process before the model is officially launched. The pre-deployment verification process includes at least the following: First, using an independent test set or out-of-time samples to verify the final effect of the trained single gradient boosting tree master model, to confirm that its ranking ability, default identification ability, and stratification effect meet preset standards; Second, performing consistency verification on the default probability, risk score, and risk level output by the model to ensure that there are no obvious logical conflicts in the scoring mapping and threshold division; Third, sampling and verifying the feature attribution explanation results to confirm that the explanation direction of key risk factors does not deviate significantly from common business sense; Fourth, comparing the model results with the existing baseline model or historical rule strategies to confirm that the method of the present invention has clear application value in real business scenarios.

[0200] In a preferred embodiment, once the pre-deployment verification results meet the preset requirements, the method of the present invention can be deployed to a real business environment. Initially, a limited-traffic verification method can be used, that is, the method of the present invention is first called in a portion of the application traffic or a portion of the sample batches, and its output results, approval success rate, and readability of the interpretation results are continuously observed. The advantage of this method is that it allows verification of the stability and applicability of the method of the present invention in a real environment without significantly affecting the operation of the entire business.

[0201] In a preferred embodiment, if after deployment, significant changes are found in the model input distribution, a marked decrease in risk stratification effectiveness, a systematic deviation between the interpretation results and business experience, or if changes in the external environment render the threshold configuration inapplicable, a model rollback or threshold callback mechanism can be triggered. The rollback mechanism may include at least one of the following: First, rollback to the previously validated single gradient boosting tree master model; second, while keeping the model unchanged, temporarily adjust the risk stratification threshold; third, before the model has completed retraining, temporarily increase the proportion of manual review to reduce business risks.

[0202] Furthermore, in a preferred embodiment, the rollback after model deployment does not alter the core methodological structure of this invention. Whether it's pre-deployment verification, limited traffic observation, threshold adjustment, or model version rollback, these are essentially application control measures of the invention's method in a real business environment. Their purpose is to ensure that the risk assessment results formed through training set sample balancing enhancement and single gradient boosting of the tree master model can stably serve the business scenario. Therefore, the aforementioned deployment verification and rollback mechanisms should be understood as further supplements to the engineering feasibility of the invention's method, without changing its essential nature as a methodological invention of "sample balancing enhancement + single gradient boosting of the tree master model + single model feature attribution explanation."

[0203] Through the above-mentioned pre-launch verification and post-launch rollback design, the method of the present invention can not only have better risk identification capabilities in the offline training and testing phases, but also have higher security, stability and controllability in real business operations, thereby further enhancing the engineering application value of the present invention.

[0204] 8. In this embodiment, the relationships between the modules in the system are as follows:

[0205] The data access layer and the feature processing layer are connected through a data interface for transmitting original application data, credit data, accounting data, and behavioral data;

[0206] The feature processing layer and the model service layer are connected through a feature vector interface, which is used to provide input features to the LightGBM model;

[0207] The model service layer and the explanation service layer are connected through a result analysis interface to generate feature contribution explanation results;

[0208] The model service layer is connected to the business decision layer through a decision interface, which is used to output risk levels and business actions;

[0209] The business decision-making layer and the monitoring and management module are connected through a log feedback interface to record scoring results, interpretation results, decision results, and model version information.

[0210] The call chain of each layer in the system is as follows:

[0211] Data access layer

[0212] →Feature Processing Layer

[0213] →Model Service Layer

[0214] →Explanation of service layer

[0215] →Business Decision-Making Level

[0216] →Monitoring and Management Module.

[0217] Through the above module relationship design, the present invention forms a complete method call chain from data input to business decision-making and then to governance feedback.

[0218] 9. In this embodiment, the overall execution logic of the method of the present invention is further explained. Unlike solutions that focus only on a single technical step, the present invention emphasizes the sequential relationship between each step and their synergistic effect in the personal credit risk assessment scenario. Specifically, the data preprocessing step is used to ensure that the original samples meet the requirements for subsequent training; the label construction step is used to transform the risk assessment target into a classification label that can be supervised for learning; the training set sample balancing enhancement step is used to alleviate the problem of insufficient learning of default minority class samples during the training stage; the single gradient boosting tree master model training step is used to learn the complex nonlinear relationship between personal credit variables; the risk probability output and risk stratification step is used to convert the model results into risk signals that can directly serve business decisions; and the feature attribution interpretation step is used to further explain the main reasons for the formation of risk judgment. The above steps are not independent and arbitrarily replaceable discrete steps, but rather a unified methodological structure formed around the technical goal of "improving the risk identification capability of the single master model and outputting interpretable results under the condition of unbalanced personal credit data".

[0219] Furthermore, the technical contribution of this invention does not lie in simply adopting an existing tree model, nor in merely introducing a sample augmentation algorithm, but rather in organically combining training set sample equalization augmentation, single gradient boosting tree master model training, risk output mapping, and single-model feature attribution interpretation to form an integrated method suitable for personal credit risk assessment scenarios. Specifically, without training set sample equalization augmentation, the single master model is prone to insufficient learning of default minority class samples; without using a gradient boosting tree master model, complex nonlinear relationships and local risk patterns are difficult to fully express; without feature attribution interpretation, the model results are difficult to meet the transparency requirements of business and regulation. Because the above steps collectively constitute a unified method chain in this invention, it possesses more complete technical effects and stronger engineering applicability compared to existing solutions that only improve local technical aspects.

[0220] Furthermore, although this invention can be deployed in various business scenarios such as pre-loan approval, credit limit assessment, mid-loan early warning, and post-loan re-scoring, its core remains a method improvement centered on a single master risk model, rather than focusing on multi-model collaboration, result fusion, or system-level branch governance. For those skilled in the art, without departing from the technical concept of this invention, adjustments can be made to the input fields, parameter ranges, threshold settings, and interpretation methods according to different business scenarios. However, as long as the basic structure of "training set sample balancing enhancement + single gradient boosting tree master model + single model feature attribution interpretation" is still followed, it should be considered a reasonable implementation of the technical solution of this invention.

[0221] Those skilled in the art should understand that this invention is not limited to the embodiments described above. Equivalent substitutions, combinations, adjustments, and conventional improvements made to the sample balancing enhancement algorithm, tree model parameters, explanation output format, risk stratification threshold, data access source, and deployment method, as long as they do not depart from the technical ideas of this invention—enhancing minority class learning ability through training set sample balancing enhancement, improving complex risk relationships in the tree master model through a single gradient, and improving model transparency through feature attribution explanation output—should all be considered to fall within the protection scope of this invention.

Claims

1. A personal credit risk assessment method based on sample balance enhancement and interpretable gradient boosting trees, characterized in that, Includes the following steps: S1, Obtain original personal credit risk sample data, which includes at least the borrower's basic attribute information, repayment history information, income and debt information, credit account usage information, and default label information; S2, preprocessing the original personal credit risk sample data, the preprocessing includes missing value imputation, outlier identification and correction, field consistency check, variable encoding and continuous feature scaling adjustment; S3, construct a target label based on the predetermined observation period. The target label is used to characterize whether the sample has experienced serious overdue payment, default, or preset credit misconduct events during the observation period. S4, divide the samples into training set, validation set and test set, where the test set retains the original class distribution; S5 performs sample equalization enhancement processing only on minority class samples in the training set to increase the representation density of default samples in the training set. S6. A gradient boosting tree risk assessment model is established based on the training set after sample equalization enhancement. S7, Perform parameter optimization and training control on the gradient boosting tree risk assessment model; S8: Input the sample to be evaluated into the trained gradient boosting tree risk assessment model to obtain the default probability output; S9. Based on a preset threshold or hierarchical mapping rule, the default probability output is converted into at least one of risk level, approval suggestion, quota control information, interest rate tier information or early warning indicator. S10, Perform feature attribution analysis on the gradient boosting tree risk assessment model to obtain the contribution results of input features to the overall output of the model and the output of individual samples, and jointly output the default probability, risk level and feature contribution description.

2. The personal credit risk assessment method according to 1, characterized in that, In step S1, the basic attribute information includes at least one of age, gender, marital status, education level, and occupation category; the income and debt information includes at least one of monthly income, income stability, debt ratio, monthly debt repayment amount, loan balance, and credit card balance; the repayment history information includes at least one of overdue number of days (30 to 59 days), overdue number of days (60 to 89 days), and overdue number of days (over 90 days); and the credit account usage information includes at least one of credit limit usage ratio, number of open credit loans, number of real estate loans, account activity, and number of credit inquiries.

3. The personal credit risk assessment method according to 1, characterized in that, In step S2, missing value imputation is performed using one or more of the following methods: median imputation, stratified median imputation, mode imputation, or model estimation imputation. Outlier identification and correction include: pruning continuous variables such as debt ratio, income, and credit limit utilization rate according to predetermined quantile thresholds; truncating count variables such as overdue number of times according to business confidence intervals; and removing or logically correcting obviously erroneous records.

4. The personal credit risk assessment method according to 1, characterized in that, The sample equalization enhancement process in step S5 adopts the synthetic minority oversampling method, which generates new synthetic minority samples by interpolating between minority samples and their nearest neighbor samples in the training set, and performs the sample equalization enhancement process only on the training set, while the validation set and test set maintain the original class distribution unchanged.

5. The personal credit risk assessment method according to 4, characterized in that, In the sample equalization enhancement process, the number of nearest neighbors for minority class samples is set to an integer value between three and ten.

6. The personal credit risk assessment method according to 1, characterized in that, The gradient boosting tree risk assessment model in step S6 is the LightGBM model, which is used to learn the nonlinear relationships and interactions between input features in the personal credit risk sample.

7. The personal credit risk assessment method according to 1, characterized in that, The parameter optimization in step S7 is achieved through cross-validation, grid search, random search or other hyperparameter optimization methods. The optimized parameters include at least one or more of the following: learning rate, tree depth, number of leaf nodes, minimum number of leaf node samples, sample sampling ratio, feature sampling ratio and regularization parameter.

8. The personal credit risk assessment method according to 1, characterized in that, In step S8, the default probability output is further mapped to a risk score. In step S9, the sample is divided into multiple risk levels based on the risk score or default probability output, and mapped to at least one business action among automatic approval, manual review, quota control, interest rate adjustment or risk warning.

9. The personal credit risk assessment method according to 1, characterized in that, The feature attribution explanation analysis in step S10 adopts the feature contribution analysis method based on cooperative game theory to obtain the global feature importance results and the local sample feature contribution results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the personal credit risk assessment method described in any one of 1 to 9.

Images