Credit risk prediction method and device and electronic equipment

Abstract

The invention provides a credit risk prediction method and device and electronic equipment, and relates to the technical field of credit risk prediction. The method comprises the steps of obtaining to-be-predicted target credit data; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; wherein the credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer comprises a plurality of feature extraction sub-models; each feature extraction sub-model is used for performing feature extraction on the target credit data to obtain a feature extraction result; and the fusion layer is used for fusing the feature extraction results of all the feature extraction sub-models to obtain a credit risk prediction result. The method is used for solving the problems that a traditional logistic regression score card method is lack of nonlinear modeling ability, depends on artificial experience and is poor in high-dimensional data adaptation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of credit risk prediction, in particular to a credit risk prediction method, a credit risk prediction device, an electronic device, a machine readable storage medium and a computer program product. BACKGROUND

[0002] Risk prediction in the field of credit risk control is of great significance to financial institutions. In existing methods, credit risk prediction is achieved only by using a traditional logistic regression score card method. The traditional logistic regression score card method takes "manual feature engineering + linear modeling" as the core, and is suitable for traditional credit scenarios such as credit cards and conventional consumer loans. However, the traditional logistic regression score card method has the problem of lacking non-linear modeling capability, which cannot adapt to complex patterns such as "ladder-like correlation between income-debt ratio and delinquency rate" and "non-linear growth relationship between credit inquiry frequency and default risk" in the credit scenario. In addition, the feature engineering in the traditional logistic regression score card method also has the problem of relying on manual experience and poor adaptation to high-dimensional data. SUMMARY

[0003] The purpose of the embodiments of the present application is to provide a credit risk prediction method, device and electronic device to solve the defects of the traditional logistic regression score card method, such as lacking non-linear modeling capability, relying on manual experience and poor adaptation to high-dimensional data.

[0004] To achieve the above-mentioned purpose, the embodiments of the present application provide a credit risk prediction method, comprising: obtaining target credit data to be predicted; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; The credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer. The feature extraction model layer includes a plurality of feature extraction sub-models. Each feature extraction sub-model is used for feature extraction on the target credit data to obtain a feature extraction result. The fusion layer is used for fusing the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result. The credit risk prediction model is trained based on a plurality of target sample credit data and a risk label associated with each target sample credit data. The plurality of target sample credit data is obtained by feature extraction on at least two of the time sequence features, text features and interaction features based on an initial sample credit data set.

[0005] Optionally, the inputting the target credit data into the credit risk prediction model to obtain the credit risk prediction result output by the credit risk prediction model comprises: input the target credit data into a plurality of feature extraction sub-models to obtain feature extraction results output by all the feature extraction sub-models; a control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficients corresponding to each of the feature extraction sub-models.

[0006] Optionally, the plurality of feature extraction sub-models include a first tree model and a second tree model; the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficients corresponding to each of the feature extraction sub-models, and calculates the credit risk prediction result by the following formula: P = 1 / (1 + e^(-(β0+β1z1+β2z2))); wherein P represents the credit risk prediction result; β0 represents a set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; and β2 represents the fusion coefficient corresponding to the second tree model.

[0007] Optionally, the plurality of feature extraction sub-models include a first tree model, a second tree model and a third tree model; the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficients corresponding to each of the feature extraction sub-models, and calculates the credit risk prediction result by the following formula: P = 1 / (1 + e^(-(β0+β1z1+β2z2+β3z3))); wherein P represents the credit risk prediction result; β0 represents a set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; β2 represents the fusion coefficient corresponding to the second tree model; z3 represents the feature extraction result of the third tree model; and β3 represents the fusion coefficient corresponding to the third tree model.

[0008] Optionally, the method further includes: generating an adjustment suggestion based on the credit risk prediction result and credit risk prediction results in a historical time period; constructing a risk report based on the credit risk prediction result and the adjustment suggestion; outputting the risk report.

[0009] Optionally, the method further includes: monitoring model performance of the credit risk prediction model; in a case where the model performance does not meet a set condition, pushing an alarm information and positioning a root cause based on a model evaluation tool. merge the new sample credit data and the plurality of target sample credit data to construct a new dataset; iteratively train model parameters of a fusion layer of the credit risk prediction model based on the new dataset.

[0010] Optionally, the credit risk prediction model is trained by the following steps: obtain an initial sample credit dataset; the initial sample credit dataset includes a plurality of initial sample credit data; perform multi-dimensional sample optimization on the plurality of initial sample credit data to obtain a plurality of intermediate sample credit data; perform feature extraction of time series features, text features, and interaction features on each of the intermediate sample credit data to obtain a plurality of candidate sample credit data; preprocess the plurality of candidate sample credit data to obtain a plurality of target sample credit data; train the credit risk prediction model based on the plurality of target sample credit data and a risk label associated with each of the target sample credit data, and determine a SHAP value of at least one feature of each of the target sample credit data; perform at least one of effect verification, efficiency verification, fairness verification, and pressure resistance verification on the credit risk prediction model.

[0011] Optionally, the multi-dimensional sample optimization on the plurality of initial sample credit data to obtain a plurality of intermediate sample credit data comprises: perform sample screening on the plurality of initial sample credit data based on business dimensions and / or compliance dimensions; perform clustering processing on the plurality of initial sample credit data that have undergone sample screening based on a clustering method; perform SMOTE oversampling, ENN outlier removal, and weight assignment based on an entropy weight method on the plurality of initial sample credit data that have undergone clustering processing; perform stratified sampling division on the plurality of initial sample credit data that have undergone weight assignment to obtain a training set, a validation set, and a test set; the training set, the validation set, and the test set serve as the plurality of intermediate sample credit data.

[0012] Optionally, the preprocessing of the plurality of candidate sample credit data to obtain a plurality of target sample credit data comprises: perform at least one of smoothing processing, feature encoding, and missing value supplementation on each candidate sample credit data; At least one of dynamic threshold filtering, redundant feature elimination and data compliance processing is performed on each candidate sample credit data after feature preprocessing, to obtain a plurality of target sample credit data.

[0013] In another aspect, the present application also provides a credit risk prediction device, comprising: An acquisition module is configured to acquire target credit data to be predicted; A prediction module is configured to input the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; The credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer comprises a plurality of feature extraction sub-models; each feature extraction sub-model is configured to perform feature extraction on the target credit data to obtain a feature extraction result; the fusion layer is configured to fuse the feature extraction results of all the feature extraction sub-models to obtain the credit risk prediction result; the credit risk prediction model is trained based on a plurality of target sample credit data and a risk label associated with each target sample credit data; the plurality of target sample credit data are obtained based on feature extraction of at least two types of features among time sequence features, text features and interaction features from an initial sample credit data set.

[0014] In another aspect, the present application also provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the credit risk prediction method described above when executing the program.

[0015] In another aspect, the present application also provides a machine-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the credit risk prediction method described above.

[0016] In another aspect, the present application also provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the credit risk prediction method described above.

[0017] Through the above technical solution, the credit risk prediction model of the present application embodiment captures the risk association of different types of features through a plurality of feature extraction sub-models; the fusion layer integrates the outputs of the plurality of feature extraction sub-models through logistic regression to improve the nonlinear modeling capability; in addition, the plurality of target sample credit data used for the credit risk prediction model in the present application embodiment are obtained based on feature extraction of at least two types of features among time sequence features, text features and interaction features from an initial sample credit data set, thereby realizing automatic mining of high-dimensional data and improving the high-dimensional data fitting degree. Thus, the present application embodiment solves the problems of lack of nonlinear modeling capability and dependence on manual experience and poor high-dimensional data fitting of existing methods.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is one of the flowcharts illustrating the credit risk prediction method provided by the present invention; Figure 2 This is the second flowchart of the credit risk prediction method provided by the present invention; Figure 3 This is the third flowchart of the credit risk prediction method provided by the present invention; Figure 4 This is the fourth flowchart of the credit risk prediction method provided by the present invention; Figure 5 This is the fifth flowchart illustrating the credit risk prediction method provided by this invention; Figure 6 This is the sixth flowchart of the credit risk prediction method provided by the present invention; Figure 7 This is the seventh flowchart of the credit risk prediction method provided by the present invention; Figure 8 This is the eighth flowchart of the credit risk prediction method provided by the present invention; Figure 9 This is a schematic diagram of the credit risk prediction device provided by the present invention; Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] Method Implementation Examples Please refer to Figure 1 This invention provides a credit risk prediction method, comprising: Step 100: Obtain the target credit data to be predicted.

[0022] Electronic devices acquire target credit data to be predicted. This target credit data can be a set of data indicating a user's creditor-debtor relationships. For example, target credit data could include a customer's loan records, basic customer information, and credit data. Loan records could include data such as loan amount and number of loans. Basic customer information could include information such as username and age.

[0023] Step 200: Input the target credit data into the credit risk prediction model to obtain the credit risk prediction result output by the credit risk prediction model.

[0024] The electronic device inputs the target credit data into a credit risk prediction model, obtaining the credit risk prediction result output by the model. The credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer. The feature extraction model layer includes multiple feature extraction sub-models; each sub-model extracts features from the target credit data to obtain a feature extraction result. Each of the multiple feature extraction sub-models can be built using a tree model. The number of feature extraction sub-models can be two or more (e.g., three). In one embodiment, the multiple feature extraction sub-models include a LightGBM model (Light Gradient Boosting Machine) and an XGBoost model (Extreme Gradient Boosting). The LightGBM model (main model) and the XGBoost model (auxiliary model) are used to capture the risk correlation of different types of features, respectively. In another embodiment, the multiple feature extraction sub-models include a LightGBM model, an XGBoost model, and a CatBoost model (Categorical Boosting). The LightGBM model (main model), XGBoost model (auxiliary model), and CatBoost model (supplementary model) are used to capture risk associations of different types of features. The fusion layer is used to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result. The fusion layer can use logistic regression to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk probability of the target credit data. The higher the credit risk probability, the greater the credit risk. In this embodiment, a credit risk threshold can be set; exceeding this threshold indicates the existence of credit risk. The model parameters (hyperparameter tuning) of the credit risk prediction model can be optimized using the Optuna tool, and can be set as follows: Search space: max_depth=[5,9] (higher dimension for cyclic features), n_estimators=[300,600]; Early stopping condition: validation set AUC value does not improve for 6 consecutive rounds (large fluctuations in cyclic samples, relaxed by 1 round); Optimal parameters: LightGBM (max_depth=8, n_estimators=420), XGBoost (reg_alpha=3.8), after fusion, validation set AUC=0.86.This invention addresses the lack of nonlinear modeling capabilities in traditional logistic regression scorecard methods by constructing an integrated architecture of "feature extraction model layer + fusion layer". The feature extraction model layer integrates multiple feature extraction sub-models (such as LightGBM model, XGBoost model, CatBoost model, etc.) to capture the risk associations of different types of features. The fusion layer integrates the output of the base model through logistic regression, breaking through the modeling limitations of traditional linear models and single tree models, and accurately capturing complex patterns such as "the stepwise association between income-debt ratio and delinquency rate" and "the nonlinear growth relationship between credit inquiry frequency and default risk", thereby improving nonlinear modeling capabilities.

[0025] The credit risk prediction model is trained based on multiple target sample credit data and the risk label associated with each target sample credit data. The multiple target sample credit data are obtained by feature extraction from an initial sample credit dataset, performing at least two of the following features: time-series features, text features, and interaction features. In other words, the multiple target sample credit data are obtained by feature extraction from an initial sample credit dataset, performing time-series features, text features, and interaction features. In one embodiment, the multiple target sample credit data are obtained by feature extraction from an initial sample credit dataset, performing time-series features, text features, and interaction features. The initial sample credit dataset includes multiple initial sample credit data. The initial sample credit data undergoes feature extraction using time-series features, text features, and interaction features to obtain the target sample credit data. The time-series features can be generated using AutoFeatureTools (based on sliding window statistics). The sliding window can be 1 month, 3 months, or 6 months, etc. For example, using a 3-month sliding window, AutoFeatureTools generates time-series features from the initial sample credit data, including "frequency of borrowing in the past 3 months" (e.g., customer A borrowed 5 times in the past 3 months); "credit limit utilization trend" (e.g., utilization rate increased from 30% to 80%, a month-on-month increase of 167%); and "interval between repayments and subsequent borrowing" (e.g., average interval decreased from 15 days to 3 days, indicating increased risk). Text feature extraction can be performed through BERT embedding, negative word frequency, and sentiment scoring text feature engineering. For example, the frequency of negative keywords such as "revolving loan delinquency" and "account status is under observation" can be extracted from the credit report of the initial sample credit data. Interactive feature extraction can combine revolving loan behavior with cross-templates (e.g., numerical × numerical, numerical / numerical) to generate interactive features such as "number of credit inquiries / borrowing frequency" (frequent inquiries but few loans, high risk) and "historical delinquency count × current credit limit utilization." Thus, this embodiment of the invention automatically generates multiple types of features from time-series features, text features, and interactive features, achieving automated mining of high-dimensional data and improving the adaptability of high-dimensional data.

[0026] Existing methods are prone to the curse of dimensionality when dealing with high-dimensional information such as customer transaction sequences and multi-dimensional credit data, resulting in a decrease in model training efficiency of over 30% and an inability to effectively extract hidden risk correlations within the data. The credit risk prediction model of this invention captures risk correlations of different types of features through multiple feature extraction sub-models; the fusion layer integrates the outputs of multiple feature extraction sub-models through logistic regression, improving nonlinear modeling capabilities; furthermore, the multiple target sample credit data used in this invention are obtained by feature extraction from at least two of the following: time-series features, textual features, and interactive features, based on the initial sample credit dataset, thereby achieving automated mining of high-dimensional data and improving high-dimensional data adaptability. Thus, this invention addresses the shortcomings of existing methods, such as lack of nonlinear modeling capabilities, reliance on human experience, and poor adaptability to high-dimensional data.

[0027] In other aspects of the present invention, the step of inputting the target credit data into a credit risk prediction model to obtain the credit risk prediction result output by the credit risk prediction model includes: inputting the target credit data into multiple feature extraction sub-models to obtain feature extraction results output by all the feature extraction sub-models; and controlling the fusion layer to determine the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model.

[0028] In one embodiment, multiple feature extraction sub-models include a first tree model and a second tree model. For example, multiple feature extraction sub-models include a LightGBM model (main model) and an XGBoost model (auxiliary model). In one embodiment, the model parameters of the LightGBM model are as follows: learning_rate = 0.04, max_depth = 8 (for more complex cyclic feature interactions), n_estimators = 450, used to focus on cyclic temporal features. The model parameters of the XGBoost model are as follows: learning_rate = 0.05, reg_alpha = 4.0 (to suppress overfitting of high-frequency features), used to enhance the learning of credit interaction features. The electronic device inputs the target credit data into the LightGBM model and the XGBoost model to obtain the feature extraction results output by the LightGBM model and the XGBoost model. Then, the control fusion layer determines the credit risk prediction result based on the feature extraction results of the LightGBM model and the XGBoost model, as well as the corresponding fusion coefficients of the LightGBM model and the XGBoost model.

[0029] In one embodiment, the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model, calculated using the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2))); Wherein, P represents the credit risk prediction result; β0 represents the set parameters; z1 represents the feature extraction result of the first tree model (e.g., the LightGBM model); β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model (e.g., the XGBoost model); and β2 represents the fusion coefficient corresponding to the second tree model. The credit risk prediction model of this embodiment is constructed based on a Stacking integrated architecture of "feature extraction model layer + fusion layer." The feature extraction model layer is constructed using two tree models. This embodiment improves the prediction accuracy and generalization ability of credit risk.

[0030] In another embodiment, the plurality of feature extraction sub-models include a first tree model, a second tree model, and a third tree model; for example, the plurality of feature extraction sub-models include a LightGBM model (main model), an XGBoost model (auxiliary model), and a CatBoost model (supplementary model). The electronic device inputs the target credit data into the LightGBM model, the XGBoost model, and the CatBoost model to obtain the feature extraction results output by the LightGBM model, the XGBoost model, and the CatBoost model. Then, the control fusion layer determines the credit risk prediction result based on the feature extraction results of the LightGBM model, the XGBoost model, and the CatBoost model, as well as the corresponding fusion coefficients of the LightGBM model, the XGBoost model, and the CatBoost model. Specifically, the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the corresponding fusion coefficients of each feature extraction sub-model, calculated using the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2+β3z3))); Wherein, P represents the credit risk prediction result; β0 represents the set parameters; z1 represents the feature extraction result of the first tree model (e.g., LightGBM model); β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model (e.g., XGBoost model); β2 represents the fusion coefficient corresponding to the second tree model; z3 represents the feature extraction result of the third tree model (e.g., CatBoost model); β3 represents the fusion coefficient corresponding to the third tree model. The credit risk prediction model of this embodiment is constructed based on a Stacking integrated architecture of "feature extraction model layer + fusion layer". The feature extraction model layer is constructed through three tree models. This embodiment further improves the prediction accuracy and generalization ability of credit risk.

[0031] This invention, through its "Stacking integrated architecture" (LightGBM main model capturing temporal / interactive features + XGBoost processing high cardinality features + CatBoost adapting missing values, and logistic regression fusing probabilities), breaks through the modeling limitations of traditional linear models and single tree models, and accurately captures complex patterns in credit scenarios such as "income-debt ratio ladder correlation" and "revolving loan frequency-risk nonlinear growth".

[0032] The beneficial effects are as follows: Quantitatively, the AUC (Area Under the Curve) of the test set is improved from 0.78-0.82 in the existing technology to over 0.85 (reaching 0.88 in the "Revolving Postal Loan" example), and KS (Kolmogorov-Smirnov) is ≥0.40; under extreme scenarios (economic downturn with a bad sample rate of +20% and a new customer influx rate of +30%), the AUC decreases by <3% (the existing single-tree model decreases by over 5%); Qualitatively, the bad debt rate is significantly reduced. In the embodiment of this invention, the M1+ delinquency rate is reduced from 5.2% to 3.8%, and the risk identification accuracy meets the business objective of "bad debt rate <4%" in consumer finance.

[0033] In other aspects of the embodiments of the present invention, the credit risk prediction model is trained through the following steps: obtaining an initial sample credit dataset; the initial sample credit dataset includes multiple initial sample credit data; performing multi-dimensional sample optimization on the multiple initial sample credit data to obtain multiple intermediate sample credit data; extracting features of time series features, text features, and interaction features from each of the intermediate sample credit data to obtain multiple candidate sample credit data; preprocessing the multiple candidate sample credit data to obtain multiple target sample credit data; training the credit risk prediction model based on the multiple target sample credit data and the risk labels associated with each target sample credit data, and determining the SHAP value of at least one feature of each target sample credit data; and performing at least one of the following verifications on the credit risk prediction model: effectiveness verification, efficiency verification, fairness verification, and resilience verification.

[0034] Please refer to Figure 2 The training steps in this embodiment of the invention include sample selection and quality optimization, feature engineering and automated mining, model construction and ensemble optimization, and model validation and fairness verification. The electronic device first acquires an initial sample credit dataset; the initial sample credit dataset includes multiple initial sample credit data sets. These initial sample credit data sets may include customer loan records, basic customer information, and credit data. The electronic device performs multi-dimensional sample optimization on the multiple initial sample credit data sets to obtain multiple intermediate sample credit data sets. For example, the electronic device can screen and optimize samples from multiple dimensions such as compliance, balance, and representativeness to solve the problems of low information density, imbalance, and outlier interference in the initial sample credit dataset, providing an unbiased, high-quality data source for model training and avoiding model overfitting or prediction bias caused by sample defects. In one embodiment, please refer to... Figure 3 The method involves multi-dimensional sample optimization of the initial sample credit data to obtain multiple intermediate sample credit data, including: screening the initial sample credit data based on business and / or compliance dimensions; clustering the initial sample credit data after sample screening using a clustering method; performing SMOTE oversampling, ENN outlier removal, and entropy weighting on the clustered initial sample credit data; and performing stratified sampling on the weighted initial sample credit data to obtain a training set, a validation set, and a test set; the training set, the validation set, and the test set serve as the multiple intermediate sample credit data.

[0035] The electronic device can perform sample screening on the multiple initial sample credit data based on business dimensions and / or compliance dimensions. In one embodiment, sample screening based on business dimensions can be based on product, customer group, or transaction cycle. In another embodiment, sample screening based on compliance dimensions can be filtering invalid accounts / test data. In yet another embodiment, the electronic device defines the initial samples according to "business dimension (product / customer group / transaction cycle) + compliance dimension (invalid account / test data filtering)" and excludes samples with low information density. For example, the electronic device performs business dimension screening: retains initial sample credit data with "two or more loan records" (core customer group of revolving products), and removes initial sample credit data with "only one loan" (420,000 transactions); for loan terms ≥7 days (covering short-cycle noise), removes abnormal loans such as "test credit limit" and "invalid cancellation" (80,000 transactions); and the electronic device performs compliance dimension screening: filters samples with "incomplete information (missing ID card / income certificate)" (150,000 transactions) to ensure compliance with the data integrity requirements of the Personal Information Protection Law. For example, there were 2.8 million initial sample credit data entries. After sample screening based on business and / or compliance dimensions, 2.15 million initial sample credit data entries were obtained, corresponding to 850,000 customers.

[0036] Electronic devices then use clustering methods to cluster the initial sample credit data after sample screening. Clustering methods include K-medoids and K-Means. For example, electronic devices can use K-Means clustering to remove outliers from the initial sample credit data. K-Means clustering calculates Euclidean distance and iteratively updates cluster centers, removing outliers exceeding three standard deviations and supplementing the removal of dormant samples. The K-Means clustering method calculates the similarity between a sample and its cluster centers, removing outliers exceeding three standard deviations, as shown in the following formula:

[0037] in: x i For the first i One sample, μ c For the first c The center of each cluster, m These are feature dimensions (such as product, customer group, and transaction cycle related features). x ik for x i The k 1 eigenvalue, μ ck for μ c The kThe K-Means clustering method can select three core features of recurring products: "loan amount / credit limit (utilization rate)", "loan interval days", and "historical overdue number". The clustering parameter can be K=5 (divided according to recurring behavior patterns such as "high frequency low amount" and "low frequency high amount"). The outlier threshold is set to "4 times the in-cluster standard deviation" (recurring products fluctuate more). The elimination results can select to remove abnormal samples (90,000 cases) that "loan amount far exceeds credit limit by 200%" and "loan interval < 1 day and more than 10 consecutive times", retaining 2.06 million data from 2.15 million initial sample credit data.

[0038] Next, the electronic device performs SMOTE oversampling, ENN (Edited Nearest Neighbors) to remove outliers from multiple initial sample credit data after clustering, and assigns weights based on entropy weighting. The electronic device synthesizes minority class samples (k=5, sampling_strategy=0.3) through SMOTE oversampling, removes outlier synthesized samples through ENN, and then assigns dynamic weights to the initial sample credit data based on entropy weighting. There may be sample imbalance among the 2.06 million initial sample credit data. For example, there are 123,600 bad samples (e.g., overdue loans) among the 2.06 million initial sample credit data, accounting for 6% (the overdue rate of revolving loans is lower than that of one-time loans). The electronic device can synthesize minority class bad samples with "high revolving frequency (≥5 loans)" through SMOTE oversampling. SMOTE oversampling (synthesizing minority class samples) alleviates sample imbalance and supplements bad samples to improve the model's risk identification ability. ;in: x new For the synthesis of new samples ，x i These are minority class (bad sample) samples. x j for x i of k Neighbors ( k =5, both are minority categories), λ ∈[0,1] is a random number. sampling _ strategy=0.3 (target minority class percentage). For example, setting k=6 (higher similarity of cyclical behavior), sampling_strategy=0.2 (target bad sample percentage 11.7%), resulting in 192,000 synthetic bad samples. The electronic device then uses ENN to remove abnormal synthetic samples. For example, using ENN to remove abnormal data (32,000 records) from multiple initial sample credit data where "the overlap between synthetic sample and good sample borrowing behavior is >60%", the final initial sample credit data volume = 206 + 192,000 - 32,000 = 2,220,000 records. The electronic device then assigns dynamic weights to the initial sample credit data based on the entropy weight method. The entropy value calculation formula (measuring the dispersion of customer group contribution) is as follows: The formula for weighting (assigning higher weight to high-contribution customer groups) is as follows: ;in: p ij For the first j The first customer group i The first indicator, for example, the second. i The indicators represent the business share (such as the percentage of loan amount / number of times). n For the total number of indicators, m The total number of customer groups is used, for example, the final weights are 1.2 for repeat borrowers, 1.0 for first-time borrowers, and 0.5 for low-value customers. In another embodiment, based on the cyclical contribution (number of loans × average amount), the entropy weight method is used to assign values ​​as follows: high-frequency cyclical customers (≥8 loans): weight 1.3 (60% business contribution); medium-frequency cyclical customers (3-7 loans): weight 1.0 (30% contribution); low-frequency cyclical customers (2 loans): weight 0.6 (10% contribution).

[0039] The electronic device uses stratified sampling to partition the dataset. The StratifiedShuffleSplit function divides the initial 2.22 million credit data samples into a training set (70%), a validation set (20%), and a test set (10%) based on "time (T-12~T-5 / T-4~T-2 / T-1~T) + customer group," ensuring that the difference in customer group distribution is <2%. For example, stratification by time: partitioning by "loan time" to avoid data leakage from multiple loans by the same customer: Training set (70%): 1.554 million transactions from January 2022 to December 2023; Validation set (20%): 444,000 transactions from January 2024 to March 2024; Test set (10%): 222,000 transactions from April 2024 to June 2024. Customer group stratification: ensuring that the proportion of each cycle frequency customer group in the three sets differs by <2%, such as the proportion of high-frequency customer groups being 45% ± 1% in the training / validation / test sets. The training set (70%), validation set (20%), and test set (10%) were used as intermediate sample credit data.

[0040] This invention addresses the issues of low sample quality and unresolved correlation in existing technologies through multi-dimensional sample processing and dynamic optimization. It uses a dual-dimensional approach of "business + compliance" to screen initial samples, combines K-Means clustering to remove outliers (such as outliers exceeding 3 standard deviations), employs SMOTE+ENN to achieve sample balancing (dynamically adjusting the proportion of minority classes), and assigns dynamic weights to customer groups (such as repeat borrowers, first-time borrowers, and low-value customers) based on entropy weighting. Simultaneously, it stratifies the dataset by "time + customer group" to ensure consistent sample distribution (difference < 2%), thus resolving the sample correlation problem in revolving credit involving "multiple loans by the same customer."

[0041] Please refer to Figure 4 The electronic device then extracts features from each of the intermediate sample credit data, including temporal features, textual features, and interactive features, to obtain multiple candidate sample credit data. This embodiment of the invention achieves automated feature generation, transforming raw data into high-value, high-dimensional data. For example, intermediate sample credit data, after feature extraction of temporal features, textual features, and interactive features, yields candidate sample credit data. Temporal feature extraction can be achieved using AutoFeatureTools (based on sliding window statistics). The sliding window can be 1 month, 3 months, or 6 months, etc. For example, using a 3-month sliding window, AutoFeatureTools generates temporal features from the loan records of the intermediate sample credit data, including "frequency of loans in the past 3 months" (e.g., customer A has borrowed 5 times in the past 3 months); "credit limit utilization trend" (e.g., utilization rate increased from 30% to 80%, a month-on-month increase of 167%); and "interval between repayments and subsequent loans" (e.g., average interval decreased from 15 days to 3 days, indicating increased risk). Textual feature extraction can be performed using BERT embedding, negative word frequency, and sentiment score text feature engineering. For example, the frequency of negative keywords such as "revolving loan delinquency" and "account status under observation" can be extracted from the credit reports of intermediate sample credit data. Interactive feature extraction can combine revolving loan behavior with cross-templates (e.g., numerical × numerical, numerical / numerical) to generate interactive features such as "credit inquiry count / borrowing frequency" (frequent inquiries but few loans indicate high risk) and "historical delinquency count × current credit limit utilization." This results in an initial feature scale of 320 (including 240 automatically generated features and 80 manually generated revolving loan-specific features). Electronic devices then preprocess the multiple candidate sample credit data to obtain multiple target sample credit data. This preprocessing includes feature preprocessing and compliance screening, transforming the raw data into high-value, non-redundant, and model-adaptable compliant features. This addresses the problems of low efficiency, missed dynamic risk signals, and poor adaptation to high-dimensional data in manual feature engineering, while also complying with regulatory requirements.

[0042] In one embodiment, the preprocessing of the multiple candidate sample credit data to obtain multiple target sample credit data includes: performing at least one feature preprocessing operation on each candidate sample credit data, including smoothing, feature encoding, and missing value imputation; and performing at least one operation on each candidate sample credit data after feature preprocessing, including dynamic threshold filtering, redundant feature removal, and data compliance processing, to obtain multiple target sample credit data. The electronic device performs one or more feature preprocessing operations on each candidate sample credit data, including smoothing, feature encoding, and missing value imputation. Specifically, the time-series features in the candidate sample credit data are smoothed using EMA (exponential moving average) (α=0.3), high-cardinality variables are encoded using target encoding (λ=10), low-cardinality variables are encoded using one-hot encoding, and missing values ​​are filled in differentiatedly for different customer groups. EMA time-series smoothing (reducing the interference of time-series feature fluctuations) highlights long-term risk trends and is suitable for fluctuation features such as transaction amount and credit limit utilization rate. The formula is as follows: ;in: EMA t For the first t Periodic smoothing value, EMA t-1 For the first t-1 Periodic smoothing value, x t For the first t Original value of the period α =0.3 (balancing the impact of recent and historical data). Feature encoding (handling high cardinality variables) avoids dimensionality inflation and retains category risk information (such as occupation and loan purpose), as shown in the following formula: ;in: enc ( c ) as category c The encoded value, n c For category c The number of samples, r c For category c bad sample rate n total The total number of samples, r total The total bad sample rate, λ=10 (smoothing coefficient, to avoid overfitting of small categories). In one embodiment, electronic devices use cyclic time-series smoothing to smooth fluctuating features such as "credit limit utilization rate" with EMA (α=0.4, focusing more on recent behavior). For example, customer B's utilization rate in the past 6 months: 50%→60%→75%→65%→80%→90%, after smoothing it becomes 50%→54%→61%→62%→69%→77%; High cardinality feature encoding: "Loan purpose" (12 categories) and "Repayment channel" (8 categories) use target encoding (λ=12, the cyclic sample categories are more dispersed), such as "shopping" purpose encoding = (number of shopping samples × shopping bad sample rate + total samples × total bad sample rate) / (number of shopping samples + total samples × 12); Missing value imputation: missing cyclic behavior (such as no history of the first loan) is imputed with "average of customers in the same credit limit range" (such as the average frequency of the first loan of customers with a credit limit of 50,000-100,000 is 1.2 times). The electronic device performs one or more operations on each candidate sample credit data after feature preprocessing, including dynamic threshold filtering, redundant feature removal, and data compliance processing, to obtain multiple target sample credit data. Specifically, the electronic device removes low-quality features based on dynamic thresholds, combines "LightGBM importance + mutual information value (MI < 0.8)" for redundancy removal, filters sensitive features, and de-sensitizes quasi-sensitive features (AES-256 encryption), ultimately retaining 120-150 features. LightGBM feature importance is used as the selection criterion, combined with correlation and clustering to remove redundant features with similar expressions. Cross-validation is then used to verify the generalization performance of the selected feature subset, preventing pseudo-correlation features with "high importance but limited contribution" from entering the final set. Redundant feature removal is achieved by calculating the mutual information value between features. The mutual information value measures the correlation between features, retaining high-value non-redundant features. The calculation formula is as follows: ;in: p ( x , y ) as a feature X and Y The joint probability density, p ( x ), p ( y ) are respectively X , YThe marginal probability density is considered redundant when it is greater than 25%. In one embodiment, the electronic device performs dynamic threshold filtering: core features with a missing rate of less than 25% (such as "frequency of loans in the past 3 months" with a missing rate of 8%) are retained, and low-value features with a missing rate of 35% such as "repayment channel" are removed; then redundancy removal is performed: the mutual information value between features is calculated, and redundant features such as "number of loans in the past 3 months" and "frequency of loans in the past 3 months" (MI=0.82) are removed, retaining 142 features; finally, compliance processing is performed: sensitive features such as "bank card number" and "detailed house number of home address" are filtered, and "monthly income" is desensitized in segments of "<5k / 5k-15k / >15k", and stored with AES-256 encryption.

[0043] This invention replaces traditional manual binning and feature design with an "automated feature generation + intelligent screening mechanism" (AutoFeatureTools generates time-series / text / interaction features, EMA smooths dynamic features, and "LightGBM importance + mutual information value" removes redundancy), adapting to high-dimensional data such as customer transaction sequences and credit reports. It solves the problems of "feature engineering relying on manual intervention and poor high-dimensional adaptability," significantly improving feature efficiency and utilization. The beneficial effects are as follows: Quantitatively: The feature engineering cycle is shortened from 2-3 weeks in existing technologies to 1-2 days, reducing manual workload by 70%; the utilization rate of high-value features increases from 60% to 90% (e.g., "Revolving Credit Loan" selects 142 effective features from 320 initial features, with no high-value time-series features omitted); Qualitatively: No manual maintenance of binning rules is required; it can automatically capture dynamic risk signals such as "revolving credit utilization trend" and "re-borrowing interval after repayment," adapting to the business characteristics of revolving credit's "high-frequency borrowing and credit reuse."

[0044] Please refer to Figure 5The electronic device trains the credit risk prediction model based on multiple target sample credit data and the risk labels associated with each target sample credit data, and determines the SHAP value of at least one feature for each target sample credit data. Based on the multiple target sample credit data and the risk labels associated with each target sample credit data, the electronic device trains the model parameters of the feature extraction model layer and the fusion layer using a supervised training method. For example, the electronic device compares the final output of the credit risk prediction model with the true risk labels and calculates the loss function value. Based on the loss function value, the gradient of each weight and bias is calculated using the backpropagation algorithm. The weights and biases in the feature extraction model layer and the fusion layer are updated using gradient descent or other optimization algorithms. Through multiple iterative training processes, the credit risk prediction model can learn effective feature representations and classification decision boundaries, thereby improving the accuracy of credit risk prediction. The electronic device can also perform interpretability analysis of SHAP values: at the global level, it calculates the feature SHAP values ​​of the target sample credit data and outputs a "TOP10 Feature-Business Interpretation Table," and at the local level, it generates a customer "Risk Factor Attribution Map" to support business decisions. SHAP (Feature Contribution Measurement for Interpretability) quantifies the impact of features on risk prediction based on game theory, supporting business interpretation. ;in: F For the entire feature set, S For those without features i A subset of, | S | is the size of the subset. v ( S ) is a subset S Predictive value (model in) S (prediction error on) i Features i The SHAP value (positive values ​​increase risk, negative values ​​decrease risk). In one embodiment, the global feature interpretation (TOP5) of SHAP value interpretability analysis is shown in Table 1: Table 1

[0045] Here's an example of local attribution: Customer C (10 revolving loans) has a risk probability of 0.72. The attribution plot shows that the main risk points are: 8 loans in the past 3 months (SHAP=0.32) and a credit utilization rate of 92% (SHAP=0.28). Manual review needs to focus on verifying the use of funds. By leveraging the interpretability of SHAP values, technical characteristics are transformed into business rules, solving the problems of weak and difficult-to-interpret single-model nonlinear modeling, and supporting compliance decisions.

[0046] Existing methods for calculating SHAP values ​​only quantify feature contributions at the technical level, failing to integrate the interpretation results with actual credit business rules and thus unable to support implementation decisions. This invention addresses this by establishing a "business-oriented SHAP value interpretability system" (globally outputting a "TOP Feature-Business Interpretation Table" and locally generating a "Customer Risk Factor Attribution Map"), transforming technical feature contributions into understandable risk rules for credit business (e.g., "Borrowing frequency > 8 times / month = tight cash flow"), and linking it to manual review processes. This solves the problems of "disconnected interpretability and high compliance risk," meeting regulatory requirements and business decision-making needs. The beneficial effects are as follows: Quantitatively: The efficiency of manual review is improved by 30% (in the example, the manual intervention time for "Revolving You Loan" is reduced from an average of 5 minutes / transaction to 3.5 minutes / transaction); the interpretability compliance rate is 100%, meeting the mandatory requirements of "risk model interpretability" in the "Interim Measures for the Management of Internet Loans by Commercial Banks"; Qualitatively: Reviewers can directly locate core risk points through attribution diagrams (such as "the customer's high risk is due to a credit line utilization rate of 92% + 8 loans in the past 3 months"), without the need for additional risk reports, and the compliance review pass rate is improved by 40%.

[0047] Please refer to Figure 6 In this embodiment of the invention, the credit risk prediction model is further validated using at least one of the following methods: effectiveness validation, efficiency validation, fairness validation, and resilience validation. The electronic device performs one or more of these validations on the credit risk prediction model. The electronic device can validate the model from four dimensions: effectiveness validation (accuracy / stability), efficiency validation, fairness validation, and resilience validation. This addresses the problem of focusing solely on accuracy while neglecting customer group discrimination and stability in extreme scenarios, ensuring the model is accurate, efficient, compliant, and non-discriminatory, and selecting the optimal version. Effectiveness validation can be performed using PSI (Population Stability Index), KS (Kolmogorov-Smirnov statistic), and AUC (Area Under the Curve). The PSI population stability index verifies model stability, measures the difference in score distribution between the training and test sets, and ensures the model adapts to new data. The formula for calculating the PSI population stability index is as follows: ;in: b =10 (number of equal frequency bins) A b For the test set b The sample percentage of boxes, E b For the training set bThe sample proportion of the boxes is considered stable if it is less than a certain value. The Kolmogorov-Smirnov statistic (KS) is used to verify the discriminative power, measure the difference in score distribution between good and bad samples, and evaluate the model's risk discrimination effect. The formula for calculating the Kolmogorov-Smirnov statistic (KS) is as follows: ;in: CDF good ( x A good sample score ≤ x The cumulative probability, CDF bad ( x ) is the score for bad samples ≤ x The cumulative probability, KS A score ≥0.40 indicates strong discriminative power. Fairness verification can be performed by calculating the customer group fairness error. The customer group fairness error avoids customer group discrimination, quantifies subgroup prediction bias, and ensures that the model's fairness meets regulatory requirements. The formula for calculating the sum of subgroup errors is as follows: .

[0048] In one embodiment, the process for electronic devices to perform effectiveness verification, efficiency verification, fairness verification, and resilience verification of the credit risk prediction model is as follows: 1. Core effectiveness verification: The test set must meet the following conditions: AUC≥0.85, KS≥0.40, PSI<0.1 (PSI=Σ(A b -E b )ln(A b / E b ), the difference in performance between different customer groups (first-time borrowers / repeat borrowers) is <3%; 2. Efficiency verification: inference response time <80ms (TensorRT quantization acceleration), training time ≤2 hours under an 8-core CPU environment; 3. Customer group fairness verification: divide into 9 subgroups according to "income / age / customer group type", and calculate the subgroup error ( ), ensuring fairness difference Δε < 5%, adjusting feature weights if it exceeds the limit; 4. Extreme scenario stress test: simulating economic downturn (bad sample rate +20%) and new customer influx (first-time loan ratio +30%), requiring AUC decrease < 3%. Verify the model's stability in cyclical scenarios, especially high-frequency borrowing and credit limit adjustment sub-scenarios. In one embodiment, electronic devices verify the stability of credit risk prediction in cyclical scenarios, especially high-frequency borrowing and credit limit adjustment sub-scenarios. Verification results are as follows: 1. Core effect verification. Test set metrics: AUC=0.85, KS=0.41, PSI=0.09 (small difference in score distribution between training set and test set); Scenario-specific effects: High-frequency borrowing (≥8 times) sub-scenarios: AUC=0.84; Credit limit adjustment (temporary credit limit increase) sub-scenarios: AUC=0.83 (meets requirements). 2. Efficiency verification. Inference Response: Single loan request response time is 82ms (meets <100ms requirement), batch processing (1000 requests / batch) takes 65ms / request; Training Time: Full training on an 8-core CPU environment takes 1.5 hours, meeting business iteration needs. 3. Fairness Verification. Subgroup Division: Divided into 9 subgroups based on "credit amount range (10,000-50,000 / 50,000-150,000 / 150,000-300,000)" and "age (22-30 / 31-45 / 46-55 years old)"; Fairness Indicator: Error difference Δε between subgroups is 4.8% < 5%, for example, the error for the 10,000-50,000 credit limit group is 4.2%, and the error for the 150,000-300,000 credit limit group is 4.5%, with no significant discrimination. 4. Extreme Scenario Stress Resistance. Scenario 1 (Credit limit exhausted): Selecting the sample of "100% utilization rate and 3 consecutive borrowings", AUC=0.82 (a decrease of 3.5%, close to the threshold); Scenario 2 (Sudden increase in high-frequency borrowing): The borrowing frequency increased from an average of 2 times per month to 5 times, AUC=0.83 (a decrease of 2.4%), and the overall stress resistance met the standard.

[0049] Existing methods fail to validate model fairness for subgroups such as low-income groups and young first-time borrowers, potentially leading to discriminatory issues like "a loan rejection rate for low-income groups exceeding that of high-income groups by more than 10% under the same risk level," posing compliance risks. This invention addresses these issues through a "subgroup stratified verification and weight adjustment mechanism" (dividing subgroups by "income / age / customer type," calculating subgroup error differences Δε, and dynamically adjusting feature weights when exceeding limits), thus preventing discriminatory decisions by the model against vulnerable groups such as low-income individuals and young first-time borrowers. This resolves the issues of "lack of fairness and customer group discrimination," mitigating compliance risks. The beneficial effects are as follows: Quantitatively, the difference in error between subgroups has been reduced from 8%-10% in the existing technology to <5% (in the example, the error for low-income customers is 4.2% and for high-income customers is 4.5%, with a difference of only 0.3%); under the same risk level, the difference between the loan rejection rate for low-income customers and that for high-income customers has been reduced from more than 10% to less than 3%; Qualitatively, there are no compliance complaints related to customer discrimination, and the "non-discrimination modeling" special inspection by regulators has been passed, meeting the requirements of the Personal Information Protection Law for "algorithmic fairness".

[0050] In other aspects of the embodiments of the present invention, the method further includes: generating adjustment suggestions based on the credit risk prediction results and credit risk prediction results within a historical time period; constructing a risk report based on the credit risk prediction results and the adjustment suggestions; and outputting the risk report.

[0051] Following model validation and fairness verification, this invention also includes report generation and automated deployment steps. Generating a compliance report, passing a three-level review, and automating deployment solves the problems of cumbersome traditional deployment processes, difficulty in tracing compliance, and high deployment risks. This enables secure model integration with business systems and rapid deployment of risk prediction capabilities. In one embodiment, please refer to... Figure 7 The process is described as follows: 1. Standardized report generation: Generate a PDF report according to regulatory standards (such as the "Interim Measures for the Management of Internet Loans by Commercial Banks"), including 6 major modules such as data compliance, sample characteristics, and model building, and simultaneously export Excel details; 2. Three-level review process: Pass compliance review (sensitive data / de-identification verification), risk review (effectiveness / fairness compliance), and business review (scoring / decision threshold alignment with targets) in sequence; 3. Automated deployment: MLflow manages the model (version number "V{year}{month}{serial number}", retaining the first 3 versions), FastAPI encapsulates the API, and after gray-scale testing (10% traffic, 7-day monitoring consistency ≥85%, call success rate ≥99.9%), it is fully launched and integrated with Prometheus monitoring. In another embodiment, report generation and automated deployment are adapted to the dual-scenario deployment of the recurring product "real-time approval + dynamic credit limit adjustment". The report generation and automated deployment process is as follows: 1. Standardized report. Key Modules: New additions include "Recurring Scenario Adaptability Analysis" (effectiveness of high-frequency loan sub-scenario) and "Credit Limit Adjustment Risk Verification"; Output format: PDF report (62 pages) + recurring feature importance dashboard (for business teams to reference when adjusting credit limits). 2. Three-tiered review. Business review focus: Confirming the model's risk identification capability for customers with "credit limit utilization > 90%" (recall rate 82%), meeting credit limit adjustment rules. 3. Automated deployment. Dual interface design: Loan approval interface ( / loan / apply): Real-time return of single loan risk probability; Credit limit adjustment interface ( / limit / adjust): Calculates customer credit limit adjustment suggestions daily at midnight (based on risk changes over the past 30 days); Gray-scale testing: 5% traffic verification for 10 days, new model and old model decision consistency 86%, interface success rate 99.97%, fully deployed. The single credit risk prediction result is the single loan risk probability. Credit limit adjustment suggestions are generated based on the single loan risk probability and the historical 30-day risk probability. For example, if the adjustment suggestion is that the risk probability gradually increases, then a reduction in the customer's credit limit is recommended. Electronic devices will output risk reports containing the probability of risk for a single loan and recommendations for adjusting the loan amount.

[0052] This invention addresses the characteristics of revolving credit—"dynamic credit limits, high-frequency borrowing, and credit limit reuse"—by designing "revolving behavior time-series characteristics," "credit limit utilization trend monitoring," and "dual-interface deployment (loan approval + credit limit adjustment)." This overcomes the limitations of existing technologies for "one-time loan" scenarios. Improved business adaptability supports complex scenarios such as revolving credit. Beneficial effects are as follows: Quantitatively, the automatic approval rate of revolving credit products increases from 70% to 90%, and manual intervention in credit limit adjustments is reduced by 40% (in the example, the dynamic credit limit adjustment approval rate for "Revolving Postal Loan" increases by 15%); Qualitatively, it can directly adapt to the "borrow and repay at will, credit limit revolving" business model, eliminating the need to develop separate models for sub-scenarios (such as credit limit adjustments and high-frequency borrowing), reducing business expansion costs by 50%.

[0053] In other aspects of the embodiments of the present invention, the method further includes: monitoring the model performance of the credit risk prediction model; pushing alarm information when the model performance does not meet the set conditions, and locating the root cause based on the model evaluation tool; merging the newly added sample credit data and the multiple target sample credit data to construct a new dataset; and iteratively training the model parameters of the fusion layer of the credit risk prediction model based on the new dataset.

[0054] This invention, following report generation and automated deployment, also includes model monitoring and incremental iteration steps. Please refer to... Figure 8The model monitoring and incremental iteration mechanism establishes a real-time monitoring (timely anomaly detection), regular evaluation, and incremental iteration (rapid optimization) system to address the problems of single monitoring dimensions, difficulty in root cause identification, and long iteration cycles. This ensures the model's long-term stability and adaptability to business changes, reducing operational costs. Electronic devices monitor the performance of the credit risk prediction model in real time. If the model performance fails to meet set conditions, alarm information is pushed, and root causes are located using model evaluation tools (such as EvidentlyAI). Model performance includes at least one of AUC, PSI, KS, subgroup error, and loan rejection rate. For example, electronic devices monitor the credit risk prediction model's AUC / PSI / subgroup error / loan rejection rate daily, and monitor API response time / success rate at the second level. If a set threshold is exceeded, Alertmanager or Prometheus pushes an alarm to assess model performance. Regular evaluation: A monthly "Model Operation Report" is generated to evaluate effectiveness / efficiency / fairness / business adaptability. Triggering conditions are AUC difference > 5%, PSI > 0.15, or Δε > 5%. Electronic devices acquire new sample credit data. After processing the new sample credit data according to the above sample selection and quality optimization steps, only the fusion layer of the credit risk prediction model is trained (the model parameters of the feature extraction model layer are fixed). The training data is "new samples + 20% historical samples", the iteration cycle is ≤1 day, and the version is updated after verification and the old version is retained for rollback. The iterative training of the model parameters of the fusion layer of the credit risk prediction model can be trained by the stochastic gradient descent method. In one embodiment, model monitoring and incremental iteration monitor the drift of cyclical behavioral features (such as sudden changes in borrowing frequency) and respond quickly to changes in risk. The model monitoring and incremental iteration process is as follows: 1. Real-time monitoring. Specific monitoring indicators: daily "changes in borrowing frequency distribution" and "credit limit utilization rate PSI", triggering alarms when abnormal; Case: In August 2024, the feature KS of "borrowing frequency in the past 3 months" was 0.23 (>0.2), and the root cause was located as "the summer consumption peak caused a sudden increase in frequency", no iteration is required. 2. Incremental iteration. Triggering condition: In October 2024, the test set AUC dropped to 0.80 (a decrease of 5%) due to "the proportion of high-frequency loans during the Double 11 pre-sale period rising to 60%"; Iteration operation: Added samples from September and October (250,000 transactions), only retrained the fusion layer, which took 1.2 hours, and the new model AUC was 0.85.

[0055] Existing methods only monitor abnormal indicators such as AUC decline and PSI exceeding the standard, failing to automatically pinpoint the root cause (e.g., feature drift, abnormal data collection); furthermore, iteration requires full retraining of the model, a cycle lasting 3-5 days, making it difficult to quickly respond to business changes. This invention replaces the existing full retraining mode with a "real-time monitoring + incremental iteration mechanism" (Prometheus monitors AUC / PSI / subgroup error, EvidentlyAI automatically locates the root cause of feature drift, and incremental updates are achieved by retraining only the Stacking fusion layer). This solves the problems of "inefficient monitoring iteration and delayed response," reducing operational costs and business risks. The beneficial effects are as follows: Quantitatively, the model iteration cycle is shortened from 3-5 days to within 1 day (in the example, the iteration of 250,000 new samples for "Revolving Postal Loan" only took 1.2 hours); the operation and maintenance cost is reduced by 40% (the time for manual anomaly investigation is shortened from an average of 8 hours / time to 2 hours / time); the API response time is <80ms, which meets the real-time approval requirements of 100,000 requests / day during the morning peak of "Revolving Postal Loan"; Qualitatively, it can quickly respond to business changes (such as "high-frequency loans during Double 11" and "summer consumption peak"), the model stability compliance rate is 98% (the PSI has been stable within 0.1 for 6 months after launch), and there is no problem of increased risk exposure due to iteration lag.

[0056] This invention addresses the shortcomings of existing methods in terms of "high accuracy, strong generalization, interpretability, compliance, and high efficiency" by proposing a machine learning-based credit risk prediction model. It overcomes the limitations of existing technologies throughout the entire process, from sample selection and feature development to model building, interpretability, and deployment iterations, achieving the comprehensive goals of "high accuracy, strong generalization, interpretability, compliance, and high efficiency" in credit risk prediction. The main core technical solutions include the following: I. Multi-dimensional Sample Processing and Dynamic Optimization Solution: To address the issues of low sample quality and unresolved correlation in existing technologies, this solution uses a dual-dimensional approach of "business + compliance" to screen initial samples. It combines K-Means clustering to remove outliers (such as outliers exceeding 3 standard deviations), employs SMOTE+ENN to achieve sample balancing (dynamically adjusting the proportion of minority classes), and assigns dynamic weights to customer groups (such as repeat borrowers / first-time borrowers / low-value customers) based on entropy weighting. Simultaneously, it stratifies the dataset by "time + customer group" to ensure consistent sample distribution (difference < 2%), thus resolving the correlation problem in revolving credit where "the same customer borrows multiple times."

[0057] II. Automated Feature Engineering and Compliance Processing Mechanism: To address the issues of traditional technologies relying on manual features and poor adaptation to high-dimensional data, the mechanism automatically generates three types of features based on AutoFeatureTools: time series (sliding window statistics, EMA smoothing), text (BERT embedding, negative word frequency), and interaction (cross-template). Redundancy is removed through "LightGBM importance + mutual information value," combined with sensitive feature filtering and quasi-sensitive feature desensitization (AES-256 encryption), ultimately retaining high-value features (120-150), thereby improving feature utilization and compliance.

[0058] III. Stacking Integrated Model Architecture and Hyperparameter Optimization: To address the weak generalization ability of single-tree models, an integrated architecture of "feature extraction model layer + fusion layer" is constructed: The feature extraction model layer integrates LightGBM (main model), XGBoost (auxiliary), and CatBoost (supplementary) to capture the risk correlation of different types of features respectively; The fusion layer integrates the output of the base model through logistic regression to improve nonlinear modeling ability; Hyperparameters (such as learning_rate and max_depth) are optimized by combining Optuna tool and early stopping mechanism to enhance the model's adaptability to extreme scenarios (economic downturn, influx of new customers).

[0059] IV. Business-Oriented Explainability and Fairness Verification System: To address the issues of explanationability being disconnected from business operations and lack of fairness, the system generates a "Global TOP Feature-Business Interpretation Table" (such as the correlation between feature impact and risk rules) and a "Local Customer Risk Factor Attribution Map" using SHAP values ​​to support manual review and decision-making. Subgroups are divided according to "income / age / customer group type," and the subgroup error (Δε < 5%) is calculated. When the error exceeds the limit, the feature weights are dynamically adjusted to meet regulatory compliance and fairness requirements.

[0060] V. Full-process automated deployment and incremental iteration mechanism: To address the issue of low monitoring and iteration efficiency, MLflow manages model versions, FastAPI encapsulates APIs (response time < 80ms), and combines gray-scale testing (10% traffic verification) to achieve automated deployment; a "real-time monitoring (AUC / PSI / subgroup error) + root cause localization (Evidently AI)" system is built, and when iteration is triggered, only the fusion layer is retrained (the model parameters of the feature extraction model layer are fixed), with an iteration cycle of ≤ 1 day, to quickly respond to business changes.

[0061] This invention addresses six core problems of existing credit risk control models: weak nonlinear modeling, low feature efficiency, poor generalization ability, disconnect from interpretability, lack of fairness, and inefficient iteration. It achieves the comprehensive goals of "high precision, strong generalization, interpretability, compliance, and high efficiency" through a full-process technical solution of "sample-feature-model-interpretation-deployment-iteration".

[0062] Device Examples Please refer to Figure 9 On the other hand, embodiments of the present invention also provide a credit risk prediction device, comprising: Module 901 is used to acquire the target credit data to be predicted; The prediction module 902 is used to input the target credit data into the credit risk prediction model and obtain the credit risk prediction result output by the credit risk prediction model. The credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer. The feature extraction model layer includes multiple feature extraction sub-models. Each feature extraction sub-model is used to extract features from the target credit data to obtain feature extraction results. The fusion layer is used to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result. The credit risk prediction model is trained based on multiple target sample credit data and the risk labels associated with each target sample credit data. The multiple target sample credit data are obtained by feature extraction from at least two of the following features: time-series features, text features, and interaction features, based on an initial sample credit dataset.

[0063] Optionally, the step of inputting the target credit data into the credit risk prediction model to obtain the credit risk prediction result output by the credit risk prediction model includes: The target credit data is input into multiple feature extraction sub-models to obtain the feature extraction results output by all the feature extraction sub-models. The control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model.

[0064] Optionally, the plurality of feature extraction sub-models include a first tree model and a second tree model; the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model, calculated using the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2))); Wherein, P represents the credit risk prediction result; β0 represents the set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; and β2 represents the fusion coefficient corresponding to the second tree model.

[0065] Optionally, the plurality of feature extraction sub-models include a first tree model, a second tree model, and a third tree model; the control fusion layer determines the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model, calculated using the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2+β3z3))); Wherein, P represents the credit risk prediction result; β0 represents the set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; β2 represents the fusion coefficient corresponding to the second tree model; z3 represents the feature extraction result of the third tree model; β3 represents the fusion coefficient corresponding to the third tree model.

[0066] Optionally, the device further includes: The output module is used to generate adjustment suggestions based on the credit risk prediction results and credit risk prediction results within the historical time period; construct a risk report based on the credit risk prediction results and the adjustment suggestions; and output the risk report.

[0067] Optionally, the device further includes: The monitoring module is used to monitor the performance of the credit risk prediction model. The control module is used to push alarm information when the model performance does not meet the set conditions, and to locate the root cause based on the model evaluation tool; The construction module is used to merge the newly added sample credit data and the multiple target sample credit data to construct a new dataset; The iteration module is used to iteratively train the model parameters of the fusion layer of the credit risk prediction model based on the newly created dataset.

[0068] Optionally, the credit risk prediction model is trained through the following steps: Obtain an initial sample credit dataset; the initial sample credit dataset includes multiple initial sample credit data sets; The initial sample credit data is optimized in multiple dimensions to obtain multiple intermediate sample credit data. Feature extraction of time-series features, text features, and interaction features is performed on each of the intermediate sample credit data to obtain multiple candidate sample credit data. The credit data of the multiple candidate samples are preprocessed to obtain the credit data of multiple target samples; The credit risk prediction model is trained based on the multiple target sample credit data and the risk label associated with each target sample credit data, and the SHAP value of at least one feature of each target sample credit data is determined. The credit risk prediction model shall be subjected to at least one of the following verifications: effectiveness verification, efficiency verification, fairness verification, and resilience verification.

[0069] Optionally, the step of performing multi-dimensional sample optimization on the multiple initial sample credit data to obtain multiple intermediate sample credit data includes: The initial sample credit data are screened based on business and / or compliance dimensions. Clustering methods are used to cluster credit data from multiple initial samples that have undergone sample screening. SMOTE oversampling, ENN outlier removal, and entropy weighting were performed on multiple initial sample credit data after clustering. Multiple initial sample credit data, after being weighted, are stratified and sampled to obtain a training set, a validation set, and a test set; the training set, the validation set, and the test set serve as the multiple intermediate sample credit data.

[0070] Optionally, the preprocessing of the multiple candidate sample credit data to obtain multiple target sample credit data includes: For each candidate sample credit data, perform at least one of the following feature preprocessing steps: smoothing, feature encoding, and missing value imputation. Perform at least one of the following operations on each candidate sample credit data that has undergone feature preprocessing: dynamic threshold filtering, redundant feature removal, and data compliance processing, to obtain multiple target sample credit data.

[0071] The credit risk prediction device includes a processor and a memory. The acquisition module 901, prediction module 902, etc. are all stored in the memory as program units. The processor executes the program units stored in the memory to realize the corresponding functions.

[0072] A processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured.

[0073] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0074] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as...Figure 10 As shown, the electronic device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communications bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communications bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute a credit risk prediction method, which includes: acquiring target credit data to be predicted; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; wherein the credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer includes multiple feature extraction sub-models; each feature extraction sub-model is used to extract features from the target credit data to obtain a feature extraction result; the fusion layer is used to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result; the credit risk prediction model is trained based on multiple target sample credit data and risk labels associated with each target sample credit data; the multiple target sample credit data are obtained by feature extraction of at least two of the following features: time-series features, text features, and interaction features, based on an initial sample credit dataset.

[0075] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a machine-readable storage medium. When the computer program is executed by a processor, the computer is able to execute a credit risk prediction method. The method includes: acquiring target credit data to be predicted; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; wherein the credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer includes multiple feature extraction sub-models; each feature extraction sub-model is used to extract features from the target credit data to obtain a feature extraction result; the fusion layer is used to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result; the credit risk prediction model is trained based on multiple target sample credit data and risk labels associated with each target sample credit data; the multiple target sample credit data are obtained by feature extraction of at least two of the following features: time-series features, text features, and interaction features, based on an initial sample credit dataset.

[0077] In another aspect, the present invention also provides a machine-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform a credit risk prediction method. The method includes: acquiring target credit data to be predicted; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; wherein the credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer includes multiple feature extraction sub-models; each feature extraction sub-model is used to extract features from the target credit data to obtain a feature extraction result; the fusion layer is used to fuse the feature extraction results of all feature extraction sub-models to obtain the credit risk prediction result; the credit risk prediction model is trained based on multiple target sample credit data and risk labels associated with each target sample credit data; the multiple target sample credit data are obtained by feature extraction from at least two of the following: temporal features, textual features, and interactive features, based on an initial sample credit dataset.

[0078] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A credit risk prediction method characterized by, The method comprises: obtaining target credit data to be predicted; inputting the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; wherein the credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer; the feature extraction model layer comprises a plurality of feature extraction sub-models; each feature extraction sub-model is used for feature extraction on the target credit data to obtain a feature extraction result; the fusion layer is used for fusing the feature extraction results of all the feature extraction sub-models to obtain the credit risk prediction result; the credit risk prediction model is trained based on a plurality of target sample credit data and a risk label associated with each target sample credit data; the plurality of target sample credit data are obtained based on initial sample credit data set through feature extraction on at least two of time sequence features, text features and interaction features.

2. The credit risk prediction method according to claim 1, characterized in that, The method comprises: inputting the target credit data into a plurality of feature extraction sub-models to obtain feature extraction results output by all the feature extraction sub-models; controlling a fusion layer to determine the credit risk prediction result based on the feature extraction results of all the feature extraction sub-models and a fusion coefficient corresponding to each feature extraction sub-model. 3.The credit risk prediction method according to claim 2, characterized in that, The plurality of feature extraction sub-models comprise a first tree model and a second tree model; the credit risk prediction result is determined based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model by the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2))); wherein P represents the credit risk prediction result; β0 represents a set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; and β2 represents the fusion coefficient corresponding to the second tree model. 4.The credit risk prediction method according to claim 2, characterized in that, The plurality of feature extraction sub-models comprise a first tree model, a second tree model and a third tree model; the credit risk prediction result is determined based on the feature extraction results of all the feature extraction sub-models and the fusion coefficient corresponding to each feature extraction sub-model by the following formula: P=1 / (1+e^(-(β0+β1z1+β2z2+β3z3))); wherein P represents the credit risk prediction result; β0 represents a set parameter; z1 represents the feature extraction result of the first tree model; β1 represents the fusion coefficient corresponding to the first tree model; z2 represents the feature extraction result of the second tree model; β2 represents the fusion coefficient corresponding to the second tree model; z3 represents the feature extraction result of the third tree model; and β3 represents the fusion coefficient corresponding to the third tree model. 5.The credit risk prediction method according to claim 1, characterized in that, The credit risk prediction model is trained by the following steps: obtaining an initial sample credit data set; the initial sample credit data set comprises a plurality of initial sample credit data; performing multi-dimensional sample optimization on the plurality of initial sample credit data to obtain a plurality of intermediate sample credit data; performing feature extraction on each of the intermediate sample credit data to obtain a plurality of candidate sample credit data; performing preprocessing on the plurality of candidate sample credit data to obtain a plurality of target sample credit data; training the credit risk prediction model based on the plurality of target sample credit data and the risk label associated with each of the target sample credit data, and determining the SHAP value of at least one feature of each of the target sample credit data; performing at least one of effect verification, efficiency verification, fairness verification and pressure resistance verification on the credit risk prediction model. 6.The credit risk prediction method according to claim 5, characterized in that, The multi-dimensional sample optimization on the plurality of initial sample credit data to obtain a plurality of intermediate sample credit data comprises: performing sample screening on the plurality of initial sample credit data based on business dimension and / or compliance dimension; performing clustering processing on the plurality of initial sample credit data subjected to sample screening based on a clustering method; performing SMOTE oversampling, ENN outlier removal and weight assignment based on entropy weight method on the plurality of initial sample credit data subjected to clustering processing; performing stratified sampling division on the plurality of initial sample credit data subjected to weight assignment to obtain a training set, a validation set and a test set; the training set, the validation set and the test set are used as the plurality of intermediate sample credit data.

7. The credit risk prediction method of claim 5, wherein, The preprocessing on the plurality of candidate sample credit data to obtain a plurality of target sample credit data comprises: performing at least one of smoothing processing, feature encoding and missing value supplement on each candidate sample credit data; performing at least one of dynamic threshold filtering, redundant feature removal and data compliance processing on each candidate sample credit data subjected to feature preprocessing to obtain a plurality of target sample credit data. 8.The credit risk prediction method according to claim 1 or 5, characterized in that, The method further comprises: generating an adjustment suggestion based on the credit risk prediction result and the credit risk prediction result in a historical time period; constructing a risk report based on the credit risk prediction result and the adjustment suggestion; outputting the risk report. 9.The credit risk prediction method according to claim 1 or 5, characterized in that, The method further comprises: monitoring the model performance of the credit risk prediction model; in the case that the model performance does not meet the set condition, pushing an alarm information, and positioning the root cause based on a model evaluation tool; merging the plurality of target sample credit data and the new sample credit data to construct a new data set; iteratively training the model parameters of the fusion layer of the credit risk prediction model based on the new data set.

10. A credit risk prediction apparatus characterized by comprising: comprises: an acquisition module, configured to acquire target credit data to be predicted; a prediction module, configured to input the target credit data into a credit risk prediction model to obtain a credit risk prediction result output by the credit risk prediction model; The credit risk prediction model is constructed based on a feature extraction model layer and a fusion layer. The feature extraction model layer includes multiple feature extraction sub-models. Each feature extraction sub-model is configured to perform feature extraction on the target credit data to obtain a feature extraction result. The fusion layer is configured to fuse the feature extraction results of all the feature extraction sub-models to obtain the credit risk prediction result. The credit risk prediction model is trained based on multiple target sample credit data and risk labels associated with each of the target sample credit data. The multiple target sample credit data are obtained by performing feature extraction on at least two of time sequence features, text features, and interaction features based on an initial sample credit data set.

11. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the credit risk prediction method of any one of claims 1 to 9 when executing the program.

12. A machine-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by the processor, implements the credit risk prediction method of any one of claims 1 to 9.

13. A computer program product comprising a computer program, characterized in that, The computer program, when executed by the processor, implements the credit risk prediction method of any one of claims 1 to 9. The computer program, when executed by the processor, implements the credit risk prediction method of any one of claims 1 to 9.

Images