Credit risk prediction method with enhanced interpretability

Abstract

The invention belongs to the technical field of credit prediction, and particularly relates to an interpretability-enhanced credit risk prediction method, which comprises the following steps of: acquiring credit-related characteristic data such as credit history, behavior characteristics and financial information of a customer, and preprocessing the characteristic data; screening effective features, and constructing derivative feature data based on the original feature data; dividing a training set and a test set, training the training set by using the integrated learning model, and evaluating model performance on the test set; inputting to-be-evaluated sample data, and outputting a credit risk prediction result; an SHAP method is introduced to carry out interpretability analysis on a model output result, and marginal contributions of feature data to credit risk scores are quantified; according to the method, the interpretability of the model is remarkably enhanced while the prediction accuracy is improved, the application risk of the black box model in financial risk control is reduced, and the method has good practical value and popularization prospect.

Description

Technical Field

[0001] This invention belongs to the field of credit forecasting technology, specifically relating to a credit risk forecasting method with enhanced interpretability. Background Technology

[0002] As the financial industry places increasing demands on risk management, more and more credit institutions are adopting machine learning algorithms for credit risk assessment. While traditional credit scoring methods (such as logistic regression and credit scoring models) offer some interpretability, their predictive performance is limited. In recent years, machine learning models (such as XGBoost and LightGBM) have been widely used in credit scoring due to their superior predictive performance. However, these "black box" models suffer from poor interpretability and opaque decision-making processes, making them difficult for business personnel, customers, and regulatory agencies to understand. This limits their application in compliance scenarios, particularly in credit review, risk monitoring, and regulatory audits. The poor interpretability and opaque decision-making processes of these "black box" models make it difficult to explain the basis for their individual credit risk scoring, leading to uncontrollable risks and insufficient compliance. Therefore, we propose a credit risk prediction method with enhanced interpretability. Summary of the Invention

[0003] The purpose of this invention is to provide a credit risk prediction method with enhanced interpretability to address the problems existing in the background art.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0005] A credit risk prediction method with enhanced interpretability, characterized by the following steps:

[0006] S1. Collect credit-related characteristic data such as customer credit history, behavioral characteristics, and financial information, and preprocess the characteristic data, using the user's default label as the target variable;

[0007] S2. Perform feature engineering on the preprocessed feature data, filter effective features, delete redundant and invalid features, and construct derived feature data based on the original feature data.

[0008] S3. Divide the feature data processed in the above steps into a training set and a test set. Use an ensemble learning model to train the training set and evaluate the model performance on the test set. The ensemble learning model includes, but is not limited to, XGBoost, LightGBM or CatBoost models.

[0009] S4. For the credit prediction model that has been trained and evaluated, input the sample data to be evaluated and output the credit risk prediction results.

[0010] S5. Introduce the SHAP method based on game theory to perform interpretability analysis of the model output at the feature data level, quantify the marginal contribution of each feature data to the credit risk score and generate visualization charts, thereby understanding the model's decision-making logic.

[0011] As a further explanation of the method of the present invention, the credit-related feature data in step S1 includes, but is not limited to, feature data such as age, gender, education level, income, credit history, debt ratio, repayment history, and whether there is a default.

[0012] As a further explanation of the method of the present invention, the preprocessing step in step S1 is specifically as follows:

[0013] S1.1 Fill missing values ​​in credit-related feature data with the mean or median;

[0014] S1.2 Remove outliers from credit-related feature data;

[0015] S1.3 Standardize the characteristics of continuous data to eliminate the influence of different units on the model.

[0016] As a further explanation of the method of the present invention, in the feature engineering process of step S2, the feature selection method includes information gain, chi-square test and SHAP feature importance ranking.

[0017] As a further explanation of the method of the present invention, step S3 further includes:

[0018] S3.1 Training / Test Set Partitioning: The dataset is divided into a training set (80%) and a test set (20%) using a stratified sampling method.

[0019] S3.2 Class Imbalance Handling: Since the proportion of fraudulent transactions in the data is extremely low, the SMOTE oversampling method is used to synthesize and amplify fraudulent samples in the training set to ensure that the ratio of the number of fraudulent samples to the number of non-fraudulent samples is consistent in both.

[0020] As a further explanation of the method of the present invention, step S4 also includes:

[0021] The performance of the trained credit prediction model is evaluated based on evaluation metrics, including but not limited to accuracy, recall, precision, and F1-score.

[0022] Accuracy represents the percentage of samples that are correctly predicted by the model in the total sample.

[0023] Recall represents the percentage of true positive samples that are correctly predicted by the model out of all true positive samples.

[0024] Precision represents the percentage of samples that are actually positive out of those predicted as positive by the model.

[0025] F1-score represents the harmonic mean of precision and recall.

[0026] As a further explanation of the method of the present invention, step S5 also includes:

[0027] The interpretability analysis of the model's prediction results mainly includes global interpretability and local interpretability;

[0028] in:

[0029] Global interpretability is used to explain the overall behavior of the model, specifically which data features have a significant impact on the prediction results of the credit model, thereby revealing the model's decision rules and providing a basis for risk control strategy formulation;

[0030] Local interpretability is used to explain the prediction results of a single sample, analyze the specific reasons why the sample is judged as high-risk or low-risk, quantify the marginal contribution of each input feature in the current prediction, and improve the transparency and understandability of the model at the individual level.

[0031] This invention proposes a comprehensive prediction scheme that integrates data preprocessing, feature engineering, model building, performance evaluation, and interpretation visualization. It solves the technical problem of the lack of interpretability in existing credit risk models, significantly enhances the interpretability of the model while improving prediction accuracy, and helps reduce the application risk of "black box" models in financial risk control. It has good practical value and prospects for promotion. Attached Figure Description

[0032] The present invention can be further illustrated by the non-limiting embodiments given in the accompanying drawings.

[0033] Figure 1 This is a schematic diagram of the process of the present invention;

[0034] Figure 2 This is a feature importance ranking diagram for global interpretability in embodiments of the present invention;

[0035] Figure 3 This is a feature distribution influence diagram illustrating the global interpretability of an embodiment of the present invention.

[0036] Figure 4 This is a feature dependency graph illustrating the global interpretability of an embodiment of the present invention.

[0037] Figure 5A waterfall diagram illustrating partial interpretability of an embodiment of the present invention;

[0038] Figure 6 This is a decision path diagram illustrating the local interpretability of an embodiment of the present invention;

[0039] Figure 7 This is an attempt to provide partial interpretability of embodiments of the present invention. Detailed Implementation

[0040] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0041] like Figure 1 As shown, a credit risk prediction method with enhanced interpretability includes the following steps:

[0042] S1. Collect credit-related characteristic data such as customer credit history, behavioral characteristics, and financial information, and preprocess the characteristic data, using the user's default label as the target variable;

[0043] In this step, credit-related characteristic data includes, but is not limited to, age, gender, education, income, credit history, debt ratio, repayment history, and whether there has been a default.

[0044] Meanwhile, the specific steps for preprocessing credit-related feature data in this step are as follows:

[0045] S1.1 Fill missing values ​​in credit-related feature data with the mean or median;

[0046] S1.2 Remove outliers from credit-related feature data;

[0047] S1.3 Standardize the characteristics of continuous data to eliminate the influence of different units on the model;

[0048] In step S1.3, the data feature annotation process shall use either of the following two methods, specifically:

[0049] ①Z-score standardization, the formula is as follows:

[0050]

[0051] Where x represents the original feature data; μ represents the mean of the original feature data; σ represents the standard deviation of the original feature data; and z represents the standardized value.

[0052] ②Min-Max normalization, the formula is as follows:

[0053]

[0054] Where x represents the original feature data; x min x max Let x represent the minimum and maximum values ​​of this feature, respectively; ' This represents the normalized value.

[0055] S2. Perform feature engineering on the preprocessed feature data, filter effective features, delete redundant and invalid features, and construct derived feature data based on the original feature data.

[0056] In this step, redundant feature data is removed using methods such as information gain, chi-square test, and correlation coefficient.

[0057] The construction of derived data features is based on the original feature data, and derived features are constructed according to business rules and data relevance, such as income to debt ratio, credit card usage rate, repayment ability ratio, etc.

[0058] S3. Divide the feature data processed in the above steps into a training set and a test set. Use an ensemble learning model to train the training set and evaluate the model performance on the test set.

[0059] In the steps, the ensemble learning model can be any one of XGBoost, LightGBM, or CatBoost;

[0060] This step also includes:

[0061] S3.1 Training / Test Set Partitioning: The dataset is divided into a training set (80%) and a test set (20%) using a stratified sampling method.

[0062] S3.2 Class Imbalance Handling: Since the proportion of fraudulent transactions in the data is extremely low, the SMOTE oversampling method is used to synthesize and amplify fraudulent samples in the training set to ensure that the ratio of the number of fraudulent samples to the number of non-fraudulent samples is consistent in both.

[0063] S4. For the credit prediction model that has been trained and evaluated, input the sample data to be evaluated and output the credit risk prediction results.

[0064] In this step, the trained credit prediction model needs to undergo performance evaluation. Only after the performance evaluation meets the standards will the model be used for prediction. The evaluation of the credit prediction model is specifically as follows:

[0065] The trained credit prediction model is evaluated on the test set based on evaluation metrics, including but not limited to accuracy, recall, precision, and F1-score.

[0066] Accuracy represents the percentage of samples that are correctly predicted by the model in the total sample.

[0067] Recall represents the percentage of true positive samples that are correctly predicted by the model out of all true positive samples.

[0068] Precision represents the percentage of samples that are actually positive out of those predicted as positive by the model.

[0069] F1-score represents the harmonic mean of precision and recall;

[0070] The specific calculation formula is as follows:

[0071]

[0072] in:

[0073] TP (TruePositive) represents the number of samples that are predicted as positive samples.

[0074] TN (TrueNegative) represents the number of samples that are predicted as negative samples.

[0075] FP(FalsePositive) represents the number of samples that are predicted as positive samples when they are negative samples;

[0076] FN(FalseNegative) represents the number of samples that are predicted as negative samples when they are positive samples.

[0077] S5. Construct a SHAP interpreter based on the trained credit risk prediction model. For the sample to be evaluated, calculate the SHAP value of each input feature, quantify the marginal contribution of each feature to the model prediction result, and complete the interpretability analysis accordingly.

[0078] In this step, the SHAP method employs an additive feature attribution model to perform feature-level interpretability analysis on the model's output. The calculation formula is as follows:

[0079]

[0080] in:

[0081] f(x) represents the predicted value of the model for the sample x to be predicted;

[0082] φ0 represents the baseline output of the model, i.e., the output when all features are omitted;

[0083] φ iThis represents the SHAP value of feature i in the sample to be predicted, i.e., the marginal contribution of this feature to the model's predicted value.

[0084] The contribution of each feature is quantified and the interpretation results are output using the SHAP method described above.

[0085] Simultaneously, this step involves interpretability analysis of the model's prediction results, primarily including global interpretability and local interpretability, where:

[0086] Global interpretability is used to explain the overall behavior of the model, specifically which data features have a significant impact on the prediction results of the credit model, thereby revealing the model's decision rules and providing a basis for risk control strategy formulation;

[0087] Local interpretability is used to explain the prediction results of a single sample, analyze the specific reasons why the sample is judged as high-risk or low-risk, quantify the marginal contribution of each input feature in the current prediction, and improve the transparency and understandability of the model at the individual level.

[0088] Example:

[0089] To verify the effectiveness of the method of the present invention, an implementation case is provided using the "Credit Card Fraud Detection" dataset (https: / / www.kaggle.com / datasets / mlg-ulb / creditcardfraud) published by Kaggle as an example. Combining the XGBoost model and the SHAP interpretability method, a complete implementation process and code are provided to illustrate the effectiveness and interpretability of the method of the present invention in credit risk prediction.

[0090] I. Experimental Environment and Tools:

[0091] Programming language: Python 3.10;

[0092] Core libraries: Pandas, Scikit-learn, LightGBM, SHAP;

[0093] Hardware environment: Intel i7 CPU, 16GB RAM;

[0094] Operating system: Windows 11 or macOS.

[0095] II. Dataset Introduction:

[0096] This dataset contains credit card transaction data from European cardholders in September 2013, totaling 284,807 transaction records, of which 492 were fraudulent transactions (approximately 0.172%), indicating a highly imbalanced data set. The raw data features have been preprocessed using PCA into anonymized variables V1–V28, and include Time (transaction time) and Amount (transaction amount) fields, with a label field for Class (whether it was a fraudulent transaction, 1 for yes, 0 for no).

[0097] The dataset is described in the table below:

[0098]

[0099] III. Implementation Steps

[0100] Step 1: Data Feature Preprocessing

[0101] Missing value handling: This dataset has no missing values;

[0102] Outlier handling: Since the data has been reduced in dimensionality using PCA, outliers are not directly removed but retained to improve outlier detection capabilities;

[0103] Training / Test Set Splitting: The data is divided into a training set (80%) and a test set (20%) using stratified sampling to ensure that the proportion of fraudulent and non-fraudulent samples is consistent in both sets.

[0104] Class Imbalance Handling: Since the proportion of fraudulent transactions in the data is extremely low, in order to solve the data imbalance problem, SMOTE (Synthetic Minority Oversampling) is used to synthesize and amplify fraudulent samples in the training set.

[0105] Step 2: Data Feature Filtering

[0106] The original features are standardized to ensure that features at different scales have the same distribution. Here, the recursive feature elimination (RFE) method is used to filter the feature subset, remove redundant or irrelevant variables, and improve model performance.

[0107] Step 3: Building a Credit Risk Prediction Model

[0108] An ensemble classification model is constructed, with XGBoost as the base classifier. Modules such as standardization, oversampling, and feature selection are integrated into the same training pipeline. The model hyperparameters are automatically optimized using a randomized search method, and the model's stability and generalization ability are ensured through five-fold cross-validation.

[0109] Step 4: Model Evaluation and Prediction:

[0110] Model evaluation:

[0111] The trained model is evaluated on the test set using a comprehensive assessment of multiple metrics, including accuracy, recall, precision, F1-score, and AUC. Accuracy represents the proportion of correctly predicted samples out of the total sample count. Recall represents the proportion of all true positive samples successfully identified by the model. Precision measures how many samples predicted as positive are actually positive. The F1-score is the harmonic mean of precision and recall, taking both into account. The formulas for each metric are shown below:

[0112]

[0113]

[0114] The evaluation results of the test set metrics are shown in the table below:

[0115]

[0116] The model performed excellently overall, with an accuracy of 1.00, indicating that the model was very reliable in overall prediction with few false positives and was able to distinguish between normal and fraudulent transactions well. The F1-Score was 0.92, and considering both precision and recall, the model showed good overall balance.

[0117] Model prediction:

[0118] When the trained credit risk prediction model is applied to a new single record, the input features of that record are first preprocessed in the same way as during training, including standardization and necessary feature filtering. Then, the processed feature vector is input into the model. Based on the learned parameters, the model calculates the probability that the record belongs to the default (or fraud) category, and finally outputs the default probability for the risk control system to perform risk assessment and decision support. Taking record 809 in the test set (a positive example) as an example, its predicted default probability is 96.24%, indicating an extremely high default risk, and the prediction is classified as "default," consistent with reality.

[0119] Step 5: Enhance model interpretability

[0120] The SHAP (SHapley Additive Explanations) framework is introduced to perform interpretability analysis on the model results, mainly including global interpretability and local interpretability.

[0121] Global interpretability is used to explain the overall behavior of the model, specifically which data features have the greatest impact on the prediction of the credit model, reveal the model's decision rules, and provide a basis for feature engineering and risk control strategies.

[0122] (1) Feature importance ranking chart:

[0123] The importance of features is displayed by their average SHAP absolute value, replacing the original "feature importance" of the model. This quantifies the average impact of each feature on the model output, and is presented in a bar chart, where each bar represents the overall importance of a feature. The feature importance ranking chart obtained in the implementation case is shown below. Figure 2 As shown in the figure, the horizontal axis represents the average influence of each feature on the model output. The larger the value, the more important it is for the prediction. The vertical axis represents the top ten feature variables. As can be seen from the figure, the most important features are V14, V4, V3, V12, V8, etc., indicating that the value changes of these features have a significant impact on fraud prediction. Features V14 and V4 are core variables, and their SHAP values ​​are much higher than those of other features. The model relies heavily on them when predicting whether a transaction is fraudulent.

[0124] (2) Feature distribution influence diagram:

[0125] Displaying the SHAP value distribution of each feature across different samples in scatter plot format can show how each feature affects the prediction in each sample, distinguishing between the directionality of "increasing risk" and "reducing risk," and helping to uncover the nonlinear relationship between complex features and the output. The color intensity indicates the feature value (dark: high, light: low), the horizontal axis represents the SHAP value, and the density represents the importance. The feature distribution influence plot obtained in the implementation case is shown below. Figure 3 As shown in the figure, the SHAP value of feature V14 has a large range, indicating that it has a significant impact on the model output at high / low values. Low values ​​tend to predict positively (fraud), while high values ​​tend to predict negatively (non-fraud). Feature V4 also has a significant impact, showing a typical two-sided expansion distribution. High values ​​tend to predict positively (fraud), while low values ​​tend to predict negatively (non-fraud). Features V3, V12, and V8 have moderate contributions and show a relatively symmetrical distribution, indicating that they may have positive / negative effects on fraud and non-fraud, reflecting the complex dependence of the model on these variables.

[0126] (3) Feature Dependency Graph

[0127] Displaying the SHAP value distribution of each data feature across different samples in scatter plot format can show how each feature affects the prediction in each sample, distinguishing between the directionality of "increasing risk" and "reducing risk," and helping to uncover the nonlinear relationship between complex features and the output. Colors represent feature values, the horizontal axis represents the SHAP value, and density indicates importance. The feature dependency plot of variable V14 obtained in the implementation case is shown below. Figure 4 As shown.

[0128] Local interpretability is used to explain the prediction results of a single sample, specifically the reasons why a sample is predicted as high-risk / low-risk.

[0129] (1) Waterfall picture:

[0130] This demonstrates how the model, starting from a baseline value, is progressively influenced by various features to determine the final predicted value. It clearly explains why a particular transaction is considered fraudulent, and whether each feature contributes positively or negatively, achieving automatic interpretation and audit compliance. The waterfall chart obtained in the implementation case is shown below. Figure 5 As shown;

[0131] (2) Decision path diagram:

[0132] The decision path diagram illustrates the path from the baseline predicted value to the final predicted value, showing the contribution of each feature. By observing the changes in the path, it's possible to understand how the model progressively builds predictions. This helps identify risk assessment patterns and individual differences among different individuals. The decision path diagram obtained in the implementation case is shown below. Figure 6 As shown;

[0133] (3) Trying to:

[0134] Visualizing the push-pull effect of each feature in a single sample on the prediction result, complex predictions are transformed into intuitive explanations. From a business perspective, it can analyze the distribution of risk drivers across a customer group, assist in stratification strategies, and explain the prediction results and rationale for each transaction to loan officers or users. The implementation case demonstrates this approach. Figure 7 As shown.

[0135] Meanwhile, the code for calculating the SHAP value in the SHAP interpreter of this application is as follows:

[0136] best_pipeline=rnd_cv.best_estimator_

[0137] best_model=best_pipeline.named_steps["model"]

[0138] original_feature_names=X.columns

[0139] X_test_scaled=best_pipeline.named_steps["standardscale"].transform(X_test)

[0140] selected_mask=best_pipeline.named_steps["feature_selction"].get_support()

[0141] selected_feature_names=original_feature_names[selected_mask]

[0142] X_test_selected=best_pipeline.named_steps["feature_selction"].transform(X_test_scaled)

[0143] explainer=shap.TreeExplainer(best_model)

[0144] shap_values=explainer.shap_values(X_test_selected);

[0145] The code for visualizing the average importance of features is as follows:

[0146] shap.plots.bar(shap.Explanation(values=shap_values,

[0147] base_values=explainer.expected_value,

[0148] data = X_test_selected,

[0149] feature_names=selected_feature_names));

[0150] The code for the feature distribution influence diagram is as follows:

[0151] shap.summary_plot(shap_values,

[0152] X_test_selected,

[0153] feature_names=selected_feature_names);

[0154] The feature dependency graph code is as follows:

[0155] shap.dependence_plot("Time",

[0156] shap_values,

[0157] X_test_selected,

[0158] feature_names=selected_feature_names);

[0159] The code for the waterfall chart is as follows:

[0160] shap.plots.waterfall(shap.Explanation(values=shap_values[0],

[0161] base_values=explainer.expected_value,

[0162] data = X_test_selected[0],

[0163] feature_names=selected_feature_names));

[0164] The decision path diagram code is as follows:

[0165] shap.decision_plot(

[0166] explainer.expected_value,

[0167] shap_values[:5], # Explain the first 5 samples

[0168] feature_names=list(selected_feature_names));

[0169] The code for the test is as follows:

[0170] shap.initjs()

[0171] shap.force_plot(explainer.expected_value,

[0172] shap_values[0],

[0173] X_test_selected[0],

[0174] feature_names=selected_feature_names).

[0175] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A credit risk prediction method with enhanced interpretability, characterized in that: Includes the following steps: S1. Collect credit-related characteristic data such as customer credit history, behavioral characteristics, and financial information, and preprocess the characteristic data, using the user's default label as the target variable; S2. Perform feature engineering on the preprocessed feature data, filter effective features, delete redundant and invalid features, and construct derived feature data based on the original feature data. S3. Divide the feature data processed in the above steps into a training set and a test set. Use an ensemble learning model to train the training set and evaluate the model performance on the test set. The ensemble learning model includes, but is not limited to, XGBoost, LightGBM or CatBoost models. S4. For the credit prediction model that has been trained and evaluated, input the sample data to be evaluated and output the credit risk prediction results. S5. Introduce the SHAP method based on game theory to perform interpretability analysis of the model output at the feature data level, quantify the marginal contribution of each feature data to the credit risk score and generate visualization charts, thereby understanding the model's decision-making logic.

2. The credit risk prediction method with enhanced interpretability according to claim 1, characterized in that: The credit-related feature data in step S1 includes, but is not limited to, features such as age, gender, education level, income, credit history, debt ratio, repayment history, and whether there has been a default.

3. The credit risk prediction method with enhanced interpretability according to claim 2, characterized in that: The preprocessing step in step S1 is specifically as follows: S1.1 Fill missing values ​​in credit-related feature data with the mean or median; S1.2 Remove outliers from credit-related feature data; S1.3 Standardize the characteristics of continuous data to eliminate the influence of different units on the model.

4. The credit risk prediction method with enhanced interpretability according to claim 1, characterized in that: In the feature engineering process of step S2, the feature selection methods include information gain, chi-square test and SHAP feature importance ranking.

5. The credit risk prediction method with enhanced interpretability according to claim 1, characterized in that: Step S3 specifically involves: S3.1 Training / Test Set Partitioning: The dataset is divided into a training set (80%) and a test set (20%) using a stratified sampling method. S3.2 Class Imbalance Handling: Since the proportion of fraudulent transactions in the data is extremely low, the SMOTE oversampling method is used to synthesize and amplify fraudulent samples in the training set to ensure that the ratio of the number of fraudulent samples to the number of non-fraudulent samples is consistent in both.

6. The credit risk prediction method with enhanced interpretability according to claim 1, characterized in that: Step S4 also includes: The performance of the trained credit prediction model is evaluated based on evaluation metrics, including but not limited to accuracy, recall, precision, and F1-score. Accuracy represents the percentage of samples that are correctly predicted by the model in the total sample. Recall represents the percentage of true positive samples that are correctly predicted by the model out of all true positive samples. Precision represents the percentage of samples that are actually positive out of those predicted as positive by the model. F1-score represents the harmonic mean of precision and recall.

7. The credit risk prediction method with enhanced interpretability according to claim 5, characterized in that: The formulas for calculating accuracy, recall, precision, and F1 score are as follows: in: TP (True Positive) represents the number of samples that are predicted as positive samples. TN (True Negative) represents the number of samples that are predicted as negative samples. FP(False Positive) represents the number of samples that are predicted as positive samples when they are negative samples; FN(False Negative) represents the number of samples that are predicted as negative samples when they are positive samples.

8. The credit risk prediction method with enhanced interpretability according to claim 1, characterized in that: In step S5, the SHAP method employs an additive feature attribution model to perform feature-level interpretability analysis on the model's output. The calculation formula is as follows: in: f(x) represents the predicted value of the model for the sample x to be predicted; φ0 represents the baseline output of the model, i.e., the output when all features are omitted; φ i This represents the SHAP value of feature i in the sample to be predicted, i.e., the marginal contribution of this feature to the model's predicted value. The contribution of each feature is quantified using the SHAP method, and the interpretation results are output.

9. The credit risk prediction method with enhanced interpretability according to claim 7, characterized in that: Step S5 also includes: The interpretability analysis of the model's prediction results mainly includes global interpretability and local interpretability; in: Global interpretability is used to explain the overall behavior of the model, specifically which data features have a significant impact on the prediction results of the credit model, thereby revealing the model's decision rules and providing a basis for risk control strategy formulation; Local interpretability is used to explain the prediction results of a single sample, analyze the specific reasons why the sample is judged as high-risk or low-risk, quantify the marginal contribution of each input feature in the current prediction, and improve the transparency and understandability of the model at the individual level.

Images