CN117473325A - Sample equalization method, device, electronic equipment and computer-readable storage medium - Google Patents

Abstract

The application provides a sample equalization method, a sample equalization device, electronic equipment and a computer readable storage medium, wherein the method comprises the following steps: predicting, for each training sample in the training sample set, a default probability for the training sample; according to the default probability of each training sample, training samples with the default probability larger than a first threshold value or smaller than a second threshold value are screened out from the training sample set, a first training sample set is constructed by using the screened training samples, and a second training sample set is constructed by using the rest training samples; the first threshold is greater than the second threshold; sample equalization processing is carried out on the first training sample set, and a third training sample set is obtained; the third training sample set comprises each training sample in the first training sample set and new training samples generated after sample expansion; and carrying out sample set merging on the third training sample set and the second training sample set to obtain a target training sample set. By the method, accuracy of the prediction model is improved.

Description

Sample equalization method, device, electronic device and computer readable storage medium

Technical Field

The present application relates to the field of computer technology, and in particular to a sample equalization method, device, electronic device, and computer-readable storage medium.

Background Art

In risk assessment, predicting the borrower's default probability is to accurately assess the borrower's risk and formulate appropriate risk control strategies. At present, the prediction of default probability relies more on traditional prediction models. In the process of training the prediction model, due to the small number of default samples and the large number of non-default samples, the training samples are unbalanced, which will lead to the low accuracy of the prediction model trained using the training samples.

In the prior art, SMOTE technology is usually used to directly perform sample balancing on training samples to expand the number of default samples. However, if sample balancing is directly performed on training samples, it is easy to cause the expanded default samples to be inaccurate. For example, there are samples with a default probability of 0.5 in the training samples (that is, the default situation of the training samples is not clear enough). If the training samples are expanded, the default situation of the expanded new samples will be inaccurate. If the accuracy of the new samples obtained after expansion is low, it will affect the training of the prediction model, resulting in an increase in the probability of misjudgment of the trained prediction model, affecting the accuracy of the prediction model.

Summary of the invention

In view of this, the purpose of the present application is to provide a sample balancing method, device, electronic device and computer-readable storage medium to improve the accuracy of the prediction model.

In a first aspect, an embodiment of the present application provides a sample equalization method, including:

For each training sample in a training sample set, predict the default probability of the training sample; wherein the training sample set includes a plurality of training samples; the sample types of the training samples include default samples and non-default samples; the sample number of the default samples is less than the sample number of the non-default samples;

According to the default probability of each of the training samples, training samples whose default probability is greater than a first threshold or less than a second threshold are screened out from the training sample set, so as to construct a first training sample set using the screened training samples, and to construct a second training sample set using the remaining training samples; the first threshold is greater than the second threshold;

Performing sample balancing processing on the first training sample set to expand the training samples whose sample types are the default samples in the first training sample set to obtain a third training sample set; the third training sample set includes each training sample in the first training sample set and new training samples generated after sample expansion;

The third training sample set and the second training sample set are merged to obtain a target training sample set.

In combination with the first aspect, the embodiment of the present application provides a first possible implementation of the first aspect, wherein, for each training sample in the training sample set, predicting the probability of default of the training sample includes:

For each training sample in the training sample set, a first default probability prediction model that has been trained is used to predict a first default probability of the training sample, and a second default probability prediction model that has been trained is used to predict a second default probability of the training sample; the default probability of the training sample includes the first default probability and the second default probability of the training sample.

In combination with the first possible implementation of the first aspect, the embodiment of the present application provides a second possible implementation of the first aspect, wherein the first default probability prediction model is trained in the following manner:

For each training sample in the training sample set, the training sample is input into a first initial default probability prediction model, a default situation of the training sample is predicted by the first initial default probability prediction model, and a first default probability of the training sample is output;

Inputting the default label of the training sample and the first default probability into a loss function to calculate the loss value of the first initial default probability prediction model;

When the loss value is greater than the preset loss value, the trainable parameters in the first initial default probability prediction model are updated according to the loss value; when the loss value is not greater than the preset loss value, the current first initial default probability prediction model is used as the first default probability prediction model.

In combination with the first possible implementation of the first aspect, the embodiment of the present application provides a third possible implementation of the first aspect, wherein the second default probability prediction model includes multiple decision trees, each decision tree includes multiple classification nodes; and the second default probability prediction model is trained in the following manner:

For each initial decision tree in the second initial default probability prediction model, each training sample in the training sample set is input into the initial decision tree to obtain a node parameter corresponding to each classification node in the initial decision tree; the node parameter is the ratio of the number of samples whose sample category is the default sample on the classification node to the total number of training samples on the classification node;

Calculating the impurity of the initial decision tree according to the node parameters corresponding to each classification node in the initial decision tree;

When the impurity is greater than the preset impurity, optimizing the structure of the initial decision tree; when the impurity is not greater than the preset impurity, determining the current initial decision tree as the decision tree after training;

When all initial decision trees in the second initial default probability prediction model are trained, the second default probability prediction model is obtained.

In combination with the first possible implementation manner of the first aspect, the embodiment of the present application provides a fourth possible implementation manner of the first aspect, wherein, according to the default probability of each of the training samples, the training samples having a default probability greater than a first threshold or less than a second threshold are screened out from the training sample set, and the first training sample set is constructed using the screened training samples, and the second training sample set is constructed using the remaining training samples, including:

Determining a first threshold and a second threshold corresponding to the first default probability prediction model according to the first default probability of each training sample;

Screening out training samples whose first default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a fourth training sample set, and using the remaining training samples to construct a fifth training sample set;

Determining a first threshold and a second threshold corresponding to the second default probability prediction model according to the second default probability of each training sample;

Screening out training samples whose second default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a sixth training sample set, and using the remaining training samples to construct a seventh training sample set; wherein the second training sample set includes the fifth training sample set and the seventh training sample set;

An intersection of the fourth training sample set and the sixth training sample set is taken as the first training sample set.

In combination with the fourth possible implementation of the first aspect, the embodiment of the present application provides a fifth possible implementation of the first aspect, wherein the performing sample balancing processing on the first training sample set to perform sample expansion on the training samples whose sample types are the default samples in the first training sample set to obtain a third training sample set includes:

For each training sample whose sample type is the default sample in the first training sample set, calculating the Euclidean distance between the training sample and other training samples; the other training samples are training samples whose sample type is the default sample other than the training sample;

According to the Euclidean distance between the training sample and other training samples, a preset number of other training samples are selected from other training samples in the order of the Euclidean distance from small to large as similar training samples of the training sample;

Selecting a target similar training sample from the similar training samples, and generating a synthetic training sample of the training sample by using the target similar training sample and the training sample;

The third training sample set is constructed according to the synthetic training samples corresponding to each training sample whose sample type is the default sample in the first training sample set and each training sample in the first training sample set.

In combination with any possible implementation manner of the first aspect to the fifth possible implementation manner, the embodiment of the present application provides a sixth possible implementation manner of the first aspect, wherein the method further includes:

The target training sample set is used to train an initial default probability prediction model to obtain a trained target default probability prediction model.

In a second aspect, an embodiment of the present application further provides a sample equalization device, including:

A prediction module, for predicting the default probability of each training sample in a training sample set; wherein the training sample set includes a plurality of training samples; the sample types of the training samples include default samples and non-default samples; the sample number of the default samples is less than the sample number of the non-default samples;

a screening module, configured to screen out training samples whose default probability is greater than a first threshold or less than a second threshold from the training sample set according to the default probability of each training sample, so as to construct a first training sample set using the screened training samples and to construct a second training sample set using the remaining training samples; the first threshold is greater than the second threshold;

a sample balancing module, configured to perform sample balancing processing on the first training sample set, so as to expand the training samples whose sample types are the default samples in the first training sample set, so as to obtain a third training sample set; the third training sample set includes each training sample in the first training sample set and new training samples generated after sample expansion;

The merging module is used to merge the third training sample set and the second training sample set to obtain a target training sample set.

In combination with the second aspect, the embodiment of the present application provides a first possible implementation of the second aspect, wherein the prediction module, when used to predict the default probability of each training sample in the training sample set, is specifically used to:

For each training sample in the training sample set, a first default probability prediction model that has been trained is used to predict a first default probability of the training sample, and a second default probability prediction model that has been trained is used to predict a second default probability of the training sample; the default probability of the training sample includes the first default probability and the second default probability of the training sample.

In combination with the first possible implementation of the second aspect, the embodiment of the present application provides a second possible implementation of the second aspect, wherein the device further includes a first training module; the first training module is used to train the first default probability prediction model in the following manner:

For each training sample in the training sample set, the training sample is input into a first initial default probability prediction model, a default situation of the training sample is predicted by the first initial default probability prediction model, and a first default probability of the training sample is output;

Inputting the default label of the training sample and the first default probability into a loss function to calculate the loss value of the first initial default probability prediction model;

When the loss value is greater than the preset loss value, the trainable parameters in the first initial default probability prediction model are updated according to the loss value; when the loss value is not greater than the preset loss value, the current first initial default probability prediction model is used as the first default probability prediction model.

In combination with the first possible implementation of the second aspect, the embodiment of the present application provides a third possible implementation of the second aspect, wherein the second default probability prediction model includes multiple decision trees, each decision tree includes multiple classification nodes; the device also includes a second training module; the second training module is used to train the second default probability prediction model in the following manner:

For each initial decision tree in the second initial default probability prediction model, each training sample in the training sample set is input into the initial decision tree to obtain a node parameter corresponding to each classification node in the initial decision tree; the node parameter is the ratio of the number of samples whose sample category is the default sample on the classification node to the total number of training samples on the classification node;

Calculating the impurity of the initial decision tree according to the node parameters corresponding to each classification node in the initial decision tree;

When the impurity is greater than the preset impurity, optimizing the structure of the initial decision tree; when the impurity is not greater than the preset impurity, determining the current initial decision tree as the decision tree after training;

When all initial decision trees in the second initial default probability prediction model are trained, the second default probability prediction model is obtained.

In combination with the first possible implementation manner of the second aspect, the embodiment of the present application provides a fourth possible implementation manner of the second aspect, wherein the screening module is used to screen out training samples having a default probability greater than a first threshold or less than a second threshold from the training sample set according to the default probability of each training sample, so as to construct a first training sample set using the screened training samples and construct a second training sample set using the remaining training samples, and is specifically used to:

Determining a first threshold and a second threshold corresponding to the first default probability prediction model according to the first default probability of each training sample;

Screening out training samples whose first default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a fourth training sample set, and using the remaining training samples to construct a fifth training sample set;

Determining a first threshold and a second threshold corresponding to the second default probability prediction model according to the second default probability of each training sample;

Screening out training samples whose second default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a sixth training sample set, and using the remaining training samples to construct a seventh training sample set; wherein the second training sample set includes the fifth training sample set and the seventh training sample set;

An intersection of the fourth training sample set and the sixth training sample set is taken as the first training sample set.

In combination with the fourth possible implementation of the second aspect, the embodiment of the present application provides a fifth possible implementation of the second aspect, wherein the sample balancing module, when used to perform sample balancing processing on the first training sample set to perform sample expansion on the training samples whose sample type is the default sample in the first training sample set to obtain the third training sample set, is specifically used to:

For each training sample whose sample type is the default sample in the first training sample set, calculating the Euclidean distance between the training sample and other training samples; the other training samples are training samples whose sample type is the default sample other than the training sample;

According to the Euclidean distance between the training sample and other training samples, a preset number of other training samples are selected from other training samples in the order of the Euclidean distance from small to large as similar training samples of the training sample;

Selecting a target similar training sample from the similar training samples, and generating a synthetic training sample of the training sample by using the target similar training sample and the training sample;

The third training sample set is constructed according to the synthetic training samples corresponding to each training sample whose sample type is the default sample in the first training sample set and each training sample in the first training sample set.

In combination with any possible implementation of the first to fourth possible implementations of the second aspect, the embodiment of the present application provides a sixth possible implementation of the second aspect, wherein the device further includes:

The third training module is used to train the initial default probability prediction model using the target training sample set to obtain a target default probability prediction model after training.

In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps in any possible implementation of the first aspect above are performed.

In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any possible implementation of the first aspect described above are executed.

The sample balancing method, device, electronic device and computer-readable storage medium provided by the embodiment of the present application, wherein, according to the default probability of each training sample in the training sample set, the training samples with the default probability greater than the first threshold or less than the second threshold are screened out from the training sample set, and the screened training samples are used to construct the first training sample set, and the remaining training samples are used to construct the second training sample set. Wherein, since the first threshold is greater than the second threshold, the first training sample set contains high probability default training samples and low probability default training samples. In this embodiment, by performing sample balancing processing on the first training sample set, that is, performing sample balancing processing on the high probability default training samples and the low probability default training samples, the default situation of the generated new training samples can be made clearer, that is, the generated new training samples are made more accurate. Therefore, the sample balancing method of this embodiment is conducive to making the generated new training samples more accurate, so that when the training samples are used to train the prediction model, it is conducive to reducing the probability of misjudgment of the prediction model and improving the accuracy of the prediction model.

In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

FIG1 shows a flow chart of a sample balancing method provided in an embodiment of the present application;

FIG2 is a schematic diagram showing some classification nodes in the initial decision tree provided in an embodiment of the present application;

FIG3 shows a schematic structural diagram of a sample equalization device provided in an embodiment of the present application;

FIG4 shows a schematic structural diagram of an electronic device provided in an embodiment of the present application.

DETAILED DESCRIPTION

To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

Considering that if the training samples in the original training sample set are directly subjected to sample balancing processing, it is easy to cause the expanded default samples to be inaccurate. Based on this, the embodiments of the present application provide a sample balancing method, device, electronic device and computer-readable storage medium to improve the accuracy of the generated new training samples, thereby reducing the probability of misjudgment of the prediction model and improving the accuracy of the prediction model, which is described below through embodiments.

Embodiment 1:

To facilitate understanding of this embodiment, a sample balancing method disclosed in the embodiment of the present application is first introduced in detail. FIG1 shows a flow chart of a sample balancing method provided in the embodiment of the present application, as shown in FIG1 , including the following steps S101-S104:

S101: For each training sample in a training sample set, predict the default probability of the training sample; wherein the training sample set includes multiple training samples; the sample types of the training samples include default samples and non-default samples; and the sample number of the default samples is less than the sample number of the non-default samples.

In this embodiment, the training samples include basic customer information, product information, customer behavior information, and historical default information. Among them, the basic customer information includes, for example, the customer's ID number, age, gender, and other information. The product can be a loan product borrowed by the customer, such as a provident fund loan. Product information includes, for example, the balance of the loan product, the deposit base, the withdrawal amount, etc. Customer behavior information includes, for example, the customer's withdrawal record, consumption record, transfer record, and other operational behaviors related to the loan.

Historical default information refers to whether a customer has ever defaulted at a historical moment. For example, if a customer has borrowed multiple loan products, or has borrowed a loan product multiple times, then as long as the customer has at least one default event (such as failure to repay on time), the customer is a default customer, and the training sample corresponding to the customer is a default sample; if the customer has never defaulted, the customer is a non-default customer, and the training sample corresponding to the customer is a non-default sample. Each customer corresponds to one training sample.

The default probability refers to the probability that the customer corresponding to the training sample has defaulted. The greater the default probability, the greater the probability that the customer in the training sample has defaulted; the smaller the default probability, the smaller the probability that the customer in the training sample has defaulted.

The training sample set includes multiple default samples (the sample type is training samples of default samples) and multiple non-default samples (the sample type is training samples of non-default samples). Among them, the sample number of default samples is less than the sample number of non-default samples. Therefore, there is a sample imbalance problem in the training samples in the training sample set.

S102: According to the default probability of each training sample, select training samples whose default probability is greater than a first threshold or less than a second threshold from the training sample set, use the selected training samples to construct a first training sample set, and use the remaining training samples to construct a second training sample set; the first threshold is greater than the second threshold.

In this embodiment, each training sample corresponds to a default probability. According to the default probability corresponding to each training sample, training samples with a default probability greater than a first threshold are screened out from the training sample set, and training samples with a default probability less than a second threshold are screened out from the training sample set. Since the first threshold is greater than the second threshold, the screened out training samples are high-probability default training samples and low-probability default training samples. The screened out training samples are used to construct the first training sample set. That is, the high-probability default training samples and the low-probability default training samples are used to construct the first training sample set. Then, the first training sample set only includes high-probability default training samples and low-probability default training samples. The remaining training samples are used to construct the second training sample set. The second training sample set includes medium-probability default training samples. The medium-probability default training samples are, for example, training samples with a default probability of 0.5.

S103: Perform sample balancing processing on the first training sample set to expand the training samples whose sample types are default samples in the first training sample set to obtain a third training sample set; the third training sample set includes each training sample in the first training sample set and new training samples generated after sample expansion.

In this embodiment, sample balancing is performed only on the first training sample set, specifically, sample expansion is performed on the training samples whose sample type is default samples in the first training sample set to obtain new training samples whose sample type is default samples. The third training sample set is constructed using the newly generated training samples and each training sample in the first training sample set. The third training sample set is the training sample set after sample balancing is performed on the first training sample set.

S104: Merge the third training sample set and the second training sample set to obtain a target training sample set.

Each training sample in the third training sample set is combined with each training sample in the second training sample set to obtain a target training sample set. Since the third training sample set includes newly generated training samples whose sample type is default samples, the number of default samples in the target training sample set is greater than the number of default samples in the training sample set. In other words, the difference between the number of default samples and the number of non-default samples in the target training sample set is less than the difference between the number of default samples and the number of non-default samples in the training sample set.

In a possible implementation manner, before executing step S101, the following steps may also be performed:

S1001: Obtain the original data set from the MySQL database; the original data set includes multiple original training samples whose sample types are default samples and multiple original training samples whose sample types are non-default samples; the number of default samples is less than the number of non-default samples; each original training sample includes basic customer information, product information, customer behavior information, and historical default information.

S1002: For each original training sample in the original data set, preprocess the original training sample to convert the text fields in the original training sample into a digital language recognized by machine learning, thereby obtaining a preprocessed original training sample.

S1003: Perform feature engineering processing on the preprocessed original training samples to obtain training samples after feature engineering processing; the feature engineering processing includes any one or more of the following: missing value filling, outlier processing, indicator derivation, variable screening, and variable correlation analysis.

Among them, for missing value filling, when there are missing values in the training sample, the variable type corresponding to the missing value can be determined first, and then the median or mode of the variable values in other training samples under the variable type can be used to fill the missing value. For example, when age is missing in training sample A, it is determined that the variable type corresponding to the missing value is age, and then the median or mode of age of other training samples is calculated, and the median or mode is used to fill the missing value of training sample A.

For outlier processing, for each variable type in the training sample, determine whether the variable value corresponding to the variable type is greater than the first variable threshold or less than the second variable threshold. If there is at least one variable type in the training sample whose variable value is greater than the first variable threshold or less than the second variable threshold, then the variable value is regarded as an outlier of the training sample; when there is an outlier in the training sample, the training sample is deleted from the original data set. Among them, the first variable threshold is greater than the second variable threshold.

For indicator derivation, at least one new variable type is derived from at least two variable types, so that each training sample in the original data set has a new variable type. For example, a new variable type "difference in telephone charges" can be generated based on the variable type "this month's telephone charges" and the variable type "last month's telephone charges".

For variable screening, each training sample in the original data set contains multiple variable types. According to the importance of each variable type, variable types whose importance is greater than a preset importance threshold are screened out from the multiple variable types, and variable types whose importance is not greater than the preset importance threshold are deleted from each training sample in the original data set.

For variable correlation analysis, for any two variable types, the correlation value between the two variable types is calculated according to the variable values of the two variable types corresponding to each training sample. When the correlation value between the two variable types is greater than the preset correlation threshold, any one of the two variable types is deleted. By eliminating one of the two variable types with high correlation, it is helpful to avoid affecting the accuracy of the model during model training.

S1004: After each original training sample in the original data set has been preprocessed and feature-engineered, an original sample set including a plurality of training samples that have been preprocessed and feature-engineered is obtained.

S1005: Randomly divide each training sample in the original sample set according to a preset ratio to obtain a training sample set and a test sample set. For example, the preset ratio is 7:3.

In a possible implementation manner, when performing step S101 to predict the default probability of each training sample in the training sample set, the following steps may be specifically performed:

S1011: For each training sample in the training sample set, use the trained first default probability prediction model to predict the first default probability of the training sample, and use the trained second default probability prediction model to predict the second default probability of the training sample; the default probability of the training sample includes the first default probability and the second default probability of the training sample.

In this embodiment, the first default probability prediction model may specifically be an XGBOOST model trained using each training sample in the training sample set; the second default probability prediction model may specifically be an RF model (random forest model) trained using each training sample in the training sample set.

In a possible implementation manner, the first default probability prediction model is trained by following the steps S1012-S1014:

S1012: for each training sample in the training sample set, input the training sample into a first initial default probability prediction model, predict the default situation of the training sample by using the first initial default probability prediction model, and output a first default probability of the training sample;

In this embodiment, steps S1012-S1014 are located before step S1011 and after step S1005. That is, after dividing the training sample set, the first initial default probability prediction model is trained using each training sample in the training sample set to obtain the first default probability prediction model after training, and then the trained first default probability prediction model is used to perform the first default probability prediction.

The first initial default probability prediction model may be an initial XGBOOST model. In this embodiment, the initial XGBOOST model may be characterized by the following function:

Among them, Obj ^(t) represents the predicted first default probability; n represents the number of training samples; _xi represents the i-th training sample input into the first initial default probability prediction model; _gi and _hi are trainable parameters; the XGBOOST model is composed of multiple decision trees, that is, the first initial default probability prediction model is composed of multiple decision trees, _ft represents the t-th decision tree; Ω is a regular term that suppresses the complexity of the model.

The first default probability represents the probability that the training sample is a default sample.

S1013: Input the default label and the first default probability of the training sample into the loss function to calculate the loss value of the first initial default probability prediction model.

In this embodiment, the loss function is as follows:

Where L represents the loss value, _yi represents the default label of training sample i, and the default label is divided into 0 and 1, where 0 represents no default and 1 represents default; represents the first default probability of training sample i;

In this embodiment, _gi is the first-order derivative of l, and _hi is the second-order derivative of l.

S1014: When the loss value is greater than the preset loss value, the trainable parameters in the first initial default probability prediction model are updated according to the loss value; when the loss value is not greater than the preset loss value, the current first initial default probability prediction model is used as the first default probability prediction model.

In this embodiment, it is determined whether the loss value is greater than the preset loss value. When the loss value is greater than the preset loss value, the trainable parameters g _i and h _i in the first initial default probability prediction model are updated according to the loss value to obtain a new first initial default probability prediction model, and then steps S1012-S1014 are re-executed until the loss value is not greater than the preset loss value, then the training is stopped, and the first initial default probability prediction model at the current moment is used as the first default probability prediction model.

In a possible implementation, the second default probability prediction model includes multiple decision trees, each decision tree includes multiple classification nodes; the second default probability prediction model is trained in the following manner:

S1015: For each initial decision tree in the second initial default probability prediction model, each training sample in the training sample set is input into the initial decision tree to obtain the node parameter corresponding to each classification node in the initial decision tree; the node parameter is the ratio of the number of samples whose sample category is default samples on the classification node to the total number of training samples on the classification node.

In this embodiment, steps S1015-S1018 are located before step S1011 and after step S1005. That is, after dividing the training sample set, the second initial default probability prediction model is trained using each training sample in the training sample set to obtain the trained second default probability prediction model, and then the trained second default probability prediction model is used to perform the second default probability prediction.

Among them, the second initial default probability prediction model can be an initial RF model (random forest model). The second initial default probability prediction model contains multiple initial decision trees to be trained, each of which contains multiple classification nodes. The classification components on the classification nodes are determined according to the variable type in the training sample. For example, when the variable type is gender, the corresponding classification components are male and female, that is, the corresponding classification nodes are classification nodes "male" and classification nodes "female". Figure 2 shows a schematic diagram of some classification nodes in the initial decision tree provided in an embodiment of the present application.

As shown in FIG2 , for example, taking the classification node “male” as an example, after inputting each training sample in the training sample set into the initial decision tree, the initial decision tree will classify the training samples according to each classification node, for example, classifying the male training samples in the training samples into the classification node “male”, and classifying the female training samples in the training samples into the classification node “female”. Assume that the number of training samples in the classification node “male” is 20. Next, if the next classification node of the classification node “male” is the classification node “age range 20-40” and the classification node “age range 40-60”, the 20 training samples are further classified according to the age of the 20 training samples on the classification node “male”. For example, 15 training samples are obtained on the classification node “age range 20-40”, and 5 training samples are obtained on the classification node “age range 40-60”.

For the classification node "Male", the total number of training samples in the classification node "Male" is 20. If 5 training samples in the classification node "Male" are default samples, then the node parameter of the classification node "Male" is 5÷20=0.25.

S1016: Calculate the impurity of the initial decision tree according to the node parameters corresponding to each classification node in the initial decision tree.

In this embodiment, the RF model is composed of multiple decision trees, which perform classification (regression) tasks by voting or averaging. The random forest model does not directly use a single objective function or loss function, but optimizes the task through the construction and integration process of decision trees. In the random forest model, the training process of each decision tree will minimize impurity to achieve better classification (regression) results. Here, impurity is usually measured using indicators such as Gini impurity or entropy;

In this embodiment, when Gini impurity is used as the impurity indicator, the impurity of the initial decision tree can be calculated by the following formula:

Among them, Gini(P) represents the Gini impurity of the initial decision tree, j represents the j-th classification node in the initial decision tree, k represents the number of classification nodes in the initial decision tree, and _Pj represents the node parameter of the j-th classification node.

When entropy is used as the impurity indicator, the impurity of the initial decision tree can be calculated by the following formula:

Where Entropy(P) represents the entropy of the initial decision tree, j represents the j-th classification node in the initial decision tree, k represents the number of classification nodes in the initial decision tree, and _Pj represents the node parameter of the j-th classification node.

S1017: When the impurity is greater than the preset impurity, optimizing the structure of the initial decision tree; when the impurity is not greater than the preset impurity, determining the current initial decision tree as the decision tree after training.

In this embodiment, it is determined whether the impurity of the initial decision tree is greater than the preset impurity. If the impurity of the initial decision tree is greater than the preset impurity, it means that the initial decision tree needs to be further optimized. At this time, the structure of the initial decision tree is optimized, and the optimized initial decision tree is used as the new initial decision tree. Steps S1015-S1017 are continued until the impurity of the initial decision tree is no greater than the preset impurity, and the current initial decision tree is determined as the decision tree after training.

S1018: When all initial decision trees in the second initial default probability prediction model are trained, a second default probability prediction model is obtained.

In this embodiment, the second default probability prediction model includes a plurality of trained decision trees.

In a possible implementation, when executing step S102, according to the default probability of each training sample, training samples whose default probability is greater than a first threshold or less than a second threshold are screened out from the training sample set, so as to construct a first training sample set using the screened training samples, and to construct a second training sample set using the remaining training samples, the following steps S1021-S1025 may be specifically performed:

S1021: Determine a first threshold and a second threshold corresponding to a first default probability prediction model according to the first default probability of each training sample.

In this embodiment, the second threshold corresponding to the first default probability prediction model is calculated in the following manner: for the first default probability of each training sample in the training sample set, the first default probability is sorted in ascending order, and the quarter digit Q1 is calculated:

Where m represents the number of first default probabilities.

When Q1 is an integer, the Q1th first default probability is used as the second threshold corresponding to the first default probability prediction model. Exemplarily, when m is 11, Q1 is 3, at this time, the third first default probability is used as the second threshold corresponding to the first default probability prediction model.

When Q1 is not an integer, the second threshold corresponding to the first default probability prediction model can be calculated according to the interpolation method. Specifically, according to the integer part z of Q1, Q1 is located between z and z+1. Therefore, the second threshold is located between the first default probability corresponding to the zth position and the first default probability corresponding to the z+1th position. The interpolation calculation formula is as follows:

C _Q1 = (1-frac(Q1))×data(z)+frac(Q1)×data(z+1)

Among them, C _Q1 represents the second threshold corresponding to the first default probability prediction model, frac(Q1) is the fractional part of Q1, data(z) is the first default probability at the z-th position, and data(z+1) is the first default probability at the z+1-th position.

For example, when m=10, Q1=2.75, at this time, z=2, z+1=3. frac(Q1)=0.75, assuming data(z)=0.2, data(z+1)=0.3, through the above formula, C _Q1 =0.275 can be calculated, at this time the second threshold is 0.275.

In this embodiment, the first threshold corresponding to the first default probability prediction model is calculated in the following manner: for the first default probability of each training sample in the training sample set, the first default probability is sorted in ascending order, and the three-quarter digit Q2 is calculated:

Where m represents the number of first default probabilities.

When Q2 is an integer, the Q2th first default probability is used as the first threshold corresponding to the first default probability prediction model. Exemplarily, when m is 11, Q2 is 9, at this time, the 9th first default probability is used as the first threshold corresponding to the first default probability prediction model.

When Q2 is not an integer, the first threshold corresponding to the first default probability prediction model can be calculated according to the interpolation method. Specifically, according to the integer part v of Q2, Q2 is located between v and v+1. Therefore, the second threshold is located between the first default probability corresponding to the vth position and the first default probability corresponding to the v+1th position. The interpolation calculation formula is as follows:

C _Q2 = (1-frac(Q2))×data(v)+frac(Q2)×data(v+1)

Among them, C _Q2 represents the first threshold corresponding to the first default probability prediction model, frac(Q2) is the fractional part of Q2, data(v) is the first default probability at the vth position, and data(v+1) is the first default probability at the v+1th position.

For example, when m=10, Q2=8.25, at this time, v=8, v+1=9. frac(Q2)=0.25, assuming data(v)=0.7, data(v+1)=0.8, through the above formula, C _Q2 =0.275 can be calculated, at this time the first threshold is 0.725.

S1022: Filter out training samples whose first default probability is greater than the first threshold or less than the second threshold from the training sample set, use the filtered training samples to construct a fourth training sample set, and use the remaining training samples to construct a fifth training sample set.

In this embodiment, according to the first default probability of each training sample in the training sample set, training samples whose first default probability is greater than the first threshold corresponding to the first default probability prediction model are screened out from the training sample set, and training samples whose first default probability is less than the second threshold corresponding to the first default probability prediction model are screened out from the training sample set, and the fourth training sample set is constructed using the screened training samples, and the fifth training sample set is constructed using the remaining training samples.

S1023: Determine a first threshold and a second threshold corresponding to a second default probability prediction model according to the second default probability of each training sample.

In this embodiment, the second threshold corresponding to the second default probability prediction model is calculated in the following manner: for the second default probability of each training sample in the training sample set, the second default probability is sorted in ascending order, and the quartile Q3 is calculated:

Wherein, r represents the number of the second default probability. In this embodiment, r=m.

When Q3 is an integer, the Q3th second default probability is used as the second threshold corresponding to the second default probability prediction model.

When Q3 is not an integer, the second threshold corresponding to the second default probability prediction model can be calculated according to the interpolation method. Specifically, according to the integer part u of Q3, Q3 is located between u and u+1. Therefore, the second threshold is located between the second default probability corresponding to the uth position and the second default probability corresponding to the u+1th position. The interpolation calculation formula is as follows:

C _Q3 = (1-frac(Q3))×data(u)+frac(Q3)×data(u+1)

Among them, C _Q3 represents the second threshold corresponding to the second default probability prediction model, frac(Q3) is the decimal part of Q3, data(u) is the second default probability at the u-th position, and data(u+1) is the second default probability at the u+1-th position.

In this embodiment, the first threshold corresponding to the second default probability prediction model is calculated in the following manner: for the second default probabilities of each training sample in the training sample set, the second default probabilities are sorted in ascending order, and the three-quarter digit Q4 is calculated:

Where r represents the number of second default probabilities.

When Q4 is an integer, the Q4th second default probability is used as the first threshold corresponding to the second default probability prediction model.

When Q4 is not an integer, the first threshold corresponding to the second default probability prediction model can be calculated according to the interpolation method. Specifically, according to the integer part d of Q4, Q4 is located between d and d+1. Therefore, the first threshold is located between the second default probability corresponding to the dth position and the second default probability corresponding to the d+1th position. The interpolation calculation formula is as follows:

C _Q4 = (1-frac(Q4))×data(d)+frac(Q4)×data(d+1)

Among them, C _Q4 represents the first threshold corresponding to the second default probability prediction model, frac(Q4) is the fractional part of Q4, data(d) is the second default probability at the d-th position, and data(d+1) is the second default probability at the d+1-th position.

S1024: Filter out training samples whose second default probability is greater than the first threshold or less than the second threshold from the training sample set, use the filtered training samples to construct a sixth training sample set, and use the remaining training samples to construct a seventh training sample set; wherein the second training sample set includes the fifth training sample set and the seventh training sample set;

In this embodiment, according to the second default probability corresponding to each training sample in the training sample set, training samples whose second default probability is greater than the first threshold corresponding to the second default probability prediction model are screened out from the training sample set, and training samples whose second default probability is less than the second threshold corresponding to the second default probability prediction model are screened out from the training sample set, and the sixth training sample set is constructed using the screened training samples, and the seventh training sample set is constructed using the remaining training samples.

The second training sample set is constructed using the fifth training sample set and the seventh training sample set.

S1025: Take the intersection of the fourth training sample set and the sixth training sample set as the first training sample set.

In a possible implementation, when performing sample balancing processing on the first training sample set in step S103 to expand the training samples whose sample types are default samples in the first training sample set to obtain the third training sample set, the following steps may be specifically performed:

S1031: For each training sample whose sample type is a default sample in the first training sample set, calculate the Euclidean distance between the training sample and other training samples; other training samples are training samples whose sample type is a default sample except the training sample.

In this embodiment, according to the sample type, the training samples whose sample type is the default sample in the first training sample set are divided into the positive sample set, and the training samples whose sample type is the non-default sample in the first training sample set are divided into the negative sample set.

For each training sample in the positive sample set, the Euclidean distance between the training sample and other training samples in the positive sample set is calculated.

In this embodiment, the Euclidean distance between the training sample B (b1, b2, b3 ... bw) and the training sample E (e1, e2, e3 ... ew) can be calculated by the following formula:

Among them, b1, b2, b3…bw are the variable values of each variable in training sample B, e1, e2, e3…ew are the variable values of each variable in training sample E; distance(B,E) represents the Euclidean distance between training sample B and training sample E.

S1032: According to the Euclidean distance between the training sample and other training samples, in the order of the Euclidean distance from small to large, select a preset number of other training samples from other training samples as similar training samples of the training sample.

In this embodiment, after calculating the Euclidean distance between the training sample B and each other training sample in the positive sample set, a preset number of other training samples are selected from the other training samples in the order of the Euclidean distance from small to large as similar training samples of the training sample. The preset number is, for example, 3 or 5.

S1033: Select a target similar training sample from the similar training samples, and use the target similar training sample and the training sample to generate a synthetic training sample of the training sample.

In this embodiment, a similar training sample is randomly selected from each similar training sample of the training sample B as a target similar training sample of the training sample B, which is recorded as B _cc .

In this embodiment, the synthetic training sample of the training sample B can be generated by the following formula:

B _newsample =B+random(0,1)×(B _cc -B)

Among them, B _newsample is the synthetic training sample of the training sample B, B _cc is the target similar training sample of the training sample B; random(0,1) represents a random value from 0 to 1.

In this embodiment, one training sample may generate one synthetic training sample, or may generate multiple synthetic training samples.

S1034: Construct a third training sample set according to the synthetic training samples corresponding to each training sample whose sample type is the default sample in the first training sample set and each training sample in the first training sample set.

In this embodiment, the third training sample set is constructed according to the synthetic training samples corresponding to each training sample in the positive sample set and each training sample in the first training sample set.

In a possible implementation manner, after executing step S104 to obtain the target training sample set, the following steps may also be performed:

The target training sample set is used to train the initial default probability prediction model to obtain the target default probability prediction model after training.

In this embodiment, the initial default probability prediction model can specifically be an initial XGBOOST model, and after the initial default probability prediction model is trained using the target training sample set after sample equalization processing, a target default probability prediction model after training is obtained. The target default probability prediction model is an XGBOOST model trained using the target training sample set.

After the target default probability prediction model is trained, the test sample set obtained in step S1005 can be used to verify the accuracy of the target default probability prediction model in predicting the default probability. When the accuracy is greater than the preset accuracy, it indicates that the training of the target default probability prediction model is completed. At this time, when the target customer wants to predict the default probability, the customer basic information, product information, customer behavior information, and historical default information of the target customer can be obtained, and this information can be input into the target default probability prediction model, and the default probability of the target customer can be predicted by the target default probability prediction model.

When the accuracy is not greater than the preset accuracy, the target default probability prediction model needs to continue training until the accuracy of the target default probability prediction model in predicting the default probability is greater than the preset accuracy, and the training stops.

Embodiment 2:

Based on the same technical concept, the present application also provides a sample balancing device. FIG3 shows a schematic diagram of the structure of a sample balancing device provided in an embodiment of the present application. As shown in FIG3 , the device includes:

Prediction module 301, for predicting the default probability of each training sample in a training sample set; wherein the training sample set includes a plurality of training samples; the sample types of the training samples include default samples and non-default samples; the sample number of the default samples is less than the sample number of the non-default samples;

A screening module 302 is used to screen out training samples whose default probability is greater than a first threshold or less than a second threshold from the training sample set according to the default probability of each training sample, so as to construct a first training sample set using the screened training samples and construct a second training sample set using the remaining training samples; the first threshold is greater than the second threshold;

The sample balancing module 303 is used to perform sample balancing processing on the first training sample set, so as to expand the training samples whose sample types are the default samples in the first training sample set, so as to obtain a third training sample set; the third training sample set includes each training sample in the first training sample set and a new training sample generated after sample expansion;

The merging module 304 is configured to merge the third training sample set and the second training sample set to obtain a target training sample set.

Optionally, when the prediction module 301 is used to predict the default probability of each training sample in the training sample set, it is specifically used to:

For each training sample in the training sample set, a first default probability prediction model that has been trained is used to predict a first default probability of the training sample, and a second default probability prediction model that has been trained is used to predict a second default probability of the training sample; the default probability of the training sample includes the first default probability and the second default probability of the training sample.

Optionally, the device further includes a first training module; the first training module is used to train and obtain the first default probability prediction model in the following manner:

For each training sample in the training sample set, the training sample is input into a first initial default probability prediction model, a default situation of the training sample is predicted by the first initial default probability prediction model, and a first default probability of the training sample is output;

Inputting the default label of the training sample and the first default probability into a loss function to calculate the loss value of the first initial default probability prediction model;

When the loss value is greater than the preset loss value, the trainable parameters in the first initial default probability prediction model are updated according to the loss value; when the loss value is not greater than the preset loss value, the current first initial default probability prediction model is used as the first default probability prediction model.

Optionally, the second default probability prediction model includes a plurality of decision trees, each decision tree includes a plurality of classification nodes; the device further includes a second training module; the second training module is used to train the second default probability prediction model in the following manner:

For each initial decision tree in the second initial default probability prediction model, each training sample in the training sample set is input into the initial decision tree to obtain a node parameter corresponding to each classification node in the initial decision tree; the node parameter is the ratio of the number of samples whose sample category is the default sample on the classification node to the total number of training samples on the classification node;

Calculating the impurity of the initial decision tree according to the node parameters corresponding to each classification node in the initial decision tree;

When the impurity is greater than the preset impurity, optimizing the structure of the initial decision tree; when the impurity is not greater than the preset impurity, determining the current initial decision tree as the decision tree after training;

When all initial decision trees in the second initial default probability prediction model are trained, the second default probability prediction model is obtained.

Optionally, when the screening module 302 is used to screen out training samples whose default probability is greater than a first threshold or less than a second threshold from the training sample set according to the default probability of each training sample, so as to construct a first training sample set using the screened training samples and construct a second training sample set using the remaining training samples, it is specifically used to:

Determining a first threshold and a second threshold corresponding to the first default probability prediction model according to the first default probability of each training sample;

Screening out training samples whose first default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a fourth training sample set, and using the remaining training samples to construct a fifth training sample set;

Determining a first threshold and a second threshold corresponding to the second default probability prediction model according to the second default probability of each training sample;

Screening out training samples whose second default probability is greater than the first threshold or less than the second threshold from the training sample set, using the screened training samples to construct a sixth training sample set, and using the remaining training samples to construct a seventh training sample set; wherein the second training sample set includes the fifth training sample set and the seventh training sample set;

An intersection of the fourth training sample set and the sixth training sample set is taken as the first training sample set.

Optionally, when the sample balancing module 303 is used to perform sample balancing processing on the first training sample set to expand the training samples whose sample types are the default samples in the first training sample set to obtain the third training sample set, it is specifically used to:

For each training sample whose sample type is the default sample in the first training sample set, calculating the Euclidean distance between the training sample and other training samples; the other training samples are training samples whose sample type is the default sample other than the training sample;

According to the Euclidean distance between the training sample and other training samples, in the order of the Euclidean distance from small to large, selecting a preset number of other training samples from other training samples as similar training samples of the training sample;

Selecting a target similar training sample from the similar training samples, and generating a synthetic training sample of the training sample by using the target similar training sample and the training sample;

The third training sample set is constructed according to the synthetic training samples corresponding to each training sample whose sample type is the default sample in the first training sample set and each training sample in the first training sample set.

Optionally, the device further comprises:

The third training module is used to train the initial default probability prediction model using the target training sample set to obtain a target default probability prediction model after training.

Embodiment three:

Figure 4 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, including: a processor 401, a memory 402 and a bus 403, wherein the memory 402 stores machine-readable instructions executable by the processor 401. When the electronic device runs the above-mentioned information processing method, the processor 401 communicates with the memory 402 through the bus 403, and the processor 401 executes the machine-readable instructions to execute the method steps described in Example 1.

Embodiment 4:

The embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method steps described in the first embodiment are executed.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices, electronic devices, and computer-readable storage media can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

In the several embodiments provided in the present application, it should be understood that the disclosed methods, devices, electronic devices, and computer-readable storage media can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims (10)

1. A sample balancing method, characterized by including: For each training sample in the training sample set, predict the default probability of the training sample; wherein the training sample set contains multiple training samples; the sample type of the training sample includes default samples and non-default samples; the default sample The sample size is less than the sample size of the non-default sample; According to the default probability of each training sample, the training samples whose default probability is greater than the first threshold or less than the second threshold are selected from the training sample set, so as to use the filtered training samples to construct the first training sample set, and use the remaining The training samples construct a second training sample set; the first threshold is greater than the second threshold; Perform sample equalization processing on the first training sample set to expand the training samples whose sample type is the default sample in the first training sample set to obtain a third training sample set; the third training sample set Containing each training sample in the first training sample set and a new training sample generated after sample expansion; The third training sample set and the second training sample set are combined to obtain a target training sample set. 2. The method according to claim 1, characterized in that, for each training sample in the training sample set, predicting the default probability of the training sample includes: For each training sample in the training sample set, use the trained first default probability prediction model to predict the first default probability of the training sample, and use the trained second default probability prediction model to predict the second default probability of the training sample ; The default probability of the training sample includes the first default probability and the second default probability of the training sample. 3. The method according to claim 2, characterized in that the first default probability prediction model is trained in the following manner: For each training sample in the training sample set, input the training sample into the first initial default probability prediction model, predict the default situation of the training sample through the first initial default probability prediction model, and output the first probability of default; Input the default label of the training sample and the first default probability into the loss function, and calculate the loss value of the first initial default probability prediction model; When the loss value is greater than the preset loss value, the trainable parameters in the first initial default probability prediction model are updated according to the loss value; when the loss value is not greater than the preset loss value, the current The first initial default probability prediction model is used as the first default probability prediction model. 4. The method according to claim 2, characterized in that the second default probability prediction model includes multiple decision trees, each decision tree contains multiple classification nodes; the second default probability prediction model is based on the following Obtained by training in this way: For each initial decision tree in the second initial default probability prediction model, input each training sample in the training sample set into the initial decision tree to obtain the node parameters corresponding to each classification node in the initial decision tree; so The node parameter is the ratio of the number of samples whose sample category is default sample on the classification node to the total number of training samples on the classification node; Calculate the impurity of the initial decision tree according to the node parameters corresponding to each classification node in the initial decision tree; When the impurity is greater than the preset impurity, optimize the structure of the initial decision tree; when the impurity is not greater than the preset impurity, determine the current initial decision tree as the decision tree after training is completed; When each initial decision tree in the second initial default probability prediction model is trained, the second default probability prediction model is obtained. 5. The method according to claim 2, characterized in that, according to the default probability of each training sample, training samples with a default probability greater than a first threshold or less than a second threshold are selected from the training sample set to Use the selected training samples to construct the first training sample set, and use the remaining training samples to construct the second training sample set, including: Determine the first threshold and the second threshold corresponding to the first default probability prediction model according to the first default probability of each training sample; Screen out training samples whose first default probability is greater than the first threshold or less than the second threshold from the training sample set, use the filtered training samples to construct a fourth training sample set, and use the remaining training samples to construct a fifth training set sample set; Determine the first threshold and the second threshold corresponding to the second default probability prediction model according to the second default probability of each training sample; Screen out training samples whose second default probability is greater than the first threshold or less than the second threshold from the training sample set, use the filtered training samples to construct a sixth training sample set, and use the remaining training samples to construct a seventh training set Sample set; wherein, the second training sample set includes the fifth training sample set and the seventh training sample set; The intersection of the fourth training sample set and the sixth training sample set is taken as the first training sample set. 6. The method according to claim 5, characterized in that, the first training sample set is subjected to sample equalization processing to perform sample balancing on the training samples whose sample type is the default sample in the first training sample set. Expand and obtain the third training sample set, including: For each training sample in the first training sample set whose sample type is the default sample, calculate the Euclidean distance between the training sample and other training samples; the other training samples are sample types other than the training sample. Training samples of the default samples; According to the Euclidean distance between the training sample and other training samples, and in the order of the Euclidean distance from small to large, select a preset number of other training samples from other training samples as similar training samples to the training sample; Select a target similar training sample from the similar training samples, and use the target similar training sample and the training sample to generate a synthetic training sample of the training sample; The third training sample set is constructed based on the synthetic training samples corresponding to each training sample whose sample type is the default sample in the first training sample set, and each training sample in the first training sample set. 7. The method according to any one of claims 1-6, characterized in that the method further includes: The target training sample set is used to train an initial default probability prediction model, and a target default probability prediction model after training is obtained. 8. A sample equalization device, characterized in that it includes: A prediction module, used to predict the default probability of each training sample in the training sample set; wherein the training sample set contains multiple training samples; the sample types of the training samples include default samples and non-default samples ;The sample number of the default sample is less than the sample number of the non-default sample; A screening module, configured to screen out training samples whose default probability is greater than the first threshold or less than the second threshold from the training sample set according to the default probability of each of the training samples, so as to construct the first training sample using the filtered training samples. set, use the remaining training samples to construct a second training sample set; the first threshold is greater than the second threshold; A sample equalization module, configured to perform sample equalization processing on the first training sample set to expand the training samples whose sample type is the default sample in the first training sample set to obtain a third training sample set; The third training sample set includes each training sample in the first training sample set and new training samples generated after sample expansion; A merging module, configured to merge the third training sample set and the second training sample set to obtain a target training sample set. 9. An electronic device, characterized in that it includes: a processor, a memory and a bus, the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the Memories communicate with each other through a bus, and when the machine-readable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are performed. 10. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, the steps of the method according to any one of claims 1 to 7 are executed.