Credit data synthesis oversampling method and system based on noiseless loss distribution migration

Abstract

This invention discloses a method and system for synthesizing oversampling credit data based on noiseless loss distribution transfer, relating to the field of imbalanced classification technology in artificial intelligence and data mining. The method includes: acquiring an original imbalanced credit dataset; filtering noisy labeled samples from the dataset, predicting the probability of each sample in the dataset using a model, and calculating its noiseless loss value; dividing the dataset into multiple loss intervals based on the range of the noiseless loss value, assigning each sample to the corresponding loss interval, and determining the noiseless loss distribution of the majority and minority classes; performing transfer based on the noiseless loss distribution to determine the distribution of synthesized samples, thereby selecting root samples and auxiliary samples, and then synthesizing minority class pseudo-samples through linear interpolation; adding the minority class pseudo-samples to the original imbalanced credit dataset to obtain a class-balanced credit dataset. The optimized balanced dataset of this invention can effectively improve the recognition accuracy of both majority and minority class samples.

Description

Technical Field

[0001] This invention relates to the field of imbalanced classification technology in artificial intelligence and data mining, and in particular to a method and system for synthesizing and oversampling credit data based on noiseless loss distribution migration. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In commercial bank lending operations (such as personal or corporate loans, general consumer loans, or emergency loans), credit default prediction is a crucial aspect of risk control. Commercial banks rely on credit default prediction models to assist in credit granting decisions. Specifically, credit approval personnel input the information of loan applicants into a credit default prediction model trained on a credit default dataset. This model then predicts the probability of default, thereby determining whether the customer will default. The credit dataset used to train this model contains two labels: default (label 1, minority class samples) and non-default (label 0, majority class samples). Due to the specific nature of financial activities, this dataset exhibits class imbalance, meaning that the number of non-default samples far exceeds the number of default samples.

[0004] In the field of artificial intelligence, modeling and training based on such imbalanced datasets falls under the category of imbalanced classification problems. Existing binary classification models, affected by the imbalance of the dataset, perform well in identifying non-defaulting customers but have low accuracy in identifying defaulting customers. This leads to insufficient risk control in credit operations and can easily result in business losses. Therefore, addressing the impact of class imbalance on credit default prediction is of significant practical importance to the credit operations of banking financial institutions.

[0005] Currently, synthetic oversampling of minority class samples to generate a certain number of minority class pseudo-samples is an important means to solve the class imbalance problem in datasets. Its core elements include: root samples (i.e., seed samples), auxiliary samples, and sampling scalar ratio. Specifically, root samples and auxiliary samples are selected from the minority class samples in the original dataset. These are then synthesized into new pseudo-samples through linear interpolation. The sampling scalar ratio determines the number of samples synthesized from each root sample, i.e., the sampling scalar ratio determines the distribution of the synthesized pseudo-samples. Classical synthetic minority oversampling techniques (SMOTE) and subsequent improvements such as adaptive synthetic oversampling (ADASYN), borderline-SMOTE, K-means clustering-based SMOTE, and support vector machine-based SVM-SMOTE have optimized root sample selection, auxiliary sample determination, and sampling scalar ratio allocation, and have been applied in financial credit risk control scenarios. At that time, current synthetic oversampling techniques still had certain problems:

[0006] (1) Focusing only on the balance of sample size without exploring the root cause of poor classification model performance due to class imbalance. In fact, the imbalance in quantity is only a superficial reason. The core reason for poor model performance due to class imbalance is that the contributions of different class samples to the model fitting are uneven, which makes the model have low performance in recognizing minority class samples (i.e. default samples).

[0007] (2) It is susceptible to the influence of noisy samples (especially noisy labeled samples). Noisy labeled samples themselves can cause errors in the classification model. When the dataset is oversampled, if noisy labeled samples are selected as root samples or auxiliary samples, the error will accumulate and the generalization performance of the model will be severely reduced.

[0008] (3) Over-reliance on the inherent information of minority class samples, assuming that certain attributes of minority class samples (such as gradient distribution) can represent the attributes of all minority class events, but this assumption is difficult to verify, which also limits the applicability and reliability of the technology. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, this invention provides a method and system for synthesizing oversampling credit data based on noiseless loss distribution transfer. Targeting an imbalanced dataset in the financial industry—a credit default dataset—it employs an improved sample distribution metric for evaluation. By filtering noisy labeled samples, it avoids the accumulation of errors caused by noise samples on the synthesis oversampling method. Furthermore, it fully utilizes the distribution characteristics of the majority class samples to determine a pseudo-sample synthesis strategy to synthesize minority class pseudo-samples. The optimized oversampling of the imbalanced credit dataset allows the classification results to maintain high majority class sample recognition accuracy while significantly improving minority class sample recognition accuracy.

[0010] In a first aspect, the present invention provides a method for synthesizing and oversampling credit data based on noiseless loss distribution migration.

[0011] A credit data synthesis oversampling method based on noiseless loss distribution transfer includes:

[0012] Obtain the original imbalanced credit dataset; where the dataset contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters;

[0013] The original imbalanced credit dataset is filtered for noisy labeled samples. Then, the model predicts the probability of each sample in the dataset and calculates the noiseless loss value for each sample.

[0014] Divide the sample into multiple loss intervals based on the range of the noiseless loss value, assign each sample to a different loss interval, and determine the noiseless loss distribution of the majority and minority class samples.

[0015] The migration is performed based on the noiseless loss distribution to determine the distribution of the synthetic samples. Then, root samples and auxiliary samples are selected based on the distribution of the synthetic samples, and minority class pseudo samples are synthesized by linear interpolation.

[0016] Adding minority class pseudo-samples to the original imbalanced credit dataset yields a class-balanced credit dataset.

[0017] A further technical solution involves filtering noisy labeled samples from the original imbalanced credit dataset, then predicting the probability of each sample in the dataset using a model, and calculating the noise-free loss value for each sample, including:

[0018] A belief learning algorithm is used to filter noisy labeled samples in the original imbalanced credit dataset to obtain a noise-free credit dataset.

[0019] Based on a noiseless credit dataset, a 5-fold cross-validation was performed on the logistic regression model to predict the probability of each sample; the predicted probability is the probability that the predicted sample is a minority class sample.

[0020] Based on the predicted probability, the noise-free loss value of each sample is calculated using the binary cross-entropy loss function; whereby the binary cross-entropy loss function is:

[0021] ;

[0022] In the above formula, To predict probabilities, The loss value. The true label of the sample; The range of values ​​is , This represents the maximum noise-free loss value for the sample.

[0023] A further technical solution involves dividing the noiseless loss value into multiple loss intervals based on its range, assigning each sample to a corresponding loss interval, and determining the noiseless loss distribution for the majority and minority classes, including:

[0024] The range of values ​​for the noise-free loss value The dataset is divided into k loss intervals on average, and the majority class and minority class samples in the noiseless credit dataset are assigned to the corresponding different loss intervals.

[0025] Count the number of majority class samples within each loss interval, and combine them to form a noiseless loss distribution vector of the majority class samples. = ,in, For the first k The number of majority class samples within each loss interval;

[0026] Count the number of minority class samples within each loss interval, and combine them to form a noiseless loss distribution vector for the minority class samples. = ,in, For the first k The number of minority class samples within each loss interval.

[0027] A further technical solution involves transferring the loss distribution based on the noiseless loss distribution to determine the distribution of the synthesized samples. This means maintaining consistency in the distribution of minority and majority class samples across the loss interval, and transferring the noiseless loss distribution of the majority class samples to at least a few classes. This includes:

[0028] Noiseless loss distribution vector for traversing the majority class samples The noiseless loss distribution vector of minority class samples The loss intervals in which the number of majority class samples is greater than the number of minority class samples are selected as the target loss intervals. The difference between the number of majority class samples and the number of minority class samples in each target loss interval is calculated. This difference is the number of minority class pseudo samples that need to be synthesized in the corresponding target loss interval.

[0029] A further technical solution involves selecting root samples and auxiliary samples based on the distribution of the synthesized samples, and then synthesizing minority class pseudo-samples through linear interpolation, including:

[0030] Select the minority class sample within each target loss interval as the root sample;

[0031] For each root sample, determine its k A neighboring loss, and from k A sample is randomly selected from the loss nearest neighbors as an auxiliary sample; the loss nearest neighbor is the sample with the smallest absolute value of the difference between the root sample and the noiseless loss value.

[0032] Based on root samples and auxiliary samples, minority class pseudo-samples are synthesized through linear interpolation; the linear interpolation formula is as follows:

[0033] ;

[0034] in, These are synthesized pseudo-samples. The root sample is given, and γ is a random number between 0 and 1. For auxiliary samples.

[0035] Further technical solutions, when synthesizing minority class pseudo-samples, also include:

[0036] When the number of root samples in the target loss interval is less than the number of pseudo samples to be synthesized, assign the same sampling factor to each root sample in that interval.

[0037] When there are no minority class samples within the target loss interval, the minority class sample whose noiseless loss value is closest to the interval is selected as the root sample.

[0038] Secondly, the present invention provides a credit data synthesis oversampling system based on noiseless loss distribution migration.

[0039] A credit data synthesis oversampling system based on noiseless loss distribution migration includes:

[0040] The data acquisition module is used to acquire the original unbalanced credit dataset, which contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters.

[0041] The preprocessing module is used to filter noisy labeled samples from the original imbalanced credit dataset, then predict the probability of each sample in the dataset through the model, and calculate the noiseless loss value of each sample.

[0042] The noiseless loss distribution calculation module is used to divide multiple loss intervals according to the range of noiseless loss values, assign each sample to the corresponding different loss intervals, and determine the noiseless loss distribution of majority and minority class samples.

[0043] The minority class pseudo-sample synthesis module is used to perform migration based on the noiseless loss distribution, determine the distribution of synthesized samples, and then select root samples and auxiliary samples according to the distribution of synthesized samples, and synthesize minority class pseudo-samples through linear interpolation.

[0044] The balanced dataset generation module is used to add minority class pseudo-samples to the original imbalanced credit dataset to obtain a class-balanced credit dataset.

[0045] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the above-described credit data synthesis oversampling method based on noiseless loss distribution migration when executing the executable instructions stored in the memory.

[0046] Fourthly, the present invention also provides a computer-readable storage medium storing executable instructions for causing a processor to execute the executable instructions to implement the above-described credit data synthesis oversampling method based on noiseless loss distribution migration.

[0047] Fifthly, the present invention also provides a computer program product comprising executable instructions stored in a computer-readable storage medium; wherein, when the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the above-mentioned credit data synthesis oversampling method based on noiseless loss distribution migration is implemented.

[0048] The above one or more technical solutions have the following beneficial effects:

[0049] 1. This invention provides a method and system for synthesizing oversampling credit data based on noiseless loss distribution transfer. This method adopts a new sample classification index, a new nearest neighbor concept, and a new synthetic sample distribution strategy. Specifically, for the imbalanced dataset of the financial industry - the credit default dataset, an improved sample distribution metric is used for evaluation. Noisy labeled samples are filtered to avoid the accumulation of errors caused by noise samples to the synthetic oversampling method. The distribution characteristics of the majority class samples are fully utilized to determine the pseudo-sample synthesis strategy to synthesize minority class pseudo-samples. The imbalanced credit dataset optimized by synthetic oversampling can maintain a high majority class sample recognition accuracy and significantly improve the minority class sample recognition accuracy, thereby optimizing the risk control effect of credit business.

[0050] 2. This invention clarifies the root cause of poor model performance due to class imbalance. It constructs loss intervals and loss distribution vectors with noiseless loss value as the core, and measures the balance of the dataset from the perspective of the contribution of samples to model fitting. It filters noisy labeled samples through confidence learning to ensure the quality of subsequently calculated noiseless loss values ​​and synthesized pseudo-samples, thereby improving the model's generalization performance. It determines the pseudo-sample synthesis strategy based on the difference in noiseless loss distribution between majority and minority class samples, without relying on the inherent information of minority class samples, making the assumptions more reasonable and the technology more applicable.

[0051] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0052] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0053] Figure 1 This is an overall flowchart of the credit data synthesis oversampling method based on noiseless loss distribution migration as described in this embodiment of the invention;

[0054] Figure 2 This is a schematic diagram of the loss intervals divided in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the noiseless loss distribution migration in an embodiment of the present invention. Detailed Implementation

[0056] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0057] Example 1

[0058] Considering the problems existing in the current synthetic oversampling schemes for imbalanced credit default datasets: (1) It only focuses on the balance in terms of the number of samples, without exploring the root cause of the unsatisfactory performance of the classification model due to class imbalance. In fact, the imbalance in quantity is only a superficial reason. The root cause of the unsatisfactory performance of the classification model due to class imbalance is that the contributions of samples of different classes to the model fitting are unbalanced, which is specifically manifested as the unbalanced distribution of the loss values ​​of the samples. In the credit default dataset, since the number of defaulted samples is far less than that of non-defaulted samples, the contribution of defaulted samples to the fitting of the classification model is also far less than that of non-defaulted samples. The classification model optimizes the model weights based on the backpropagation of the loss function. That is, when the loss value of a sample is large, it will play a greater role in optimizing the model weights. For binary credit datasets, the disadvantage in the number of minority class samples makes the loss value of all samples much smaller than the loss value of all non-default samples, thus resulting in low model recognition performance for default samples. (2) Current synthetic oversampling techniques that can be used for credit default prediction are easily affected by noise samples, especially noise labels. Noise label samples themselves are prone to causing errors in the classification model. When synthetic oversampling is performed on the dataset, regardless of whether the noise label sample is selected as the root sample or the auxiliary sample, it will accumulate the error generated by the noise, which seriously affects the generalization performance of the classification model. (3) Most current synthetic oversampling techniques utilize the inherent information of minority class samples, believing that they have certain attributes of minority class samples and can represent the attributes of all minority class events. For example, it is believed that the gradient distribution of minority class samples can represent the gradient distribution of all minority class samples in the real world. However, these premises or assumptions are difficult to prove. To solve the above problems, this embodiment proposes a credit data synthetic oversampling method based on noiseless loss distribution transfer, such as Figure 1 As shown, the method specifically includes the following steps:

[0059] Step S1: Obtain the original imbalanced credit dataset. This dataset contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters.

[0060] Specifically, taking a credit default dataset containing a certain amount of sample data from real-world lending transactions as an example, such as the existing Chinese Personal Loan dataset, there are 8371 non-default samples and only 1683 default samples. The number of credit data with normal repayments (non-default samples) far exceeds the number of credit data with abnormal repayments (default samples), with a ratio of approximately 5:1, thus causing a class imbalance. Furthermore, non-default samples are called majority class samples or negative class samples, and are usually labeled 0, while default samples are called minority class samples or positive class samples, and are usually labeled 1.

[0061] Step S2: Filter noisy labeled samples from the original imbalanced credit dataset, then predict the probability of each sample in the dataset using the model, and calculate the noiseless loss value for each sample.

[0062] Specifically, due to factors such as anonymization, transmission, and manual annotation in the credit default dataset, noisy labels may exist in the dataset. These noisy labels affect the loss measurement of normal samples, resulting in a noisy loss distribution. Therefore, in this embodiment, noisy samples are first filtered out using confidence learning to remove them from the original imbalanced credit dataset, resulting in a noise-free credit dataset. Then, based on the noise-free credit dataset, a 5-fold cross-validation is performed on the logistic regression model to predict the probability of each sample. This probability represents the probability that the sample is a minority class sample. Finally, based on this probability... The loss value for each sample is calculated using the binary cross-entropy loss function. This loss value is the noise-free loss value.

[0063] The binary classification cross-entropy loss function is:

[0064] ;

[0065] In the above formula, To predict probabilities, The loss value. The true label of the sample; The range of values ​​is Since the model has already been fitted, the loss value for each sample is very small and will not tend to positive infinity. Let the maximum loss value for the sample be... Then the range of loss values ​​for all samples is: .

[0066] Step S3: Divide the sample into multiple loss intervals according to the range of the noiseless loss value, assign each sample to the corresponding different loss interval, and determine the noiseless loss distribution of the majority class and minority class samples.

[0067] like Figure 2 As shown, the range of values ​​for the noise-free loss value is... Divided into average k The majority and minority class samples in the noisy credit dataset are assigned to different loss intervals. Based on this, the number of majority class samples within each loss interval is counted, and these are combined to form the noisy loss distribution vector of the majority class samples. = ,in For the first kThe number of majority class samples within each loss interval is counted; similarly, the number of minority class samples within each loss interval is counted, and these are combined to form a noiseless loss distribution vector for the minority class samples. = ,in For the first k The number of minority class samples within each loss interval.

[0068] Step S4: Based on the noiseless loss distribution, perform migration to determine the distribution of the synthesized samples, and then select root samples and auxiliary samples according to the distribution of the synthesized samples. Synthesize minority class pseudo samples through linear interpolation.

[0069] Specifically, firstly, let and These are the noiseless loss distribution vectors calculated based on the credit dataset. Both vectors are k-dimensional. Assume the number of majority class samples in the credit dataset is k. The number of minority class samples is The number of samples needed to achieve class balance is then... Iterate through the noiseless loss distribution vectors of the majority class samples respectively. The noiseless loss distribution vector of minority class samples If the loss interval for the majority class samples is greater than that for the minority class samples, then the majority class samples and the minority class samples each obtain a statistical vector of the same dimension. and That is, the loss interval in which the number of majority class samples is greater than the number of minority class samples is selected as the target loss interval. Then, the difference between the number of majority class samples and the number of minority class samples in each target loss interval is calculated. This difference is the number of minority class pseudo samples that need to be synthesized in the corresponding target loss interval.

[0070] like Figure 3 As shown, traversal and Then, in the first, second, and fourth intervals, the number of majority class samples is greater than the number of minority class samples. and These are the three-dimensional statistical vectors for the majority and minority classes, respectively. The difference in the number of samples in the corresponding dimension represents the number of samples to be synthesized within each minority class sample interval. This ensures that the distributions of the minority and majority class samples are consistent across the loss interval, transferring the noiseless loss distribution of the majority class samples to the minority class samples. This noiseless loss distribution transfer determines the synthetic sample distribution of the synthetic oversampling method.

[0071] Secondly, this embodiment also defines the selection strategy for root samples and auxiliary samples, specifically as follows:

[0072] The root sample selection strategy is as follows: within certain loss intervals, the minority class samples are fewer than the majority class samples; these samples within the loss intervals are selected as root samples. The root samples are determined by the aforementioned statistical vector. When it is necessary to synthesize samples for a certain loss interval, the samples within that interval are used as root samples. This sample selection strategy does not depend on the spatial characteristics of the samples, thus avoiding the problem of distance metrics failing in high-dimensional samples.

[0073] The auxiliary sample selection strategy is as follows: Define the concept of loss nearest neighbors as the sample whose loss value has the smallest absolute difference from that sample's loss value. Then, for a given sample... The nearest neighbors of the loss, i.e., the ones with the smallest absolute difference between their loss values ​​and that of the sample. A sample, and then, the auxiliary samples of a certain root sample are from its... A random auxiliary sample is selected from the loss nearest neighbors, and the selection of this auxiliary sample does not depend on the spatial features of the sample.

[0074] Therefore, in this embodiment, minority class samples within each target loss interval are selected as root samples, and for each root sample, its... k A neighboring loss, and from k A random auxiliary sample is selected from the nearest neighbors of the loss sample. This nearest neighbor is the sample with the smallest absolute difference from the root sample's noiseless loss value.

[0075] Finally, based on the root samples and auxiliary samples, minority class pseudo-samples are synthesized through linear interpolation. The linear interpolation formula is as follows:

[0076] ;

[0077] in, These are synthesized pseudo-samples. The root sample is given, and γ is a random number between 0 and 1. For auxiliary samples.

[0078] As one implementation method, the process of synthesizing minority class pseudo-samples as described above further includes:

[0079] When the number of root samples in the target loss interval is less than the number of pseudo samples to be synthesized, assign the same sampling factor to each root sample in that interval.

[0080] When there are no minority class samples within the target loss interval, the minority class sample whose noiseless loss value is closest to the interval is selected as the root sample.

[0081] Step S5: Add minority class pseudo-samples to the original imbalanced credit dataset to obtain a class-balanced credit dataset.

[0082] To further verify the significant advancements of the synthetic oversampling method based on noiseless loss transfer provided in this embodiment compared to existing technologies, this embodiment compares the performance improvement of the proposed method with that of eight other existing synthetic oversampling methods on six metrics, using three imbalanced financial datasets. The existing synthetic oversampling methods employed include: Synthetic Minority Oversampling Technique (SMOTE), Boundary Sample-Based Synthetic Oversampling Technique (BL-SMOTE), Support Vector Machine-Based Synthetic Oversampling Technique (SVM-SMOTE), Adaptive Synthetic Oversampling Technique (ADASYN), Nearest Neighbor-Based Hybrid Oversampling Technique (SMOTEN), Neighbor Association-Based Hybrid Oversampling Technique (SMOTE-Tomek), and K-Nearest Neighbor-Based Hybrid Oversampling Technique (SMOTE-ENN). In addition, six metrics were set: F1 score, recall, area under the curve (AUC), geometric mean (G-Mean), Matthews correlation coefficient (MCC), and Case statistic (KS). These metrics best reflect the performance of imbalanced classification models. The specific experimental results are shown in Tables 1-3 below.

[0083] Table 1 Experimental results using the Australian dataset

[0084]

[0085] As shown in Table 1 above, on the existing Australian dataset, this method achieves the best values ​​for all metrics in the LR model. Specifically, the Recall metric is improved by 2.47 percentage points compared to the second-best KM-SMOTE, and the F1-Score is also improved by 1.88 percentage points compared to KM-SMOTE. This method enables LGBM to achieve the best Recall, Recall, and MCC metrics on this dataset.

[0086] Table 2 Experimental results on the German dataset

[0087]

[0088] Similar to its performance on the Australian dataset, as shown in Table 2 above, this method also achieves optimal values ​​for all metrics for the LR classifier on the existing German dataset, while also enabling the LGBM model to achieve optimal F1-Score, Recall, and MCC values.

[0089] Table 3 Experimental results of the Credit Card Fraud Detection dataset

[0090]

[0091] On the existing CreditCardFraudDetection dataset, as shown in Table 3 above, this method achieves the best F1-Score, G-mean, MCC, and KS metrics for the LR model, while achieving the best F1-Score, AUC, and MCC metrics for the LGBM model.

[0092] The experiments on the three imbalanced credit datasets above demonstrate that, on small-scale datasets, the proposed method can improve the identification accuracy of both minority and majority class samples. On large-scale datasets, this method can significantly improve the model's F1-Score while maintaining a recall metric similar to other methods. In other words, this method consistently demonstrates the best performance in identifying high-quality customers in credit transactions without significantly reducing its ability to identify defaulting customers.

[0093] The synthetic oversampling method for imbalanced classification of credit datasets proposed in this embodiment synthesizes pseudo-samples of the minority class by making full use of the noiseless loss distribution information of the majority class samples. Experiments have shown that the method proposed in this embodiment improves the classification model's performance in identifying defaulting and non-defaulting users by optimizing the quality of credit data.

[0094] Example 2

[0095] This embodiment provides a credit data synthesis oversampling system based on noiseless loss distribution migration, including:

[0096] The data acquisition module is used to acquire the original unbalanced credit dataset, which contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters.

[0097] The preprocessing module is used to filter noisy labeled samples from the original imbalanced credit dataset, then predict the probability of each sample in the dataset through the model, and calculate the noiseless loss value of each sample.

[0098] The noiseless loss distribution calculation module is used to divide multiple loss intervals according to the range of noiseless loss values, assign each sample to the corresponding different loss intervals, and determine the noiseless loss distribution of majority and minority class samples.

[0099] The minority class pseudo-sample synthesis module is used to perform migration based on the noiseless loss distribution, determine the distribution of synthesized samples, and then select root samples and auxiliary samples according to the distribution of synthesized samples, and synthesize minority class pseudo-samples through linear interpolation.

[0100] The balanced dataset generation module is used to add minority class pseudo-samples to the original imbalanced credit dataset to obtain a class-balanced credit dataset.

[0101] Example 3

[0102] This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0103] Example 4

[0104] This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0105] Example 5

[0106] This embodiment provides a computer program product including executable instructions, which are computer instructions; the executable instructions are stored in a computer-readable storage medium. When the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the electronic device performs the method described in this embodiment.

[0107] The steps and methods involved in Embodiments 2 to 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0108] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0109] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. A method for synthesizing and oversampling credit data based on noiseless loss distribution migration, characterized in that, include: Obtain the original imbalanced credit dataset; where the dataset contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters; The original imbalanced credit dataset is filtered for noisy labeled samples. Then, the model predicts the probability of each sample in the dataset and calculates the noise-free loss value for each sample, including: A confidence learning algorithm is used to filter noisy labeled samples in the original imbalanced credit dataset to obtain a noise-free credit dataset. Based on the noise-free credit dataset, a 5-fold cross-validation is performed on the logistic regression model to predict the probability of each sample. The predicted probability is the probability that the predicted sample is a minority class sample. Divide the sample into multiple loss intervals based on the range of the noiseless loss value, assign each sample to a different loss interval, and determine the noiseless loss distribution of the majority and minority class samples. Based on the noiseless loss distribution, the transfer is performed to determine the distribution of the synthetic samples, that is, to maintain the consistency of the distribution of minority class samples and majority class samples in the loss interval, and to transfer the noiseless loss distribution of majority class samples to at least a few classes of samples, including: Noiseless loss distribution vector for traversing the majority class samples The noiseless loss distribution vector of minority class samples The loss intervals in which the number of majority class samples is greater than the number of minority class samples are selected as target loss intervals. The difference between the number of majority class samples and the number of minority class samples within each target loss interval is calculated; this difference represents the number of minority class pseudo-samples to be synthesized within the corresponding target loss interval. Then, based on the distribution of the synthesized samples, root samples and auxiliary samples are selected, and minority class pseudo-samples are synthesized through linear interpolation, including: Select the minority class sample within each target loss interval as the root sample; For each root sample, determine its k nearest neighbors, and randomly select a sample from the k nearest neighbors as an auxiliary sample; the nearest neighbor is the sample with the smallest absolute value of the difference between the root sample and the noiseless loss value. Adding minority class pseudo-samples to the original imbalanced credit dataset yields a class-balanced credit dataset.

2. The credit data synthesis oversampling method based on noiseless loss distribution migration as described in claim 1, characterized in that, The process involves filtering noisy labeled samples from the original imbalanced credit dataset, predicting the probability of each sample in the dataset using a model, and calculating the noise-free loss value for each sample. It also includes: Based on the predicted probability, the noise-free loss value of each sample is calculated using the binary cross-entropy loss function; whereby the binary cross-entropy loss function is: ； In the above formula, To predict probabilities, The loss value. The true label of the sample; The range of values ​​is , This represents the maximum noise-free loss value for the sample.

3. The credit data synthesis oversampling method based on noiseless loss distribution migration as described in claim 1, characterized in that, Based on the range of noise-free loss values, multiple loss intervals are divided, and each sample is assigned to a corresponding different loss interval to determine the noise-free loss distribution of the majority and minority classes, including: The range of values ​​for the noise-free loss value The dataset is divided into k loss intervals on average, and the majority class and minority class samples in the noiseless credit dataset are assigned to the corresponding different loss intervals. Count the number of majority class samples within each loss interval, and combine them to form a noiseless loss distribution vector of the majority class samples. = ,in, The number of majority class samples within the k-th loss interval; Count the number of minority class samples within each loss interval, and combine them to form a noiseless loss distribution vector for the minority class samples. = ,in, This represents the number of minority class samples within the k-th loss interval.

4. The credit data synthesis oversampling method based on noiseless loss distribution migration as described in claim 1, characterized in that, Root samples and auxiliary samples are selected based on the distribution of the synthesized samples. Then, minority class pseudo-samples are synthesized through linear interpolation. This also includes: Based on root samples and auxiliary samples, minority class pseudo-samples are synthesized through linear interpolation; the linear interpolation formula is as follows: ； in, These are synthesized pseudo-samples. The root sample is given, and γ is a random number between 0 and 1. For auxiliary samples.

5. The credit data synthesis oversampling method based on noiseless loss distribution migration as described in claim 4, characterized in that, The synthesis of minority class pseudo-samples also includes: When the number of root samples in the target loss interval is less than the number of pseudo samples to be synthesized, assign the same sampling factor to each root sample in that interval. When there are no minority class samples within the target loss interval, the minority class sample whose noiseless loss value is closest to the interval is selected as the root sample.

6. A credit data synthesis oversampling system based on noiseless loss distribution migration, characterized in that, include: The data acquisition module is used to acquire the original unbalanced credit dataset, which contains minority class samples labeled as defaulters and majority class samples labeled as non-defaulters. The preprocessing module filters noisy labeled samples from the original imbalanced credit dataset, then predicts the probability of each sample in the dataset using a model, and calculates the noise-free loss value for each sample, including: A confidence learning algorithm is used to filter noisy labeled samples in the original imbalanced credit dataset to obtain a noise-free credit dataset. Based on the noise-free credit dataset, a 5-fold cross-validation is performed on the logistic regression model to predict the probability of each sample. The predicted probability is the probability that the predicted sample is a minority class sample. The noiseless loss distribution calculation module is used to divide multiple loss intervals according to the range of noiseless loss values, assign each sample to the corresponding different loss intervals, and determine the noiseless loss distribution of majority and minority class samples. The minority class pseudo-sample synthesis module is used for transfer based on the noiseless loss distribution, determining the distribution of synthesized samples, that is, maintaining the consistency of the distribution of minority class samples and majority class samples in the loss interval, and transferring the noiseless loss distribution of majority class samples to at least a few classes of samples, including: Noiseless loss distribution vector for traversing the majority class samples The noiseless loss distribution vector of minority class samples The loss intervals in which the number of majority class samples is greater than the number of minority class samples are selected as target loss intervals. The difference between the number of majority class samples and the number of minority class samples within each target loss interval is calculated; this difference represents the number of minority class pseudo-samples to be synthesized within the corresponding target loss interval. Then, based on the distribution of the synthesized samples, root samples and auxiliary samples are selected, and minority class pseudo-samples are synthesized through linear interpolation, including: Select the minority class sample within each target loss interval as the root sample; For each root sample, determine its k nearest neighbors, and randomly select a sample from the k nearest neighbors as an auxiliary sample; the nearest neighbor is the sample with the smallest absolute value of the difference between the root sample and the noiseless loss value. The balanced dataset generation module is used to add minority class pseudo-samples to the original imbalanced credit dataset to obtain a class-balanced credit dataset.

7. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the credit data synthesis oversampling method based on noiseless loss distribution migration as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The system stores executable instructions that, when executed by a processor, implement the credit data synthesis oversampling method based on noiseless loss distribution migration as described in any one of claims 1-5.

9. A computer program product, characterized in that, The computer program product includes executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the credit data synthesis oversampling method based on noiseless loss distribution migration as described in any one of claims 1-5.

Images