Coevolution type fair machine learning model integrated training method oriented to class imbalance
By acquiring and filtering training datasets that violate fairness in computation, and training and optimizing neural network models, the contradiction between accuracy and fairness in class-imbalanced data is resolved, enabling fair prediction in bank credit approval.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHERN UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2025-12-18
- Publication Date
- 2026-04-28
AI Technical Summary
Existing machine learning models face a trade-off between accuracy and fairness when dealing with imbalanced training data, especially when sensitive attributes are intertwined, leading to errors in predictions and uneven distribution of opportunities among different groups.
By acquiring a training dataset that has undergone fair violation computation processing, a first-target neural network set is trained, a fair machine learning model is selected, and applied to the personal credit approval scenario of a bank. The accuracy and fairness of the model are optimized by utilizing a co-evolutionary mechanism.
This improved the accuracy and fairness of the model, enabling equitable predictions in bank credit approval and reducing prediction errors and unequal opportunity allocation among different groups.
Smart Images
Figure CN121936633A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of machine learning technology. More specifically, this application relates to a co-evolutionary fair machine learning model ensemble training method for class imbalance. Background Technology
[0002] In real-world machine learning system applications (such as credit systems or fraud detection systems), training data often exhibits severe class imbalance. Furthermore, training data is frequently intertwined with sensitive attributes (such as gender, race, etc.), leading to errors in predictions or unequal distribution of opportunities for different groups even when the overall accuracy of the machine learning model is acceptable, resulting in significant unfairness.
[0003] Therefore, existing technologies need to be improved and refined. Summary of the Invention
[0004] The purpose of this application is to provide a co-evolutionary fair machine learning model ensemble training method for class imbalance, which can balance the accuracy and fairness of model predictions. This application is mainly achieved through the following technical solutions: A first aspect of this application provides a method for ensemble training of class-imbalanced, co-evolutionary, fair machine learning models, comprising: Obtain a training dataset, wherein the worst-case validation subset of the training dataset is a set that has undergone fair violation computation. The first target neural network set is trained using the training dataset to obtain the second target neural network set; The second target neural network set is filtered according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
[0005] According to one embodiment of this application, the step of filtering the second target neural network set according to preset rules to obtain a fair machine learning model includes: Obtain the initial set of neural networks; The initial neural network set and the second target neural network set are merged to obtain the third target neural network set; Multiple third-objective neural networks are selected from the set of third-objective neural networks as the fair machine learning model.
[0006] According to one embodiment of this application, prior to the step of selecting multiple third-objective neural networks from the set of third-objective neural networks as the fair machine learning model, the ensemble training method for class imbalance-oriented co-evolutionary fair machine learning models further includes: The third target neural network set is evaluated to obtain the evaluation result.
[0007] According to one embodiment of this application, the step of evaluating the third target neural network set to obtain the evaluation result includes: The third target neural network set is subjected to a balanced accuracy evaluation process to obtain the target balanced accuracy. The third target neural network set is subjected to equalization opportunity evaluation processing to obtain target equalization opportunities; The target equalization accuracy and the target equalization opportunity are used as the evaluation results.
[0008] According to one embodiment of this application, the step of evaluating the third target neural network set to obtain the evaluation result further includes: The geometric mean score is evaluated on the third target neural network set to obtain the target geometric mean score; The target equilibrium accuracy, the target equalization chance, and the target geometric mean score are used as the evaluation results.
[0009] According to one embodiment of this application, the method for ensemble training of class-imbalanced co-evolutionary fair machine learning models further includes: Set the validation dataset of the training dataset as the initial evaluation dataset; The initial evaluation dataset is subjected to evolutionary processing to obtain the target evaluation dataset; The training dataset is updated based on the target evaluation dataset.
[0010] According to one embodiment of this application, the step of performing evolutionary processing on the initial evaluation dataset to obtain the target evaluation dataset includes: The initial evaluation dataset is expanded to obtain the evaluation dataset to be processed; The dataset to be evaluated is subjected to mutation and evolution processing to obtain the target evaluation dataset.
[0011] A second aspect of this application provides an ensemble training apparatus for class imbalance-oriented co-evolutionary fair machine learning models, comprising: The acquisition module is used to acquire the training dataset, wherein the worst-validation subset of the training dataset is a set that has undergone fairness violation computation. The training module is used to train the first target neural network set using the training dataset to obtain the second target neural network set. The filtering module is used to filter the second target neural network set according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
[0012] A third aspect of this application provides a method for predicting personal credit approval, comprising: Obtain users' credit transaction records; The credit business record information is input into a fair machine learning model for prediction processing to obtain personal credit approval results.
[0013] A fourth aspect of this application provides a terminal device, including a processor and a memory, the memory being used to store a computer program, and the processor being used to call and run the computer program stored in the memory to perform the steps of the class imbalance-oriented co-evolutionary fair machine learning model ensemble training method provided in the first aspect of this application.
[0014] A fifth aspect of this application provides a computer-readable storage medium for storing a computer program that causes a computer to perform the steps of the class-imbalanced co-evolutionary fair machine learning model ensemble training method provided in the first aspect of this application.
[0015] The beneficial effects of the embodiments of this application include: This application embodiment obtains a training dataset, where the worst-case validation subset of the training dataset is a set processed by fairness violation calculation; it then trains a first target neural network set using the training dataset to obtain a second target neural network set; finally, it filters the second target neural network set according to preset rules to obtain a fair machine learning model. This fair machine learning model is applied to a bank's personal credit approval application scenario to predict a user's credit status. Compared with the prior art, this application embodiment sets the worst-case validation subset of the training dataset as a set processed by fairness violation calculation. Therefore, this application embodiment can improve the accuracy and fairness of the model, achieving the goal of balancing the accuracy and fairness of model prediction. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 The flowcharts for some embodiments of the class imbalance-oriented co-evolutionary fair machine learning model integration training method of this application are shown below. Figure 2 A schematic diagram of the principle block diagram of the integrated training device for class-imbalanced co-evolutionary fair machine learning model integration in some embodiments of this application; Figure 3 This is a schematic block diagram of the terminal device of this application in some embodiments. Detailed Implementation
[0018] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.
[0019] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0020] The terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0021] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are expressly listed, but may include other steps or units that are not expressly listed or that are inherent to such process, method, product, or apparatus.
[0022] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items.
[0023] The specific embodiments of this application will be further described below with reference to the accompanying drawings.
[0024] refer to Figure 1 The diagram shown is a flowchart of a method for ensemble training of class-imbalanced, co-evolutionary, fair machine learning models, provided in the first aspect of an embodiment of this application. Figure 1 The method for ensemble training of class-imbalanced co-evolutionary fair machine learning models includes: S1. Obtain the training dataset, wherein the worst-case validation subset of the training dataset is a set that has undergone fair violation computation.
[0025] The worst validation subset is the worst performance on the validation subset of the training dataset.
[0026] The fairness violation is a quantifiable numerical indicator (such as the difference rate and / or Gini coefficient) used to reflect the fairness performance of the model on sensitive attributes.
[0027] Further, the step of obtaining the worst-case validation subset in the training dataset includes: performing a difference rate calculation on each validation data in the original worst-case validation subset of the training dataset to obtain a target difference rate for each validation data; taking the target difference rate that does not exceed a first set tolerance threshold among all target difference rates as the final difference rate; and constructing the worst-case validation subset in the training dataset from all the validation data corresponding to the final difference rates.
[0028] Furthermore, the step of obtaining the worst-case validation subset in the training dataset further includes: performing Gini coefficient calculation processing on each validation data in the original worst-case validation subset to obtain the target Gini coefficient corresponding to each validation data; taking the target Gini coefficients that do not exceed a second set tolerance threshold among all target Gini coefficients as the final Gini coefficients; and constructing the worst-case validation subset in the training dataset from the validation data corresponding to all final Gini coefficients and the validation data corresponding to all final difference rates.
[0029] The specific values of the first set tolerance threshold and the second set tolerance threshold can be set by those skilled in the art according to actual needs.
[0030] The step of obtaining the worst validation subset from the training dataset is a fairness constraint process and can occur during the evaluation phase of the fair machine learning model.
[0031] The fairness constraint can also be the fairness value of the fair machine learning model. The fairness constraint is selected using a random sorting strategy to balance the objective and the constraint.
[0032] In this application, the embodiments aim for "accuracy + diversity" under the definition of the worst-case verification subset, while incorporating hard fairness constraints and using random sorting for constrained selection.
[0033] In other implementations, random sorting can be replaced by an adaptive penalty function, multi-armed gambling-style feasibility sampling, etc., as long as it can achieve "preserving exploration near the constraint boundary".
[0034] S2. The first target neural network set is trained using the training dataset to obtain the second target neural network set.
[0035] Further, the step of obtaining the first target neural network set includes: using Multiple neural networks are selected from the initial neural network set as the first target neural network set.
[0036] It is a Neural Architecture Search (NAS) algorithm.
[0037] The initial neural network set can be expressed as: The initial neural network set includes multiple primitive neural networks, each of which can be selected from multilayer perceptrons, convolutional neural networks, long short-term memory networks, gated recurrent units, temporal convolutional networks, or residual networks. All primitive neural networks in the initial neural network set can be the same type of network or different networks. The parameters of each primitive neural network can be determined according to the type of problem being solved. Each primitive neural network can also be understood as a base learner.
[0038] In other embodiments, each original neural network may also be other neural networks, which can be specifically set by those skilled in the art according to actual needs.
[0039] The first target neural network set can be described as follows The first target neural network set contains A descendant model, that is A neural network.
[0040] In other implementations, using The step of selecting multiple neural networks from the initial neural network set as the first target neural network set can be understood as adding Gaussian noise to each weight of each neural network in the initial neural network set to form the first target neural network set. The Gaussian noise is a value with a mean of 0 and a standard deviation of σ. Each weight of each neural network in the first target neural network set can be expressed as... ,in, It is the first neural network in the first target neural network set. Each weight; It is the first neural network in the initial neural network set. Each weight; It is the Gaussian noise.
[0041] Furthermore, the training dataset can be used to partially train the first target neural network set to obtain the second target neural network set.
[0042] The second target neural network set can be described as follows: .
[0043] Furthermore, the step of using the training dataset to partially train the first target neural network set to obtain the second target neural network set can be understood as follows: for each neural network in the first target neural network set, calculate the cross-entropy loss function value on the training dataset, adjust the parameters of each neural network in the first target neural network set based on the loss function value, and perform gradient update on each neural network in the first target neural network set to obtain the second target neural network set.
[0044] S3. The second target neural network set is filtered according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
[0045] The fair machine learning model computes accuracy and fairness on each dataset in the initial evaluation dataset. The accuracy value of the fair machine learning model is the worst-case accuracy performance across all datasets in the initial evaluation dataset. The fairness value of the fair machine learning model is the worst-case fairness performance across all datasets in the initial evaluation dataset.
[0046] Further, the step of filtering the second target neural network set according to preset rules to obtain a fair machine learning model includes: obtaining an initial neural network set; merging the initial neural network set and the second target neural network set to obtain a third target neural network set; and selecting multiple third target neural networks from the third target neural network set as the fair machine learning model.
[0047] Further, the step of selecting multiple third-objective neural networks from the set of third-objective neural networks as the fair machine learning model includes: using Multiple third-objective neural networks are selected from the set of third-objective neural networks as the fair machine learning model.
[0048] In some implementations, using The step of selecting multiple third-target neural networks from the set of third-target neural networks as the fair machine learning model can be understood as adding Gaussian noise to each weight of each third-target neural network in the set of third-target neural networks to form the fair machine learning model.
[0049] Through the above implementation methods, the worst validation subset of the training dataset is set as a set that has undergone fairness violation calculation. Therefore, the present application embodiment can improve the accuracy and fairness of the model and achieve the purpose of balancing the accuracy and fairness of the model prediction.
[0050] In some implementations, prior to the step of selecting multiple third-target neural networks from the set of third-target neural networks as the fair machine learning model, the ensemble training method for class imbalance-oriented co-evolutionary fair machine learning models further includes: evaluating the set of third-target neural networks to obtain evaluation results.
[0051] Further, the step of evaluating the third target neural network set to obtain the evaluation result includes: performing an equalization accuracy evaluation on the third target neural network set to obtain a target equalization accuracy; performing an equalization chance evaluation on the third target neural network set to obtain a target equalization chance; and using the target equalization accuracy and the target equalization chance as the evaluation result.
[0052] Balanced Accuracy (BA) is used to encourage balanced performance across the four subgroups (true labels × sensitive attributes).
[0053] Equalized Odds (EO) measures the equality of chance under population conditions by the difference between the true positive rate and the false positive rate.
[0054] Furthermore, the step of evaluating the third target neural network set to obtain the evaluation result further includes: calculating the geometric mean score of the third target neural network set (i.e., The evaluation process is performed to obtain the target geometric mean score; the target equilibrium accuracy, the target equalization chance, and the target geometric mean score are used as the evaluation results.
[0055] Furthermore, embodiments of this application provide a formula for calculating the unfairness index: ; ; ; ; .
[0056] The above formula can be understood as an indicator of the degree of unfairness of an algorithm when comparing two groups (such as men and women), determining which group it is "better" and which group it is "worse." In the above formula, Represents the true situation; This means that the user "should have been approved" in the first place (e.g., they actually have the ability to repay). This means that the user in question "should not have been approved" in the first place. This represents the prediction result of the fair machine learning model (i.e., the personal credit approval result). The representative approved it; It represents sensitive attributes, such as gender, race, etc. Representing women; Representing the male population; The true rate representing the female population; The true rate representing the male population; The false positive rate representing the female population; The false positive rate representing the male population. The representative's approval was not granted. The closer the value is to 0, the closer the two groups are in terms of "good people pass rate" and "bad people are mistakenly passed rate", and the fairer the system is. The larger the value, the greater the difference in treatment between the two groups, and the more unfair the treatment.
[0057] The true pass rate refers to the percentage of men or women who should have passed the test but were correctly judged as approved by the fair machine learning model.
[0058] The false positive rate refers to the percentage of men or women who should not have passed the application but were correctly judged as failing the application by the fair machine learning model.
[0059] In practical applications, the above formula can be used in automated decision-making systems to predict whether to grant a loan (1 = grant, 0 = no), or whether to admit a student, or whether further review is required.
[0060] In multiple evaluations, the target equivalence opportunity of the first-place candidate was approximately higher than that of the second-place candidate. The target geometric mean of the first-place team can be improved by approximately [percentage missing] compared to the target geometric mean of the second-place team. The target equilibrium accuracy of the first-place player is approximately [missing information] compared to the target equilibrium accuracy of the second-place player. The fluctuations, therefore, reflect "a significant improvement in fairness and a basic maintenance of accuracy".
[0061] Furthermore, the step of evaluating the third target neural network set and obtaining the evaluation result also includes evaluating the third target neural network set using an objective function to obtain the evaluation result.
[0062] The objective function is evaluated as "the minimum accuracy in the current evaluation set family" plus "a diversity term (based on the idea of negative correlation learning, a measure of the difference between the base model and the current ensemble prediction)".
[0063] In some implementations, the method for ensemble training of class-imbalanced co-evolutionary fair machine learning models further includes: setting the validation dataset of the training dataset as the initial evaluation dataset; performing evolutionary processing on the initial evaluation dataset to obtain the target evaluation dataset; and updating the training dataset based on the target evaluation dataset.
[0064] The initial evaluation dataset can be expressed as: .
[0065] Furthermore, the step of performing evolutionary processing on the initial evaluation dataset to obtain the target evaluation dataset includes: expanding the initial evaluation dataset to obtain a dataset to be processed; and performing mutation evolution processing on the dataset to be processed to obtain the target evaluation dataset. These steps can be understood as a co-evolutionary mechanism for the evaluation dataset. This co-evolutionary mechanism uses a non-fixed validation set and adaptively generates a fair key subset through "fragment exchange mutation + worst-case scoring" to continuously expose the weak areas of the current model set.
[0066] The purpose of performing mutation and evolution processing on the evaluation dataset to be processed in this embodiment is to gradually move the "validation set family" toward a subset that can most clearly expose the unfairness of the current model.
[0067] This application embodiment also employs a dataset scoring function to calculate "" for each validation subset on the current base model set. The maximum value of "" is used as the worst supervision signal for the model for that subset. In other implementations, the "" of DataEval can also be used. "Replacing it with other conservative risk measures (such as upper quantiles or conditional risk), or introducing other fairness indicators (such as error rate difference) for compounding, can still maintain the idea of 'worst-case assessment driven'."
[0068] Furthermore, the step of expanding the initial evaluation dataset to obtain the evaluation dataset to be processed is to expand the length of the initial evaluation dataset to a first preset value, thereby obtaining the evaluation dataset to be processed.
[0069] The first preset value is expressed as The specific value of the first preset value can be set by those skilled in the art according to actual needs.
[0070] Further, the step of performing mutation and evolutionary processing on the dataset to be processed to obtain the target dataset includes: The initial iteration count is 1; If the number of iterations is less than or equal to a second preset value, the following steps are executed cyclically: Select an evaluation data point from the initial evaluation dataset as the evaluation data to be mutated; mutate the evaluation data to be mutated to obtain mutated evaluation data; perform calculations on the evaluation data to be mutated and the initial neural network set using a preset framework to obtain a first calculation result; perform calculations on the mutated evaluation data and the initial neural network set using the preset framework to obtain a second calculation result; if the second calculation result is greater than the first result, perform calculations on the evaluation data to be mutated, the mutated evaluation data, and the initial evaluation dataset using a first preset algorithm to obtain a modified initial evaluation dataset; use the modified initial evaluation dataset as the initial evaluation dataset; and increment the number of iterations by 1. If the number of iterations exceeds the second preset value, the above loop ends, and the initial evaluation dataset modified by the above loop is used as the target evaluation dataset.
[0071] The preset framework is the DataEval function.
[0072] The specific value of the second preset value can be set by those skilled in the art according to actual needs.
[0073] The data to be evaluated for mutation is a set.
[0074] Further, the step of performing mutation processing on the data to be mutated to obtain mutated evaluation data includes: randomly dividing the data to be mutated into... Multiple target folds are obtained by folding; one fold is selected from the multiple target folds as the fold to be processed; the fold to be processed is calculated and processed using a second preset algorithm to obtain the post-mutation evaluation data.
[0075] The multiple target folds can be expressed as: .
[0076] Furthermore, the calculation formula for the step of using the second preset algorithm to process the fold to be processed and obtaining the post-mutation evaluation data is as follows: ; in, These are the post-mutation evaluation data; This is data for the evaluation of variants; It is the fold to be processed, that is, the fold among the multiple target folds. One fold; It is a validation dataset.
[0077] In addition to folded exchange, the embodiments of this application may also employ stratified sampling, priority sampling of difficult samples, and subset replacement folded exchange based on uncertainty.
[0078] Furthermore, the calculation formula for the step of using the first preset algorithm to process the data to be mutated, the mutated evaluation data, and the initial evaluation dataset to obtain the modified initial evaluation dataset is as follows: ; in, This is the modified initial evaluation dataset; This is the initial evaluation dataset; It is the data to be evaluated for the variation; These are the post-mutation evaluation data.
[0079] In other embodiments, the step of "using a preset framework to perform calculations on the data to be mutated and the fair machine learning model to obtain a first calculation result" can be used instead of the step of "using a preset framework to perform calculations on the data to be mutated and the initial neural network set to obtain a first calculation result".
[0080] In other embodiments, the step of "using the preset framework to perform calculations on the post-mutation evaluation data and the fair machine learning model to obtain a second calculation result" can be used instead of the step of "using the preset framework to perform calculations on the post-mutation evaluation data and the initial neural network set to obtain a second calculation result".
[0081] Furthermore, the step of updating the training dataset based on the target evaluation dataset includes: modifying the initial evaluation dataset in the training dataset to the target evaluation dataset to form a new training dataset.
[0082] In other embodiments, if the verification dataset is not part of the initial evaluation dataset, the first calculation result can be replaced by the verification dataset in the above loop.
[0083] The embodiments of this application employ the alternating collaboration of two links: dataset evolution improves the discriminative power of the "evaluation target," and model evolution moves towards a more robust direction under fairness constraints. The two reinforce each other in a cyclical manner, taking into account both the improvement of fairness and the maintenance of accuracy.
[0084] The embodiments of this application can also use data resampling, reweighting, or post-processing threshold adjustment to approximate the fairness index.
[0085] The embodiments of this application can maximize accuracy in the sense of the "worst validation subset" through alternating co-evolution, while encouraging prediction differences among base models with diversity regularization terms and restricting group differences with fairness constraints, ultimately obtaining an ensemble model with optimal accuracy-fairness balance on the test set.
[0086] refer to Figure 2 The diagram shown is a principle block diagram of an ensemble training device for class imbalance-oriented co-evolutionary fair machine learning models, provided in the second aspect of an embodiment of this application. Figure 2 The ensemble training device 100 for class-imbalanced co-evolutionary fair machine learning models includes: The acquisition module 101 is used to acquire the training dataset, wherein the worst validation subset of the training dataset is a set that has undergone fairness violation computation. Training module 102 is used to train the first target neural network set using the training dataset to obtain the second target neural network set; The filtering module 103 is used to filter the second target neural network set according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
[0087] A third aspect of this application provides a method for predicting personal credit approval, comprising: obtaining a user's credit business record information; inputting the credit business record information into a fair machine learning model for prediction processing to obtain a personal credit approval result.
[0088] The personal loan approval result is either approved or rejected.
[0089] A fourth aspect of this application provides a terminal device, the schematic diagram of which can be as follows: Figure 3As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface of the terminal device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an ensemble training method for class imbalance-oriented co-evolutionary fair machine learning models. The display screen can be a liquid crystal display (LCD) or an e-ink display. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.
[0090] When the computer program is executed by the processor, it can also implement a predictive method for personal credit approval.
[0091] Those skilled in the art will understand that Figure 3 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.
[0092] In some embodiments, this application provides a terminal device including a processor and a memory for storing computer programs. The processor is used to call and run the computer programs stored in the memory to perform the steps of the class imbalance-oriented co-evolutionary fair machine learning model ensemble training method provided in the first aspect of this application.
[0093] The processor can call and run the computer program stored in the memory, and can also execute the steps of the prediction method for personal credit approval provided in the third aspect of the embodiments of this application.
[0094] A fifth aspect of this application provides a computer-readable storage medium for storing a computer program that causes a computer to perform the steps of the class-imbalanced co-evolutionary fair machine learning model ensemble training method provided in the first aspect of this application.
[0095] The computer program can also cause the computer to perform the steps of the prediction method for personal credit approval provided in the third aspect of the embodiments of this application.
[0096] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0097] The technical features of the above embodiments can be combined without changing the basic principles of this application. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0098] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the patent protection scope of this application should be determined by the appended claims.
Claims
1. A method for ensemble training of class-imbalanced, co-evolutionary, fair machine learning models, characterized in that, include: Obtain a training dataset, wherein the worst-case validation subset of the training dataset is a set that has undergone fair violation computation. The first target neural network set is trained using the training dataset to obtain the second target neural network set; The second target neural network set is filtered according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
2. The ensemble training method for class-imbalanced co-evolutionary fair machine learning models according to claim 1, characterized in that, The steps for filtering the second target neural network set according to preset rules to obtain a fair machine learning model include: Obtain the initial set of neural networks; The initial neural network set and the second target neural network set are merged to obtain the third target neural network set; Multiple third-objective neural networks are selected from the set of third-objective neural networks as the fair machine learning model.
3. The ensemble training method for class-imbalanced, co-evolutionary, fair machine learning models as described in claim 2, characterized in that, Prior to the step of selecting multiple third-objective neural networks from the set of third-objective neural networks as the fair machine learning model, the ensemble training method for class imbalance-oriented co-evolutionary fair machine learning models further includes: The third target neural network set is evaluated to obtain the evaluation result.
4. The ensemble training method for class-imbalanced, co-evolutionary, fair machine learning models as described in claim 3, characterized in that, The steps for evaluating the third target neural network set and obtaining the evaluation results include: The third target neural network set is subjected to a balanced accuracy evaluation process to obtain the target balanced accuracy. The third target neural network set is subjected to equalization opportunity evaluation processing to obtain target equalization opportunities; The target equalization accuracy and the target equalization opportunity are used as the evaluation results.
5. The ensemble training method for class-imbalanced, co-evolutionary, fair machine learning models as described in claim 4, characterized in that, The step of evaluating the third target neural network set to obtain the evaluation result further includes: The geometric mean score is evaluated on the third target neural network set to obtain the target geometric mean score; The target equilibrium accuracy, the target equalization chance, and the target geometric mean score are used as the evaluation results.
6. The ensemble training method for class-imbalanced, co-evolutionary, fair machine learning models as described in claim 1, characterized in that, The co-evolutionary fair machine learning model ensemble training method for class imbalance also includes: Set the validation dataset of the training dataset as the initial evaluation dataset; The initial evaluation dataset is subjected to evolutionary processing to obtain the target evaluation dataset; The training dataset is updated based on the target evaluation dataset.
7. The ensemble training method for class-imbalanced co-evolutionary fair machine learning models according to claim 6, characterized in that, The steps for performing evolutionary processing on the initial evaluation dataset to obtain the target evaluation dataset include: The initial evaluation dataset is expanded to obtain the evaluation dataset to be processed; The dataset to be evaluated is subjected to mutation and evolution processing to obtain the target evaluation dataset.
8. A co-evolutionary fair machine learning model ensemble training device for class imbalance, characterized in that, include: The acquisition module is used to acquire the training dataset, wherein the worst-validation subset of the training dataset is a set that has undergone fairness violation computation. The training module is used to train the first target neural network set using the training dataset to obtain the second target neural network set. The filtering module is used to filter the second target neural network set according to preset rules to obtain a fair machine learning model. The fair machine learning model is applied to the personal credit approval application scenario of the bank to predict the credit status of users.
9. A terminal device, characterized in that, include: A processor and a memory, the memory for storing a computer program, the processor for calling and running the computer program stored in the memory, to perform the steps of the class imbalance-oriented co-evolutionary fair machine learning model ensemble training method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the steps of the ensemble training method for class imbalance-oriented co-evolutionary fair machine learning models as described in any one of claims 1 to 7.