A Fairness-Enhanced Classification Method Based on Generative Adversarial Networks

Abstract

The present invention provides a fairness-enhanced classification method based on a generative adversarial network, comprising the following steps: Step 1: Preprocess the original data set to obtain a preprocessed data set; Step 2: Construct a latent factor data set; Step 3: Establish a generative adversarial network for enhancing fair classification; Step 4: Synthesize fair data and output a fair classifier. Applying this technical solution can not only generate fair data while ensuring good data utility, but also achieve classification fairness while ensuring good classification utility.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a fairness-enhanced classification method based on a generative adversarial network. Background Art

[0002] In the past few decades, with the popularization of AI and machine learning and their prevalence in different applications, security and fairness constraints have become a huge problem faced by researchers and engineers. It is used in different medical fields, child welfare systems, and autonomous vehicles. All these applications have a direct impact on our lives, and considering fairness, if not properly designed, they may harm our society. The ways in which these applications and these AI systems affect our daily lives through their inherent biases, such as AI chatbots, employment matching, flight routes, and the biases existing in advertising display position algorithms. Therefore, for researchers and engineers, when modeling an algorithm or system, they must pay attention to downstream applications and their potential harmful effects.

[0003] The prediction of bias stems from the hidden or ignored biases in the data or algorithms. The reason for the existence of such social biases is that the training sets used by machine learning models are either sampled from real life or synthetic data that simulates real samples. These data implicitly contain people's preferences or biases towards certain groups to varying degrees, and the evaluation metrics used during training will cause the model to amplify these preferences or biases, resulting in the prediction models trained from these data sets exhibiting biased behaviors and giving different prediction tendencies to different populations.

[0004] Recently, generative adversarial networks (GANs) have shown satisfactory results in simulating real data distributions and generating high-quality synthetic data similar to real data. Synthetic data can be used for model training, solving the problem of data shortage. However, due to the high similarity between real data and synthetic data, if the real data is biased, the predictive analysis based on synthetic data will also be biased. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a fairness-enhanced classification method based on a generative adversarial network, which can not only generate fair data while ensuring good data utility, but also achieve fairness in classification while ensuring good classification utility.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: A fairness-enhanced classification method based on a generative adversarial network, comprising the following steps:

[0007] Step 1: Preprocess the original data set to obtain a preprocessed data set;

[0008] Step 2: Construct the latent factor dataset:

[0009] Step 3: Establish a generative adversarial network for enhancing fair classification:

[0010] Step 4: Synthesize fair data and output a fair classifier.

[0011] In a preferred embodiment, the specific steps of Step 1 include:

[0012] Step 11: Clean the dataset, including handling missing values, null values, and removing redundancy;

[0013] Step 12: Perform one-hot encoding on the categorical attributes in the dataset obtained after the processing of Step 11;

[0014] Step 13: Normalize the continuous attributes in the dataset obtained after the processing of Step 11.

[0015] In a preferred embodiment, the specific steps of Step 2 include:

[0016] Step 21: First, screen out the attribute groups in the original dataset that are helpful for label prediction;

[0017] Step 22: Then, select the attributes with relatively low correlation with the sensitive attributes from the attribute groups in Step 21 to form the latent factors c = {a1, a2... a n};

[0018] Step 23: Finally, retain the dimensions of the latent factors, and form the remaining dimensions with random noise z to construct the latent factor dataset.

[0019] In a preferred embodiment, the specific steps of Step 3 include:

[0020] Step 31: Construct a generator G and a discriminator D, and then construct the adversarial loss function V(G, D) for these two:

[0021]

[0022] Among them, x represents the non-sensitive attribute group, y represents the data label, s represents the sensitive attribute, and P real (x|y, s) represents the probability distribution of the data on the real dataset, (x|y, s) ~ P real (x|y, s) means that the data (x|y, s) follows the probability distribution P real (x|y, s), represents calculating the function expectation for each training data (x|y, s) that follows the distribution P real (x|y, s); x g represents the fake data synthesized by the generator, and P g(x|y,s) represents the probability distribution of data on the synthetic dataset, (x g |y,s) ~ P g (x|y,s) represents that the synthetic data (x g |y,s) follows the probability distribution P g (x|y,s), represents calculating the function expectation for each synthetic data (x g |y,s) that follows the distribution P g |y,s); D(*) represents the probability that the discriminator discriminates * as true;

[0023] Step 32: Construct the adversary A and the classifier C, and then construct the adversarial loss function V(C,A) of these two:

[0024]

[0025] where s is regarded as binary, taking values in {0,1}; y c represents the predicted label of the classifier, and A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s;

[0026] Step 33: Construct the adversarial loss function V(G,A) of the adversary A and the generator G:

[0027]

[0028] where A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s;

[0029] Step 34: Construct the collaborative optimization function L(G,C) of the generator G and the classifier C:

[0030]

[0031] where C(*) represents the probability that the classifier predicts * as the correct class label y;

[0032] Step 35: According to the adversarial loss function V(G,D) of the generator G and the discriminator D, the adversarial loss function V(C,A) of the adversary A and the classifier C, and the adversarial loss function V(G,A) of the adversary A and the generator G, construct the global optimization objective, that is, the generative adversarial network for enhancing fair classification: V(G,C,D,A) = V(G,D) + λV(G,A) + μV(C,A) + L(G,A); where λ and μ are relative importance parameters that control V(G,C,D,A).

[0033] In a preferred embodiment, the specific steps of step 4 include:

[0034] Set the maximum number of iterations as N and the iteration stop threshold as k. Iteratively train the generative adversarial network constructed in step 3 using the original dataset and the latent factor dataset until the function value of the current iteration is less than the threshold k or the maximum number of iterations N is reached, and output the generated fair data and the fair classifier.

[0035] In a preferred embodiment, the random noise z in step 23 is generated by a Gaussian / Normal distribution N(0, 1).

[0036] Compared with the prior art, the present invention has the following beneficial effects: 1. In the adversarial training of the discriminator, generator, and classifier, by simultaneously learning the generative model and the classifier, the performance of the model is superior to that of an independent generative model or an independent classifier. The generator can generate fair data while ensuring good data utility; the classifier can achieve classification fairness while ensuring good classification utility. 2. Instead of directly encoding the original dataset fairly or using random noise as the model input to generate highly imitated fair samples, the method of using non-sensitive attributes that are helpful for label prediction and have little correlation with sensitive attributes as latent factors and filling the remaining dimensions with random noise is adopted, which can reduce the bias brought to the model by directly encoding the original dataset and can also improve the model accuracy and training effect. At the same time, it is easier to balance classification fairness and accuracy during the model training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flowchart of the implementation of the preferred embodiment of the present invention;

[0038] Figure 2 is a schematic diagram of the overall structure of the network of the preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs.

[0041] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] A fairness-enhanced classification method based on generative adversarial networks, referring to Figures 1 to 2 , mainly including the following steps:

[0043] Step 1: Preprocess the original dataset to obtain the preprocessed dataset;

[0044] (1.1) Clean the dataset, including handling missing values, null values, and removing redundancy;

[0045] (1.2) Perform one-hot encoding on the categorical attributes in the dataset obtained after processing in step (1.1);

[0046] (1.3) Normalize the continuous attributes in the dataset obtained after processing in step (1.1);

[0047] Step 2: Construct the latent factor dataset:

[0048] (2.1) First, screen out the attribute groups in the original dataset that are helpful for label prediction;

[0049] (2.2) Then, select the attributes with low correlation with sensitive attributes from the attribute groups in (2.1) to form the latent factor c = {a1, a2…a n};

[0050] (2.3) Finally, retain the dimension of the latent factor, and the remaining dimensions are composed of random noise z to construct the latent factor dataset. The random noise z is generated by the Gaussian / Normal distribution N(0, 1).

[0051] Step 3: Establish a generative adversarial network for enhancing fair classification:

[0052] (3.1) Construct the generator G and the discriminator D, and then construct the adversarial loss function V(G, D) for these two:

[0053]

[0054] Among them, x represents the non-sensitive attribute group, y represents the data label, s represents the sensitive attribute, P real (x|y, s) represents the probability distribution of the data on the real dataset, (x|y, s) ~ P real (x|y, s) represents that the data (x|y, s) follows the probability distribution P real (x|y, s), represents calculating the function expectation for each training data (x|y, s) that follows the distribution P real (x|y, s); x g represents the fake data synthesized by the generator, P g (x|y, s) represents the probability distribution of the data on the synthetic dataset, (xg (x|y,s) ~ P g (x|y,s) represents the synthetic data (x g |y,s) that follows the probability distribution P g (x|y,s), represents calculating the expected value of the function for each synthetic data (x g |y,s) that follows the distribution P; D(*) represents the probability that the discriminator classifies * as true. g |y,s) calculates the expected value of the function; D(*) represents the probability that the discriminator classifies * as true.

[0055] (3.2) Construct the adversary A and the classifier C, and then construct the adversarial loss function V(C,A) for these two:

[0056]

[0057] where, for ease of representation, s is regarded as binary and takes values in {0,1}. y c represents the predicted label of the classifier, and A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s.

[0058] (3.3) Construct the adversarial loss function V(G,A) for the adversary A and the generator G:

[0059]

[0060] where, A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s.

[0061] (3.4) Construct the collaborative optimization function L(G,C) for the generator G and the classifier C:

[0062]

[0063] where, C(*) represents the probability that the classifier classifies * as the correct class label y.

[0064] (3.5) Based on the adversarial loss function V(G,D) of the generator G and the discriminator D, the adversarial loss function V(C,A) of the adversary A and the classifier C, and the adversarial loss function V(G,A) of the adversary A and the generator G, construct the global optimization objective, i.e., the generative adversarial network for enhancing fair classification: V(G,C,D,A) = V(G,D) + λV(G,A) + μV(C,A) + L(G,A). Here, λ and μ are relative importance parameters that control V(G,C,D,A).

[0065] Step 4, Synthesize fair data and output a fair classifier:

[0066] Set the maximum number of iterations to N and the iteration stop threshold to k. Iteratively train the generative adversarial network constructed in step (3) using the original dataset and the latent factor dataset until the function value of the current iteration is less than the threshold k or the maximum number of iterations N is reached, and output the synthesized fair data and the fair classifier.

[0067] By combining the fair training of the classification model and the generative model, the generator and the classifier are co-trained so that the model can better achieve fair classification. In model training, through the mutual game between the generator and the discriminator, the encoded samples do not contain sensitive information, forming fair data; through the game training between the classifier and the discriminator, the predicted labels of the classifier do not encode sensitive information, achieving universal classification fairness. To reduce the impact of adding fairness constraints on the model accuracy and eliminate potential biases caused by the correlation between attributes, the input data of the model is preprocessed, and non-sensitive attributes that are helpful for label prediction and have little correlation with sensitive information are retained to form latent factors, and the remaining dimensions are filled with random noise.

[0068] Refer to Appendix Figures 1-2 The fair-enhanced classification method based on the generative adversarial network proposed by the present invention mainly includes four independent network structures: a generator, a discriminator, a classifier, and a discriminator.

[0069] Generator: Given a random noise z and a latent factor c, each synthesized sample (x g |y,s) has a corresponding sensitive attribute s ∼ P real (s) and a label y ∼ P real (y), that is, the synthesized data is (x g |y,s) ∼ P g (x|y,s). The goal of the generator is to ensure that the synthesized data (x g |s,y) is close to the real data and can be correctly classified as y, but the sensitive attribute s of the user cannot be inferred from it. Among them, the latent factor c is composed of non-sensitive attributes with low correlation with sensitive attributes, which reduces the probability of encoding sensitive information in the generated data while ensuring good classification accuracy.

[0070] Discriminator: mainly distinguish the real data from P real (x|y,s) and the synthesized data from P g (x|y,s).

[0071] Classifier: mainly perform fair and correct label classification on the synthesized data (x g |y,s) and the real data (x|y,s).

[0072] Discriminator: from the synthesized data (x g |y,s) and the prediction result y of the classifierc derive the user's true sensitive attribute s.

[0073] In the collaborative adversarial training among the discriminator, generator, and classifier, improving the utility of the generator can enhance the utility of the classifier in making correct predictions. Improving the classifier can enhance the performance of the generator in generating more realistic samples for each class. In the adversarial training among the adversary, generator, and classifier, improving the fairness of the generator can enhance the fair classification ability of the classifier. Improving the fairness of the classifier can enhance the ability of the generator to generate fair data.

[0074] The following combines experimental examples to further illustrate the effects of the present invention.

[0075] Fairness measurement criteria:

[0076] For a labeled dataset D(X, Y, S), it contains a set of non-sensitive attributes X ∈ R n , class labels Y ∈ {0, 1}, and sensitive attributes S ∈ {0, 1}.

[0077] ∈-Fairness: Given a data distribution D(X, Y, S), for any classifier C: X → S, if BER(C(X), S) > ∈, then the data distribution can be considered to satisfy ∈-Fairness. ∈-Fairness is used to measure the potential bias caused by the correlation between non-sensitive attributes X and sensitive attributes S. Among them, BER is the average class-conditional error of the data distribution (X, Y, S). It can be expressed as the following equation:

[0078] BER(C(X, S)) = [P(C(X) = 0|S = 1) + P(C(X) = 1|S = 0)] / 2

[0079] Demographic Parity: For a classifier C trained based on the data distribution D(X, Y, S), if the classification result C(X) is independent of the sensitive attribute S, that is, P(C(X) = 1|S = 1) = P(C(X) = 1|S = 0), then the classifier C can be considered to satisfy Demographic Parity. It can be expressed as the following inequality:

[0080] |P(C(X) = 1|S = 1) - P(C(X) = 1|S = 0)| ≤ τ

[0081] where the τ threshold is used as a fairness constraint.

[0082] For a classifier C trained based on the data distribution D(X, Y, S), if the classification result C(X) is conditionally independent of the sensitive attribute S given the label Y, that is, P(C(X) = 1|Y = y, S = 1) = P(C(X) = 1|Y = y, S = 0), then the classifier C can be considered to satisfy Equalized Odds. It can be expressed as the following inequality:

[0083] |P(C(X) = 1|Y = y, S = 1) - P(C(X) = 1|Y = y, S = 0)| ≤ τ, y ∈ {0, 1}

[0084] where the τ threshold is used as a fairness constraint.

[0085] Dataset preprocessing:

[0086] Experimental dataset 1 is sourced from UCI Adult Income, which has severe gender bias. This dataset records the basic information of US residents at that time and their corresponding annual incomes, and can be used to predict whether a person's annual salary exceeds 50K. The dataset contains 48,842 (45,222 after data cleaning) samples, each sample consisting of 14 attributes. The sensitive attribute is gender, where males are the protected group and females are the unprotected group. The positive class label is that a person's annual income exceeds 50K.

[0087] In the data preprocessing stage, attributes that are helpful for label prediction and have a low correlation with the sensitive attribute need to be selected to form latent factors. Therefore, the F-value of each attribute with respect to the label is calculated using analysis of variance, and the attributes are sorted according to the magnitude of their F-values. The F-value of an attribute can be used to judge the help of the attribute for predicting the class. The larger the F-value, the stronger the predictive ability of the attribute and the greater its correlation with the label. Secondly, the Kendall correlation coefficient matrix is calculated to measure the correlation between attributes.

[0088] For the Adult Income dataset, first, based on the F-value and correlation matrix, select the attribute groups that are helpful for label prediction, which are: {'educational-num','relationship', 'age', 'hours-per-week', 'capital-gain','marital-status', 'capital-loss', 'education', 'race'}. Then, remove the attributes in the attribute group whose correlation with the sensitive attribute is greater than or equal to its correlation with the label, which are: {'relationship', 'hours-per-week', 'race'}. Finally, select the latent factor c composed of 6 non-sensitive attributes, that is, c = {'age', 'education-num', 'capital-gain', 'capital-loss', 'education','marital-status'}.

[0089] Before model training, since each sample in the dataset contains continuous and categorical attributes, one-hot dimensionality expansion encoding is performed on each sample attribute in the original dataset: one-hot encoding is performed on each categorical attribute, while the continuous attribute is normalized to limit the value range to the interval [0, 1]. Similarly, for the latent factor dataset, one-hot encoding is performed on the attributes of the corresponding dimensions of the samples in the original dataset according to the attributes in the latent factor c, and the remaining dimensions are composed of random noise z.

[0090] Experimental configuration:

[0091] In the experiment, the generator of the adversarial network model uses two hidden layers, with 128 neurons in each hidden layer; the discriminator and the adversary each use one hidden layer, with 256 neurons; the classifier uses two hidden layers, with 64 neurons in the first layer and 32 neurons in the second layer; the ReLU function is used as the activation function for the hidden layers, and the Sigmoid function is used as the activation function for the output layer. The number of iterations epoches for model training is 3000 times, the number of samples batchsize used each time is 128, and the learning rate is set to e-4. For the value of each hyperparameter, we conduct 10 experimental trainings, remove the best and worst cases, and take the average of the results. The software and hardware configurations used in the experiment are shown in Table 1.

[0092] Table 1 Software and hardware configurations

[0093] Central Processing Unit Memory Graphics Card Operating System CUDA Programming Language Deep Learning Framework Inteli5 32GB RTX3080 Win10 Cuda11.7 Python3.9 Pytorch

[0094] Experimental results and conclusions:

[0095] To test the quality of the fair data generated by this method, in this experiment, the Euclidean distance between the synthetic dataset and the real dataset is calculated to measure the utility of the synthetic dataset, and the calculation method is as follows:

[0096] ED(X,S) = ||P real (X,S) - P g (X,S)||2.

[0097] In addition, ∈-Fairness is used as the measurement standard for the fairness of the dataset, and the calculation method is as follows:

[0098] BER(C(X,S)) = [P(C(X)=0|S = 1) + P(C(X)=1|S = 0)] / 2.

[0099] The experimental results are shown in Table 2. The proposed method in the table is the method used in the present invention.

[0100] Table 2 Performance of Synthetic Datasets by Different Methods

[0101]

[0102]

[0103] From the above experimental results, it can be concluded that: on the premise of maintaining good data utility, the data generated by the method proposed in the present invention has a higher error rate than the FairGAN + method when predicting the sensitive attribute s. This is because the model eliminates the sensitive attribute and the non-sensitive attributes related to the sensitive attribute during input, so the probability of the generated data encoding the sensitive attribute is lower and the fairness is better. In addition, in terms of the accuracy of classifying the generated data, since the generator of the method proposed in the present invention retains the non-sensitive attributes that are helpful for label prediction during input, it has a higher classification accuracy than the FairGAN + method that uses noise as the input of the generated data.

[0104] To test the classification effect of this method, in this experiment, the prediction accuracy Accuracy of the trained model on the real dataset and the fairness criteria Demographic Parity and Equalized Odds are used as the measurement standards. Among them, the measurement standards for the fair classifier are calculated as follows:

[0105] (1) DP(C) = |P(C(X)=1|S = 1) - P(C(X)=1|S = 0)|

[0106] (2) Eos1(C) = |P(C(X)=1|Y = 1,S = 1) - P(C(X)=1|Y = 1,S = 0)|

[0107] (3) Eos2(C) = |P(C(X) = 1|Y = 0, S = 1) - P(C(X) = 1|Y = 0, S = 0)|

[0108] The experiment uses the classifier trained on the dataset without any processing as the Baseline, and conducts multiple experimental result comparisons with the FairGan + method. The experimental results are shown in Table 3. The proposed method in the table is the method used in the present invention.

[0109] Table 3 Performance of Classifiers Trained by Different Methods

[0110]

[0111]

[0112] From the above experimental results, it can be concluded that: the method proposed in the present invention has better classification fairness than the FairGAN + method as a whole, and in terms of classification accuracy, both can maintain good classification utility.

[0113] The analysis of all the above experimental data proves the correctness and effectiveness of the method proposed in the present invention.

[0114] The beneficial effects of the present invention are as follows: 1. In the adversarial training of the discriminator, generator and classifier, by simultaneously learning the generative model and the classifier, the performance of the model is better than that of an independent generative model or an independent classifier. The generator can generate fair data while ensuring good data utility; the classifier can achieve classification fairness while ensuring good classification utility. 2. Instead of directly performing fair encoding on the original dataset or using random noise as the model input to generate highly imitated fair samples, the method of taking non-sensitive attributes that are helpful for label prediction and have little correlation with sensitive attributes as latent factors and filling the remaining dimensions with random noise is adopted, which can reduce the bias brought to the model by directly encoding the original dataset, and can also improve the model accuracy and training effect. At the same time, it is easier to balance classification fairness and accuracy during the model training process.

[0115] The above specific implementation manners further elaborate in detail the purpose, technical solution and beneficial effects of the present invention. It should be understood that the above are only the specific implementation manners of the present invention, and do not limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fairness-enhanced classification method based on generative adversarial networks, characterized in that, It includes the following steps: Step 1: Preprocess the original dataset to obtain the preprocessed dataset; Step 2: Construct the latent factor dataset: Step 3: Establish a generative adversarial network for enhancing fair classification: Step 4: Synthesize fair data and output a fair classifier; The specific content of Step 2 includes: Step 21: Screen out the attribute groups in the original dataset that are helpful for label prediction; Step 22: Select attributes with relatively low correlation with sensitive attributes from the attribute group in Step 21 to form the implicit factor c = {a1, a2…a n}; Step 23: Retain the dimensions of the latent factors, and the remaining dimensions are composed of random noise z to construct the latent factor dataset; The specific content of Step 3 includes: Step 31: Construct a generator G and a discriminator D, and then construct the adversarial loss function V(G,D) of these two: Among them, x represents the non-sensitive attribute group, y represents the data label, s represents the sensitive attribute, and P real (x|y,s) represents the probability distribution of data in the real data set, (x|y,s)~P real (x|y,s) means that the data (x|y,s) follows the probability distribution P real (x|y,s), Represents the distribution P real Calculate the function expectation for each training data (x|y,s); g represents the fake data synthesized by the generator, P g (x|y,s) represents the probability distribution of data on the synthetic dataset, (x g |y,s)~P g (x|y,s) represents the synthetic data (x g |y,s) obeys the probability distribution P g (x|y,s), Represents the distribution P g Each synthetic data (x|y,s) g |y,s) calculates the expected function; D(*) represents the probability that the discriminator classifies * as true; Step 32: Construct an adversary A and a classifier C, and then construct the adversarial loss function V(C,A) of these two: Among them, s is regarded as binary, taking values in {0, 1}; y c represents the predicted label of the classifier, and A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s; Step 33: Construct the adversarial loss function V(G,A) of the adversary A and the generator G: Among them, A(*) represents the probability that the adversary A predicts * as the true sensitive attribute s; Step 34: Construct the collaborative optimization function L(G,C) of the generator G and the classifier C: Among them, C(*) represents the probability that the classifier predicts * as the correct class label y; Step 35: According to the adversarial loss function V(G,D) of the generator G and the discriminator D, the adversarial loss function V(C,A) of the adversary A and the classifier C, and the adversarial loss function V(G,A) of the adversary A and the generator G, construct a global optimization objective, that is, the generative adversarial network for enhancing fair classification: V(G,C,D,A)=V(G,D)+λV(G,A)+μV(C,A)+L(G,A); where λ and μ are relative important parameters that control V(G,C,D,A).

2. The fair-enhanced classification method based on a generative adversarial network according to claim 1, wherein The specific content of Step 1 includes: Step 11: Clean the dataset, including handling missing values, null values, and removing redundancy; Step 12: Perform one-hot encoding on the categorical attributes in the dataset obtained after Step 11; Step 13: Normalize the continuous attributes in the dataset obtained after Step 11.

3. The fair-enhanced classification method based on a generative adversarial network according to claim 1, wherein, The specific content of Step 4 includes: Set the maximum number of iterations to N and the stop iteration threshold to k. Iteratively train the generative adversarial network constructed in Step 3 using the original dataset and the latent factor dataset until the function value of the current iteration is less than the threshold k or reaches the maximum number of iterations N, and output the generated fair data and the fair classifier.

4. The fair-enhanced classification method based on a generative adversarial network according to claim 1, wherein, The random noise z in Step 23 is generated by the Gaussian / normal distribution N(0,1).

Images