Conditional table GAN establishment method based on interpretability
By introducing interpretability models and dynamic weights in the generative adversarial network, a conditional table GAN was established, which solved the problem of traditional models ignoring data relationships and lacking explanatory when generating tabular data, and achieved higher quality and interpretable data generation.
Patent Information
- Application Number
- CN202411855650.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-16
- Publication Date
- 2025-04-11
AI Technical Summary
When generating tabular data, the traditional generative adversarial network model ignores the complex relationships and potential dependencies between data features, resulting in the limitation of the quality and reliability of the generated data, and lacks explanatory power of the model output, making it difficult to evaluate the reliability of the generated data.
A conditional table GAN establishment method based on interpretability is adopted to form a conditional table GAN by constructing and training an interpretability pre-trained model and embedding it into the basic GAN framework. This method uses an interpretability model to calculate data characteristics and improves generation efficiency and data quality through dynamic weighting and high-quality noise.
It improves the quality and diversity of generated data, enhances the interpretability of data, reduces the distance correlation coefficient between generated data and real data (about 40%), and reaches a convergence state in a shorter time, improving data generation efficiency.
Smart Images
Figure CN119808845A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of data generation, and in particular relates to a method for establishing a conditional table GAN based on interpretability. Background Art
[0002] In recent years, Generative Adversarial Networks (GANs) have been widely used in the field of data generation and enhancement, especially in dealing with small samples and unbalanced data. However, when generating tabular data, traditional GAN models often ignore the complex relationships and potential dependencies between data features, which limits the quality and reliability of generated data, and the generated data samples cannot be correctly matched with the labels, which has a negative impact on the subsequent use of the generated data.
[0003] In addition, existing table generation methods usually lack the interpretability of model outputs, making it difficult for researchers and users to deeply understand the quality of generated data and its correspondence with actual data. This not only limits the evaluation of the reliability of generated data, but also increases the difficulty of promoting the model in practical applications, especially in fields that require high transparency and strong interpretability, such as medical, finance, and scientific research. Summary of the invention
[0004] In view of this, in order to solve the problems raised in the above background technology, the purpose of the present invention is to provide a conditional table GAN establishment method based on interpretability to solve the existing problems.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A method for establishing a conditional table GAN based on interpretability, comprising:
[0007] Build and train interpretable pre-trained models;
[0008] A basic GAN framework is constructed, and the trained pre-trained model is embedded into the basic GAN framework to obtain a conditional table GAN.
[0009] Furthermore, the conditional table GAN is trained using the original data set.
[0010] Preferably, the pre-training model is constructed based on a fuzzy clustering model.
[0011] Preferably, the training of the interpretable pre-trained model includes:
[0012] The features of each sample in the original data set are calculated through the interpretable model and aggregated to form a feature data set;
[0013] The pre-trained model is trained using the feature data set.
[0014] Preferably, the step of training the pre-trained model using the feature data set includes:
[0015] Dividing the feature data set into a training set and a test set;
[0016] Using the training set to train the pre-trained model;
[0017] The pre-trained model is tested using the test set.
[0018] Preferably, the step of calculating the features of each sample in the original data set by using an interpretable model includes:
[0019] Constructing a tree model about the original data set based on a machine learning algorithm;
[0020] Building a tree-type interpreter for the interpretability model based on the tree-type model;
[0021] The tree interpreter is used to calculate the features of each sample in the original data set.
[0022] Preferably, the basic GAN framework includes a generator and a discriminator.
[0023] Preferably, embedding the trained pre-trained model into the basic GAN framework comprises:
[0024] Constructing two paths for transmission from the generator to the discriminator;
[0025] The pre-trained model is embedded in one of the paths.
[0026] Preferably, the pre-trained model is used to assign dynamic weights to the discriminator according to the output of the generator.
[0027] Preferably, embedding the trained pre-trained model into the basic GAN framework further comprises:
[0028] The generator, discriminator and pre-trained model are respectively assigned balancing weights λ1, λ2, λ3, and λ1+λ2+λ3=1.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] (1) Improving the quality of generated data: By using an interpretable model, a machine learning algorithm, and a basic GAN framework, the conditional table GAN of the present invention can better control the categories of generated samples and improve the diversity of generated samples, thereby improving the quality of the final generated table data. The experimental results show that the distance correlation coefficient between the generated data of the conditional table GAN of the present invention and the real data is reduced by 40%.
[0031] (2) Enhanced data interpretability: The interpretability model calculates the contribution of each feature to data prediction based on game theory. The present invention uses the interpretability model to calculate the training data set of the pre-trained model, thereby increasing the interpretability of the conditional table GAN.
[0032] (3) Improving data generation efficiency: Through the interpretable dynamic weights and high-quality uniform random noise provided by the pre-trained model, the conditional table GAN of the present invention can reach a convergence state in a shorter time, thereby reducing the data generation time and improving the data generation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is the training flow chart of the pre-trained model in the present invention
[0034] Figure 2 This is a structural diagram of the basic GAN framework in the present invention;
[0035] Figure 3 This is a structural diagram of the conditional table GAN in the present invention;
[0036] Figure 4 This is an application architecture diagram of the conditional table GAN in the present invention;
[0037] Figure 5 Comparison chart of data generation experiments for basic GAN and conditional table GAN. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] A method for establishing a conditional table GAN based on interpretability, comprising:
[0040] S1. Build and train interpretable pre-trained models;
[0041] (1) Build a pre-trained model
[0042] The pre-training model is constructed based on a fuzzy clustering model.
[0043] Regarding the fuzzy clustering model, the present invention adds fuzzy properties to the clustering model, so that the output of the model can fully consider the uncertainty between data relationships, thereby improving the robustness of the model.
[0044] The fuzzy clustering model is based on multi-objective optimization, maximizing the separation between clusters of "good samples" and "bad samples" and minimizing the distance within clusters of samples of the same category. Such a setting will be beneficial to the clustering effect and thus provide a higher quality weight value.
[0045] (2) Training the pre-trained model
[0046] The training steps are as follows Figure 1 As shown:
[0047] Constructing a tree model about the original data set A based on a machine learning algorithm;
[0048] Building a tree-type interpreter for the interpretability model based on the tree-type model;
[0049] The tree interpreter is used to calculate the features of each sample in the original data set A, and the features are collected to form a feature data set D;
[0050] Divide the feature data set into a training set T and a test set V;
[0051] Using the training set T to train the pre-trained model;
[0052] The pre-trained model is tested using the test set V.
[0053] About the Explanable Model (SHAP): The Explanable Model is an explanation method based on game theory, which provides explanations by calculating the contribution of each feature to the prediction results. The Explanable Model assigns an importance score to each feature, which helps understand how the prediction results of the data are affected by different features. The Explanable Model can also improve the interpretability and transparency of data predictions, and can help users better understand the prediction results, both globally and locally, with high explanation accuracy and consistency.
[0054] S2. Construct a basic GAN framework, and embed the trained pre-trained model into the basic GAN framework to obtain a conditional table GAN.
[0055] like Figure 2As shown in the figure, the basic GAN framework includes a generator and a discriminator. Specifically, the random noise vector Z and the label vector y as conditional information are regarded as the input of the generator, and the generator generates a pseudo sample (x*|y,y) that can match the label y. The discriminator receives the labeled pseudo sample (x*|y,y) generated by the generator and the labeled real sample (x,y) in the training set T, and learns to distinguish whether the sample matches its label.
[0056] like Figure 3 As shown, embedding the trained pre-trained model into the basic GAN framework includes: constructing two paths for transmission from the generator to the discriminator, and embedding the pre-trained model into one of the paths. In this way, the pre-trained model used is used to assign dynamic weights to the discriminator according to the output of the generator.
[0057] In addition, balancing weights λ1, λ2, λ3 are respectively assigned to the generator, discriminator and pre-trained model, and λ1+λ2+λ3=1.
[0058] S3. Use the original dataset A to train the conditional table GAN.
[0059] Divide the original data set A into training set A_T and test set A_V;
[0060] (31) The training steps include:
[0061] Use the quantum random number generation library to generate a random noise vector Z, and jointly represent the label vector y and the noise vector Z and input them into the generator;
[0062] The generator generates a labeled fake sample (x*|y,y) based on the input and obtains the generator loss L (G) ;
[0063] The generator is output by two paths: one path outputs to the pre-trained model, which generates dynamic weights and pre-trained model loss L according to each different output from the generator (S) ; Another path outputs to the discriminator;
[0064] The discriminator receives the forged samples output by the generator, the dynamic weights provided by the pre-trained model, and the real samples (x, y) provided by the training set A_T, and obtains the discriminator loss L (D) ;
[0065] According to the balance weights λ1, λ2, λ3, the generator loss, the discriminator loss and the pre-trained model loss are combined to obtain the joint loss L (U) =λ1L (G) +λ2L (D) +λ3L (S), and λ1+λ2+λ3=1;
[0066] Iterate training until the number of iterations or the joint loss meets the threshold.
[0067] (32) The test steps include:
[0068] Use the trained conditional table GAN to generate tabular data about the training set A_T;
[0069] The generated tabular data is compared with the test set A_V (for example, data comparison is performed through indicators such as average Wasserstein distance, average JS divergence and correlation distance).
[0070] like Figure 4 As shown, the present invention also provides the following application examples:
[0071] ① Determine the target of generating data (such as housing price data, disease data, etc.), then collect the original data, and use the K-fold validation method to divide the obtained data set into multiple training sets and test sets;
[0072] ②Build and train interpretable pre-trained models;
[0073] ③ Construct a basic GAN framework, and embed the trained pre-trained model into the basic GAN framework to obtain a conditional table GAN;
[0074] ④ Through the K-fold validation method, the conditional table GAN with the best test evaluation index is selected to complete the generation of tabular data.
[0075] Using the heart disease dataset in the UCI dataset as an example, a single-round basic GAN and multiple-round interpretable conditional table GAN experiments were conducted. The experimental results are as follows:
[0076] like Figure 5 As shown in a, the horizontal axis represents the influence of the feature value obtained based on the interpretability model (SHAP) on the output of the conditional table GAN, where the vertical line from point 0 represents the positive and negative influence of the specific feature value on the output of the conditional table GAN; the vertical axis color on the right side of the figure represents the size of the feature value; the vertically arranged features in the figure are arranged from large to small according to their influence, and the features at the top have a greater overall influence on the output of the conditional table GAN, while the features at the bottom have a smaller influence;
[0077] like Figure 5 As shown in Figure b, after multiple rounds of interpretable conditional table GAN training, the distance correlation coefficient of the generated data is reduced by nearly 40%, indicating that the generated data is closer to the distribution of the real data, that is, the quality of the generated data is improved; and the average training time of the model is reduced by about 45%, that is, the efficiency of data generation is improved.
[0078] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for establishing a conditional table GAN based on interpretability, characterized in that: include: Build and train interpretable pre-trained models; A basic GAN framework is constructed, and the trained pre-trained model is embedded into the basic GAN framework to obtain a conditional table GAN.
2. The method for establishing a conditional table GAN based on interpretability according to claim 1, characterized in that: Also includes: The conditional table GAN is trained using the original dataset.
3. The method for establishing a conditional table GAN based on interpretability according to claim 2, characterized in that: The pre-training model is constructed based on a fuzzy clustering model.
4. The method for establishing a conditional table GAN based on interpretability according to claim 3, characterized in that: The training of the pre-trained model with interpretability includes: The features of each sample in the original data set are calculated through the interpretable model and aggregated to form a feature data set; The pre-trained model is trained using the feature data set.
5. The method for establishing a conditional table GAN based on interpretability according to claim 4, characterized in that: The using the feature data set to train the pre-trained model includes: Dividing the feature data set into a training set and a test set; Using the training set to train the pre-trained model; The pre-trained model is tested using the test set.
6. The method for establishing a conditional table GAN based on interpretability according to claim 5, characterized in that: The features of each sample in the original data set calculated by the interpretable model include: Constructing a tree model about the original data set based on a machine learning algorithm; Building a tree-type interpreter for the interpretability model based on the tree-type model; The tree interpreter is used to calculate the features of each sample in the original data set.
7. The method for establishing a conditional table GAN based on interpretability according to claim 6, characterized in that: The basic GAN framework includes a generator and a discriminator.
8. The method for establishing a conditional table GAN based on interpretability according to claim 7, characterized in that: The embedding the trained pre-trained model into the basic GAN framework comprises: Constructing two paths for transmission from the generator to the discriminator; The pre-trained model is embedded in one of the paths.
9. The method for establishing a conditional table GAN based on interpretability according to claim 8, characterized in that: The pre-trained model is used to assign dynamic weights to the discriminator according to the output of the generator.
10. The method for establishing a conditional table GAN based on interpretability according to claim 9, characterized in that: The embedding the trained pre-trained model into the basic GAN framework further comprises: The generator, discriminator and pre-trained model are respectively assigned balancing weights λ1, λ2, λ3, and λ1+λ2+λ3=1.