Method and device for generating data

Abstract

The invention relates to a method and a device for generating data. The method and the device can generate new data under the condition that characteristic information unique to table type data is restrained. The method includes: generating a constraint vector variable specifying a constraint specific to table type data; obtaining the generated data by applying the constraint vector variable and the potential vector variable to the generator; discriminating whether the generated data is true data or false data by applying the raw data and the generated data to a discriminator; predicting whether the generated data satisfies a constraint specific to table type data; and generating a prediction constraint vector variable based on the prediction result.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority and benefit to Korean Patent Application No. 10-2024-0061069, filed with the Korean Intellectual Property Office on May 9, 2024, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to a method and apparatus for generating data. Background Technology

[0004] Generative models can learn to generate new data from given data. Specifically, a generative model can learn the distribution of a given dataset and generate new data samples based on the learned distribution. Examples of generative models can include Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or diffusion models. Generative models can be used to generate various types of data, such as images, speech, and text. In recent years, methods for limiting generation to specific types of data using generative models have been investigated. Summary of the Invention

[0005] This disclosure provides a method and apparatus for generating data, which can generate new data with regard to the characteristic information specific to constraint table type data.

[0006] This disclosure also provides a method and apparatus for generating data, which can utilize each of high-dimensional and low-dimensional information as its constraints to generate new data.

[0007] According to an embodiment, a method for generating data is provided. The method generates new data while considering characteristic information specific to table-type data. The method includes: generating constraint vector variables specifying constraints specific to the table-type data; obtaining the generated data by applying the constraint vector variables and latent vector variables to a generator; identifying whether the generated data is real or fake by applying the original data and the generated data to a discriminator; predicting whether the generated data satisfies constraints specific to the table-type data; and generating predictive constraint vector variables based on the prediction result.

[0008] Acquiring the generated data may include: converting the constraint vector variables into low-dimensional vectors through a first multilayer perceptron (MLP) embedding, and may include: converting the latent vector variables into low-dimensional vectors through a second MLP embedding.

[0009] Acquiring the generated data may further include: applying the feature vector variables together with the constraint vector variables to the generator. Converting the constraint vector variables into low-dimensional vectors through a first MLP embedding may include: converting the constraint vector variables and feature vector variables into low-dimensional vectors through a first MLP embedding.

[0010] The acquisition of the generated data may further include: expanding information about the low-dimensional vectors obtained from the first MLP embedding and the second MLP embedding, and may include: ensuring the connectivity between the latent vector variables and the constrained vector variables.

[0011] Acquiring the generated data may further include: computing the attention of each of the multiple heads with different focusing aspects to the vector with extended information via a multi-head attention network (MHAN), and may include: outputting the final result by combining the attention results with each other.

[0012] Acquiring the generated data may further include: generating a specific type of result by analyzing the final result; and acquiring that specific type of result as the generated data.

[0013] The identification process may further include performing frame reconstruction on the generated data.

[0014] The identification may further include: calculating the attention of each of the multiple heads with different focusing aspects to the reconstructed generated data via MHAN, and may include: outputting a final result by combining the attention results with each other.

[0015] The identification can further include: performing feature extraction and transformation through MLP; and identifying whether the generated data is real or fake based on the results of feature extraction and transformation.

[0016] The method may further include performing error measurement on the generated data and the prediction constraint vector variables using an error function based on mean squared error (MSE).

[0017] According to an embodiment, an apparatus for generating data is provided. The apparatus generates new data with respect to characteristic information specific to table-type data, and executes program code loaded in at least one memory device via at least one processor. The program code is executed to: generate constraint vector variables specifying constraints specific to the table-type data; obtain the generated data by applying the constraint vector variables and potential vector variables to a generator; identify whether the generated data is true or false by applying the original data and the generated data to a discriminator; predict whether the generated data satisfies the constraints specific to the table-type data; and generate a prediction constraint vector variable based on the prediction result.

[0018] Acquiring the generated data may include: converting the constraint vector variables into low-dimensional vectors through a first MLP embedding, and may also include: converting the latent vector variables into low-dimensional vectors through a second MLP embedding.

[0019] Acquiring the generated data may include: applying the feature vector variables together with the constraint vector variables to the generator, and converting the constraint vector variables into low-dimensional vectors through a first MLP embedding may include: converting the constraint vector variables and feature vector variables into low-dimensional vectors through a first MLP embedding.

[0020] The acquisition of the generated data may further include: expanding information about the low-dimensional vectors obtained from the first MLP embedding and the second MLP embedding, and may include: ensuring the connectivity between the latent vector variables and the constrained vector variables.

[0021] Acquiring the generated data may further include: calculating the attention of each of the multiple heads with different focusing aspects to the vector with extended information via MHAN, and may include: outputting the final result by combining the attention results with each other.

[0022] Acquiring the generated data may further include: generating a specific type of result by analyzing the final result; and acquiring that specific type of result as the generated data.

[0023] The identification process may further include performing frame reconstruction on the generated data.

[0024] The identification may further include: calculating the attention of each of the multiple heads with different focusing aspects to the reconstructed generated data via MHAN, and may include: outputting a final result by combining the attention results with each other.

[0025] The identification can further include: performing feature extraction and transformation through MLP; and identifying whether the generated data is real or fake based on the results of feature extraction and transformation.

[0026] The program code can perform error measurement on the generated data and prediction constraint vector variables using an error function based on mean squared error (MSE). Attached Figure Description

[0027] Figure 1 This is a block diagram depicting an apparatus for generating data according to an embodiment.

[0028] Figure 2 This is a flowchart of a method for generating data according to an embodiment.

[0029] Figure 3 This is a diagram illustrating an implementation example of a device for generating data according to an embodiment.

[0030] Figure 4 This is a diagram illustrating an implementation example of a device for generating data according to an embodiment.

[0031] Figures 5 to 7 This is an illustration of an example of data generation from a device for generating data according to an embodiment.

[0032] Figure 8 This is a diagram of a computing device according to an embodiment. Detailed Implementation

[0033] Embodiments of this disclosure are described in detail below with reference to the accompanying drawings to enable those skilled in the art to readily implement this disclosure. However, embodiments of this disclosure may be implemented in various different forms and are not limited to those described herein. Furthermore, in the drawings, portions irrelevant to the description have been omitted to more clearly illustrate aspects of this disclosure, and similar portions are indicated by similar reference numerals throughout the specification.

[0034] Throughout the specification and claims, unless expressly stated otherwise, an element described as "comprising," "having," or "including" any component should be understood as implying the possible inclusion of another component, rather than excluding it. Terms including ordinal numbers such as "first," "second," etc., may be used to describe a variety of components. However, these components are not limited by these terms. These terms are used only to distinguish one component from another.

[0035] When a component, apparatus, element, module, etc., of this disclosure is described as having a purpose or performing an operation, function, etc., it shall be regarded herein as being "configured" to satisfy that purpose or perform that operation or function. Each component, apparatus, element, module, etc., may be implemented individually or may include a processor and memory such as a non-transitory computer-readable medium as part of a device or apparatus.

[0036] Terms such as “~part,” “-device / machine,” and “module” used in this specification can refer to a unit capable of performing at least one function or operation described herein, which can be implemented as hardware, circuitry, software, or a combination of hardware, circuitry, and software. Furthermore, at least some components or functions of the method and apparatus for generating data according to the embodiments described below can be implemented as a program or software, and such program or software can be stored on a computer-readable medium.

[0037] Figure 1 It is a block diagram depicting and used to explain an apparatus for generating data according to an embodiment.

[0038] Reference Figure 1According to an embodiment, the data generation apparatus 10 can execute program code loaded in at least one memory device via at least one processor. For example, the data generation apparatus 10 can be implemented as described below. Figure 8 The computing device 50 is described above. In this case, at least one processor may correspond to processor 510 of computing device 50, and at least one memory device may correspond to memory 530 of computing device 50. Program code may be executed by at least one processor to generate new data in the presence of characteristic information specific to constraint table type data, or to generate new data using each of high-dimensional and low-dimensional information as its constraints. In this specification, the term "module" is used to logically distinguish functions performed by program code.

[0039] The data generation apparatus 10 according to the embodiment can generate new data while constraining the characteristic information specific to table-type data. Table-type data can store various information in a structured form, including sparse characteristic information and variables or columns, such as those with mostly zero values ​​or missing values. The data generation apparatus 10 can generate new data while constraining the aforementioned sparse characteristic information. To this end, the data generation apparatus 10 can execute program code including a constraint vector variable generation module 110, a generator network 120, a discriminator network 130, and a constraint prediction network 140.

[0040] The constraint vector variable generation module 110 can generate constraint vector variables that specify constraints specific to table-type data. Constraint vector variables can be condition variables used to control the characteristics to be constrained when generating data. For example, constraint vector variables can be set to generate data by setting constraints on specific locations, rows, columns, or ranges of certain table-type data.

[0041] The generator network 120 can perform learning on the generated data based on the constraint vector variables and latent vector variables generated by the constraint vector variable generation module 110, until the generated data is identified as real data by the discriminator network 130. Here, the latent vector variables can be initial random variables with randomness added to the data, which are transformed into specific dimensions.

[0042] The discriminator network 130 can receive generated data and original data provided by the generator network 120, and can distinguish whether the generated data is real or fake. The discriminator network 130 can perform learning until high discrimination accuracy is achieved in determining the authenticity of the generated data.

[0043] The constraint prediction network 140 can predict whether the data generated by the discriminator network 130 satisfies constraints specific to table-type data. Furthermore, the constraint prediction network 140 can generate a prediction constraint vector variable based on the prediction results.

[0044] According to this embodiment, apparatus 10 can generate new data while constraining feature information specific to table-type data, and can utilize each of high-dimensional and low-dimensional information as its constraints to generate new data. For example, some conventional data generation models may be limited to generating images while constraining high-dimensional feature information. However, using conventional models can be challenging because constraining low-dimensional information such as one or a few pixels can lead to model overfitting. On the other hand, apparatus 10, a newly disclosed embodiment of this disclosure by adjusting the architecture of the generation model, can generate data by constraining high-dimensional information such as implicit semantic information in table-type data, and simultaneously constraining sparse feature information in table-type data, i.e., low-dimensional information.

[0045] Figure 2 This is a flowchart used to explain the method for generating data according to an embodiment.

[0046] Reference Figure 2 The method for generating data according to an embodiment may include: generating constraint vector variables that specify constraints specific to table-type data (step S201). The method may further include: obtaining the generated data by applying the constraint vector variables and latent vector variables to a generator (step S202). The method may further include: identifying whether the generated data is real or fake data by applying the original data and the generated data to a discriminator (step S203). The method may further include: predicting whether the generated data satisfies constraints specific to table-type data (step S204). The method may further include: generating predictive constraint vector variables based on the prediction result (step S205).

[0047] A more detailed description of the methods used to generate the data can be found in the embodiments described in this specification. Therefore, redundant descriptions are omitted here.

[0048] Figure 3 This is a diagram illustrating an example of an apparatus for generating data according to an embodiment.

[0049] Reference Figure 3 The apparatus for generating data according to the embodiments may include a generator network 30, a discriminator network 31, and a constraint prediction network 32.

[0050] In this implementation example, V1 can be configured with a latent vector variable Z, and V2 can be configured with a constraint vector variable c and an eigenvector variable y that specify constraints specific to table-type data. The latent vector variable Z, the constraint vector variable c, and the eigenvector variable y can be input into the generator network 30.

[0051] The generator network 30 can output generated data G(z) based on the latent vector variable Z, the constraint vector variable c, and the feature vector variable y. The generated data G(z) and the feature vector variable y can be input into an error function for learning by the generator network 30. In some embodiments, the error function can be an error function based on the mean squared error (MSE). MSE can be a method of calculating the average of the squares of all differences to quantify the difference between the actual and predicted values.

[0052] The discriminator network 31 can receive the original data x and the generated data G(z), and can distinguish whether the generated data G(z) is real data or fake data.

[0053] The constraint prediction network 32 can predict whether the generated data G(z) satisfies the constraints specific to table-type data, and can generate the prediction constraint vector variable V3 from the prediction results.

[0054] Figure 4 This is a diagram illustrating an example of an apparatus for generating data according to an embodiment.

[0055] Reference Figure 4 In the apparatus for generating data according to the embodiment, the generator network 30 and the discriminator network 31 may be implemented using multiple fully connected networks.

[0056] In this embodiment, the generator network 30 may include a first multilayer perceptron (MLP) embedding 301, a second MLP embedding 302, an inverse compressed fully connected network (ICFCN) 303, a multi-head attention network (MHAN) 304, and a data determination fully connected network (DDFCN) 305.

[0057] The generator network 30 can transform the constraint vector variable c and the feature vector variable f into a low-dimensional vector through the first MLP embedding 301. On the other hand, the generator network 30 can transform the latent vector variable z into a low-dimensional vector through the second MLP embedding 302.

[0058] Next, the generator network 30 can extend the information about the low-dimensional vectors obtained from the first MLP embedding 301 and the second MLP embedding 302 through the ICFCN 303. The generator network 30 can ensure the connectivity between the latent vector variable z and the constraint vector variable c.

[0059] Next, generator network 30 can compute the attention of each of the multiple heads with different focusing aspects to the vector with extended information through MHAN 304. Generator network 30 can output the final result by combining the attention results.

[0060] Next, the generator network 30 can generate specific types of results by analyzing the final results through DDFCN 305, and can output specific types of results as the generated data G(c, f, z). Error measurement can be performed on the generated data G(c, f, z) and the constraint vector variable c using an error function based on MSE. The learning of the generator network 30 can be performed based on the results of the error measurement.

[0061] In addition, the discriminator network 31 may include ICFCN 311, MHAN 312, MLP 313 and Conditional Predictive Fully Connected Network (CFFCN) 314.

[0062] The discriminator network 31 can perform frame reconstruction on the original data x and the generated data G(c, f, z) through ICFCN 311.

[0063] Next, the discriminator network 31 can compute the attention of each of the multiple heads with different focusing aspects to the reconstructed original data x and the generated data G(c, f, z) via MHAN. The discriminator network 31 can output the final result by combining the attention results.

[0064] Next, the discriminator network 31 can perform feature extraction and transformation through MLP 313, and can identify whether the generated data G(c, f, z) is real data or fake data based on the results of feature extraction and transformation.

[0065] Furthermore, the discriminator network 31 can predict whether the generated data G(c, f, z) satisfies the constraints specific to table-type data using CFFCN 314. The discriminator network 31 can generate a predicted constraint vector variable c' based on the prediction result. Error measurement can be performed on the constraint vector variable c and the predicted constraint vector variable c' using an error function based on MSE. The learning of the discriminator network 31 can be performed based on the results of the error measurement.

[0066] Figures 5 to 7 This is a diagram illustrating an example of data generation using a device for generating data according to an embodiment.

[0067] Reference Figure 5 and Figure 6As an example of data generation by a device for generating data according to an embodiment, this device can generate data while constraining specific feature values ​​in table-type data. This configuration overcomes the fact that conventional generative models typically handle image data. Therefore, applying conventional models to the table-type data shown above or generating data while constraining specific feature values ​​is challenging.

[0068] Next, refer to Figure 7 As can be seen, the device 10 can generate data that satisfies predictable information through data analysis such as stiffness or weight, given specific characteristic values ​​in constraint table type data.

[0069] Figure 8 This is a diagram used to explain the computing device according to the embodiments.

[0070] Reference Figure 8 The method and apparatus for generating data according to the embodiments can be implemented using computing device 50.

[0071] The computing device 50 may include at least one of a processor 510, a memory 530, a user interface input device 540, a user interface output device 550, and a storage device 560, each of which performs its communication via a bus 520. The computing device 50 may also include a network interface 570 electrically connected to a network 40. The network interface 570 can send or receive signals with another entity via the network 40.

[0072] Processor 510 can be implemented in any of the following types: microcontroller unit (MCU), application processor (AP), central processing unit (CPU), graphics processing unit (GPU), neural processing unit (NPU), or quantum processing unit (QPU). Processor 510 can be any semiconductor device that executes instructions stored in memory 530 or storage device 560. Processor 510 can implement the above-mentioned... Figures 1 to 7 The functions and methods described.

[0073] The memory 530 and storage device 560 may include various types of volatile or non-volatile storage media. For example, the memory may include read-only memory (ROM) 531 and random access memory (RAM) 532. In some embodiments, the memory 530 may be disposed inside or outside the processor 510 and may be connected to the processor 510 by various known means.

[0074] In some embodiments, at least some components or functions of the method and apparatus for generating data according to the embodiments may be implemented as a program or software executed by computing device 50. The program or software may be stored in a computer-readable medium. Specifically, the computer-readable medium according to the embodiments may be a program recorded on a computer including processor 510 for performing the steps included in the method and apparatus for generating data according to the embodiments, the processor 510 executing programs or instructions stored in memory 530 or storage device 560.

[0075] In some embodiments, at least some components or functions of the method and apparatus for generating data according to the embodiments may be implemented using the hardware or circuitry of the computing device 50, or may be implemented using separate hardware or circuitry that may be electrically connected to the computing device 50.

[0076] As described above, the methods and apparatus for generating data according to embodiments can utilize each of high-dimensional and low-dimensional information as constraints to generate new data. For example, conventional data generation models may be limited to generating images while constraining high-dimensional feature information. However, using conventional models can be challenging because constraining low-dimensional information, such as one or a few pixels, can lead to model overfitting. On the other hand, the methods and apparatus newly disclosed according to embodiments of this disclosure, by adjusting the architecture of the generative model, can generate data by constraining high-dimensional information, such as implicit semantic information, in tabular data, and by simultaneously constraining sparse feature information, i.e., low-dimensional information, in tabular data.

[0077] While embodiments of the present disclosure have been described in detail above, the scope of the disclosure is not limited thereto. In other words, various modifications and alterations made by those skilled in the art using the basic concepts of the disclosure as defined in the appended claims also fall within the scope of the disclosure.

Claims

1. A method for generating data, the method comprising: Generate constraint vector variables that specify constraints for table-type data; The generated data is obtained by applying the constraint vector variables and latent vector variables to the generator. The generated data is identified as either real or fake by applying the original data and the generated data to the discriminator. Predict whether the generated data satisfies constraints specific to the table type. as well as Generate prediction constraint vector variables based on the prediction results. The method generates new data while constraining the characteristic information specific to the table type data.

2. The method according to claim 1, wherein, The generated data includes: The constraint vector variables are transformed into low-dimensional vectors through a first multilayer perceptron embedding, i.e., a first MLP embedding; and The latent vector variables are transformed into low-dimensional vectors through a second MLP embedding.

3. The method according to claim 2, wherein, Obtaining the generated data further includes: applying the feature vector variables together with the constraint vector variables to the generator; and Converting the constraint vector variables into low-dimensional vectors through the first MLP embedding includes: converting the constraint vector variables and the feature vector variables into low-dimensional vectors through the first MLP embedding.

4. The method according to claim 2, wherein, Obtaining the generated data further includes: Expanding information about the low-dimensional vectors obtained from the first MLP embedding and the second MLP embedding; and Ensure the connectivity between the potential vector variables and the constraint vector variables.

5. The method according to claim 4, wherein, Obtaining the generated data further includes: The attention of each of the multiple heads with different focusing aspects to the vector with extended information is computed through a multi-head attention network, namely MHAN; and The final result is output by combining the attention results.

6. The method according to claim 5, wherein, Obtaining the generated data further includes: Specific types of results are generated by analyzing the final results; and The specific type of result is obtained as the generated data.

7. The method according to claim 1, wherein, Further identification includes: The generated data is then reconstructed using a frame.

8. The method according to claim 7, wherein, Further identification includes: The attention of each of the multiple heads with different focusing aspects to the reconstructed data is calculated using a multi-head attention network, namely MHAN; and The final result is output by combining the attention results.

9. The method according to claim 8, wherein, Further identification includes: Feature extraction and transformation are performed using a multilayer perceptron (MLP); and The generated data can be identified as real or fake based on the results of feature extraction and transformation.

10. The method of claim 1, further comprising: Error measurement is performed on the generated data and the prediction constraint vector variables using an error function based on mean squared error (MSE).

11. An apparatus for generating data, wherein, The device executes program code loaded in at least one memory device via at least one processor, and the program code is executed to: Generate constraint vector variables that specify constraints for table-type data; The generated data is obtained by applying the constraint vector variables and latent vector variables to the generator. The generated data is identified as either real or fake by applying the original data and the generated data to the discriminator. Predict whether the generated data satisfies constraints specific to the table type; and Generate prediction constraint vector variables based on the prediction results. The device generates new data while constraining the characteristic information specific to the table type data.

12. The apparatus according to claim 11, wherein, The generated data includes: The constraint vector variables are transformed into low-dimensional vectors through a first multilayer perceptron embedding, i.e., a first MLP embedding; and The latent vector variables are transformed into low-dimensional vectors through a second MLP embedding.

13. The apparatus according to claim 12, wherein, Obtaining the generated data includes: further applying the feature vector variables together with the constraint vector variables to the generator; and Converting the constraint vector variables into low-dimensional vectors through the first MLP embedding includes: converting the constraint vector variables and the feature vector variables into low-dimensional vectors through the first MLP embedding.

14. The apparatus according to claim 12, wherein, Obtaining the generated data further includes: Expanding information about the low-dimensional vectors obtained from the first MLP embedding and the second MLP embedding; and Ensure the connectivity between the potential vector variables and the constraint vector variables.

15. The apparatus according to claim 14, wherein, Obtaining the generated data further includes: The attention of each of the multiple heads with different focusing aspects to the vector with extended information is computed through a multi-head attention network, namely MHAN; and The final result is output by combining the attention results.

16. The apparatus according to claim 15, wherein, Obtaining the generated data further includes: Specific types of results are generated by analyzing the final results; and The specific type of result is obtained as the generated data.

17. The apparatus according to claim 11, wherein, Further identification includes: The generated data is then reconstructed using a frame.

18. The apparatus according to claim 17, wherein, Further identification includes: The attention of each of the multiple heads with different focusing aspects to the reconstructed data is calculated using a multi-head attention network, namely MHAN; and The final result is output by combining the attention results.

19. The apparatus according to claim 18, wherein, Further identification includes: Feature extraction and transformation are performed using a multilayer perceptron (MLP); and The generated data can be identified as real or fake based on the results of feature extraction and transformation.

20. The apparatus according to claim 11, wherein, The program code performs error measurement on the generated data and the prediction constraint vector variables using an error function based on mean square error (MSE).

Images