Method and system for data augmentation for learning from tabular data

Abstract

The present invention relates to a method and a system for data augmentation for learning from tabular data. More specifically, the present invention relates to a method and a system for expression level data augmentation for improving the performance of self-supervised learning (SSL) in tabular data.

Description

Data Augmentation Method and System for Learning Tabular Data

[0001] The present invention relates to a data augmentation method and system for learning tabular data. More specifically, the present invention relates to a representation-level data augmentation method and system for improving the performance of Self-Supervised Learning (SSL) in tabular data.

[0002] Tabular data consists of rows and columns and is used in various fields such as finance, medical, manufacturing, healthcare, marketing, and research. Despite this universality, deep learning research has paid relatively less attention to tabular data compared to fields such as computer vision or natural language processing.

[0003] Recently, Self-Supervised Learning (SSL) has garnered attention as a promising pre-training method for tabular data. Through self-supervised learning, models can extract meaningful features and patterns from unlabeled data and apply them to various downstream tasks. Such self-supervised learning often relies on contrastive learning, a method that exposes the model to various transformations of the input data through data augmentation. This approach enables the model to retain important information while maintaining invariance to specific transformations, thereby supporting the learning of robust and generalizable representations.

[0004] In this regard, applying self-supervised learning to tabular data is challenging due to the difficulties involved in designing effective data augmentation. Images and text possess clear structures, making it relatively easy to generate augmentations (e.g., cropping or rotating images, manipulating the semantic structure of text). In contrast, tabular data is heterogeneous and lacks a clear structure. That is, tabular data lacks intrinsic location information or spatial relationships, and the heterogeneous mix of data types, arbitrary column order, and correlations between complex features further complicate this issue.

[0005] Despite the lack of inherent relational structures in tabular data, most conventional research has focused on augmentation definitions at the input level. For example, this may include methods that modify or randomly remove selected features. While these augmentation methods are simple to implement, they carry the risk of distorting important patterns or generating unrealistic data. Consequently, input-level augmentation can hinder the performance of self-supervised learning (SSL) in tabular domains.

[0006] Accordingly, the present invention proposes a method that overcomes the limitations of conventional data augmentation and learning methods and enables an artificial intelligence model to learn effectively even from tabular data.

[0007] The present invention is intended to provide a data augmentation method and system for learning tabular data that enables an artificial intelligence model to better understand tabular data.

[0008] More specifically, the present invention is intended to provide a data augmentation method and system for learning tabular data that enables an artificial intelligence model to learn an information-rich representation of tabular data.

[0009] In addition, the present invention is intended to provide a data augmentation method and system for learning tabular data that can effectively identify and maintain important patterns and relationships within tabular data.

[0010] More specifically, the present invention aims to provide a data augmentation method and system for learning tabular data that can generate various augmented representations while preserving important information of the tabular data.

[0011] Furthermore, the present invention aims to provide a data augmentation method and system for learning tabular data that can maximize the effects of self-supervised learning (SSL) in tabular data.

[0012] To solve the problem described above, a data augmentation method for learning tabular data according to the present invention, which is computerized, may include the following steps: inputting tabular data into at least one encoder; obtaining a first embedding vector corresponding to the tabular data as the output of the at least one encoder; extracting a specific parameter from the at least one encoder from which the first embedding vector is output, based on a preset extraction criterion; updating at least one other encoder based on the specific parameter extracted according to the preset extraction criterion; inputting the tabular data into the updated at least one other encoder; obtaining an augmented second embedding vector corresponding to the tabular data from the at least one other encoder; and training a target model using the first embedding vector and the second embedding vector.

[0013] In an embodiment, the method may further include the step of inputting the specific parameter extracted from the at least one encoder into a pre-set specific module and the step of performing decomposition of the specific parameter using a pre-set decomposition method in the specific module.

[0014] In an embodiment, the specific module decomposes the specific parameter into at least one of a first singular vector matrix containing first singular vectors, a second singular vector matrix containing second singular vectors, and a diagonal matrix containing singular values ​​corresponding to each of the first singular vectors and the second singular vectors, and in the step of updating the at least one other encoder, the at least one other encoder can be updated using the output of the specific module for the specific parameter.

[0015] In an embodiment, the method further includes the steps of specifying at least one singular value among the singular values ​​that satisfies a preset criterion and generating a parameter to be used for updating the at least one other encoder based on the specified singular value, wherein the parameter corresponds to the output of the specific module for the specified parameter and can be used for updating the at least one other encoder.

[0016] In an embodiment, the specific module may specify at least one of a specific value among the specific values ​​according to a preset standard, a first specific vector and a second specific vector corresponding to the specific specific value, and generate the parameter based on at least one of the specified specific value, the specified first specific vector and the specified second specific vector.

[0017] In an embodiment, in the step of updating the at least one other encoder, the at least one other encoder can be updated using the parameter output from the specific module.

[0018] In an embodiment, the step of updating the at least one other encoder is to input the parameter to the at least one other encoder and set the parameter of the at least one other encoder to the parameter output from the specific module, and the updated at least one other encoder may be configured to generate an augmented second embedding vector corresponding to table data based on the set parameter.

[0019] In an embodiment, the step of updating the at least one other encoder may be a step of updating the at least one other encoder using the output of the specific module to generate the augmented second embedding vector for the first embedding vector obtained from the at least one encoder.

[0020] In an embodiment, the step of applying a pre-set normalization technique to at least one of the at least one encoder and at least one other encoder may be further included.

[0021] In an embodiment, when the preset normalization technique is applied to at least one of the at least one encoder and at least one other encoder, probabilistic variability may be imparted to at least one of the first embedding vector corresponding to the table data obtained from the at least one encoder and the second embedding vector corresponding to the table data obtained from the at least one other encoder.

[0022] In an embodiment, for training the target model, the method may further include the step of defining a first loss function and a second loss function, respectively, using at least one of the first embedding vector and the second embedding vector.

[0023] In an embodiment, the method further includes the steps of inputting the first embedding vector and the second embedding vector, respectively, to a projection head, and obtaining from the projection head a first projection vector corresponding to the first embedding vector and a second projection vector corresponding to the second embedding vector, wherein the first loss function can be defined using the first projection vector corresponding to the first embedding vector and the second projection vector corresponding to the second embedding vector.

[0024] In an embodiment, the first loss function may be defined to learn in a direction that maximizes the similarity between the first embedding vector and the second embedding vector.

[0025] In an embodiment, the method further includes the step of inputting the first embedding vector to at least one decoder and the step of reconstructing the first embedding vector to correspond to the table data in the at least one decoder to obtain a reconstructed output value, and the second loss function can be defined using the table data and the reconstructed output value.

[0026] In an embodiment, the second loss function may be defined to learn the relationship between the table data and the reconstructed output value.

[0027] In an embodiment, the method further includes the step of defining a final loss function using the first loss function and the second loss function, and can train the artificial intelligence model to be trained using the final loss function.

[0028] In an embodiment, the method may further include the step of performing binning on the table data, the step of obtaining a plurality of binning table data as a result of performing binning on the table data, and the step of training the target model using the plurality of binning table data.

[0029] In an embodiment, based on the acquisition of the plurality of segmented table data, the plurality of segmented table data is input into at least one encoder, and as the output of the at least one encoder, a plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained; based on the preset extraction criteria, a specific parameter is extracted from the at least one encoder from which the plurality of embedding vectors are output, and based on the specific parameter extracted according to the preset extraction criteria, at least one other encoder is updated, the plurality of segmented table data is input into the updated at least one other encoder, and augmented multiple embedding vectors corresponding to each of the plurality of segmented table data are obtained from the at least one other encoder, and a target model can be trained using the plurality of embedding vectors obtained from the at least one encoder and the augmented multiple embedding vectors obtained from the at least one other encoder.

[0030] A data augmentation system for learning tabular data according to the present invention, comprising a memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, inputs tabular data into at least one encoder and obtains a first embedding vector corresponding to the tabular data as the output of the at least one encoder, extracts a specific parameter from the at least one encoder from which the first embedding vector is output based on a preset extraction criterion, updates at least one other encoder based on the specific parameter extracted according to the preset extraction criterion, inputs the tabular data into the updated at least one other encoder, obtains an augmented second embedding vector corresponding to the tabular data from the at least one other encoder, and can train a target model using the first embedding vector and the second embedding vector.

[0031] A program according to the present invention is a program that is executed by one or more processes in an electronic device and can be stored on a computer-readable recording medium, and may include instructions for performing the steps of: inputting table data into at least one encoder; obtaining a first embedding vector corresponding to the table data as the output of the at least one encoder; extracting a specific parameter from the at least one encoder from which the first embedding vector is output, based on a preset extraction criterion; updating at least one other encoder based on the specific parameter extracted according to the preset extraction criterion; inputting the table data into the updated at least one other encoder; obtaining an augmented second embedding vector corresponding to the table data from the at least one other encoder; and training a target model using the first embedding vector and the second embedding vector.

[0032] A data augmentation method for learning tabular data according to the present invention, which is computerized and includes the following steps: inputting table data into a first encoder; obtaining a first embedding vector corresponding to the table data as the output of the first encoder; extracting a specific parameter from the first encoder from which the first embedding vector is output, based on a preset extraction criterion; updating a second encoder based on the specific parameter extracted according to the preset extraction criterion; inputting the table data into the updated second encoder; obtaining an augmented second embedding vector corresponding to the table data from the second encoder; and training a target model using the first embedding vector and the second embedding vector.

[0033] In an embodiment, the method may further include the step of inputting the specific parameter extracted from the first encoder into a pre-set specific module and the step of performing decomposition of the specific parameter using a pre-set decomposition method in the specific module.

[0034] In an embodiment, the specific module decomposes the specific parameter into at least one of a first singular vector matrix including a first singular vector, a second singular vector matrix including a second singular vector, and a diagonal matrix including singular values ​​corresponding to each of the first singular vector and the second singular vector, and in the step of updating the second encoder, the second encoder can be updated using the output of the specific module for the specific parameter.

[0035] In an embodiment, the method further includes the step of specifying at least one singular value among the singular values ​​that satisfies a preset criterion, and the step of generating a parameter to be used for updating the second encoder based on the specified singular value, wherein the parameter corresponds to the output of the specific module for the specified parameter and can be used for updating the second encoder.

[0036] In an embodiment, the specific module may specify at least one of a specific value among the specific values ​​according to a preset standard, a first specific vector and a second specific vector corresponding to the specific specific value, and generate the parameter based on at least one of the specified specific value, the specified first specific vector and the specified second specific vector.

[0037] In an embodiment, in the step of updating the second encoder, the second encoder can be updated using the parameter output from the specific module.

[0038] In an embodiment, the step of updating the second encoder is to input the parameter into the second encoder and set the parameter of the second encoder to the parameter output from the specific module, and the updated second encoder may be configured to generate an augmented second embedding vector corresponding to table data based on the set parameter.

[0039] In an embodiment, the step of updating the second encoder may be a step of updating the second encoder using the output of the specific module to generate the augmented second embedding vector for the first embedding vector obtained from the first encoder.

[0040] In an embodiment, the step of applying a normalization technique previously set to at least one of the first encoder and the second encoder may be further included.

[0041] In an embodiment, when the preset normalization technique is applied to at least one of the first encoder and the second encoder, probabilistic variability may be imparted to at least one of the first embedding vector corresponding to the table data obtained from the first encoder and the second embedding vector corresponding to the table data obtained from the second encoder.

[0042] In an embodiment, for training the artificial intelligence model to be trained, the method may further include the step of defining a first loss function and a second loss function, respectively, using at least one of the first embedding vector and the second embedding vector.

[0043] In an embodiment, the method further includes the steps of inputting the first embedding vector and the second embedding vector, respectively, to a projection head, and obtaining a first projection vector corresponding to the first embedding vector and a second projection vector corresponding to the second embedding vector from the projection head, and the first loss function can be defined using the first projection vector corresponding to the first embedding vector and the second projection vector corresponding to the second embedding vector.

[0044] In an embodiment, the first loss function may be defined to learn in a direction that maximizes the similarity between the first embedding vector and the second embedding vector.

[0045] In an embodiment, the method further includes the step of inputting the first embedding vector to a decoder and the step of reconstructing the first embedding vector to correspond to the table data in the decoder to obtain a reconstructed output value, and the second loss function can be defined using the table data and the reconstructed output value.

[0046] In an embodiment, the second loss function may be defined to learn the relationship between the table data and the reconstructed output value.

[0047] In an embodiment, the method further includes the step of defining a final loss function using the first loss function and the second loss function, and can train the target model using the final loss function.

[0048] In an embodiment, the method may further include the step of performing binning on the table data, the step of obtaining a plurality of binning table data as a result of performing binning on the table data, and the step of training the target model using the plurality of binning table data.

[0049] In an embodiment, based on the acquisition of the plurality of segmented table data, the plurality of segmented table data is input into a first encoder, and as the output of the first encoder, a plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained; based on the preset extraction criteria, a specific parameter is extracted from the first encoder from which the plurality of embedding vectors are output, and based on the specific parameter extracted according to the preset extraction criteria, a second encoder is updated, the plurality of segmented table data is input into the updated second encoder, and augmented multiple embedding vectors corresponding to each of the plurality of segmented table data are obtained from the second encoder, and a target model can be trained using the plurality of embedding vectors obtained from the first encoder and the augmented multiple embedding vectors obtained from the second encoder.

[0050] A data augmentation system for learning tabular data according to the present invention, comprising a memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, inputs tabular data into a first encoder and obtains a first embedding vector corresponding to the tabular data as the output of the first encoder, extracts a specific parameter from the first encoder from which the first embedding vector is output based on a preset extraction criterion, updates a second encoder based on the specific parameter extracted according to the preset extraction criterion, inputs the tabular data into the updated second encoder, obtains an augmented second embedding vector corresponding to the tabular data from the second encoder, and can train a target model using the first embedding vector and the second embedding vector.

[0051] A program according to the present invention is a program that is executed by one or more processes in an electronic device and can be stored on a computer-readable recording medium, and may include instructions for performing the steps of: inputting table data into a first encoder; obtaining a first embedding vector corresponding to the table data as the output of the first encoder; extracting a specific parameter from the first encoder from which the first embedding vector is output, based on a preset extraction criterion; updating a second encoder based on the specific parameter extracted according to the preset extraction criterion; inputting the table data into the updated second encoder; obtaining an augmented second embedding vector corresponding to the table data from the second encoder; and training a target model using the first embedding vector and the second embedding vector.

[0052] As described above, according to the data augmentation method and system for learning tabular data of the present invention, by combining a pre-set normalization technique and a pre-set decomposition method, various augmented representations can be generated while preserving important features of the tabular data in the embedding space. Through this, the present invention can improve the quality of the augmented representation by maintaining important information and promote robust feature learning of the model through variability.

[0053] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, an encoder is updated based on encoder parameters extracted according to preset extraction criteria, and an augmented representation corresponding to the table data can be generated using the updated encoder. Through this, the present invention can secure diversity of augmented representations while maintaining the core information of the table data. That is, the present invention can generate new augmented representations while maintaining the structure of the table data by using the updated encoder.

[0054] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, the diversity of the representation space can be increased by applying a pre-set normalization technique to the encoder. That is, the present invention can induce the generation of various variations by randomly removing some of the parameters of the encoder to which the pre-set normalization technique is applied. Through this, the present invention enables the model to learn generalized representations and prevents overfitting to specific patterns. In other words, by applying a pre-set normalization technique to the encoder, the present invention enables the model to learn more robust representations and to experience various data variations.

[0055] Accordingly, the data augmentation method and system for learning tabular data according to the present invention allows for the maximization of augmentation effects while maintaining the intrinsic structure of the tabular data by performing augmentation and transformation within the learned embedding space (i.e., structured representation space) itself using a pre-set normalization technique and a pre-set decomposition method. Through this, more robust representation learning is possible even with tabular data, and model learning can be performed stably. In other words, the present invention can improve the learning performance of a model by minimizing information loss while preserving important information.

[0056] FIG. 1 is a conceptual diagram illustrating a data augmentation system for learning tabular data according to the present invention.

[0057] FIG. 2 is a flowchart illustrating a data augmentation method for learning tabular data according to the present invention.

[0058] FIGS. 3a, FIGS. 3b, FIGS. 4a, and FIGS. 4b are conceptual diagrams illustrating a data augmentation method for learning tabular data according to the present invention.

[0059] FIGS. 5, FIGS. 6, FIGS. 7, FIGS. 8 and FIGS. 9 are formulas and algorithms related to a data augmentation method for learning tabular data according to the present invention.

[0060] FIGS. 10 and 11 are tables showing an example of the performance of a data augmentation method according to the present invention and the learning results of an artificial intelligence model learned using a learning method according to the present invention.

[0061] FIGS. 12a, FIGS. 12b, and FIGS. 12c are conceptual diagrams for explaining a method of training a model by performing segmentation on table data according to an embodiment of the present invention.

[0062] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components are assigned the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not have distinct meanings or roles in themselves. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the present invention.

[0063] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0064] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0065] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0066] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0067] The present invention relates to a data augmentation method and system for learning tabular data that enables an artificial intelligence model to better understand tabular data. More specifically, the present invention relates to a data augmentation method and system for learning tabular data that enables an artificial intelligence model to learn information-rich representations of tabular data.

[0068] In particular, the present invention can perform augmentation in the embedding space of a model (or encoder) to maximize the effectiveness of Self-Supervised Learning (SSL) on tabular data (or tables, tabular data, tabular data, tabular data, etc.). More specifically, the present invention can perform augmentation in the embedding space by utilizing a pre-configured decomposition method (e.g., Singular Value Decomposition (SVD), Truncated SVD, etc.) to learn an information-rich representation of tabular data. In this case, the present invention can effectively reflect the characteristics of tabular data by applying a method that preserves important information (or core information) contained in the tabular data while performing augmentation in the embedding space based on the pre-configured decomposition method. The present invention performs augmentation by modifying the weight matrix of a specific layer of the encoder, and by applying the pre-configured decomposition method during the augmentation process, it ensures diversity of representation while preserving important information of the tabular data. In one embodiment, the layer included in the model (or encoder) according to the present invention may include at least one of a Fully Connected Layer (Dense Layer) of a Multi-Layer Perceptron (MLP) based or Transformer based encoder. In this specification, the term "layer" may also be referred to as "layer," "hierarchy," or "block."

[0069] Here, the embedding space can refer to the space remaining after input data has been embedded through an encoder. Unlike the input space, the embedding space can be understood as a space containing more structured and meaningful information. As a multidimensional space where embeddings transformed (or generated) by the encoder exist, the embedding space can reflect semantic and / or structural relationships between data. In other words, the embedding space is a space where embedding vectors generated by an artificial intelligence model (or encoder) are distributed (or located), and it can also be understood as a vector space containing representations learned by the model (or encoder). For example, in the embedding space, each vector represents a single data sample, and in contrastive learning, learning can be performed based on similarity between samples (e.g., cosine similarity, distance, etc.) within the embedding space. In this specification, the term "embedding space" may be used interchangeably with "representation space," "latent space," "latent representation space," "feature space," "hidden space," "vector space," or "internal representation space." Additionally, in this specification, the term "embedding vector" may be used interchangeably with "representation," "embedding representation," "latent representation," "feature representation," "hidden representation," "embedding," "vector," "latent vector," "feature vector," "hidden vector," or "intermediate representation."

[0070] In the present invention, performing augmentation in the embedding space of an artificial intelligence model may mean performing augmentation on embedding vectors generated from an encoder or on the parameters (or parameter matrices, weights, weight matrices, etc.) of the encoder from which the embedding vectors are generated. This may mean applying modifications to the embedding vectors or parameters generated by the encoder during training, rather than directly modifying the input data (e.g., tabular data). For example, in the present invention, the parameters of the final layer of the encoder generating embedding vectors are updated (or adjusted, modified, changed, etc.), and various augmented embedding vectors can be generated for the input data using the updated encoder. In other words, this can be understood as imparting diversity to the structure of the encoder itself that forms the embedding space (e.g., an encoder to which SVD is applied).

[0071] Accordingly, the data augmentation method and system for learning tabular data according to the present invention may also be referred to as the “Representation-space augmentation for Tabular data (RaTab) method and system.”

[0072] Meanwhile, the present invention can be usefully utilized in various situations. More specifically, the data augmentation method and system for learning tabular data according to the present invention can be applied to various industries and services and utilized effectively. For example, in the manufacturing industry, major decisions such as precision optimization, quality control, equipment maintenance, and production forecasting are made based on tabular data composed of sensor data, equipment logs, process variables, etc. The present invention can be usefully applied to the manufacturing industry that utilizes the vast amount of tabular data generated from such manufacturing processes and inspections. However, the fields in which the present invention is utilized are not necessarily limited to the fields mentioned above; it is obvious that it can also be applied and utilized in various other fields (e.g., the medical field, the financial field, the scientific field, urban analysis, etc.).

[0073] Hereinafter, the present invention will be examined in more detail with reference to the attached drawings. FIG. 1 is a conceptual diagram illustrating a data augmentation system for learning tabular data according to the present invention. FIG. 2 is a flowchart illustrating a data augmentation method for learning tabular data according to the present invention, and FIG. 3a, FIG. 3b, FIG. 4a, and FIG. 4b are conceptual diagrams illustrating a data augmentation method for learning tabular data according to the present invention. FIG. 5, FIG. 6, FIG. 7, FIG. 8, and FIG. 9 are formulas and algorithms related to a data augmentation method for learning tabular data according to the present invention. FIG. 10 and FIG. 11 are tables showing the performance of the data augmentation method according to the present invention and an example of the learning result of an artificial intelligence model learned using the learning method according to the present invention. Furthermore, FIG. 12a, FIG. 12b, and FIG. 12c are conceptual diagrams illustrating a method for learning a model by performing segmentation on table data according to an embodiment of the present invention.

[0074] Meanwhile, as illustrated in FIG. 1, a data augmentation system for learning tabular data according to the present invention (hereinafter referred to as the “data augmentation system (100)”) may include at least one of an input unit (110), an output unit (120), a communication unit (130), a storage unit (140), a control unit (150), and an artificial intelligence model (160).

[0075] A data augmentation system (100) according to the present invention may include at least one processor and at least one memory including computer program code. In this case, the memory may perform the role of the storage unit (140) above. In the present invention, the memory and the program code may cooperate with the processor to perform a series of processes described below.

[0076] Although not illustrated, the data augmentation system (100) according to the present invention may include one or more processors, and such processors may include one or more general-purpose processors and / or one or more special-purpose processors (e.g., digital signal processor, tensor processing unit (TPU), graphics processing unit (GPU), neural network processing unit (NPU), application integrated circuit, application semiconductor (ASIC), field programmable gate array (FPGA), quantum processing unit (or quantum processor, QPU), etc.). One or more processors may be configured to execute instructions, computer-readable instructions, and / or other instructions described herein that are stored (or included) in the storage unit (140). The data augmentation method and system according to the present invention may perform the data processing described below in cooperation with memory and at least one processor. The processor may perform a series of operations and data processing using data and information stored in memory. In this case, memory may be a component of the storage unit (140).

[0077] In addition, the data augmentation system (100) according to the present invention can perform data processing and computation processes using quantum gates, quantum entanglement, and quantum superposition states, taking into consideration implementation in a quantum computer environment. For example, the present invention can perform parallel computations based on qubits, and such quantum computations can operate complementarily with existing classical computers.

[0078] Such quantum computers may include parallel computation using qubits and high-speed data processing devices utilizing quantum entanglement, and hardware-based computational optimization using FPGAs and ASICs is possible. In addition, quantum computers may utilize quantum processors capable of qubit-based parallel computation, and data processing efficiency can be improved through a hybrid structure with existing classical computers.

[0079] Meanwhile, the input unit (110) can be configured in various ways as a means of data input. For example, the input unit (110) can be configured to receive user input. The input unit (110) can be configured to receive user input from a user terminal (10). Here, “receiving input” may mean receiving an input signal (or selection signal) corresponding to the user’s input based on input made by the user through the input unit configuration provided in the user terminal (10).

[0080] Here, the user terminal (10) may include at least one of a mobile phone, a smartphone, a notebook computer, a laptop computer, a slate PC, a tablet PC, an ultrabook, a desktop computer, a digital broadcasting terminal, a PDA (personal digital assistants), a PMP (portable multimedia player), a navigation device, and a wearable device (e.g., a smartwatch, a smart glass, a head-mounted display).

[0081] In addition, the input unit (110) in the present invention does not necessarily mean a hardware means, but can be understood as a channel for receiving input from a user.

[0082] The input unit (110) may also be referred to as a user interface module. The input unit (110) may include a touch screen, a computer mouse, a keyboard, a keypad, a touchpad, a trackball, a joystick, a voice recognition module, or other similar devices. However, the present invention does not limit the type of input unit (110).

[0083] Here, user input may include documents, text, images (or videos), voice, etc. In this case, the data augmentation system (100) may further include a module that converts voice into text.

[0084] Next, the output unit (120) can output information through an output unit configuration (e.g., a display unit, a touch screen, a speaker, etc.) provided in a user terminal (10) linked to the data augmentation system (100) according to the present invention. For example, the output unit (120) can output a page (or service page, 1000) linked to the data augmentation system (100) according to the present invention to the display unit of the user terminal. In addition, the output unit (120) does not necessarily mean a hardware means, but can be understood as a channel for outputting results to a user.

[0085] Next, the communication unit (130) may be connected via a wireless or wired network to a user terminal, a server (e.g., a central server, an external server, etc.), a device, and at least one network, etc., to receive or transmit overall data and information necessary for the operation of the data augmentation system (100) according to the present invention.

[0086] The communication unit (130) can support various communication methods depending on the communication standard of the communicating device.

[0087] For example, the communication unit (130) may be configured to communicate with a communication target using at least one of the following technologies: WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), 5G (5th Generation Mobile Telecommunication), Bluetooth (Bluetooth™ RFID (Radio Frequency Identification), Infrared Communication (Infrared Data Association; IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, and Wireless USB (Wireless Universal Serial Bus).

[0088] Next, the storage unit (140, or memory) serves to store various data related to the present invention and may include one or more non-transient computer-readable storage media that can be read and / or accessed by at least one of one or more processors.

[0089] One or more computer-readable storage media may include volatile and / or non-volatile storage components, such as optical, magnetic, organic, or other memory or disk storage devices. In some examples, the storage unit (140) may be implemented using a single physical device (e.g., one optical, magnetic, organic, or other memory or disk storage device), whereas in other examples, the storage unit (140) may be implemented using two or more physical devices.

[0090] The storage unit (140) may include computer-readable instructions and additional data. The storage unit (140) may include a storage necessary to perform at least some of the methods, scenarios, and techniques described herein and / or at least some of the functions of the device and network.

[0091] Furthermore, at least a portion of the storage unit (140) may be a cloud storage or a cloud server. At least a portion of the data corresponding to user input received from the input unit (110) (e.g., table data (300)) and training data may be stored in the storage unit (140).

[0092] That is, the storage unit (140) is sufficient as a space where information necessary for the operation of the data augmentation system (100) according to the present invention is stored, and it can be understood that there are no restrictions on the physical space.

[0093] Furthermore, the storage unit (140) may store a computer program including computer program instructions. Furthermore, the storage unit (140) may store a computer program including computer program instructions that control the operation of the system (100) or control the operation of the control unit (150) when loaded into the processor of the system (100).

[0094] Next, the control unit (150) can perform the role of controlling the overall operation of the data augmentation system (100) related to the present invention. The control unit (150) can process signals, data, information, etc. that are input or output through the components of the data augmentation system (100) described above, or perform a series of data processing to provide or process appropriate information and functions to the user. The control unit (150) can be physically implemented by the processor described above.

[0095] Meanwhile, the present invention aims to provide a data augmentation method and system for learning tabular data that enables an artificial intelligence model to better understand tabular data. More specifically, the present invention aims to provide a data augmentation method and system for learning tabular data that enables an artificial intelligence model to learn information-rich representations of tabular data. Below, we will examine in more detail a method for generating various augmented representations while preserving important information in tabular data, and for training a target artificial intelligence model based thereon.

[0096] In the present invention, table data is input into a first encoder (S210), and a process of obtaining a first embedding vector corresponding to the table data as the output of the first encoder may be performed (S220). The first encoder may be at least one first encoder, and in this specification, it may also be expressed as “at least one encoder.” For convenience, in this specification, the at least one encoder is named and described as the “first encoder.” At this time, the first encoder may be implemented with a single or multiple components.

[0097] As illustrated in FIGS. 3a and 3b, the data augmentation method for learning tabular data according to the present invention may include at least one of an encoder (161), a projection head (163), and a decoder (164). In this case, at least one of the encoder (161), the projection head (163), and the decoder (164) may be a component included in the artificial intelligence model to be learned (hereinafter referred to as the “model to be learned”). Alternatively, at least one of the encoder (161), the projection head (163), and the decoder (164) may be a component separate from the model to be learned. The present invention is not limited to any one of these.

[0098] At this time, the first encoder (161) in the present invention may be configured to include a pre-set normalization technique. In the present invention, to improve the diversity of the representation (or augmented representation, embedding, embedding representation, embedding vector, etc.) generated from the first encoder (161), a pre-set normalization technique may be applied to the first encoder (161).

[0099] For example, a pre-configured regularization technique may include dropout. Dropout is a technique that probabilistically disables or randomly removes (e.g., sets to zero) some neurons while the model is being trained, and it can be a regularization technique to prevent overfitting and improve the model's generalization performance.

[0100] In this invention, dropout is used to allow the same input to be embedded with different representations, thereby introducing noise or variability into the embedding space. This process can be understood as being performed because the structure of table data is fixed, making it difficult to augment in various forms like general images or text, and because it is important to generate different samples (or variations, views, vectors, representations, etc.) in the embedding space.

[0101] In this regard, the size (or degree) of the dropout applied to the first encoder (161) in the present invention may be set (or applied) in various ways by the data augmentation system (100) or the manager of the data augmentation system (100). For example, the dropout rate of the first encoder (161) may be set to 0.1, but the rate set is not necessarily limited thereto.

[0102] Specifically, when the first encoder (161) receives table data (300) as input, it can be trained to generate an embedding vector corresponding to the table data (300). At this time, the control unit (150) can apply a normalization technique set in advance to the first encoder (161) to impart probabilistic variability to the embedding vector generated (or obtained) from the first encoder (161). In the present invention, the imparting of probabilistic variability to the embedding vector generated from the first encoder can also be understood as imparting probabilistic variability to the embedding vector generated from the updated encoder after the first encoder (161) is updated.

[0103] Here, stochastic variability refers to the characteristic that, due to the application of dropout, the output result is not constant for the same input but varies diversely according to a specific probability distribution. In other words, it means that the result is not probabilistically fixed but can differ according to a specific probability distribution, and that values ​​or states can change or shift depending on variability.

[0104] As an example, dropout may be applied to the entire layer or a specific layer of the first encoder (161). Here, the specific layer may include at least one of the final (or last) layer and / or the layer immediately preceding the final layer among the layers included in the first encoder (161). The final layer of the first encoder (161) serves as a location (or layer) for converting input data into a vector in the most compressed and meaningful form, and the final layer may have a structured representation through learning. In this case, the present invention utilizes dropout to randomly remove a portion of the parameters (or weights, weight matrices, parameter matrices, etc.) included in the first encoder (161), thereby enabling the generation of various data transformations.

[0105] When table data (300) is input to a first encoder (161) to which dropout is applied, the first encoder (161) can generate different embedding vectors for table data (300) corresponding to the same input as some neurons are probabilistically deactivated during learning. That is, even if the same input is input to the first encoder (161), various embedding vectors can be generated due to dropout.

[0106] In this way, the present invention applies dropout to the encoder, thereby providing natural diversity to the output generated from the encoder. This can contribute to securing various augmented representations in the embedding space without separate data augmentation. That is, the present invention secures diversity of representations through dropout, and thereby enables learning more robust and generalizable embedding representations, thereby improving the performance of the entire model.

[0107] Furthermore, the present invention can contribute to improving the model's learning performance by effectively constructing positive pairs during the subsequent contrastive learning process. For example, in contrastive learning, different embedding vectors are generated from the same input data and configured into positive pairs to perform learning. At this time, if the difference between representations is too small or too large, learning becomes unstable or inefficient, so diversity in augmentation is required. As discussed above, the present invention applies dropout even to the same input data to enable the encoder to generate various embedding vectors, and by utilizing these in contrastive learning, the model can perform more flexible and generalized learning.

[0108] Furthermore, the control unit (150) may receive table data (300) input from the user terminal (10) and input the received table data (300) into the first encoder (161). As another example, the control unit (150) may input table data (300) stored in a specified storage (e.g., storage unit (140) or memory, etc.) into the first encoder (161). Here, the table data (300) corresponding to the input data may be represented as shown in FIG. 5 (a), and the first encoder (161) may be represented as shown in FIG. 5 (b).

[0109] Table data (300) may be configured to include rows (or records, tuples, samples, etc.) and columns (or columns, fields, features, characteristics, attributes, variables, predictors, etc.). For example, a table refers to a table-shaped data structure composed of rows and columns, and may be a unit for storing and managing data in a relational database. A table may be used to systematically organize and manage data, such as an Excel sheet or a database table. Here, a column refers to a set of data arranged vertically in a table, and each column may represent a specific attribute (or field) of the table. Additionally, a row refers to a line of data arranged horizontally in a table, and each row may represent a single record or tuple in the table. Additionally, a column refers to a set of data arranged vertically in a table, and each column may represent a specific attribute (or field) of the table. In other words, each row represents a single record in the table, and each column can represent a specific attribute of that record.

[0110] The control unit (150) can obtain a first embedding vector (301) corresponding to table data (300) from the first encoder (161). For example, the first encoder (161) can generate a first embedding vector (301) corresponding to table data (300) by performing encoding on the table data (300) through parameters of multiple (or multiple) layers.

[0111] For example, the first embedding vector (301) may include at least one of i) structural information of the table data (300) (location and arrangement information of rows and columns, relationships between cells and hierarchical structure within the table data (300) (e.g., header, subheader distinction), etc.), ii) content information of the cells (content such as text values, numeric data, symbols, etc. contained in each cell, type information of the cells (e.g., characteristics of data types such as numbers, strings, dates, etc.), iii) context information (contextual relationships from other cells surrounding the cell, location-based information necessary to interpret the meaning of a specific cell (e.g., header-data mapping), etc.), iv) interaction information between rows and columns (semantic associations between columns or rows (e.g., a specific column represents a summary for a certain purpose), information from which patterns or correlations can be inferred at the column or row level), and v) an overall structural or semantic summary of the table data (300). That is, the first embedding vector (301) can be understood as a high-dimensional representation (or vector, vector value, etc.) that compresses (or summarizes) structural patterns, cell contents, interactions between features (or characteristics, features, attributes, etc.) included in the table data (300), surrounding context information, etc. In the present invention, the first embedding vector (301) may also be named as “first representation,” “first latent vector,” “first latent representation vector,” “first feature vector,” “first hidden vector,” “first vector,” “first representation vector,” “original representation,” “original vector,” etc.

[0112] Meanwhile, in the present invention, a process may be performed to extract a specific parameter from a first encoder from which a first embedding vector is output based on a preset extraction criterion (S230), and to update a second encoder based on the specific parameter extracted according to the preset extraction criterion (S240, see FIG. 2). The second encoder may be at least one other encoder and may also be expressed as “at least one other encoder.” For convenience, the at least one other encoder is referred to as the “second encoder” in this specification, and the second encoder may be implemented with a singular or multiple components. In the present invention, the “first encoder” is a term referring to “at least one encoder” for convenience, and the “second encoder” is a term referring to “at least one other encoder” for convenience.

[0113] Here, “another encoder” refers to an encoder that is not identical to the first encoder (a physically separate component or a logically distinct module). Accordingly, in this specification, the expression “first encoder” refers to “at least one encoder,” and the expression “second encoder” refers to “at least one other encoder,” and the two terms are not confused with each other. Additionally, each encoder may be implemented as a singular or plural component. Specifically, the encoders mentioned in the present invention may be implemented as a singular (1) or as a plural (e.g., 2, 3, 4 or more). That is, for convenience, they are described in this specification as the “first encoder” and the “second encoder,” respectively, but these encoders may be physically separated modules or multiple functional units within a single piece of hardware.

[0114] As illustrated in FIG. 3b, the control unit (150) can extract a specific parameter (161a) of the first encoder (161) based on a preset extraction criterion. The control unit (150) can extract a specific parameter (161a) of any one of a plurality of parameters included in the first encoder (161) trained to generate the first embedding vector (301) based on a preset extraction criterion. Alternatively, the control unit (150) can extract a specific parameter (161a) from the first encoder (161) to which a preset normalization technique is applied. In this case, it can also be understood as extracting a specific parameter (161a) from the final layer of the first encoder (161) to which a preset normalization technique is applied.

[0115] In the present invention, extracting parameters can also be understood as extracting a parameter matrix or a weight or a weight matrix (W). A specific parameter (161a) is a learned parameter of the final layer to generate a first embedding vector (301) and can perform the role of converting input features to the next layer or output. In this specification, the specific parameter (161a) may also be referred to as the “first parameter.” This specific parameter (161a) may be represented as shown in (d) of FIG. 5.

[0116] Here, the preset extraction criteria may be related to extracting parameters (or weights, weight matrices, etc.) of the final layer (e.g., MLP layer) of the first encoder (161). The control unit (150) may extract specific parameters (161a) from the final layer among the layers of the first encoder (161) according to the preset extraction criteria.

[0117] In the present invention, extracting a specific parameter (161a) from the final layer of the first encoder (161) may mean extracting the last linear transformation coefficient used to generate a learned representation (i.e., a first embedding vector). The specific parameter (161a) of the final layer, compared to the initial layer of the first encoder (161), may include high-dimensional relationships between input features and abstracted information. This specific parameter (161a) may be a core structure (or component) of information compression and abstraction involved in the first encoder (161) converting table data (300) into a first embedding vector (301). That is, extracting a specific parameter (161a) from the final layer of the first encoder (161) can be understood as securing the most important embedding space structure learned by the first encoder (161), thereby enabling a balanced representation augmentation between information preservation and diversity based thereon.

[0118] As illustrated in FIGS. 3a and 3b, the control unit (150) can input the extracted specific parameter (161a) into a specific module (400) that has been set up. The specific module (400) may be included in the configuration of the model to be learned, or it may be configured separately from the model to be learned (in this case, it is included in the configuration of the data augmentation system (100)). In this specification, the specific module (400) may also be named as an “augmentation module,” “data augmentation module,” “parameter update module,” “encoder update module,” “representation-space augmentation module,” “representation-space augmentation model,” “RaTab (Representation-space augmentation for Tabular data) augmentation module,” or “RaTab augmentation model,” etc.

[0119] A specific module (400) may be a module configured to perform decomposition for a specific parameter (161a) using a pre-configured decomposition method. For example, the pre-configured decomposition method may include at least one of Singular Value Decomposition (SVD) and Truncated Singular Value Decomposition (Truncated SVD).

[0120] Here, Singular Value Decomposition (SVD) is a method of decomposing a matrix into three matrix products; for example, it can be understood as decomposing a weight matrix into three matrix products. SVD serves to separate structural patterns of information by dividing the weight matrix into three interpretable components. Additionally, Truncated SVD is a method of approximating the original weight matrix by selecting only the most important subsets of the SVD results; for example, it can be understood as a method that enables dimensionality reduction and information compression by selecting only the top subsets (k) of all singular values ​​and singular vectors produced through SVD.

[0121] A specific parameter (161a) extracted from the first encoder (161) may be a parameter optimized for learning a representation derived from input data (i.e., table data (300)). The specific parameter (161a) may implicitly reflect complex information such as the variance, correlation, direction of important features (or characteristics, features, properties, etc.) of the table data (300), and the interaction structure between variables of the table data (300). That is, the specific parameter (161a) is a learning result that reflects the statistical and / or structural characteristics of the table data (300). In this case, since singular value decomposition is a process of extracting important directions (or patterns) among the structures, it may reflect important relationships of the table data (300).

[0122] A specific module (400) can decompose a parameter (161a, or weight, weight matrix, etc.) extracted from the first encoder (161) using singular value decomposition. A plurality of matrices decomposed through singular value decomposition for such a specific parameter (161a) can be represented as in (e) of 5.

[0123] More specifically, as illustrated in FIGS. 4a and 4b, a specific module (400) can decompose a specific parameter (161a) into a plurality (three) matrix products. The specific module (400) can decompose the specific parameter (161a) into at least one of a first singular vector matrix (or left singular vector matrix, 401) containing a first singular vector, a second singular vector matrix (or right singular vector matrix, 403) containing a second singular vector, and a diagonal matrix (402) containing singular values ​​corresponding to each of the first singular vector and the second singular vector. Here, the first singular vector matrix (401) can be represented as in FIG. 5 (f), the diagonal matrix (402) containing singular values ​​can be represented as in FIG. 5 (g), and the second singular vector matrix (403) can be represented as in FIG. 5 (h).

[0124] Singular value decomposition is a decomposition of the structure contained in a specific parameter (161a) corresponding to the result of the first encoder (161) learning the table data (300), and the multiple matrices obtained through singular value decomposition may each include elements that explain various information contained in the table data (300) (e.g., structure between variables, correlation, important axis, representation dimension, etc.).

[0125] The first singular vector matrix (401) can also be understood as an orthogonal matrix configured to contain the first singular vector. For example, the first singular vector matrix (401) can indicate which output direction (e.g., hidden dimension) important information is projected when the table data (300) is transformed. The first singular vector matrix (401) can indicate which direction of the representation space each row (or sample) is summarized in. That is, the first singular vector matrix (401) can indicate which meaningful vector direction a particular sample is projected in. This first singular vector matrix (401) can be understood as a central direction axis for generating row-based embedding vectors.

[0126] The diagonal matrix (402) is a matrix in which the singular values ​​are sorted in descending order (or the singular values ​​are placed along the diagonal), and each singular value may represent the magnitude in the direction of the important axis of a specific parameter (161a). The diagonal matrix (402) may contain information expressing, in quantitative terms, how much feature information each axis (or basis direction) explains (i.e., the order of importance of the column structure). In this case, the larger the singular value, the better the direction (or basis) explains the table data (300). For example, if the first singular value is large, it may indicate that most of the variance of the table data (300) is concentrated on the first basis axis. That is, it may quantitatively indicate which column combinations in the table data (300) explain the main information or are important.

[0127] The second singular vector matrix (403) is the transpose matrix of the second singular vector and may contain axial information of the input feature space. The second singular vector matrix (403) may form the basic axes in singular value decomposition. For example, the second singular vector matrix (403) may contain information on how each column (or feature) included in the table data (300) is combined in a certain direction to form a major information axis. Here, the axis is a direction in the vector space and may be formed by a combination of column directions (or feature space) (i.e., the axis is a significant directionality created by combining features). That is, the second singular vector matrix (403) may contain information indicating correlations, combination structures, statistical structures, dependencies, etc., between the columns of the table data (300).

[0128] In one embodiment, the columns of table data (300) may determine which combinations are important through a diagonal matrix (402) and a second singular vector matrix (403), and the rows of table data (300) may determine how they are projected onto the important combinations through a first singular vector matrix (401). Singular value decomposition may be a process of summarizing the structure of columns (or variables) and projecting rows into a new embedding space based thereon.

[0129] A specific module (400) can perform a low-rank approximation based on singular value decomposition on a specific parameter (161a) extracted from the first encoder (161) so that only singular values ​​satisfying a preset standard (or preset ratio standard) and vectors corresponding to said singular values ​​are retained. For example, low-rank approximation is a technique for approximating high-dimensional data into a low-dimensional form, and can be used to reduce the size of the data while maintaining the core structure and information of the original data. In the case of a matrix, by approximating it with a low-rank matrix, the main information of the original matrix can be preserved while reducing storage space and computational costs. That is, low-rank approximation can remove noise and simplify the data into a form that is better understood by finding a low-rank approximation for the original data. In this process, the dimensionality of the data can be reduced while important information can be preserved as much as possible.

[0130] Specifically, a specific module (400) may specify (or select) at least one specific singular value that satisfies a preset standard, and generate a new parameter to be used for updating the second encoder (162) based on the specified (or selected) singular value.

[0131] Here, the pre-set criteria may be related to satisfying the top-k. In the present invention, k can be set as a certain percentage (p%) relative to the total rank. As an example, the value of p can be selected from a first value (50%), a second value (60%), and a third value (70%). If p = 0.6, only 60% of the total singular values ​​can be retained, and if p = 0.7, only 70% of the total singular values ​​can be retained. In this case, k can also be understood as a hyperparameter (see FIG. 6 (e)), and such a hyperparameter can determine the rank to be retained. The hyperparameter can control to what extent important information within the latent representation is preserved. It is obvious that these pre-set ratio criteria are not limited to any one and can be varied by the data augmentation system (100) or the manager of the data augmentation system (100).

[0132] As seen above, the singular values ​​included in the diagonal matrix (402) can be sorted in descending order. The larger the value of a singular value, the greater the amount of information (or variance) that the corresponding axis (singular vector direction) occupies in the table data (300). That is, the largest singular value explains a large amount of variance in the table data (300), which may correspond to an important pattern or structure. On the other hand, a small singular value may explain an unimportant axis direction (e.g., noise or fine patterns) within the table data (300). Therefore, identifying (or selecting) a specific singular value that satisfies a pre-set criterion can also be understood as preserving the most important information (principal component direction).

[0133] A specific module (400) may specify a specific value (or specific specific value) satisfying a preset standard and a first specific vector and a second specific vector corresponding to the specific value satisfying the preset standard in order to maintain a specific value among the specific values ​​that satisfy a preset standard. A specific module (400) may specify a specific value satisfying a preset standard among the specific values ​​corresponding to each of the first specific vector and the second specific vector, and specify a first specific vector (or specific first specific vector) and a second specific vector (or specific second specific vector) corresponding to the specified specific value in order to maintain the specified specific value and the specified first specific vector and second specific vector.

[0134] Alternatively, a specific module (400) may reduce the dimensions of the first singular vector matrix (401), the diagonal matrix (402), and the second singular vector matrix (403) (e.g., reduced to k) so that only the specified singular value and the first singular vector and the second singular vector corresponding to the specified singular value are retained.

[0135] For example, a specific module (400) can identify a first singular vector among the first singular vectors included in the first singular vector matrix (401) that satisfies a pre-set criterion. Then, the specific module (400) can reduce the dimension of the first singular vector matrix (401) to maintain the identified first singular vector (or the top k first singular vectors) as being continuously included in the first singular vector matrix (401) and to remove (or exclude) the remaining first singular vectors excluding the identified first singular vector. The dimensionally reduced low-dimensional first singular vector matrix (or the left singular vector matrix corresponding to the top k singular values, the first singular vector matrix composed of the identified first singular vector, etc., 411) can be represented as in FIG. 4a and FIG. 6 (b).

[0136] As another example, a specific module (400) can identify a singular value among the singular values ​​included in the diagonal matrix (402) that satisfies a pre-set criterion. Then, the specific module (400) can reduce the dimension of the diagonal matrix (402) to maintain the identified singular value (or the top k singular values) in the diagonal matrix (402) and to remove the remaining singular values ​​excluding the identified singular value. The lower-dimensional diagonal matrix with reduced dimension (or a diagonal matrix containing the top k singular values, a diagonal matrix configured to include a specific singular value, etc., 412) can be represented as in FIG. 4a and FIG. 6 (c).

[0137] As another example, a specific module (400) can identify a second singular vector among the second singular vectors included in the second singular vector matrix (403) that satisfies a pre-set criterion. Then, the specific module (400) can reduce the dimension of the second singular vector matrix (403) to maintain the identified second singular vector (or the top k second singular vectors) in the second singular vector matrix (403) and to remove the remaining second singular vectors excluding the identified second singular vector. The dimensionally reduced low-dimensional second singular vector matrix (or the right singular vector matrix corresponding to the top k singular values, the second singular vector matrix composed of the identified second singular vector, etc., 413) can be represented as in FIG. 4a and FIG. 6 (d).

[0138] Furthermore, a specific module (400) can generate parameters based on at least one of a specified singular value, a specified first singular vector, and a second singular vector. The specific module (400) can generate (or output) new parameters to be used for updating the second encoder (162) based on at least one of a low-dimensional first singular vector matrix (411), a low-dimensional diagonal matrix (412), and a low-dimensional second singular vector matrix (413).

[0139] In this case, the process of generating a new parameter to be used for the update may be a process of performing a low-dimensional approximation based on singular value decomposition on a specific parameter (161a) extracted from the first encoder (161) to approximate it into a low-dimensional parameter (or weight, weight matrix, parameter matrix, etc.) that preserves only the components satisfying a preset criterion. Alternatively, it may be a process of transforming the specific parameter (161a) into a new parameter (Wk, 162a) based on at least one of a specific singular value, a specific first singular vector, and a second singular vector while preserving important information. Here, preserving important information means that the singular value decomposition is performed centered on a direction that can explain the table data (300) well, and the said direction may be the direction with the least variance. The specific module (400) may preserve important information based on these directions with low variance.

[0140] That is, in order to preserve important information (or key information) of table data (300), a specific module (400) can approximate a low-dimensional parameter that preserves only the singular values ​​satisfying a preset standard and the first singular vector and second singular vector corresponding to said singular values ​​after performing singular value decomposition on a specific parameter (161a) of the final layer of the first encoder (161).

[0141] For example, as illustrated in FIGS. 4a and 4b, a parameter (162a) generated from a specific module (400) can be generated through the multiplication of a low-dimensional first singular vector matrix (411), a low-dimensional diagonal matrix (412), and a low-dimensional second singular vector matrix (413). The parameter (162a) corresponding to the output of this specific module (400) may also be named “updated parameter,” “dimensionally reduced parameter,” “information-reduced parameter,” “low-dimensional parameter,” “second parameter,” “truncation parameter,” “transformed parameter,” “updated weight matrix,” “dimensionally reduced weight matrix,” “low-dimensional weight matrix,” “second weight matrix,” “truncation weight matrix,” “transformed weight matrix,” or “information-reduced weight matrix,” etc. Such a parameter (162a) can be represented as in FIG. 6 (a).

[0142] Meanwhile, the control unit (150) can update the second encoder (162) using the output of a specific module (400) for a specific parameter (161a). The parameter (162a) corresponding to the output from the specific module (400) can be used to update the second encoder (162).

[0143] As illustrated in FIGS. 4b to 5b, the control unit (150) can update the second encoder (162) using the parameter (162a) to obtain (or generate) the second embedding vector (302).

[0144] Here, the process of updating the second encoder (162) may be a process of inputting a parameter (162a) output from a specific module (400) into the second encoder (162) to set the parameter of the second encoder (162) to the parameter (162a) output from the specific module (400). Alternatively, it may be a process of updating a specific parameter (161a) of the final layer of the first encoder (161) to the parameter (162a) output from the specific module (400). Alternatively, it may be a process of updating a specific parameter (161a) of the final layer of the first encoder (161) to the parameter (162a) output from the specific module (400). Alternatively, the process may involve inputting a parameter (162a) output from a specific module (400) into a first encoder (161), and setting a specific parameter (161a) of the first encoder (161) to the parameter (162a) output from the specific module (400) to newly configure an updated second encoder (162).

[0145] Furthermore, in the present invention, table data is input into an updated second encoder (S250), and an augmented second embedding vector corresponding to the table data is obtained from the second encoder (S260).

[0146] The update process of the second encoder (162) described above may be a process of updating the second encoder (162) using the output of a specific module (400) to generate a new embedding vector (or second embedding vector) augmented for the first embedding vector (301) obtained from the first encoder (161) (or for the first embedding vector (301), from the first embedding vector (301), etc.). Alternatively, it may be a process of updating the second encoder (162) using the output of a specific module (400) to generate a positive representation (or positive view, or positive sample, positive variation, etc.) for the first embedding vector (301) obtained from the first encoder (161).

[0147] This updated second encoder (162) may be configured to generate an augmented second embedding vector corresponding to the table data (300) based on the set parameter (162a). This updated second encoder (162) may be represented as (f) in FIG. 6.

[0148] In this case, the updated second encoder (162) may be an encoder with dropout applied. As seen above, since the second encoder (162) is updated based on the parameter (161a) extracted from the first encoder (161) with dropout applied, the updated second encoder (162) may be configured to include the dropout applied to the first encoder (161).

[0149] When the second encoder (162) receives table data (300) as input, it can be trained to generate an embedding vector (e.g., a second embedding vector) corresponding to the table data (300). At this time, if a normalization technique set in advance is applied to the second encoder (162), probabilistic variability may be imparted to the embedding vector generated (or obtained) from the second encoder (162). For example, when table data (300) is input to the second encoder (162) to which dropout is applied, the second encoder (161) may generate different embedding vectors for the same input table data (300) as some neurons are probabilistically deactivated during training. That is, even if the same input is input to the second encoder (162), various embedding vectors may be generated due to dropout. That is, the second encoder (162) to which the pre-set decomposition method and pre-set normalization technique are applied can generate various embedding vectors even for table data (300) corresponding to the same input.

[0150] Meanwhile, as illustrated in FIGS. 3a and 3b, the control unit (150) inputs table data (300) into the second encoder (162) and obtains a second embedding vector (302) corresponding to the table data (300) from the second encoder (162). These second embedding vectors (302) can be represented as shown in (g) and (h) of FIG. 6. For example, the second embedding vector (302) may include i) structural information of the table data (300) (structural attributes such as row and column arrangement, correlation between columns, data type (numerical, categorical, etc.), representation of the hierarchical structure of the table data (300) (header / subheader, column grouping, etc.), ii) content information at the cell and feature level (embedded representations of numeric / categorical values ​​included in each feature (i.e., semantically reconstructed vectors rather than raw values), iii) semantic relationships in latent space (representations compressed or refined based on the axis with the highest information content (principal component) in the table through the main axis direction maintained by singular value decomposition, diversity within the same class, representations suitable for boundary learning between different classes, etc.), iv) interaction information between multiple features (non-linear relationships and covariance structures between different columns in the table, high-dimensional It may include at least one of the important relationship axes between features in space, etc. That is, the second embedding vector (302) can be understood as a high-dimensional representation (or vector, vector value, etc.) that possesses diversity and distinctiveness of representation while maintaining the structure, content, relationships, patterns, and semantic axes of the table data (300), rather than raw data for the table data (300). In the present invention, the second embedding vector (302) may also be named as “second latent vector,” “second latent representation vector,” “second feature vector,” “second hidden vector,” “second vector,” “second representation vector,” “augmented representation,” “augmented vector,” or “second representation,” etc.

[0151] This second embedding vector (302) contains principal component-centered compressed features compared to the first embedding vector (301) examined above, thereby improving diversity and generalization performance in the representation space. For example, the second embedding vector (302) is a representation generated through a second encoder (162) configured based on a truncated singular value decomposition result in which singular value decomposition is performed on a specific parameter (161a) of the final layer of the first encoder (161) corresponding to the first embedding vector (301), and only the upper subset number (upper k) of singular values ​​calculated therefrom and the corresponding principal axis direction vectors (principal components) are preserved. At this time, since the truncated singular value decomposition is configured to selectively preserve important structural information centered on the principal component direction with high variance within the embedding space (or entire representation space) of the first encoder (161), the second embedding vector (302) has the characteristic of having a reduced amount of information compared to the first embedding vector (301), while appearing in a form that emphasizes key information and semantic features within the embedding space.

[0152] Additionally, the diversity of representations of the second embedding vector (302) can be enhanced by the stochastic variability of dropout applied to the encoder. For example, since the second embedding vector (302) is generated from the second encoder (162) to which dropout is applied, it is possible to generate various representations as the configuration of neuron deactivation within the encoder varies probabilistically for the same input. This stochastic variability increases representation diversity in the latent space, which can contribute to expanding the distribution width and intra-class variance of representations while maintaining semantic similarity between positive pairs during self-supervised learning. Consequently, this enables more robust and generalizable representation learning in contrastive learning.

[0153] Subsequently, the control unit (150) can configure (or set) the first embedding vector (301) and the second embedding vector (302) as a single positive pair, and optimize in parallel the contrast loss that maximizes the similarity between the two vectors and the reconstruction loss that reconstructs the input data based on the original representation.

[0154] More specifically, as seen above, the control unit (150) can generate different embedding vectors (i.e., an augmented second embedding vector (302) different from the first embedding vector (301)) by using a second encoder (162) updated based on specific parameters (161a) of the first encoder (161) that generates the first embedding vector (301). That is, different first embedding vectors (301) and second embedding vectors (302) are generated through augmentation even for the same input table data (300), and at this time, the second embedding vector (302) may be a vector expressed differently from the first embedding vector (301) due to the parameter (162a) structure and dropout of the updated second encoder (162). That is, the first embedding vector (301) and the second embedding vector (302) generated for the same table data (300) can be mapped to different embedding spaces (or locations), and the distance between the first embedding vector (301) and the second embedding vector (302) can be appropriately maintained for contrast learning. Here, the phrase "the distance is appropriately maintained" can be understood as meaning that since the first embedding vector (301) and the second embedding vector (302) are vectors generated for the same input data, they are kept close enough to imply the same meaning (i.e., they may belong to the same label).

[0155] As such, the present invention utilizes different first embedding vectors (301) and second embedding vectors (302) as positive pairs to maximize the SSL learning effect and simultaneously ensure diversity of representation and robustness. More specific details regarding this will be described later.

[0156] Thus, the present invention generates meaningful augmentations in the embedding space by augmenting data based on singular value decomposition, and can augment (or generate) various representations while preserving important information of the table data. In particular, the present invention, when combined with diversity resulting from dropout, can further enhance the robust and generalizable representation learning of the model.

[0157] In other words, to preserve important structures during the data augmentation process, the present invention maintains important information by focusing on the most important eigenspace directions through truncated singular value decomposition, and can generate various representations by introducing variability in the augmentation process through the probabilistic properties of dropout.

[0158] Meanwhile, in the present invention, a process of training a target model using a first embedding vector and a second embedding vector may be performed (S270, see FIG. 2).

[0159] The control unit (150) can define a first loss function and a second loss function, respectively, using at least one of a first embedding vector and a second embedding vector for learning a model (or module) to be learned.

[0160] First, as illustrated in FIGS. 3a and 3b, the control unit (150) can input a first embedding vector (301) and a second embedding vector (302) to a projection head (163), respectively. At this time, the control unit (150) can form a positive pair (or pair) of the first embedding vector (301) and the second embedding vector (302), and train a target model using the positive pair that includes the first embedding vector (301) and the second embedding vector (302) (or the first embedding vector (301) and the second embedding vector (302) forming a pair). The first embedding vector (301) and the second embedding vector (302) generated for the same table data (300) belong to the same label (e.g., positive label) but may represent different views. Here, different views mean that the way the data is represented is different. For example, the first embedding vector (301) may be a representation generated through the original weights of the encoder based on the same table data (300), and the second embedding vector (302) may be a representation generated with some modification of the weights by applying SVD (Singular Value Decomposition) or dropout to the last layer of the encoder based on the same table data (300). Thus, the two vectors generated from the same data but differing in the generation process have different representation methods (different views), which allows them to exist in different locations in the representation space while maintaining the same meaning.

[0161] The projection head (163) can convert the first embedding vector (301) and the second embedding vector (302) into a lower-dimensional (or lower dimension, smaller dimension, etc.) space, and as a result of projecting the first embedding vector (301) and the second embedding vector (302) respectively, output (or generate) a first projection vector (311) corresponding to (or about) the first embedding vector (301) and a second projection vector (312) corresponding to the second embedding vector (302).

[0162] The control unit (150) can obtain a first projection vector (311) corresponding to a first embedding vector (301) and a second projection vector (312) corresponding to a second embedding vector (302) from the projection head (163). Then, the control unit (150) can define a first loss function for contrast learning using the first projection vector (311) and the second projection vector (312). The first loss function may also be named a “contrast loss function” or a “contrast loss function,” etc. Such a first loss function may be represented as shown in (f) of FIG. 7. Here, the first projection vector (311) can be represented as in FIG. 8 (a), the second projection vector (312) can be represented as in FIG. 8 (b), and the projection head (163) can be represented as in FIG. 8 (c). Additionally, the temperature scaling coefficient in the first loss function can be represented as in FIG. 8 (d).

[0163] The first loss function may be defined to learn the similarity between the first embedding vector (301) and the second embedding vector (302). The control unit (150) may use the first loss function to train at least one of the first encoder (161) and the projection head (163) to distinguish (or identify, differentiate, detect, etc.) the difference between features.

[0164] The control unit (150) can calculate a contrast loss (or first loss, contrast loss value, first loss value, etc.) based on the similarity between the first projection vector (311) and the second projection vector (312) using a first loss function, and update at least one parameter of the first encoder (161) and the projection head (163) based on the calculated contrast loss.

[0165] That is, at least one of the encoder (161) and the projection head (163) can calculate the similarity between the first projection vector (311) and the second projection vector (312) using a first loss function, and the similarity between the first projection vector (311) and the second projection vector (312) can be learned in a direction that maximizes it.

[0166] That is, the first encoder (161) or projection head can learn the similarity between the first embedding vector (301) corresponding to the table data (300) and the augmented second embedding vector (302). This can be understood as learning to maximize (make closer) the similarity between the first embedding vector (301) and the augmented second embedding vector (302) generated for the same table data (300), and to minimize (make further apart) the similarity between the table data (300) and the embedding vector generated for a different table data.

[0167] Next, as illustrated in FIGS. 3a and 3b, the control unit (150) can input a first embedding vector (301) to the decoder (164). The decoder (164) can perform the role of restoring the first embedding vector to correspond to the original input data form. The decoder (164) can reconstruct the first embedding vector (301) to correspond to table data (300) and generate a reconstructed output value (321). This can be understood as an output value for determining similarity with the table data (300) corresponding to the input.

[0168] The control unit (150) obtains a reconstructed output value (321) from the decoder (164) and can define a second loss function for reconstructed learning using table data (300) and the reconstructed output value. The second loss function may also be named a reconstructed loss function. This second loss function is intended to model the relationship within the sample and can be represented as in FIG. 7 (a). Here, the data batch can be represented as in FIG. 7 (b), the decoder (164) can be represented as in FIG. 7 (c), the number of batch samples in the data can be represented as in FIG. 7 (d), and the reconstructed output value (321) through the decoder (164) can be represented as in FIG. 7 (e).

[0169] The second loss function may be defined to learn the relationship between the table data (300) and the reconstructed output value (321). The control unit (150) may use the second loss function to train at least one of the first encoder (161) and the decoder (164).

[0170] The control unit (150) can calculate a reconstruction loss (or second loss, reconstruction loss value, second loss value, etc.) between table data (300) and a reconstructed output value (321) using a second loss function, and update at least one parameter of the first encoder (161) and decoder (164) based on the calculated reconstruction loss.

[0171] That is, at least one of the first encoder (161) and the decoder (164) can learn the relationship between the table data (300) and the reconstructed output value (321) using the second loss function. Here, learning the relationship between the table data (300) and the reconstructed output value (321) may mean that at least one of the first encoder (161) and the decoder (164) is trained to understand and express the structure and features well enough to restore the table data (300) input as input data. That is, through the process of converting the table data (300) into the first embedding vector (301) and then restoring it, the structure of the original data can be identified and the structure can be reconstructed.

[0172] Furthermore, the control unit (150) can define a final loss function (or total loss function) using a first loss function and a second loss function, and train a target model using the final loss function. In this case, the process of training a target model using the final loss function can also be understood as a process of updating at least one parameter among the first encoder (161), projection head (163), and decoder (164) so ​​that the difference between the first loss value and the second loss value calculated by each calculating the first loss function and the second loss function included in the final loss function is minimized.

[0173] The final loss function can be defined as a weighted combination (or weighted sum) of a first loss value calculated using a first loss function and a second loss value calculated using a second loss function. Here, the final loss function can be represented as shown in (e) of FIG. 8, and the weighting coefficients that control the importance of the first loss and the second loss can be represented as shown in (f) of FIG. 8.

[0174] After the pre-training process (e.g., SSL) using the final loss function is completed, the control unit (150) can fine-tun the first encoder (161) after removing at least one of the remaining projection head (163) and decoder (164), excluding the first encoder (161). In the fine-tuning process, a data set containing (or assigned) labels (or labels) can be used to perform refinement for task-specific performance improvement. In this fine-tuning process, the first encoder (161) is set to a retrainable state and trained with a new target model, so that the parameters of the first encoder (161) can be fine-tuned to better adapt to the specific task performed by the target model. For example, the trained first encoder (161) can be connected to the target model and used to train the target model, in which case the control unit (150) can train the target model based on the trained first encoder (161). At this time, the target model may be a model separately defined to perform a pre-set task (e.g., prediction or classification or regression, etc.) and may be configured as a separate structure from the trained first encoder (161).

[0175] For example, the model to be trained may include at least one of a classification model (or classification module, classifier, etc.) and a regression model (or regression module) with an encoder (e.g., the first encoder (161)) as a backbone to perform downstream tasks such as prediction, classification, or regression on tabular data. Alternatively, the model to be trained may include at least one of a Multi-layer Perceptron (MLP) based model, an FT-Transformer based model, or a T2G-Former based model. However, the types of models to be trained are not limited to the examples mentioned above and may include various other models.

[0176] In one embodiment, when the learned first encoder (161) is connected to the model to be trained, a first embedding vector corresponding to table data and an augmented second embedding vector are obtained from the learned first encoder (161), and the model to be trained can be trained using the obtained first embedding vector and second embedding vector.

[0177] Finally, the artificial intelligence model (160) trained through the method described above can effectively detect complex feature interactions and relationships between samples by integrating the goals of reconstruction learning and contrastive learning, and the trained artificial intelligence model (160) can be applied to industries and services utilizing various tabular data and utilized effectively. In addition, since the learning method according to the present invention is applied only to the final layer of the encoder, it is compatible with any encoder structure.

[0178] Meanwhile, the process for the data augmentation method and system according to the present invention and the learning method of the artificial intelligence model (160) described above can also be represented by the algorithm of FIG. 9. For example, after initializing the parameters of the encoder, decoder, and projection head, iterative learning is performed during the entire learning epoch. In each epoch, data is loaded in batch units, and an encoder is applied to each batch to embed the input data. Subsequently, weights existing in the last layer of the encoder are extracted, and Singular Value Decomposition (SVD) is performed on the corresponding weights to identify the main structure. Information of the higher rank among the SVD results is selected to reconstruct the weights, and the weights of the last layer of the encoder are updated to the corresponding values. Then, the same input batch is embedding again through the updated encoder to obtain an augmented representation. The original representation and the augmented representation are each passed through a decoder to generate a restoration result, and the Mean Squared Error (MSE) between the original input and the restored input is calculated. In addition, the original embedding and the augmented embedding are passed through the projection head, respectively, to calculate the contrast loss between the two representations. Furthermore, the final loss is composed of a weighted sum of the reconstruction loss and the contrast loss, and the parameters of the encoder, decoder, and projection head are updated to minimize the reconstruction loss and the contrast loss values. By repeating this process for all batches and epochs, an encoder is pre-trained that preserves various structural information and ensures representation diversity.

[0179] In one embodiment, as illustrated in FIG. 10, it can be seen that the data augmentation method according to the present invention described above exhibits superior performance compared to other methods, including all tasks and data sets. That is, the data augmentation method according to the present invention consistently demonstrates superior performance across various tasks and data sets, which indicates that the augmentation method according to the present invention is effective.

[0180] In another embodiment, as illustrated in FIG. 11, it can be seen that the data augmentation method according to the present invention exhibits very high performance, surpassing conventional methods by a significant margin in all data sets. That is, since the data augmentation method according to the present invention performs transformations within the learned representation space itself, it can maximize the effect of augmentation while maintaining the intrinsic structure of the data.

[0181] Meanwhile, the data augmentation method and system for learning tabular data according to the present invention described above can perform binning on the tabular data to efficiently learn the tabular data.

[0182] More specifically, the control unit (150) performs segmentation by assigning empty values ​​to various data values ​​included in the table data set according to a preset standard, and can train a target model (e.g., encoder, autoencoder, etc.) using the segmentation data obtained as a result of performing segmentation on the table data set. At this time, since the extraction of encoder parameters, the application of a preset normalization technique (e.g., dropout), the performance of a preset decomposition method (e.g., Singular Value Decomposition (SVD), Truncated SVD, etc.), encoder updates, contrast learning, and reconstruction learning have been explained in more detail above, they will be explained only briefly below to avoid duplication of explanation.

[0183] For example, table data can be structured in a table form that includes row information representing information on various multiple heterogeneous features of individual samples and column information representing information on one feature of multiple samples.

[0184] Tabular data can possess both categorical features, such as gender and country, which are difficult to represent numerically and can be expressed using combinations of characters, nouns, or specific words, and numerical features, such as height, weight, and age, which can be expressed numerically. As a result, tabular data can have heterogeneous features.

[0185] In order for an artificial intelligence model to efficiently learn the characteristics of table data, there is a need to sufficiently handle the heterogeneous characteristics of such table data. To this end, the present invention generates binning data by performing binning on table data, and by training a target model based on table data and binning data, an artificial intelligence model (160) capable of effectively extracting features that appropriately reflect the heterogeneous characteristics of table data can be constructed.

[0186] Here, interval data may refer to specific numerical values ​​assigned to data included in table data. For example, interval data may be data corresponding to bin values, such as representative values ​​assigned to receiving data within a predetermined range or specific numerical values ​​assigned to categorical data. As such, interval data generated by performing intervals on table data may be data to which inductive bias has been applied to the table data.

[0187] Below, we will examine in more detail a method for performing intervals on table data according to one embodiment of the present invention and training a target model based thereon.

[0188] As illustrated in FIGS. 12a and FIGS. 12c, the control unit (150) can obtain table data (1200).

[0189] The control unit (150) can receive (or acquire) table data (1200) by receiving input from a user terminal (10) or from an external source. For example, the control unit (150) may receive table data (1200) directly from an external source, or may extract table data (1200) from document data (e.g., papers, books, patent documents, reports, etc.) received from an external source. The control unit (150) can extract table data structured in a table form from input document data based on a machine learning model and / or a document understanding model included in the data augmentation system (100).

[0190] As an example, tabular data relates to personal information such as the age, height, weight, and gender of multiple people, and may include information in which various characteristic information regarding multiple samples is structured in a table format. Here, multiple samples may correspond to multiple people.

[0191] Alternatively, tabular data may contain various types of information related to various industrial fields. In this case, the tabular data may include numerical information that can be expressed as continuous numbers, and categorical information that is difficult to express numerically and can be expressed as a combination of characters, nouns, or specific words.

[0192] For example, table data is information related to products held by a store, and may include numerical information such as the product's year, weight, and length, as well as categorical information such as the product's brand name and color.

[0193] As another example, table data is information related to items produced in a manufacturing plant, and may include numerical information such as the height, width, and depth of the produced items, as well as categorical information such as manufacturing process line information and manufacturing process management supervisor information.

[0194] As such, table data may include at least one of numeric information and categorical information, and may be data of information structured in the form of a table containing various types of information regarding multiple samples. Additionally, table data may include information regarding various types of samples related to various industrial fields.

[0195] Such tabular data may exhibit heterogeneous features by simultaneously including numerical and categorical information. Additionally, since information regarding multiple samples included in the tabular data can have irregular values ​​regardless of specific patterns, the tabular data may possess irregularities.

[0196] The control unit (150) can perform segmentation on the table data (1200) to obtain a plurality of segmented table data (1201, 1202, 1203). In the present invention, the plurality of segmented table data may also be named as “a plurality of segmented data,” “a plurality of segmented data,” “a plurality of segmented table data,” or “a plurality of segmented data sets,” etc.

[0197] Specifically, the control unit (150) can perform segmentation on table data (1200) using a segmentation module (500) and obtain a plurality of segmented table data (1201, 1202, 1203) output (or generated) from the segmentation module (500). In this case, the segmentation module (500) may be a component included in the learning target model. Alternatively, the segmentation module (500) may be a component separate from the learning target model (e.g., included in the data augmentation system (100)). The present invention is not limited to either one.

[0198] Segmentation may refer to assigning a single representative value to information included within a predetermined (or at least partial) numerical range among the various information contained in table data. Here, the representative value may be referred to as a bin value. In this case, segmented table data may refer to data concerning multiple bin values ​​generated by performing segmentation on the table data. The bin value, which serves as the representative value to be assigned to the various information contained in the table data, may be set irregularly.

[0199] The segmentation module (500) may be a module (or model, encoder, autoencoder, etc.) trained to output a plurality of segmented table data (1201, 1202, 1203) generated by performing segmentation on the table data (1200) when table data (1200) is input. The segmentation module (500) may be a module trained to effectively extract features including irregularities in the table data. As a result of performing segmentation through the segmentation module (500), the table data (1200) may be divided into a plurality of sub-table data. In the present invention, this may be named as segmented table data (1201, 1202, 1203). Furthermore, the plurality of sub-table data may also be named as a plurality of table data. As a result of performing segmentation through the segmentation module (500), a plurality of segmented table data (1201, 1202, 1203) may be obtained.

[0200] Additionally, the segmentation module (500) can group data samples with similar values ​​by performing segmentation on the table data (1200). Through this, the segmentation module (500) can extract data features that are robust against errors that may occur due to the analysis of ungrouped individual data.

[0201] That is, the segmentation module (500) can be configured to effectively extract features of the table data (1200) based on segmentation of the table data (1200).

[0202] In one embodiment, the segmentation module (500) may assign different bin values ​​to different categorical data for a plurality of samples included in the table data (1200). In this case, by assigning multiple bin values ​​to different categorical data included in the table data (1200), the categorical data may be converted into numerical data.

[0203] In another embodiment, the table data (1200) may include a plurality of column information, which are raw feature values ​​related to features of a plurality of samples, and the segmentation module (500) may convert the raw feature values ​​of the table data into a plurality of segmented table data (1201, 1202, 1203). The raw feature values ​​in the form of column information included in the table data may be converted into a plurality of segmented table data (1201, 1202, 1203), in which case the raw feature values ​​may each be included in a plurality of arbitrarily divided segments, and for each of the raw feature values, an empty value, which is a representative value corresponding to the segment, may be assigned.

[0204] Furthermore, the number of multiple empty values ​​assigned to multiple numeric data included in the table data (1200) can be set in various ways. For example, the number of multiple empty values ​​may be from 2 to 100. However, the number of multiple empty values ​​is not necessarily limited to this, and the number of multiple empty values ​​may be set to be less than 2 or more than 100.

[0205] In this way, among multiple numerical data for multiple samples included in the table data, the same single bin value may be assigned to some numerical data within a predetermined range. Additionally, different bin values ​​may be assigned to each of multiple categorical data for multiple samples included in the table data. Multiple bin values ​​assigned to multiple samples may also be understood as intervalized table data.

[0206] Accordingly, the control unit (150) can obtain from the segmentation module (500) segmented table data in which a plurality of numeric data included in the table data are replaced with empty values ​​of the corresponding range, or segmented table data in which a plurality of categorical data are replaced with empty values ​​of numeric form.

[0207] Since such intervald table data is generated by replacing numerical or categorical data included in the existing table data with empty values ​​instead of their original values, intervald data may be data to which inductive bias has been applied to the table data.

[0208] Meanwhile, as illustrated in FIG. 12a and FIG. 12c, the control unit (150) can input a plurality of segmented table data (1201, 1202, 1203) corresponding to the table data (1200) obtained from the segmentation module (500) to the first encoder (161), respectively. Then, the control unit (150) can obtain embedding vectors (1201a, 1202a, 1203a) corresponding to each of the plurality of segmented table data (1201, 1202, 1203) generated (or output) from the first encoder (161).

[0209] The first encoder (161) can be trained to generate an embedding vector (1201a) corresponding to each of the multiple segmented table data (1201, 1202, 1203) when each of the multiple segmented table data (1201, 1202, 1203) is input. In this case, as the multiple segmented table data (1201, 1202, 1203) are input to the first encoder (161), multiple embedding vectors corresponding to the multiple segmented table data (1201, 1202, 1203) can be generated from the first encoder (161).

[0210] For example, the first encoder (161) can generate a first embedding vector (1201a, 1202a, 1203a) corresponding to the segmented first table data (1201). The control unit (150) can obtain the first embedding vector (1201a) generated for the segmented first table data (1201) from the first encoder (161).

[0211] As another example, the first encoder (161) can generate a second embedding vector (1202a) corresponding to the segmented second table data (1202). The control unit (150) can obtain the second embedding vector (1202a) generated for the segmented second table data (1202) from the first encoder (161).

[0212] As another example, the first encoder (161) can generate a third embedding vector (1203a) corresponding to the segmented third table data (1203). The control unit (150) can obtain the third embedding vector (1203a) generated for the segmented third table data (1201) from the first encoder (161).

[0213] Additionally, the control unit (150) can extract a specific parameter (161a) of the first encoder (161) based on a preset extraction criterion. Based on the preset extraction criterion, the control unit (150) can extract a specific parameter (161a) among a plurality of parameters included in the first encoder (161) which is trained to generate a plurality of embedding vectors (1201a, 1202a, 1203a) corresponding to each of the plurality of segmented table data (1201, 1202, 1203). The control unit (150) can extract a specific parameter (161a) from the final layer among the layers of the first encoder (161) according to the preset extraction criterion.

[0214] Next, the control unit (150) can input the extracted specific parameter (161a) into the specific module (400). As seen above, the specific module (400) can be configured to perform singular value decomposition for the specific parameter (161a) using a pre-set decomposition method, and then generate (or output) a new parameter to be used for updating the second encoder (162).

[0215] For example, a specific module (400) can approximate a low-dimensional parameter that preserves only the singular values ​​satisfying a preset standard and the first singular vector and second singular vector corresponding to the singular values, after performing singular value decomposition on a specific parameter (161a) of the final layer of the first encoder (161) to preserve important information (or key information) of a plurality of segmented table data (1201, 1202, 1203) for table data (1200).

[0216] The control unit (150) can update the second encoder (162) using the output of a specific module (400) for a specific parameter (161a). The parameter (162a) corresponding to the output from the specific module (400) can be used to update the second encoder (162). The control unit (150) can update the second encoder (162) using the parameter (162a) obtained from the specific module (400).

[0217] Furthermore, as illustrated in FIG. 12a and FIG. 12c, the control unit (150) inputs a plurality of segmented table data (1201, 1202, 1203) to the updated second encoder (162), and obtains an augmented embedding vector (1201b, 1202b, 1203b) corresponding to each of the plurality of segmented table data (1201, 1202, 1203) from the updated second encoder (162).

[0218] The update process of the second encoder (162) described above may be a process of updating the second encoder (162) using the output of a specific module (400) to generate a new embedding vector augmented for each of the multiple embedding vectors (1201a, 1202a, 1203a) obtained from the first encoder (161) (or for each of the multiple embedding vectors (1201a, 1202a, 1203a), from each of the multiple embedding vectors (1201a, 1202a, 1203a), etc.). Alternatively, it may be a process of updating the second encoder (162) using the output of a specific module (400) to generate a positive representation (or positive view, or positive sample, positive variation, etc.) for each of the multiple embedding vectors (1201a, 1202a, 1203a) obtained from the first encoder (161).

[0219] This updated second encoder (162) may be configured to generate augmented embedding vectors (1201b, 1202b, 1203b) corresponding to each of the plurality of segmented table data (1201, 1202, 1203) based on the set parameter (162a). In this case, as the plurality of segmented table data (1201, 1202, 1203) are input to the second encoder (162), augmented multiple embedding vectors (1201b, 1202b, 1203b) corresponding to the plurality of segmented table data (1201, 1202, 1203) may be generated from the second encoder (162).

[0220] For example, the second encoder (162) can generate a first embedding vector (1201b) corresponding to the segmented first table data (1201). The control unit (150) can obtain the augmented first embedding vector (1201b) generated for the segmented first table data (1201) from the second encoder (162).

[0221] As another example, the second encoder (162) can generate a second embedding vector (1202b) corresponding to the segmented second table data (1202). The control unit (150) can obtain the second embedding vector (1202b) generated for the segmented second table data (1202) from the second encoder (162).

[0222] As another example, the second encoder (162) can generate a third embedding vector (1203b) corresponding to the segmented third table data (1203). The control unit (150) can obtain the third embedding vector (1203b) generated for the segmented third table data (1203) from the second encoder (162).

[0223] Meanwhile, the control unit (150) can train a target model using a plurality of embedding vectors (1201a, 1202a, 1203a) obtained from the first encoder (161) and an augmented plurality of embedding vectors (1201b, 1202b, 1203b) obtained from the second encoder (162).

[0224] As illustrated in FIGS. 12a and 12c, the control unit (150) can combine embedding vectors corresponding to each of the plurality of segmented table data (1201, 1202, 1203) to generate a combined embedding vector.

[0225] The control unit (150) can combine a plurality of embedding vectors (1201a, 1202a, 1203a) corresponding to each of a plurality of segmented table data (1201, 1202, 1203) obtained from the first encoder (161) using a combining module (600). Additionally, the control unit (150) can combine a plurality of augmented embedding vectors (1201b, 1202b, 1203b) corresponding to each of a plurality of segmented table data (1201, 1202, 1203) obtained from the second encoder (162) using a combining module (600).

[0226] For example, the combining module (600) can combine a plurality of embedding vectors (a first embedding vector (1201a), a second embedding vector (1202a), and a third embedding vector (1203a), etc.) corresponding to a plurality of segmented table data (1201, 1202, 1203) obtained from the first encoder (161) to generate a combined first embedding vector (1211) corresponding to the table data (1200).

[0227] As another example, the combining module (600) can combine a plurality of augmented embedding vectors (a first embedding vector (1201b), a second embedding vector (1202b), and a third embedding vector (1203b)) corresponding to a plurality of segmented table data (1201, 1202, 1203) obtained from a second encoder (162) to generate a combined second embedding vector (1212) corresponding to the table data (1200).

[0228] Meanwhile, the control unit (150) can train the target model using the combined first embedding vector (1211) and the combined second embedding vector (1212). Since the method of training the target model is the same as described above, it will be explained only briefly to avoid duplication of explanation.

[0229] First, as illustrated in FIGS. 12a and 12c, the control unit (150) can input the combined first embedding vector (1211) and the combined second embedding vector (1212) to the projection head (163), respectively. At this time, the control unit (150) can form the combined first embedding vector (1211) and the combined second embedding vector (1212) into a positive pair (or pair, positive pair), and train a target model using the positive pair that includes the combined first embedding vector (1211) and the combined second embedding vector (1212) (or the combined first embedding vector (1211) and the combined second embedding vector (1212) form a pair).

[0230] The control unit (150) can obtain, respectively, a first projection vector (1211a) corresponding to a combined first embedding vector (1211) and a second projection vector (1212a) corresponding to a combined second embedding vector (1212) from the projection head (163). Then, the control unit (150) can define a first loss function for contrast learning using the first projection vector (1211a) and the second projection vector (1212a).

[0231] The first loss function may be defined to learn the similarity between the combined first embedding vector (1211) and the combined second embedding vector (1212). The control unit (150) may use the first loss function to train at least one of the first encoder (161) and the projection head (163). For example, the control unit (150) may use the first loss function to calculate a contrast loss based on the similarity between the first projection vector (1211a) and the second projection vector (1212a), and update at least one parameter of the first encoder (161) and the projection head (163) based on the calculated contrast loss.

[0232] Next, the control unit (150) can input the combined first embedding vector (1211). The decoder (164) can reconstruct the combined first embedding vector (1211) to correspond to the original table data (1200) to generate a reconstructed output value (1221).

[0233] The control unit (150) obtains a reconstructed output value (1221) from the decoder (164) and can define a second loss function for reconstruction learning using the reconstructed output value (1221) and table data (1200).

[0234] The second loss function may be defined to learn the relationship between the table data (1200) and the reconstructed output value (1221). The control unit (150) may use the second loss function to train at least one of the first encoder (161) and the decoder (164). For example, the control unit (150) may use the second loss function to calculate the reconstruction loss between the table data (1200) and the reconstructed output value (1221), and update at least one parameter of the first encoder (161) and the decoder (164) based on the calculated reconstruction loss.

[0235] Furthermore, the control unit (150) can define a final loss function (or total loss function) using a first loss function and a second loss function, and train a target model using the final loss function. In this case, the process of training a target model using the final loss function can also be understood as a process of updating at least one parameter among the first encoder (161), projection head (163), and decoder (164) so ​​that the difference between the first loss value and the second loss value calculated by each calculating the first loss function and the second loss function included in the final loss function is minimized.

[0236] Meanwhile, in the process described above, embedding vectors corresponding to segmented table data are combined through a combination module (600), and the combined embedding vector is input to the projection head (163) and decoder (164), respectively; however, the present invention is not necessarily limited thereto.

[0237] In one embodiment, as illustrated in FIG. 12b, the control unit (150) may input embedding vectors (1201a, 1202a, 1203a) corresponding to a plurality of segmented table data (1201, 1202, 1203) obtained from the first encoder (161) to the projection head (163), respectively. In this case, the first embedding vector (1201a) corresponding to the segmented first table data (1201) is input to the projection head (163), and the first projection vector (1211a) corresponding to the first embedding vector (1201a) can be obtained from the projection head (163). Additionally, a second embedding vector (1202a) corresponding to the segmented second table data (1202) is input to the projection head (163), and a second projection vector (not shown) corresponding to the second embedding vector (1202a) can be obtained from the projection head (163). Additionally, a third embedding vector (1203a) corresponding to the segmented third table data (1203) is input to the projection head (163), and a third projection vector (not shown) corresponding to the third embedding vector (1203a) can be obtained from the projection head (163).

[0238] In another embodiment, the control unit (150) may input augmented embedding vectors (1201b, 1202b, 1203b) corresponding to a plurality of segmented table data (1201, 1202, 1203) obtained from the updated second encoder (161) to the projection head (163), respectively. In this case, an augmented first embedding vector (1201b) corresponding to the segmented first table data (1201) is input to the projection head (163), and a first projection vector (1212a) corresponding to the first embedding vector (1201b) can be obtained from the projection head (163). Additionally, an augmented second embedding vector (1202b) corresponding to the segmented second table data (1202) is input to the projection head (163), and a second projection vector (not shown) corresponding to the second embedding vector (1202b) can be obtained from the projection head (163). Additionally, an augmented third embedding vector (1203b) corresponding to the segmented third table data (1203) is input to the projection head (163), and a third projection vector (not shown) corresponding to the third embedding vector (1203b) can be obtained from the projection head (163).

[0239] In this case, the control unit (150) may define a first loss function using a first projection vector (1211a) for a first embedding vector (1201a) corresponding to segmented first table data (1201) and a first projection vector (1212a) for an augmented first embedding vector (1201b) corresponding to segmented first table data (1201) during a contrast learning process, and may train at least one of a first encoder (161) and a projection head (163) using the first loss function. Additionally, the control unit (150) may define a first loss function using a second projection vector for a second embedding vector (1202a) corresponding to segmented second table data (1202) and a second projection vector for an augmented second embedding vector (1202b) corresponding to segmented second table data (1202) during a contrast learning process, and the At least one of the first encoder (161) and the projection head (163) can be trained using the first loss function.

[0240] Furthermore, in the contrast learning process, the control unit (150) can define a first loss function using a third projection vector for a third embedding vector (1203a) corresponding to segmented third table data (1203) and a third projection vector for an augmented third embedding vector (1203b) corresponding to segmented third table data (1203), and learn at least one of a first encoder (161) and a projection head (163) using the first loss function.

[0241] In another embodiment, the control unit (150) may input a plurality of embedding vectors (a first embedding vector (1201a), a second embedding vector (1202a), and a third embedding vector (1203a), etc.) corresponding to a plurality of segmented table data (1201, 1202, 1203) obtained from a first encoder (161) to a decoder (164). The decoder (164) may reconstruct the plurality of embedding vectors (1201a, 1202a, 1203a) to correspond to the original table data (1200) and generate a reconstructed output value (1221).

[0242] The control unit (150) obtains a reconstructed output value (1221) from the decoder (164), defines a second loss function for reconstruction learning using the reconstructed output value (1221) and table data (1200), and then trains at least one of the encoder (161) and the decoder (164) using the second loss function.

[0243] In this way, according to a method for performing binning on table data according to an embodiment of the present invention and training a target model based thereon, by performing binning on the table data, an appropriate inductive bias can be applied to the table data so that an artificial intelligence model can effectively learn an irregular function. In the present invention, by training a binning module to output binning data by performing binning on table data, the binning module can be enabled to effectively extract features including irregularities of the table data. That is, in the present invention, by combining a binning module trained to effectively extract features of a table data set with various types of downstream task models, an artificial intelligence model capable of effectively performing various types of tasks based on a table data set can be provided.

[0244] As described above, according to the data augmentation method and system for learning tabular data of the present invention, by combining a pre-set normalization technique and a pre-set decomposition method, various augmented representations can be generated while preserving important features of the tabular data in the embedding space. Through this, the present invention can improve the quality of the augmented representation by maintaining important information and promote robust feature learning of the model through variability.

[0245] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, an encoder is updated based on encoder parameters extracted according to preset extraction criteria, and an augmented representation corresponding to the table data can be generated using the updated encoder. Through this, the present invention can secure diversity of augmented representations while maintaining the core information of the table data. That is, the present invention can generate new augmented representations while maintaining the structure of the table data by using the updated encoder.

[0246] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, the diversity of the representation space can be increased by applying a pre-set normalization technique to the encoder. That is, the present invention can induce the generation of various variations by randomly removing some of the parameters of the encoder to which the pre-set normalization technique is applied. Through this, the present invention enables the model to learn generalized representations and prevents overfitting to specific patterns. In other words, by applying a pre-set normalization technique to the encoder, the present invention enables the model to learn more robust representations and to experience various data variations.

[0247] Thus, according to the data augmentation method and system for learning tabular data of the present invention, by performing transformations within the learned embedding space itself using a pre-set normalization technique and a pre-set decomposition method, the effect of augmentation can be maximized while maintaining the intrinsic structure of the tabular data. Through this, more robust representation learning is possible even in tabular data, and model learning can be performed stably. In other words, the present invention can improve the learning performance of a model by minimizing information loss while preserving important information.

[0248] Meanwhile, the present invention described above can be implemented based on a quantum computer. The present invention implemented based on a quantum computer may include a qubit-based quantum processor and quantum memory, and may include software and hardware interfaces optimized for quantum computation.

[0249] Quantum processors in quantum computers utilize qubits to efficiently process complex operations through parallel computation, quantum entanglement, and quantum superposition, which cannot be performed by the binary bits of classical computers. Quantum processors process data using quantum gates and can provide exponential speed improvements for specific problems.

[0250] Meanwhile, the present invention described above can be implemented as a program that is executed by one or more processes on a computer and can be stored on a computer-readable medium (or recording medium).

[0251] Furthermore, the present invention described above can be implemented as computer-readable code or instructions on a medium on which a program is recorded. That is, the present invention can be provided in the form of a program.

[0252] Meanwhile, computer-readable media include all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0253] Furthermore, the computer-readable medium may be a server or cloud storage that includes a storage and is accessible to an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage via wired or wireless communication.

[0254] A computer program may reach the system (100) through various suitable transmission mechanisms. The transmission mechanism may be, for example, a computer-readable storage medium, a computer program product, a memory device, a recording medium such as a CD-ROM or DVD, or a product that tangibly embodies the computer program. The transmission mechanism may be a signal configured to reliably transmit the computer program through air or an electrical connection. The system (100) may propagate or transmit the computer program as a computer data signal.

[0255] Furthermore, references to 'computer-readable storage media,' 'computer program products,' 'computer programs embodied in a tangible form,' etc., or to 'controller,' 'computer,' 'processor,' etc., should be understood to include not only computers with various architectures such as single / multi-processor architectures and sequential (Von Neumann) / parallel architectures, but also specialized circuits such as Field-Programmable Gate Arrays (FPGAs), Application Specific Circuits (ASICs), signal processing units, and other devices. References to computer programs, instructions, code, etc., should be understood to include software for programmable processors or firmware, such as programmable content for hardware devices, whether it is instructions for a processor or configuration settings for a fixed-function device, gate array, or programmable logic device.

[0256] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, namely a CPU (Central Processing Unit), and no special limitations are placed on its type.

[0257] Meanwhile, the above detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

1. In a computerized method comprising the following, A step of inputting table data into at least one encoder; A step of obtaining a first embedding vector corresponding to the table data as the output of the at least one encoder; A step of extracting specific parameters from the at least one encoder from which the first embedding vector is output, based on a preset extraction criterion; A step of updating at least one other encoder based on the specific parameter extracted according to the above-mentioned preset extraction criteria; A step of inputting the above table data into the updated at least one other encoder; A step of obtaining an augmented second embedding vector corresponding to the table data from at least one other encoder; and A data augmentation method for learning tabular data, characterized by including the step of training a target model using the first embedding vector and the second embedding vector.

2. In Paragraph 1, The step of inputting the specific parameter extracted from the at least one encoder into a pre-set specific module; and A data augmentation method for learning tabular data, characterized by further including the step of performing decomposition for the specific parameter using a decomposition method pre-set in the specific module.

3. In Paragraph 2, The above specific module is, The above specific parameter is decomposed into at least one of a first singular vector matrix including first singular vectors, a second singular vector matrix including second singular vectors, and a diagonal matrix including singular values ​​corresponding to each of the first singular vectors and the second singular vectors, and In the step of updating at least one other encoder mentioned above, A data augmentation method for learning tabular data, characterized by updating at least one other encoder using the output of the specific module for the specific parameter.

4. In Paragraph 3, A step of specifying at least one singular value among the above singular values ​​that satisfies a preset criterion; and The method further includes the step of generating parameters to be used for updating at least one other encoder based on the specified singular value, and The above parameters are, A data augmentation method for learning tabular data, characterized by corresponding to the output of the specific module for the specific parameter and being used for updating at least one other encoder.

5. In Paragraph 4, The above specific module is, Among the above singular values, at least one of the singular value specified according to the above-previously set standard, and the first singular vector and the second singular vector corresponding to the specified singular value is specified, and A data augmentation method for learning tabular data characterized by generating the parameter based on at least one of the specified singular value, the specified first singular vector, and the specified second singular vector.

6. In Paragraph 4, In the step of updating at least one other encoder mentioned above, A data augmentation method for learning tabular data, characterized by updating at least one other encoder using the parameter output from the specific module.

7. In Paragraph 6, The step of updating at least one other encoder mentioned above is, The step of inputting the above parameter into the at least one other encoder and setting the parameter of the at least one other encoder as the parameter output from the specific module, and The above-mentioned updated at least one other encoder is, A data augmentation method for learning tabular data, characterized by generating an augmented second embedding vector corresponding to the table data based on the parameters set above.

8. In Paragraph 2, The step of updating at least one other encoder mentioned above is, A data augmentation method for learning tabular data, characterized by the step of updating at least one other encoder using the output of the specific module to generate the augmented second embedding vector for the first embedding vector obtained from at least one encoder.

9. In Paragraph 1, A data augmentation method for learning tabular data, characterized by further including the step of applying a normalization technique pre-set to at least one of the at least one encoder and at least one other encoder.

10. In Paragraph 9, When the pre-set normalization technique is applied to at least one of the above at least one encoder and the above at least one other encoder, A data augmentation method for learning tabular data, characterized in that probabilistic variability is applied to at least one of the first embedding vector corresponding to the table data obtained from at least one encoder and the second embedding vector corresponding to the table data obtained from at least one other encoder.

11. In Paragraph 1, A data augmentation method for learning tabular data, characterized by further including the step of defining a first loss function and a second loss function, respectively, using at least one of the first embedding vector and the second embedding vector, for the learning of the above-mentioned learning target model.

12. In Paragraph 11, A step of inputting the first embedding vector and the second embedding vector, respectively, to a projection head; and The method further includes the step of obtaining a first projection vector corresponding to the first embedding vector and a second projection vector corresponding to the second embedding vector from the projection head. The above first loss function is, A data augmentation method for learning tabular data, characterized by being defined using the first projection vector corresponding to the first embedding vector and the second projection vector corresponding to the second embedding vector.

13. In Paragraph 12, The above first loss function is, A data augmentation method for learning tabular data, characterized by being defined to learn in a direction that maximizes the similarity between the first embedding vector and the second embedding vector.

14. In Paragraph 11, The step of inputting the first embedding vector to at least one decoder; and The above at least one decoder further includes the step of reconstructing the first embedding vector to correspond to the table data and obtaining a reconstructed output value. The above second loss function is, A data augmentation method for learning tabular data, characterized by being defined using the above table data and the above reconstructed output value.

15. In Paragraph 14, The above second loss function is, A data augmentation method for learning tabular data, characterized by being defined to learn the relationship between the above table data and the above reconstructed output value.

16. In Paragraph 11, The method further includes the step of defining a final loss function using the first loss function and the second loss function. A data augmentation method for learning tabular data characterized by training the target artificial intelligence model using the above-mentioned final loss function.

17. In Paragraph 1, A step of performing binning on the above table data; A step of obtaining a plurality of segmented table data as a result of performing the segmentation on the table data; and A data augmentation method for learning tabular data, characterized by further including the step of training a target model using the plurality of segmented table data.

18. In Paragraph 17, Based on the acquisition of the above-mentioned plurality of segmented table data, the above-mentioned plurality of segmented table data are input into at least one encoder, and As the output of the at least one encoder, a plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained, and Based on the above-mentioned preset extraction criteria, specific parameters are extracted from the at least one encoder from which the plurality of embedding vectors are output, and Based on the specific parameter extracted according to the above-mentioned preset extraction criteria, at least one other encoder is updated, and Input the above-mentioned plurality of segmented table data into the updated at least one other encoder, and From the above at least one other encoder, an augmented plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained, and A data augmentation method for learning tabular data, characterized by training a target model using the plurality of embedding vectors obtained from at least one encoder and the augmented plurality of embedding vectors obtained from at least one other encoder.

19. A system comprising memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, The above system is, Input table data into at least one encoder, and As the output of the above at least one encoder, a first embedding vector corresponding to the table data is obtained, and Based on a preset extraction criterion, specific parameters are extracted from the at least one encoder from which the first embedding vector is output, and Based on the specific parameter extracted according to the above-mentioned preset extraction criteria, at least one other encoder is updated, and Input the above table data into the updated at least one other encoder, and From the above at least one other encoder, an augmented second embedding vector corresponding to the table data is obtained, and A data augmentation system for learning tabular data characterized by training a target model using the first embedding vector and the second embedding vector.

20. A program that is executed by one or more processes in an electronic device and stored on a computer-readable recording medium, The above program is, A step of inputting table data into at least one encoder; A step of obtaining a first embedding vector corresponding to the table data as the output of the at least one encoder; A step of extracting specific parameters from the at least one encoder from which the first embedding vector is output, based on a preset extraction criterion; A step of updating at least one other encoder based on the specific parameter extracted according to the above-mentioned preset extraction criteria; A step of inputting the above table data into the updated at least one other encoder; A step of obtaining an augmented second embedding vector corresponding to the table data from at least one other encoder; and A program stored on a computer-readable recording medium characterized by including instructions for performing a step of training a target model using the first embedding vector and the second embedding vector.

Images