Method and system for data augmentation for learning tabular data

Abstract

The present invention relates to a method and system for data augmentation for learning tabular data. More specifically, the present invention relates to a method and system for data augmentation based on a self-attention mechanism for contrastive learning of 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 self-attention mechanism-based data augmentation method and system for contrastive learning of 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 generally relies on Contrastive Learning, a method that exposes the model to various variations of the input data through data augmentation.

[0004] Contrastive learning is emerging as a powerful self-supervised learning framework and is achieving success in various fields. One of the key elements of contrastive learning is the generation of positive samples through data augmentation, which aims to introduce changes while preserving the intrinsic characteristics of the original data. However, despite these possibilities, applying contrastive learning to tabular data presents a relatively under-explored and challenging task.

[0005] Specifically, tabular data contains hundreds of features and complex interactions, and failure to process them effectively can lead to errors in critical decision-making processes. In particular, due to the nature of tabular data—which lacks a spatial or sequential structure unlike images or text—applying conventional data augmentation techniques without modification poses a risk of distorting important relationships or compromising their meaning. For instance, conventional augmentation methods based on randomness may overlook interactions between important features in tabular data, potentially leading to a decline in model performance.

[0006] Therefore, considering the structured and heterogeneous nature of tabular data, a new augmentation technique is required to effectively apply contrastive learning that can produce meaningful results while preserving the core structure of the 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 aims to provide a data augmentation method and system for learning tabular data that enables an artificial intelligence model to learn various data through augmentation of tabular data.

[0009] In particular, the present invention is intended to provide a data augmentation method and system for learning tabular data that can generate various augmented representations while maintaining important structures and features of the tabular data.

[0010] Furthermore, the present invention aims to provide a data augmentation method and system for learning tabular data that can achieve high learning performance while maintaining the structural characteristics of the tabular data.

[0011] More specifically, the present invention aims to provide a data augmentation method and system for learning tabular data that can significantly improve the performance of contrastive learning while preserving important structures of tabular data.

[0012] To solve the problem described above, a data augmentation method for learning computerized tabular data according to the present invention, comprising the following steps, may include: specifying table data configured to include a plurality of different columns and at least one record having a value corresponding to each of the plurality of columns; calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data; augmenting the table data based on the calculated importance to generate augmented table data; inputting the table data and the augmented table data into an encoder, respectively; obtaining a first embedding vector corresponding to the table data from the encoder; obtaining a second embedding vector corresponding to the augmented table data from the encoder; and training a target model using the first embedding vector and the second embedding vector.

[0013] In an embodiment, to calculate the importance of each of the plurality of columns included in the table data, the method further includes the step of inputting the table data into a pre-configured specific module, and in the step of generating the augmented table data, the importance of each of the plurality of columns is calculated using attention scores for each of the plurality of columns obtained from the specific module, and the table data is augmented based on the importance of each of the plurality of columns calculated through the attention scores to generate the augmented table data.

[0014] In an embodiment, the specific module may be configured to generate the augmented table data by calculating an attention score for each of the plurality of columns that serves as a criterion for calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data, and selectively performing augmentation on at least some of the plurality of columns based on the importance of each of the plurality of columns calculated through the calculated attention score.

[0015] In the embodiment, the importance can be calculated by averaging the attention scores calculated from the specific module to calculate the importance of each of the plurality of columns.

[0016] In an embodiment, the method further includes the step of selecting at least some specific columns among the plurality of columns to be augmented based on the importance of each of the plurality of columns, and the specific module may be configured to randomly select one of a plurality of preset augmentation techniques and perform augmentation on the specific columns among the plurality of columns using the selected augmentation technique to generate the augmented table data.

[0017] In an embodiment, in the step of selecting the specific column, based on the importance of each of the plurality of columns, first columns satisfying a first criterion and second columns satisfying a second criterion among the plurality of columns are each specified, and at least some of the first columns and the second columns can be selected as the specific column.

[0018] In an embodiment, the specific column includes at least some of the second columns satisfying the second criterion, and the specific module can generate the augmented table data by performing augmentation on at least some of the second columns satisfying the second criterion using the selected augmentation technique among the plurality of augmentation techniques.

[0019] In an embodiment, in the step of generating the augmented table data, at least some of the second columns satisfying the second criterion are selected as the specific columns to be augmented based on a preset selection ratio criterion, and augmentation is performed on the specific columns selected according to the preset selection ratio criterion to generate the augmented table.

[0020] In an embodiment, the step of generating the augmented table data may be a step of generating the augmented table data by performing augmentation on the specific column selected based on the attention score.

[0021] In an embodiment, the step of generating the augmented table data may be a step of generating the augmented table data by performing augmentation on at least some of the second columns having the second importance in order to maintain the structure of the first columns having the first importance.

[0022] In an embodiment, for each training epoch of the model to be trained, any one of the plurality of augmentation techniques may be randomly selected and applied to perform augmentation on the specific column to generate the augmented table data.

[0023] In an embodiment, the specific module may randomly select one of the plurality of augmentation techniques for each training epoch of the target model and perform augmentation on the specific column using the selected augmentation technique to generate the augmented table data.

[0024] In an embodiment, for training the target model, the method may further include the step of defining a loss function using at least one of the first embedding vector corresponding to the table data and the second embedding vector corresponding to the augmented table data.

[0025] 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 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.

[0026] In an embodiment, the loss function is defined to learn in a direction that maximizes the similarity between the first embedding vector corresponding to the table data and the second embedding vector corresponding to the augmented table data, and in the step of training the target model, the target model can be subjected to contrastive learning using the loss function.

[0027] In an embodiment, based on the contrast learning, the method may further include the steps of obtaining a contrast-learned model with the loss function and performing fine-tuning on the contrast-learned model.

[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 association between a plurality of columns included in table data, the importance of each of the plurality of columns is calculated, and based on the calculated importance, the table data is augmented to generate augmented table data; the plurality of segmented table data and the augmented table data are each input into an encoder; a plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained from the encoder; an embedding vector corresponding to the augmented table data is obtained from the encoder; and a target model can be trained using the plurality of embedding vectors corresponding to each of the plurality of segmented table data and the embedding vector corresponding to the augmented table data.

[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, specifies table data configured to include a plurality of different columns and at least one record having a value corresponding to each of the plurality of columns, calculates the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data, augments the table data based on the calculated importance to generate augmented table data, inputs the table data and the augmented table data into an encoder, obtains a first embedding vector corresponding to the table data from the encoder, obtains a second embedding vector corresponding to the augmented table data from the 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 steps such as: specifying table data configured to include a plurality of different columns and at least one record having a value corresponding to each of the plurality of columns; calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data; augmenting the table data based on the calculated importance to generate augmented table data; inputting the table data and the augmented table data into an encoder, respectively; obtaining a first embedding vector corresponding to the table data from the encoder; obtaining a second embedding vector corresponding to the augmented table data from the encoder; and training a target model using the first embedding vector and the second embedding vector.

[0032] As described above, according to the data augmentation method and system for learning tabular data of the present invention, the importance of each of the plurality of columns included in the table data is calculated based on the correlation between the plurality of columns, and the original table data can be augmented based on the calculated importance. Through this, the present invention selectively performs augmentation only on columns with low importance, while maintaining columns with high importance. That is, the present invention can introduce changes in augmentation while preserving the core structure and correlation of the table data. Through this, the quality of the augmented table data is improved, enabling more powerful representation learning in contrastive learning.

[0033] In addition, according to the data augmentation method and system for learning tabular data of the present invention, various augmentation techniques can be randomly selected and applied at each learning epoch of the target model. That is, the present invention enables the transformation of various tabular data through various augmentation techniques without using a fixed augmentation technique, thereby allowing the model to learn various aspects of the tabular data. Through this, the generalization performance of the model is improved, and adaptability to various problems can be increased.

[0034] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, by calculating the importance of each of a plurality of columns using an attention mechanism and performing augmentation on the original table data based on the calculated importance, it is possible to perform augmentation while preserving important structures, features, and interactions within the table data. In other words, the present invention can perform augmentation that sufficiently reflects the characteristics of the table data and considers the specificity of the table data.

[0035] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, meaningful changes can be introduced while maintaining important relationships between the original tabular data and the augmented tabular data. This enables the model to learn various representations, thereby contributing to the learning of more powerful and generalized representations. In particular, the present invention can maximize the performance of contrastive learning by appropriately adjusting the relationship between positive and negative samples.

[0036] FIG. 1 is a conceptual diagram illustrating a data augmentation system for learning tabular data according to the present invention.

[0037] FIG. 2 is a flowchart illustrating a data augmentation method for learning tabular data according to the present invention.

[0038] 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.

[0039] FIGS. 5, FIGS. 6 and FIGS. 7 are formulas related to a data augmentation method for learning tabular data according to the present invention.

[0040] FIGS. 8 and 9 are tables showing an example of the performance of the data augmentation method according to the present invention and the learning results of an artificial intelligence model learned using the learning method according to the present invention.

[0041] FIG. 10 illustrates an example of a block diagram of a computing system in which the present invention can be implemented.

[0042] FIG. 11 illustrates an example of a block diagram of a computing device that may be included in a user computing device, a server computing system, and a training computing system, as an embodiment of a computing system in which the present invention can be implemented.

[0043] FIG. 12 illustrates an example of a block diagram from another perspective of a computing device, which is one of the components of a computing system.

[0044] 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.

[0045] 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.

[0046] 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.

[0047] A singular expression includes a plural expression unless the context clearly indicates otherwise.

[0048] 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.

[0049] 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 various data through the augmentation of tabular data.

[0050] In particular, the present invention aims to maximize the performance of contrastive learning by performing meaningful augmentation while preserving the unique characteristics of tabular data (or tables, tabular data, tabular data, tabular data, etc.). The present invention may utilize a Transformer Architecture to apply augmentation while maintaining important features of tabular data that include both categorical and numerical features.

[0051] Here, the Transformer is a deep learning model architecture configured to efficiently learn the relationships between elements (or tokens, features, etc.) within input data (or input sequence), and may include a self-attention mechanism as a core component. Alternatively, the Transformer is a deep learning model architecture configured to process correlations between multiple input data in parallel, and the Transformer can compute interactions between all input elements in parallel without sequential data processing.

[0052] For example, a transformer is configured to effectively extract a representation that reflects important relationships within the entire input space by calculating a weighted sum for each input feature, taking into account its interaction with all other features. In other words, the transformer calculates the correlation (or relationship) between each input element and other elements, captures the interactions between input elements, and generates a representation (e.g., embedding, embedding vector, etc.) based on this.

[0053] Such transformers can be configured as encoder-decoder structures or encoder-only (or single) structures, and can be applied to various domains such as time-series data, natural language, and images. In the process of augmenting (or transforming) tabular data to generate augmented (or transformed) data, the present invention utilizes a transformer-based module (or model, encoder, etc.) to calculate self-attention scores for each feature included in the tabular data, and measures the relative importance of the features based on the calculated self-attention scores. These self-attention scores can be used as a criterion to maintain important features during the tabular data augmentation process and to apply augmentation only to features of low importance. Here, each column of the tabular data may refer to a single feature. In this specification, the term "feature" may be used interchangeably with terms such as "column," "row," "feature," "characteristic," "attribute," "variable," or "predictor variable." In addition, in the present invention, the term "self-attention score" may be used interchangeably with "attention score," "attention score," "importance score," or "importance score."

[0054] Furthermore, a self-attention mechanism may be a computational method that produces a representation reflecting the relative importance of an element by dynamically calculating the association (or relevance) between each element within a single input data (or input sequence) or a multidimensional input vector and all other elements within the same input.

[0055] For example, a self-attention mechanism generates vectors such as Query, Key, and Value from input data, calculates the similarity between each element based on the dot product between the Query and Key, and determines (or calculates) attention weights by performing normalization (or softmax) on them. Subsequently, the attention weights are applied to the Values ​​to derive the final output (or output value). In this case, the final output can be derived through the weighted sum of the attention weights and the Values.

[0056] This self-attention mechanism is extended into a multi-head structure, enabling parallel learning of correlations in different representation spaces and thereby more effectively reflecting interactions between complex features. In this invention, by utilizing the self-attention mechanism of a Transformer to quantify the importance of each feature included in tabular data and performing data augmentation on features of relatively low importance, the performance of contrastive learning can be significantly improved while preserving important structures of the tabular data.

[0057] Accordingly, the present invention aims to provide a data augmentation method and system capable of effectively performing contrastive learning while preserving the structural characteristics of tabular data and the interactions between features.

[0058] Existing data augmentation techniques are primarily designed for non-tabular data such as images or text; consequently, applying them directly to tabular data leads to problems such as distorted correlations between columns or the degradation of important features. These issues result in reduced training stability, limitations in generalization performance, and a decrease in the model's expressiveness.

[0059] Therefore, the present invention aims to provide a technology that simultaneously improves learning efficiency and representation performance while maintaining the consistency of the data structure by quantitatively evaluating the associations between columns of tabular data to calculate importance and selectively augmenting only columns with low importance. Furthermore, another objective of the present invention is to secure high reproducibility and generalization performance in a self-supervised learning environment by utilizing an attention-based importance calculation mechanism to precisely control the selectivity of augmentation and improving the quality of positive samples during the contrastive learning process.

[0060] To achieve the above objective, the present invention may include the following configuration. A data augmentation method for learning tabular data according to the present invention specifies table data comprising a plurality of different columns and at least one record having a value corresponding to each column, calculates the importance of each column by calculating an attention score based on the association between the plurality of columns included in the table data, and can generate augmented table data by selectively augmenting only the columns with low importance based on the calculated importance.

[0061] In this case, a specific module (e.g., a transformer-based encoder) can generate an attention matrix that reflects the interactions between features within the table data and quantify column importance by averaging it.

[0062] Columns of high importance are maintained in their original form, while columns of low importance may be modified through one of a plurality of augmentation techniques (masking, shuffling, cutmix, etc.) selected at random. Subsequently, the original table data and the augmented table data are input into an encoder to obtain a first embedding vector and a second embedding vector, respectively, and contrastive learning is performed by defining a loss function to maximize the similarity between the two vectors. Additionally, the generalization performance of the model is enhanced through various transformations by randomly selecting an augmentation technique for each training epoch, and parallel learning with binning data may be performed if necessary. This method may be implemented in a system or a recording medium, and the system may be configured to perform the same function by including a processor, memory, an encoder, and an augmentation module.

[0063] According to the present invention, since correlations between columns are maintained through attention score-based importance calculation, the structural consistency of tabular data can be preserved. Furthermore, according to the present invention, unnecessary noise is removed due to importance-based selective augmentation, thereby improving learning stability and convergence speed. Since augmentation techniques are randomly combined at each epoch and the model experiences various data transformations, the generalization ability of contrastive learning results can be improved. Furthermore, according to the present invention, the quality of positive samples generated through attention-based importance evaluation is improved, which enhances feature separability in the representation space and ultimately strengthens the model's expressive power.

[0064] Therefore, the present invention solves the problems of structural distortion and reduced efficiency associated with existing random augmentation techniques and can achieve technical effects by realizing a self-supervised learning framework specialized for processing tabular data.

[0065] 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.).

[0066] 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, and FIG. 7 are formulas related to a data augmentation method for learning tabular data according to the present invention. Furthermore, FIG. 8 and FIG. 9 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.

[0067] 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)”, 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).

[0068] 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.

[0069] 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).

[0070] 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.

[0071] 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.

[0072] 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).

[0073] 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).

[0074] 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.

[0075] 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).

[0076] 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.

[0077] 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.

[0078] 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.

[0079] The communication unit (130) can support various communication methods depending on the communication standard of the communicating device.

[0080] 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).

[0081] 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.

[0082] 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.

[0083] 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.

[0084] 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).

[0085] 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.

[0086] 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).

[0087] 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.

[0088] 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 various data through the augmentation of tabular data. Below, we will examine in more detail a method for generating various augmented representations while maintaining the important structure and features of tabular data, and for training a target artificial intelligence model based thereon.

[0089] In the present invention, table data configured to include a plurality of different columns and at least one record having a value corresponding to each of the plurality of columns is specified (S210), and a process of calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data can be performed (S220, see FIG. 2).

[0090] As discussed above, the present invention can utilize a Transformer Architecture to apply augmentation while maintaining important features of table data that include both categorical and numerical features.

[0091] Here, attention scores can be an indicator that quantitatively represents the interrelationships (or interactions) and / or importance between each input feature during the self-attention process of a Transformer-based module (or model, encoder, etc.). An attention score is a numerical value representing how much the model pays attention to each input feature, indicating how much attention a specific input feature receives out of the total number of input features. These attention scores are calculated through the self-attention mechanism of the Transformer structure and can reflect the degree of association between each input feature and other features. In other words, the attention score represents the relative importance that other inputs have in relation to a specific input, which can indicate how high or low the association is between a specific input feature and other features.

[0092] In this regard, let us examine an embodiment of the self-attention mechanism of the Transformer. The core component of the Transformer is the self-attention mechanism, which allows the model to calculate the importance of each feature and reflect the attention score. For each feature, the attention mechanism calculates a weighted sum of all other features, wherein the weights can be dynamically learned based on the input data. Input processing can be performed using different embedding methods depending on the feature type. For example, in the case of categorical features, a learnable embedding table may be used, which can be represented as in FIG. 5(e). Here, Xc represents the categorical input, and Emded may represent the embedding function. Additionally, in the case of numerical features, a simple linear transformation is applied, which can be represented as in FIG. 5(f). Subsequently, these embeddings (e.g., a first embedding vector in which categorical features are embedded (or vectorized) through an embedding table, a second embedding vector in which numerical features are embedded (or vectorized) through a linear transformation, etc.) can be combined (or linked) and the combined (or linked) embedding vector can be processed as input to a self-attention module (see Fig. 5(g)). Then, the self-attention mechanism can operate on the combined embedding (E) as shown in Fig. 5(h). Here, the query matrix, key matrix, and value matrix can each be represented as shown in Fig. 5(i), and can be generated from the input embedding (E). In addition, the dimension of the Key can be represented as shown in (j) of Fig. 5, and the output of the attention mechanism (or output value, attention value, attention output value, etc.) can be represented as shown in (k) of Fig. 5.

[0093] As such, self-attention mechanisms enable models to learn complex relationships between features, which can be effective for various types of data, including tabular data. By combining categorical and numerical embeddings and applying attention operations, models can capture interactions between features. Attention mechanisms help understand the relative importance of different features, and the feedforward layers (or layers, strata, etc.) of the Transformer architecture can further process the features to which attention has been applied.

[0094] Accordingly, the present invention utilizes attention scores derived from a self-attention mechanism during the augmentation process, and selectively performs augmentation on features with low attention scores. According to one embodiment of the present invention related thereto, the attention scores of each of the multiple columns included in table data are calculated, and the importance of each column can be calculated by averaging them. At this time, columns with relatively high attention scores are identified as key information that the model focuses on, while conversely, columns with low attention scores are identified as information of low importance and can become targets for selective data augmentation. More specific details regarding this will be examined below.

[0095] 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 (162), a multi-layer perceptron (MLP, 163), and a pre-configured specific module (400). In this case, at least one of the encoder (161), the projection head (162), the multi-layer perceptron (MLP, 163), and the pre-configured specific module (400) may be a configuration 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), projection head (162), multi-layer perceptron (MLP, 163), and a specific pre-configured module (400) may be configured separately from the model to be learned (in this case, included as part of the configuration of the data augmentation system (100)). The present invention is not limited to any one of these.

[0096] The control unit (150) can specify at least one table data (300) input (or received) from the user terminal (10). Alternatively, the control unit (150) can specify at least one table data (300) among a plurality of table data (or learning data sets) stored in a previously specified storage (e.g., storage unit (140) or memory, etc.). The method by which the table data (300) is specified in the present invention is not necessarily limited to this, and it is obvious that it can be specified through various methods other than the method mentioned above.

[0097] 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.

[0098] That is, the table data (300) may be configured to include at least one (or multiple) record having different columns and a value corresponding to each of the multiple columns. In this case, each record includes a value for (or corresponding to) each of the multiple columns, and in this specification, the term record may be used interchangeably with “row,” “sample,” “tuple,” “record value,” “or case,” “example,” “observation,” “pattern,” etc.

[0099] The control unit (150) may input the table data (300) into a pre-configured specific module (400) to calculate (or specify, determine, calculate, measure, etc.) the importance of each of the multiple columns included in the table data (300). The specific module (400) is a module based on a transformer structure and may be configured to generate augmented table data by calculating an attention score for each of the multiple columns that serves as a criterion for calculating the importance of each of the multiple columns based on the association between the multiple columns included in the table data (300), and selectively performing augmentation on at least some of the multiple columns based on the importance of each of the multiple columns calculated through the calculated attention score. In this specification, the specific module (400) may also be named an “augmentation module,” an “augmentation model,” a “self-attention model,” a “self-attention module,” a “transformer model,” or a “transformer module.”

[0100] A specific module (400) can quantify the interaction between multiple columns included in the input table data (300) using a self-attention mechanism and calculate the relative importance between the multiple columns based on this. In this case, the attention score calculated (or output, etc.) from the specific module (400) can be used as a criterion for determining the augmentation target.

[0101] Specifically, a specific module (400) can embed each of the multiple columns included in the input table data (300) to generate an embedding vector corresponding to each of the multiple columns, and analyze the interaction between the multiple columns through a self-attention operation.

[0102] For example, let us assume that the table data (300) includes a number of columns, including a numeric column and a categorical column. The categorical column can be converted into an embedding vector (or a first embedding vector) that can be learned through an embedding table (see FIG. 5 (e)). Additionally, the numeric column can be converted into an embedding vector (or a second embedding vector) through a linear transformation of a linear layer (see FIG. 5 (f)). Then, the first embedding vector corresponding to the categorical column and the second embedding vector corresponding to the numeric column can be combined (or connected) to form a single input embedding vector (or combined embedding vector, connected embedding vector, etc.) (see FIG. 5 (g)). When a specific module (400) receives the combined embedding vector, it can calculate the interaction between the multiple columns using a self-attention mechanism. In this case, the self-attention mechanism for the combined embedding vector can operate as shown in (h) of FIG. 5. Here, the query matrix, key matrix, and value matrix can each be represented as shown in (i) of FIG. 5, and can be generated from the combined input embedding (E). Additionally, the dimension of the key can be represented as shown in (j) of FIG. 5, and the output of the attention mechanism (or output value, attention value, attention output value, multi-head attention value, etc.) can be represented as shown in (k) of FIG. 5.

[0103] And, as illustrated in FIGS. 4a and 4b, a specific module (400) can calculate an attention score for each of the multiple columns that serves as a criterion for calculating the importance of each of the multiple columns based on the association between the multiple columns (e.g., “Column 1,” “Column 2,” “Column 3,” etc.) included in the table data (300). In the final (or last) layer of the specific module (400), the importance of each of the multiple columns can be quantified to calculate (or output) an attention score for each of the multiple columns.

[0104] In this case, the present invention can quantify the importance of interactions between multiple columns by using attention scores for each of multiple columns calculated (or obtained) from the final layer of a specific module (400). That is, the present invention can calculate the importance of each of multiple columns by using attention scores for each of multiple columns obtained from a specific module (400).

[0105] In one embodiment, an attention matrix (or attention score tensor, attention score, etc.) of dimensions Batch (B) x Head (H) x Feature (F) x Feature (F) can be generated using attention values ​​calculated in the final layer of a specific module (400). Here, for a given batch (B, or batch size), number of heads (H, or number of heads), and number of columns (F, or number of columns, number of features, number of features, etc.), the attention matrix (or attention score tensor, attention score, etc., see FIG. 6 (a)) can be defined as FIG. 6 (b). Here, the batch (or batch size) can be represented as FIG. 6 (c), the number of heads (or number of heads) can be represented as FIG. 6 (d), and the number of columns (or number of features) can be represented as FIG. 6 (e).

[0106] In the present invention, the importance of each of the plurality of columns can be calculated by averaging the attention scores calculated for each of the plurality of columns. In the present invention, the importance of each of the plurality of columns can be calculated by averaging the attention scores for each of the plurality of columns calculated through a specific module (400).

[0107] A specific module (400) can produce (or derive, calculate, etc.) a column importance matrix (e.g., feature importance matrix) through the average per head. The specific module (400) calculates a column importance matrix (or feature importance matrix, average attention matrix, etc.) by averaging the attention scores over all heads, and the column importance matrix can be represented as shown in (f) of FIG. 6.

[0108] Furthermore, a specific module (400) can calculate an attention score representing the importance of each of the multiple columns based on a column importance matrix representing the association between multiple columns. The specific module (400) can calculate an average attention score for each of the multiple columns. From the attention scores, the average score for each of the multiple columns can be calculated by averaging the interactions between all columns. In this case, the overall importance of each of the multiple columns can be calculated as the average attention value that a specific column (e.g., a first feature) receives from other columns (e.g., a second column, a third column, etc.), which can be represented as in (h) of FIG. 6. This FIG. 6 (h) can be used to generate a vector representing the overall importance of each of the multiple columns (i.e., a feature-specific importance vector). That is, the specific module (400) can generate (or obtain) a feature-specific importance vector by calculating the average importance of each of the multiple columns (see FIG. 6 (i)). This refers to the average received attention value per column and can be understood as a vector representing the overall (or general) importance of each of multiple features. Alternatively, it can be understood as a vector (or scalar vector) of scalar values ​​representing (or expressing) the relative importance (or amount of information) of each of multiple features.

[0109] Meanwhile, in the present invention, a process of generating augmented table data by augmenting the table data based on the calculated importance may be carried out (S230, see FIG. 2).

[0110] In the present invention, the table data (300) can be augmented to generate augmented table data based on the importance of each of the plurality of columns calculated through the attention score for each of the plurality of columns included in the table data (300).

[0111] A specific module (400) can specify a specific column to be augmented among a plurality of columns included in the table data (300) based on a pre-set criterion. More specifically, the specific module (400) can specify a specific column to be augmented according to a pre-set criterion so that augmentation is selectively performed (or made) only on the columns that require augmentation among the plurality of columns included in the table data (300).

[0112] For example, the pre-set criteria may include at least one of a first criterion (or a first importance criterion) or a second criterion (or a second importance criterion). A specific module (400) may specify, among a plurality of columns, first columns satisfying the first criterion and second columns satisfying the second criterion, respectively.

[0113] A specific module (400) may select (or specify, determine, etc.) at least some specific columns among the multiple columns that are to be augmented, based on the importance of each of the multiple columns. The first columns satisfying the first criterion described above are columns with high attention scores and may include columns that contain important information or have a large amount of information and must be maintained without being modified. On the other hand, the second columns satisfying the second criterion are columns with low attention scores and may include columns with low importance or a small amount of information.

[0114] Alternatively, the importance may include at least one of a different first importance and a second importance calculated through an attention score. For example, the first importance may be understood as having a high importance because the attention score is high, and the second importance may be understood as having a low importance because the attention score is low. In this case, a plurality of columns included in the table data (300) may include at least one of first columns having a first importance (or satisfying a first criterion) and second columns having a second importance (or satisfying a second criterion). A specific module (400, or control unit (150)) may specify the first columns having a first importance and the second columns having a second importance, respectively, based on the attention score for each of the plurality of columns.

[0115] In one embodiment, as illustrated in FIGS. 4a and 4b, a column among a plurality of columns with an attention score of “0.4” or higher may be designated as a first column having a first importance. Alternatively, a column among a plurality of columns with an attention score of less than “0.4” may be designated as a column having a second importance.

[0116] In another embodiment, a specific module (400) can specify first columns having a first importance and second columns having a second importance among a plurality of columns using a feature-specific importance vector representing the overall importance of each of a plurality of columns.

[0117] That is, the column to be augmented in the present invention may include at least some of the second columns satisfying a second criterion of a low attention score. Accordingly, at least some of the specific columns to be augmented in the present invention may include at least some of the second columns satisfying the second criterion.

[0118] Specifically, a specific module (400) may select at least some of the second columns satisfying the second criterion as specific columns to be augmented based on a pre-set selection ratio criterion. At least some of the second columns satisfying the second criterion may be selected as specific columns to be augmented according to the pre-set selection ratio criterion. For example, the pre-set selection ratio criterion may be set to select the ratio of columns to be augmented based on the attention score for each of the multiple columns. That is, the pre-set selection ratio criterion may indicate how much (e.g., what percentage) of the second columns satisfying the second criterion with a low attention score will be augmented. In the present invention, the pre-set selection ratio criterion may be set to “40%”. However, the numerical value of the pre-set selection ratio criterion is not necessarily limited to the case mentioned, and may be changed (or set) in various ways by the administrator of the data augmentation system (100) or the data augmentation system (100). These pre-set selection ratio criteria (or column selection ratio, feature selection ratio, etc.) can be represented as shown in (g) of FIG. 6.

[0119] Additionally, a specific module (400) can generate augmented table data by performing augmentation on a specific column using any one of a plurality of preset augmentation techniques. Any one of the plurality of preset augmentation techniques may be applied to a specific column with a low attention score.

[0120] In the present invention, a plurality of enhancement techniques may include various enhancement techniques. For example, a plurality of enhancement techniques may include at least one of a first enhancement technique (e.g., Masking), a second enhancement technique (e.g., Shuffling), and a third enhancement technique (e.g., CutMix).

[0121] Masking can be a method of replacing the value of a selected specific column with the average of the dataset or a fixed value. For example, masking can remove the influence of a specific column during the model training process by replacing the value of a specific column with a low attention score with a constant (the average value of the column) or a mask token. At the same time, important columns (first columns satisfying the first criterion) can be maintained as they are. Such masking can be represented as shown in FIG. 5 (a).

[0122] Shuffling can be a method of replacing the values ​​of a selected specific column with the values ​​of other samples within a batch. For example, shuffling can resolve correlations between columns and create new combinations by collecting (or acquiring) values ​​of a specific column with a low attention score from other samples within the same batch and mixing them (collecting values ​​randomly regardless of attention scores). In other words, shuffling is a method of mixing by randomly collecting (or acquiring) one of the other sample values ​​within the same specific column; since it mixes information within the distribution of the same column, it prevents the model from overfitting to a specific value. Such shuffling can be represented as shown in Fig. 5(b).

[0123] CutMix is ​​a method that combines two samples to generate a single new sample, and it can generate the values ​​of a selected specific column as a mixture of the two samples. Here, mixing can also be understood as changing the position. For example, CutMix mixes the data between two samples for a specific column with a low attention score, while maintaining important columns with high attention scores (first columns satisfying the first criterion) as they are. In other words, CutMix creates a new combination by mixing specific columns with low attention scores from different samples within a batch, while maintaining important columns and inducing the model to generalize more to less important columns. This can be effective in obtaining meaningful diversity without compromising the relationships between key columns. Such CutMix can be represented as shown in Fig. 5(c). At this time, the randomly generated binary mask can be represented as shown in Fig. 5(d).

[0124] A specific module (400) can augment table data (300) by randomly selecting one of a plurality of augmentation techniques. More specifically, as illustrated in FIGS. 4a and 4b, the specific module (400) can generate augmented table data (310) by randomly selecting one of a plurality of augmentation techniques and performing augmentation on a specific column using the selected augmentation technique. For example, the specific module (400) can generate augmented table data (310) by performing augmentation on (or on) a specific column selected according to a preset ratio criterion (i.e., at least some of the second columns satisfying the second criterion) using any one of the randomly selected augmentation techniques.

[0125] Furthermore, the table data (300) and the augmented table data (310) may be configured as positive pairs for contrast learning. The control unit (150) may configure a positive pair in which the table data (300) and the augmented table data (310) form a pair (or the table data (300) and the augmented table data (310) form a pair).

[0126] As discussed above, the process of augmenting table data in the present invention may include a process of generating augmented table data by performing augmentation on specific columns selected based on attention scores for each of a plurality of columns. Through this, the model can learn interactions between meaningful features while maintaining the core structure of the table data.

[0127] In this case, the present invention aims to focus on modifying columns determined to have low importance according to an attention mechanism during the process of augmenting table data, while simultaneously preserving the important features or structure of columns determined to have high importance. Accordingly, the process of augmenting table data can also be understood as a process of generating augmented table data by performing augmentation on at least some of the second columns satisfying a second criterion in order to maintain the structure of the first columns having a first importance.

[0128] Meanwhile, in the present invention, an augmentation technique can be randomly selected and applied for each training epoch (or each batch, each iteration, etc.). Instead of applying a fixed augmentation technique for all training epochs, in the present invention, any one of a plurality of augmentation techniques can be randomly selected and applied for each training epoch of the model to be trained. Here, an epoch may refer to the number of times the entire training data set is completely trained once in model training. That is, 1 epoch may refer to the process of sequentially training the model with all data included in the training data set one by one. For example, if there are 100 table data items and the epoch is set to 10, it can be understood that the entire table data is trained 10 times.

[0129] In one embodiment, a set of augmentation functions (e.g., A = {Masking, Shuffling, CutMix}) is defined, and the augmentation functions can be represented as (j) in FIG. 6. That is, the augmentation technique can be randomly selected from the set (A) each time.

[0130] In this regard, for each training epoch of the model to be trained, one of a plurality of augmentation techniques may be randomly selected and applied to generate augmented table data by performing augmentation on a specific column.

[0131] In this case, a specific module (400) may randomly select one of a plurality of augmentation techniques for each training epoch of the model to be trained, and perform augmentation on a specific column using the selected augmentation technique to generate augmented table data.

[0132] As such, the present invention enables more diverse data transformations compared to conventional methods using fixed augmentation techniques by randomly selecting and applying various augmentation techniques at every epoch, and supports the model in learning various aspects of the data. Through this, the model's generalization performance is improved, and its adaptability to various problems can be enhanced. Furthermore, by exposing the model to various augmented data, robustness is increased while the core structure of the table data is preserved. In other words, the present invention maintains a balance between introducing diversity and preserving statistical characteristics by focusing on columns with low information content and applying augmentation techniques randomly. This augmentation process ensures that the augmentation is meaningful while preserving the structure of the table data, and allows the relationships between important columns in the table data to be maintained without being compromised.

[0133] Meanwhile, in the present invention, the table data and augmented table data are each input to an encoder (S240), a first embedding vector corresponding to the table data is obtained from the encoder (S250), and a second embedding vector corresponding to the augmented table data is obtained from the encoder (S260).

[0134] As illustrated in FIGS. 3a and 3b, the control unit (150) may input table data (300) and augmented table data (310) obtained from a specific module (400) into the encoder (161), respectively. Alternatively, the control unit (150) may input table data (300) and augmented table data (310) included in a positive pair into the encoder (161), respectively. In the present invention, the table data (300) may also be named "first table data" or "original table data," and the augmented table data (310) may also be named "second table data."

[0135] When the encoder (161) receives table data (300) as input, it can be trained to generate a first embedding vector (311) corresponding to the table data (300). The control unit (150) can obtain the first embedding vector (311) corresponding to the table data (300) from the encoder (161) to which the table data (300) is input. For example, the encoder (161) can generate the first embedding vector (311) corresponding to the table data (300) by performing encoding on the table data (300) through parameters of multiple (or multiple) layers.

[0136] In one embodiment, 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. included in each cell, type information of the cells (e.g., characteristics regarding 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.

[0137] Additionally, the encoder (161) can be trained to generate a second embedding vector (312) corresponding to the augmented table data (310) when it receives the augmented table data (310) as input. The control unit (150) can obtain the second embedding vector (312) corresponding to the augmented table data (310) from the encoder (161) to which the augmented table data (300) is input. For example, the encoder (161) can generate the second embedding vector (312) corresponding to the augmented table data (310) by performing encoding on the augmented table data (310) through parameters of multiple (or multiple) layers.

[0138] In one embodiment, the second embedding vector (312) comprises: i) augmented structural information (modified structural patterns regarding the configuration, arrangement, or relationships between some cells of a row or column partially modified through an augmentation technique, etc.); ii) augmented cell content information (values ​​changed due to masking, shuffling between samples, or cutmix applied to cells with low attention scores; characteristics where the data type is maintained but the semantics are partially modified (e.g., mean substitution, application of other sample values, etc.); iii) modified context information (local contextual relationships affected by changes in values ​​surrounding the cell and partial adjustment of location-based information for semantic interpretation of the cell, etc.); iv) reconstructed interaction information (new correlations or combinations between features derived from augmentation on columns or rows with low attention scores. This may include semantic connections in a manner different from the original data); v) structural / semantic representations containing intentional diversity (semantics introduced through augmentation along with core structural patterns maintained through the attention mechanism). It may include at least one of the following: a high-dimensional embedding composed of variations that are present, i.e., representational diversity is added to improve the generalization performance of the model while preserving core information. That is, the second embedding vector (312) can be understood as a high-dimensional representation (or vector, vector value, etc.) generated by reflecting selective variations that occurred in parts with low attention scores, while maintaining as much as possible the structural information and semantic relationships contained in the original table data (300). This is used to construct a positive pair in contrastive learning, allowing the model to learn a balance between information preservation and representational diversity.In the present invention, the second embedding vector (312) may also be named “augmented second embedding vector” or “second latent vector” or “second latent representation vector” or “second feature vector” or “second hidden vector” or “second vector” or “second representation vector” or “augmented representation” or “augmented vector” or “second representation”, etc.

[0139] 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).

[0140] The control unit (150) can define a loss function using at least one of a first embedding vector (311) corresponding to table data (300) and a second embedding vector (312) corresponding to augmented table data (310) for training a model to be trained.

[0141] As illustrated in FIGS. 3a and 3b, the control unit (150) can input a first embedding vector (311) and a second embedding vector (312) to a projection head (162), respectively. At this time, the control unit (150) can form the first embedding vector (311) and the second embedding vector (312) into a positive pair (or pair, positive pair) and train a target model using the positive pair containing the first embedding vector (311) and the second embedding vector (312) (or the first embedding vector (311) and the second embedding vector (312) forming a pair). The first embedding vector (311) and the second embedding vector (312) generated for different inputs may belong to the same label (e.g., positive label) but may represent different views. Here, different views mean that the way data is represented is different. For example, the first embedding vector (101) may be a representation generated based on the original table data (300), and the second embedding vector (312) may be a representation generated based on augmented table data (310) generated by augmenting the original table data (300). Thus, 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.

[0142] The projection head (162) can convert the first embedding vector (311) and the second embedding vector (312) into a lower-dimensional (or lower dimension, smaller dimension, etc.) space, and as a result of projecting the first embedding vector (311) and the second embedding vector (312) respectively, output (or generate) a first projection vector (311a) corresponding to (or about) the first embedding vector (311) and a second projection vector (312a) corresponding to the second embedding vector (312).

[0143] The control unit (150) can obtain a first projection vector (311a) corresponding to a first embedding vector (311) and a second projection vector (312a) corresponding to a second embedding vector (312) from the projection head (162). Then, the control unit (150) can define a loss function for contrast learning using the first projection vector (311a) and the second projection vector (312a). For example, for a given anchor sample (see FIG. 7(a)), a corresponding positive pair (see FIG. 7(b)), and all negative samples (see FIG. 7(c)), the loss function can be represented as FIG. 7(d). Here, the similarity function can be represented as FIG. 7(e), and the temperature parameter can be represented as FIG. 7(f) and set to 0.7. In the present invention, the loss function may also be named a “contrast loss function” or a “contrast loss function,” etc.

[0144] This loss function can be defined to learn in a way that maximizes the similarity between the first embedding vector (311) corresponding to the table data (300) and the second embedding vector (312) corresponding to the augmented table data (310).

[0145] Furthermore, the control unit (150) can perform contrastive learning on the target model using a loss function. The control unit (150) can perform contrastive learning on the target model using a loss function to maximize (make closer) the similarity between the first embedding vector (311) included in the positive pair and the augmented second embedding vector (312), and minimize (make further apart) the similarity with other embedding vectors not included in the positive pair.

[0146] That is, the control unit (150) can train the model to form a strong embedding space by inducing the model to position positive pairs (table data (300) and augmented table data (310)) closer together in the embedding space and voice samples further apart.

[0147] Meanwhile, the control unit (150) can obtain a model that has been learned through comparison as a loss function based on comparison learning, and perform fine-tuning on the model that has been learned through comparison.

[0148] After contrastive learning (or pre-training) is completed, the model can be fine-tuned for subtasks (such as classification or regression). In this process, predictive performance can be optimized using the original data, and the representations obtained from contrastive learning can be adjusted to suit the actual task. During the fine-tuning process, the model is further trained with labeled data, and the learned embeddings can be adapted to the corresponding task. For example, cross-entropy loss can be used for classification tasks, and mean squared error can be applied for regression tasks.

[0149] In one embodiment, as illustrated in FIG. 3a, when a first embedding vector (311) is generated by passing table data (300) through an encoder (161), the first embedding vector (311) can be input to a multilayer perceptron (163) to which a ReLU activation function is applied. The multilayer perceptron (163) can output a final predicted value (321) from the first embedding vector (311). Then, the control unit (150) can fine-tune the model using the final predicted value (321) and a value (322) corresponding to the ground truth data. This fine-tuning process supports the model in optimizing performance in various tasks.

[0150] Meanwhile, based on the data augmentation method and system for learning tabular data according to the present invention described above, we will examine below 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.

[0151] It can be confirmed that the artificial intelligence model (160) trained using the data augmentation method according to the present invention achieves superior performance compared to existing models in terms of accuracy and mean squared error (MSE). For example, as shown in FIG. 8, the present invention demonstrates a performance record that surpasses conventional augmentation techniques.

[0152] Additionally, as illustrated in FIG. 9, the artificial intelligence model (160) has achieved higher performance than conventional models that have shown strengths in tabular data, and has proven its excellence by demonstrating performance that surpasses recently presented artificial intelligence algorithms.

[0153] Meanwhile, the present invention can be applied to various industries and services and utilized effectively. For example, the data augmentation system (100) according to the present invention can be utilized very effectively in manufacturing and process management, such as in appearance inspection and defective product detection. In particular, by introducing meaningful data transformations while maintaining important information in tabular data containing various sensor data of a manufacturing line, the accuracy of process optimization and defect detection can be greatly improved.

[0154] In this case, the data augmentation system (100) can apply self-attention, which is a value indicating how important each variable of the manufacturing process (e.g., temperature, pressure, vibration, etc.) is. The data augmentation system (100) can use a specific module (400) to calculate the importance of each variable as a self-attention score, maintain important variables without modification, and apply the augmentation technique only to variables of low importance.

[0155] For example, in the case of masking, when detecting defects, unnecessary noise in prediction can be removed by ignoring low-importance features or replacing them with constants to induce the model to learn only the key variables. As another example, in the case of shuffling, randomly mixing the values ​​of low-importance variables within a batch prevents the model from overfitting to specific variables and allows it to learn new data combinations. As yet another example, in the case of cutmixing, mixing two samples allows the model to learn various defect types by maintaining important variables while applying modifications only to less important ones.

[0156] In this regard, the data augmentation system (100) can calculate an attention score for each variable by inputting sensor data from a manufacturing process into a specific module (400) based on a transformer through data preprocessing and attention score calculation. Then, the data augmentation system (100) can perform augmentation on variables with low attention scores by randomly selecting and applying one of a plurality of preset augmentation techniques. At this time, the augmentation technique applied at each epoch is different so that the model can learn various data transformations. Furthermore, the data augmentation system (100) uses the original sample and the augmented sample as positive samples and learns to make similar samples close and different samples far apart using a loss function for contrast learning. Through this, the model can learn strong embeddings in defect detection. Finally, the model is fine-tuned and optimized using actual data for defect detection. As such, the present invention can be utilized to improve detection performance by learning various defect types while preserving important structures during the appearance inspection or defect detection process in manufacturing.

[0157] In addition, the data augmentation system (100) according to the present invention can be very usefully utilized in the field of new material research and development. In the field of new materials, the present invention can be applied to predict the properties of new materials or to detect outliers. Experimental data of new materials generally consists of various physical / chemical characteristic data, and the goal is to improve the generalization performance of the model by modifying only less important characteristics while preserving important characteristics.

[0158] In this case, the data augmentation system (100) inputs physical / chemical characteristic data of each material into a specific module (400) based on a transformer during the new material data preparation and preprocessing process to calculate an attention score, and during the attention score-based augmentation application process, it can randomly select and apply augmentation techniques such as Masking, Shuffling, and CutMix to characteristics of low importance.

[0159] In addition, the data augmentation system (100) can induce the model to learn the relationship between important features by using the original sample and the augmented sample as positive samples during the contrastive learning process. Furthermore, model fine-tuning can fine-tune the model with actual data to optimize the new material prediction performance.

[0160] As such, the present invention can enhance the performance of new material property prediction and anomaly detection through various data modifications while maintaining important characteristics in new material development. Through this, the present invention can contribute to automating the property evaluation of new materials and efficiently selecting high-performance materials.

[0161] 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.

[0162] More specifically, the control unit (150) performs intervals to assign 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 interval data obtained as a result of performing intervals on the table data set. At this time, since the details regarding attention score calculation, calculation of importance of each of multiple columns, table data augmentation, and model comparison learning have been explained in more detail above, they will be explained only briefly below to avoid duplication of explanation.

[0163] 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.

[0164] 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.

[0165] 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.

[0166] 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.

[0167] 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.

[0168] The control unit (150) can receive (or acquire) table data (300) by receiving input from a user terminal (10) or from an external source. For example, the control unit (150) may receive table data (300) directly from an external source, or may extract table data (300) 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).

[0169] 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.

[0170] 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.

[0171] 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.

[0172] 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.

[0173] 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.

[0174] 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.

[0175] The control unit (150) can perform segmentation on the table data (300) to obtain a plurality of segmented table data (e.g., segmented first table data, segmented second table data, segmented third table data, etc.). In the present invention, the plurality of segmented table data may also be named as “multiple segmented data,” “multiple segmented data,” “multiple segmented table data,” or “multiple segmented data sets,” etc.

[0176] Although not explicitly described, the data augmentation system (100) according to the present invention may further include a segmentation module. Alternatively, the segmentation module may also be composed of a model to be learned. The present invention is not limited to either of these.

[0177] The control unit (150) can perform segmentation on table data (300) using a segmentation module and obtain a plurality of segmented table data (e.g., segmented first table data, segmented second table data, segmented third table data, etc.) output (or generated) from the segmentation module (500).

[0178] Segmentation refers to assigning a single representative value to information within a predetermined numerical range among the various pieces of 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.

[0179] The segmentation module may be a module (or model, encoder, autoencoder, etc.) trained to output a plurality of segmented table data generated by performing segmentation on the table data (300) when table data (300) is input. The segmentation module may be a module trained to effectively extract features including irregularities in the table data.

[0180] Additionally, the segmentation module can perform segmentation on the table data (300) to group data samples with similar values. Through this, the segmentation module can extract data features that are robust against errors that may occur due to the analysis of ungrouped individual data.

[0181] That is, the segmentation module can be configured to effectively extract features of the table data (300) based on segmentation of the table data (300).

[0182] In one embodiment, the segmentation module may assign different bin values ​​to different categorical data for a plurality of samples included in the table data (300). In this case, by assigning multiple bin values ​​to different categorical data included in the table data (300), the categorical data may be converted into numerical data.

[0183] In another embodiment, the table data (300) may include a plurality of column information, which are raw feature values ​​related to features of a plurality of samples, and the segmentation module may convert the raw feature values ​​of the table data into a plurality of segmented table data. 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, 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.

[0184] Furthermore, the number of multiple empty values ​​assigned to multiple numeric data included in the table data (300) 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.

[0185] 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.

[0186] Accordingly, the control unit (150) can obtain from the segmentation module 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.

[0187] 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.

[0188] Meanwhile, the control unit (150) can input segmented table data obtained from the segmentation module and augmented table data (310) generated by performing augmentation on the table data (300) in a specific module (400) into the encoder (161).

[0189] For example, the encoder (161) can be trained to generate an embedding vector corresponding to each of the multiple segmented table data. The control unit (150) can obtain an embedding vector corresponding to each of the multiple segmented table data from the encoder (161). In this case, as the multiple segmented table data is input to the encoder (161), multiple embedding vectors corresponding to the multiple segmented table data can be generated from the encoder (161).

[0190] As another example, the encoder (161) can be trained to generate an embedding vector corresponding to the augmented table data (310). The control unit (150) can obtain the embedding vector corresponding to the augmented table data (310) from the encoder (161).

[0191] Furthermore, although not explicitly described, the data augmentation system (100) according to the present invention may further include a combination module. Alternatively, the combination module may also be composed of a model to be learned. The present invention is not limited to any one of these.

[0192] Meanwhile, the control unit (150) can train a target model using a plurality of embedding vectors corresponding to a plurality of segmented table data and an embedding vector corresponding to augmented table data (310).

[0193] The control unit (150) can combine a plurality of embedding vectors corresponding to a plurality of segmented table data obtained from a first encoder using a combining module, and obtain a combined embedding vector from the combining module. Then, the control unit (150) can use the combined embedding vector for training a target model.

[0194] However, multiple embedding vectors corresponding to multiple segmented table data are not necessarily combined and used for learning; instead, the multiple embedding vectors may each be processed as inputs to the projection head (162) along with the embedding vector corresponding to the augmented table data (310) and used for learning the target model. In this case, multiple embedding vectors corresponding to multiple segmented table data and embedding vectors corresponding to the augmented table data (310) are input to the projection head (162), and the target model can be trained using multiple projection vectors corresponding to the multiple embedding vectors obtained from the projection head and projection vectors corresponding to the embedding vectors of the augmented table data (310).

[0195] Finally, the control unit (150) can train the target model (e.g., contrastive training) using the combined embedding vector and the embedding vector corresponding to the augmented table data (310). Since the method of training the target model is the same as described above, it will be omitted to avoid duplication of explanation.

[0196] 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.

[0197] Furthermore, the data augmentation system for learning tabular data according to the present invention can be implemented through the computing device described below and can perform data processing related to the data augmentation method for learning tabular data described above.

[0198] Meanwhile, FIG. 10 illustrates an example of a block diagram of a computing system in which the present invention can be implemented.

[0199] Referring to FIG. 10, a computing system (10000) that performs a self-attention mechanism-based data augmentation method for contrastive learning of tabular data according to one embodiment of the present invention may include at least one computing device. At this time, the at least one computing device may be a single processor or a multi-processor computing device.

[0200] The components of at least one computing device of the present invention may include various hardware components such as one or more processors, memory, other hardware, and a system bus (not shown) that connects various system components so that they can transmit and receive data to and from each other (e.g., telecommutatively connected, physically connected, electrically connected), and the components of at least one computing device are not limited thereto and may be very diverse.

[0201] Meanwhile, at least one computing device included in a computing system (10000) that performs a data augmentation method based on a self-attention mechanism for comparative learning of tabular data may be connected to communicate via a network (1070). For example, at least one computing device included in the computing system (10000) may be clustered or part of a local area network (LAN). Additionally, at least one computing device may be part of a wide area network (WAN) or connected to at least one of a client-server network and a peer-to-peer network within the cloud.

[0202] Meanwhile, when at least one computing device is used in at least one of a network environment and a cloud computing environment, the at least one computing device may be connected to at least one of a public and private network through a network interface or adapter. In one embodiment, other communication connection devices, such as a modem, may be used to establish communication through the network. The modem may be at least one of an internal modem and an external modem, and may be connected to a system bus through a network interface or a specific mechanism, etc. A wireless network component consisting of an interface and an antenna may be coupled to the network through a device such as an access point, a peer computer, etc. In the present invention, the method of connecting at least one computing device to communicate through the network (1070) is not limited, and it may be connected to communicate in a manner different from the described example.

[0203] Furthermore, other computer-type devices and / or systems not shown in FIG. 10 may also interact technically with at least one computing device or other system through one or more connections to the network (1070) via a network interface. Here, the network interface may include network interface equipment such as a physical network interface controller (NIC) or a virtual network interface (VIF).

[0204] The network (1070) of the present invention may include various forms such as the Internet, 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 Data Association (IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, Wireless USB (Wireless Universal Serial Bus), etc., and in the present invention, data transmission may be performed based on standard communication protocols such as TCP / IP, HTTP, SSL, etc.

[0205] A computing system (10000) for performing a data augmentation method based on a self-attention mechanism for comparative learning of tabular data according to the present invention may include at least one of a user computing device (1010), a training computing system (1050), and a server computing system (1030).

[0206] A user computing device (1010) according to the present invention may be understood as a computing device comprising at least one processor (1011) and a memory (1012) for performing a data augmentation method based on a self-attention mechanism for comparative learning of tabular data. For example, the user computing device (1010) may include at least one computing device among a smartphone, a smart TV, a laptop computer, a desktop computer, a digital broadcasting terminal, a PDA (personal digital assistants), a PMP (portable multimedia player), a navigation device, a slate PC, a tablet PC, an ultrabook, a wearable device (e.g., a smartwatch, a smart glass, and a head-mounted display).

[0207] At least one processor (1011) constituting the user computing device (1010) may include one or more general-purpose processors and / or one or more special-purpose processors. For example, at least one processor (1011) constituting the user computing device (1010) may be composed of at least one of a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a neural network processing unit (NPU), an arithmetic logic unit (ALU), a floating-point arithmetic unit (FPU), an application integrated circuit, an application semiconductor (ASIC), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, and / or electrical units for performing other functions, or a plurality of electrically connected processors.

[0208] Furthermore, at least one processor (1011) may be configured to execute computer-readable instructions contained in memory (1012) and / or other instructions described herein.

[0209] The memory (1012) constituting the user computing system (1010) according to the present invention may include volatile memory, non-volatile memory, fixed media, removable media, magnetic media, optical media, semiconductor media and / or other types of physically durable storage media.

[0210] For example, memory (1012) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), EEPROM, EPROM, flash memory device, magnetic disk, and combinations thereof, and may include web storage of a server performing memory storage functions on the internet. Such memory (1012) may store data and instructions necessary for the at least one processor (1011) to perform the operation of an application that augments data based on a self-attention mechanism for comparative learning of tabular data.

[0211] A user computing device (1010) may include one or more user input components (1021) that detect user input. For example, the user input component (1021) may also be referred to as a user interface module. The user input component (1021) 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 user input component (1021). In this case, the user input component (1021) in the present invention does not necessarily mean a hardware means, but can be understood as a channel for receiving input from a user. Meanwhile, the user of the present invention may refer to an automated agent, script, playback software, etc., that operates on behalf of one or more people.

[0212] A user can interact with a computing system (10000) including at least one computing device through input text, touch, voice, movement, computer vision, gestures and / or other forms of input / output using a user input component (1021). For example, the user input component (1021) may include one or more of a command line interface (CLI), a graphical user interface (GUI), a natural user interface (NUI), a voice command interface and / or other user interface (UI) representations.

[0213] Between the user input component (1021) and the user computing device (1010), one or more application programming interface (API) calls may be made based on user input received from the user interface and / or network.

[0214] Here, the expression "based on" may be interpreted to include cases where it is based on the use of a specific configuration, modified from, derived from, influenced by, dependent on, or otherwise derived from a specific configuration. In some embodiments, an API call may be configured for a specific API, which may be interpreted or converted into an API call configured for another API. Here, an API may refer to a defined interface or connection between computers or between computer programs.

[0215] In one embodiment, the user computing device (1010) may store at least one machine learning model (1020). For example, the user computing device (1010) may be various machine learning models, such as a plurality of neural networks (e.g., deep neural networks) that perform data augmentation based on a self-attention mechanism for contrast learning of tabular data, or other types of machine learning models including non-linear models and / or linear models, and may be composed of a combination thereof.

[0216] According to an embodiment of the present invention, a user computing device (1010) may perform a self-attention mechanism-based data augmentation method for comparative learning of tabular data using a local or / and external machine learning model (1020). Alternatively, the user computing device (1010) may perform a self-attention mechanism-based data augmentation method for comparative learning of tabular data using a machine learning model (1040) provided by a server.

[0217] Additionally, according to another embodiment of the present invention, a server computing system (1030) communicating with a user computing device (1010) may provide original table data and table data augmented with original table data to the user computing device (1010) on an application or / and the web in accordance with a request from a user received through the user computing device (1010).

[0218] In addition, according to another embodiment of the present invention, by linking at least a part of a user computing device (1010) and a server computing system (1030) with each other to perform a data augmentation method based on a self-attention mechanism for comparative learning of tabular data, original table data and table data augmented from the original table data can be provided to the user.

[0219] Additionally, according to various embodiments of the present invention, a user computing device (1010) and / or a server computing system (1030) may learn machine learning models (1020, 1040) performed in a self-attention mechanism-based data augmentation method for comparative learning of tabular data through interaction with a training computing system (1050) that is communicatedly connected via a network (1070). In this case, the training computing system (1050) may be a computing system separate from the server computing system (1030). Alternatively, in some embodiments, the training computing system (1050) may be part of the server computing system (1030) or part of the user computing device (1010).

[0220] Meanwhile, the server computing system (1030) may include at least one processor (1031) and memory (1032). Here, the processor (1031) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a neural network processing unit (NPU), an application integrated circuit, an application semiconductor (ASIC), an arithmetic logic unit (ALU), a floating-point arithmetic unit (FPU), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions. For example, at least one processor (1031) may include a circuit and a transistor configured to execute instructions from memory (1032).

[0221] The memory (1032) constituting the server computing system (1030) according to the present invention may include volatile memory, non-volatile memory, fixed media, removable media, magnetic media, optical media, semiconductor media, and / or other types of physically durable storage media. For example, the memory (1032) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), EEPROM, EPROM, flash memory device, magnetic disk, etc., and combinations thereof, and may include web storage of a server that performs the storage function of memory over the internet. Additionally, the server computing system (1030) may further include a data storage (data store). For example, the data storage may be composed of at least one of a relational database, a NoSQL database, a data warehouse, and a local file system.

[0222] In the memory (1032) constituting the server computing system (1030) according to the present invention, data and instructions necessary for the at least one processor (1031) to perform the operation of an application that augments data based on a self-attention mechanism for comparative learning of tabular data may be stored.

[0223] In one embodiment, the server computing system (1030) may be composed of a single device or a plurality of computing devices, and may be configured to operate according to a sequential or parallel computing architecture. Additionally, a distributed processing system may be configured with a plurality of networked devices.

[0224] Meanwhile, the training computing system (1050) may include at least one processor (1051) and memory (1052). The model trainer (1060) is a logical component that executes the training of at least one machine learning model (1020, 1040) and may be implemented in the form of hardware, firmware, or software. For example, the model trainer (1060) may be executed by the processor (1051) after loading training data (1061) stored in a storage device into memory (1052). For example, the model trainer (1060) may be configured to execute one or more operations (e.g., model training, model reconstruction, model validation, model testing) on ​​at least one machine learning model.

[0225] The machine learning model of the present invention may include at least one of a statistical model, an algorithm, a neural network (NN), a convolutional neural network (CNN), a generative neural network (GNN), a Word2Vec model, a Bag of Words model, a TF-IDF (document frequency-inverse document frequency) model, a GPT (Generative Pre-trained Transformer) model (or other autoregressive models), a PPO (Proximal Policy Optimization) model, a nearest neighbor model (e.g., a k-nearest neighbor model), a linear regression model, a K-means clustering model, a Q-learning model, a TD (Temporal Difference) model, a Deep Adversarial Network model, and all other types of models further described herein.

[0226] Specifically, the model trainer (1060) may execute operations to train a machine learning model, and said operations may include at least one of adding, removing, and modifying model parameters. At this time, the training of the machine learning model may be at least one of supervised learning, semi-supervised learning, and unsupervised learning. In one embodiment, the training of the machine learning model may include the step of repeatedly inputting training data (1061) based on epochs and repeatedly performing the machine learning model training process configured in this way. Here, an epoch may refer to a unit in which the entire set of training data (1061) undergoes forward and backpropagation processing once. In some implementations, different levels of training methods (e.g., supervised learning, semi-supervised learning, unsupervised learning) may be used for different epochs.

[0227] The training data (1061) of the present invention may include input data and / or data previously output from at least one machine learning model (e.g., recursive learning feedback).

[0228] At least one parameter of a machine learning model may include at least one of a seed value, a model node, a model layer, an algorithm, a function, connections between different machine learning models, connections between parameters, machine learning model constraints, and other digital components that influence the output of the machine learning model. In this case, model connections between different machine learning models may include or represent relationships between model parameters and / or models, which may be dependent or interdependent, hierarchical, and / or static or dynamic. The combinations and configurations of model parameters described herein may be too complex to be maintained or utilized by human cognitive abilities.

[0229] In the present invention, the machine learning parameters described according to the embodiments are not limited, and a single machine learning model may further include a plurality of model parameters.

[0230] Meanwhile, FIG. 11 illustrates an example of a block diagram of a computing device (1100) that may be included in a user computing device (1010), a server computing system (1030), and a training computing system (1050), as an embodiment of a computing system (10000) in which the present invention can be implemented.

[0231] As illustrated in FIG. 11, the computing device (1100) may include at least one application (e.g., Application 1 to Application N), and each of the at least one application may include a machine learning library and a model execution environment for performing a self-attention mechanism-based data augmentation method for comparative learning of machine learning-based tabular data. The at least one application included in the computing device (1100) may communicate with the sensor, context manager, device state manager, or additional component(s) within the computing device (1100) via an Application Programming Interface (API). In one embodiment, the at least one application may interface with device components, such as receiving sensor data or state data or transmitting prediction results to an output device via a public or private API.

[0232] Meanwhile, FIG. 12 illustrates an example of a block diagram in another aspect of a computing device (1200), which is one of the components of a computing system (10000) that performs a data augmentation method based on a self-attention mechanism for comparative learning of tabular data according to an embodiment of the present invention.

[0233] A computing device (1200) according to the present invention may include at least one application (e.g., Application 1 to Application N), and at least one application may communicate with a central intelligence layer (1210). Each application may interact with a shared model within the central intelligence layer (1210) through an API (e.g., a common API).

[0234] The central intelligence layer (1210) includes one or more machine learning models and may share them among multiple applications or provide them independently to each. In one embodiment, the central intelligence layer (1210) may be integrated as part of an operating system or implemented as a separate logical layer.

[0235] Additionally, the central intelligence layer (1210) can communicate with the central device data layer (1220). The central device data layer (1220) can store table data stored within the computing device (1200) and embedding vectors corresponding to said table data, and provide them as input data required for data augmentation based on a self-attention mechanism for comparative learning of tabular data. Each device component (e.g., sensor, state manager, etc.) can communicate with the central device data layer (1220) through a private API, etc.

[0236] The technology described herein may be composed of a single or multiple computing devices, and a machine learning model that performs a self-attention mechanism-based data augmentation method for comparative learning of tabular data may be executed sequentially or in parallel on one component or multiple distributed components. The data storage, machine learning model, and application may be distributed and operated locally or over a network, and these configurations can be flexibly applied to various system architectures.

[0237] The present invention relates to a data augmentation method, a system, and a program for performing the same, for improving the efficiency and generalization performance of contrastive learning while maintaining the consistency of the data structure during the learning process of tabular data.

[0238] According to the present invention, table data consisting of a plurality of different columns and one or more records having values ​​corresponding to each column is received as input, and the importance of each column is calculated based on the association between columns within the table data. At this time, a specific Transformer-based module calculates an attention score to reflect the strength of the interaction between columns, thereby quantifying the relative importance of each column within the data.

[0239] Based on the calculated importance, only low-importance columns are selected as targets for augmentation, and one of various augmentation techniques, such as masking, shuffling, or cutmix, is randomly selected and applied to the selected columns. On the other hand, high-importance columns are maintained in their original state, thereby preserving the core relationships of the data structure. In addition, since an augmentation technique is randomly selected and applied at each epoch of the training process, diversity in data transformation is ensured and overfitting of the model is prevented.

[0240] Thus, the augmented table data and the original table data are input into an encoder, respectively, and converted into a first embedding vector and a second embedding vector. Subsequently, a loss function is defined to maximize the representation similarity between the two embedding vectors, and InfoNCE-based contrast learning is performed. In addition, the trained model can be optimized for downstream tasks (defect detection, new material prediction, anomaly detection, etc.) through fine-tuning as needed.

[0241] According to one embodiment, by performing binning on table data in parallel to generate multiple binning table data and inputting these data in parallel along with augmented data, representational diversity and learning stability of the model can be simultaneously secured. This parallel learning structure provides the effect of correcting imbalances in data distribution and finely reflecting data characteristics by segment.

[0242] The configuration of the present invention may also be implemented in the form of a system and a program. The system includes a processor and memory, and the processor sequentially executes the steps described above—specification of data, calculation of importance, performance of augmentation, encoding, generation of embeddings, and performance of learning. Furthermore, a program that performs the same functions may be stored on a computer-readable recording medium, thereby providing consistent learning efficiency in server environments, cloud systems, or embedded devices.

[0243] Accordingly, the present invention provides the following technical effects. First, structural consistency of tabular data is preserved because correlations between columns are maintained through importance calculation based on attention scores. Second, unnecessary data deformation is reduced and the quality of training data is improved through importance-based selective augmentation. Third, the model can respond to various data deformations and improve generalization performance through the application of random augmentation combinations and changes at the epoch level. Fourth, representation consistency between positive samples is ensured through the definition of loss based on contrastive learning, thereby improving feature separability in the embedding space. Fifth, diversity of data distribution is ensured and the stability of learning is strengthened through a segmented parallel learning structure.

[0244] In this way, the present invention solves the problems of structural distortion and learning inefficiency associated with existing random augmentation methods, thereby realizing a self-supervised learning framework optimized for the characteristics of tabular data and achieving technological advancements that improve data representation, learning efficiency, and generalization performance.

[0245] A data augmentation method for learning tabular data according to an embodiment of the present invention may begin with the step of specifying table data configured to include a plurality of different columns and at least one record having a value corresponding to each of the plurality of columns. The table data may be structured data collected from an external database, experimental logs, or sensors, and may be configured through preprocessing steps such as missing value processing, normalization, and categorical encoding.

[0246] The system according to the present invention can calculate the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data. To this end, a specific module may utilize a Transformer-based Self-Attention structure and derive an attention score by calculating the inter-column interaction between each column.

[0247] The above attention scores can be averaged across multiple heads to calculate the relative importance of each column. Since this configuration can reflect non-linear correlations more accurately than simple statistics-based importance calculation methods, importance can be estimated while maintaining the structural characteristics of the table data.

[0248] Based on the calculated importance, the table data can be augmented to generate augmented table data. The augmentation may be performed selectively on only columns with low importance, and for the selected columns (specific columns), one of a plurality of predefined augmentation techniques may be randomly selected and applied. The plurality of augmentation techniques may include masking, shuffling, cutmix, noise injection, etc. The specific columns may be identified as columns that satisfy a second threshold (threshold_2) with a preset importance value, and columns with high importance (columns satisfying the first threshold) may be maintained in their original state.

[0249] This selective augmentation method can induce various data transformations during the learning process while preserving the core structure of the data.

[0250] The above table data and the above augmented table data can each be input into an encoder.

[0251] The above encoder may be implemented as a transformer encoder, a multilayer perceptron (MLP), or a graph-based encoder, etc. The encoder can convert the input data into an embedding vector, thereby obtaining a first embedding vector corresponding to table data and a second embedding vector corresponding to augmented table data.

[0252] Subsequently, a target model can be trained using the first embedding vector and the second embedding vector. The training process can be performed based on contrastive learning, and a loss function can be defined in a direction that maximizes the similarity between the embedding vectors.

[0253] According to one embodiment, the loss function may apply an InfoNCE structure and may be designed to maximize the proximity between positive samples based on the cosine similarity of the first and second embedding vectors.

[0254] Learning according to the present invention can be performed in units of epochs, and in each epoch, one of the plurality of augmentation techniques can be randomly selected and applied. This configuration can improve the generalization performance of the model and prevent overfitting to a specific augmentation pattern by repeatedly training various augmentation combinations on the same dataset.

[0255] According to another embodiment, binning can be performed on the table data. The binning process can generate multiple binning table data by dividing a continuous feature into multiple segments.

[0256] The generated multiple segmented table data can be input into the encoder in parallel along with the augmented table data, thereby ensuring diversity in data distribution and improving learning stability.

[0257] The present invention may also be implemented as a system and a program for executing the method. The system may include one or more processors and memory, and the processor may execute the data augmentation, importance calculation, encoding, and contrast learning steps. Additionally, a program performing the same function may be stored on a computer-readable recording medium and may be executed in a server, cloud, or edge device environment.

[0258] As described above, according to the data augmentation method and system for learning tabular data of the present invention, the importance of each of the plurality of columns included in the table data is calculated based on the correlation between the plurality of columns, and the original table data can be augmented based on the calculated importance. Through this, the present invention selectively performs augmentation only on columns with low importance, while maintaining columns with high importance. That is, the present invention can introduce changes in augmentation while preserving the core structure and correlation of the table data. Through this, the quality of the augmented table data is improved, enabling more powerful representation learning in contrastive learning.

[0259] In addition, according to the data augmentation method and system for learning tabular data of the present invention, various augmentation techniques can be randomly selected and applied at each learning epoch of the target model. That is, the present invention enables the transformation of various tabular data through various augmentation techniques without using a fixed augmentation technique, thereby allowing the model to learn various aspects of the tabular data. Through this, the generalization performance of the model is improved, and adaptability to various problems can be increased.

[0260] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, by calculating the importance of each of a plurality of columns using an attention mechanism and performing augmentation on the original table data based on the calculated importance, it is possible to perform augmentation while preserving important structures, features, and interactions within the table data. In other words, the present invention can perform augmentation that sufficiently reflects the characteristics of the table data and considers the specificity of the table data.

[0261] Furthermore, according to the data augmentation method and system for learning tabular data of the present invention, meaningful changes can be introduced while maintaining important relationships between the original tabular data and the augmented tabular data. This enables the model to learn various representations, thereby contributing to the learning of more powerful and generalized representations. In particular, the present invention can maximize the performance of contrastive learning by appropriately adjusting the relationship between positive and negative samples.

[0262] 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.

[0263] 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.

[0264] 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).

[0265] 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.

[0266] 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.

[0267] 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.

[0268] 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.

[0269] 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.

[0270] 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.

[0271] 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 specifying table data configured to include multiple different columns and at least one record having a value corresponding to each of the multiple columns; A step of calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data; A step of generating augmented table data by augmenting the table data based on the importance calculated above; A step of inputting the table data and the augmented table data into an encoder, respectively; A step of obtaining a first embedding vector corresponding to the table data from the encoder; A step of obtaining a second embedding vector corresponding to the augmented table data from the 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, To calculate the importance of each of the plurality of columns included in the table data, the method further includes the step of inputting the table data into a pre-configured specific module. In the step of generating the augmented table data mentioned above, Using attention scores for each of the plurality of columns obtained from the specific module above, the importance of each of the plurality of columns is calculated, and A data augmentation method for learning tabular data, characterized by augmenting the table data to generate the augmented table data based on the importance of each of the plurality of columns calculated through the attention score.

3. In Paragraph 2, The above specific module is, Based on the association between the plurality of columns included in the table data above, an attention score for each of the plurality of columns is calculated to serve as a criterion for calculating the importance of each of the plurality of columns, and A data augmentation method for learning tabular data, characterized by generating augmented table data by selectively performing augmentation on at least some of the plurality of columns based on the importance of each of the plurality of columns calculated through the above-mentioned attention score.

4. In Paragraph 3, The calculation of the above importance is, A data augmentation method for learning tabular data characterized by averaging the attention scores calculated from the above specific module to calculate the importance of each of the above plurality of columns.

5. In Paragraph 4, Based on the importance of each of the plurality of columns, the method further includes the step of selecting at least some specific columns among the plurality of columns that are to be augmented. The above specific module is, Randomly select one of the multiple pre-set augmentation techniques, and A data augmentation method for learning tabular data, characterized by generating augmented table data by performing augmentation on a specific column among a plurality of columns using the selected augmentation technique.

6. In Paragraph 5, In the step of selecting the specific column mentioned above, Based on the importance of each of the plurality of columns above, first columns satisfying a first criterion and second columns satisfying a second criterion among the plurality of columns are each specified, and A data augmentation method for learning tabular data characterized by selecting at least some of the first columns and the second columns as the specific columns.

7. In Paragraph 6, The specific column above includes at least some of the second columns satisfying the second criterion, and The above specific module is, A data augmentation method for learning tabular data, characterized by generating augmented table data by performing augmentation on at least some of the second columns satisfying the second criterion using the selected augmentation technique among the plurality of augmentation techniques.

8. In Paragraph 7, In the step of generating the augmented table data mentioned above, Based on a pre-set selection ratio criterion, at least some of the second columns satisfying the second criterion are selected as the specific columns to be augmented, and A data augmentation method for learning tabular data, characterized by generating an augmented table by performing augmentation on a specific column selected according to the above-mentioned pre-set selection ratio criteria.

9. In Paragraph 5, The step of generating the augmented table data above is, A data augmentation method for learning tabular data, characterized by the step of performing augmentation on a specific column selected based on the attention score to generate the augmented table data.

10. In Paragraph 6, The step of generating the augmented table data above is, A data augmentation method for learning tabular data, characterized by the step of generating augmented table data by performing augmentation on at least some of the second columns satisfying the second criteria in order to maintain the structure of the first columns satisfying the first criteria.

11. In Paragraph 6, A data augmentation method for learning tabular data, characterized in that, for each learning epoch of the model to be learned, one of the plurality of augmentation techniques is randomly selected and applied to perform augmentation on the specific column to generate the augmented table data.

12. In Paragraph 11, The above specific module is, For each training epoch of the above-mentioned target model, any one of the above-mentioned plurality of augmentation techniques is randomly selected, and A data augmentation method for learning tabular data, characterized by generating augmented table data by performing augmentation on a specific column using the selected augmentation technique.

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

14. In Paragraph 13, 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 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.

15. In Paragraph 13, The above loss function is, It is defined to learn in a direction that maximizes the similarity between the first embedding vector corresponding to the table data and the second embedding vector corresponding to the augmented table data, and In the step of training the above-mentioned target model, A data augmentation method for learning tabular data characterized by performing contrastive learning on the target model using the above loss function.

16. In Paragraph 15, Based on the above contrast learning, a step of obtaining a contrast-learned model with the above loss function; and A data augmentation method for learning tabular data, characterized by further including the step of performing fine-tuning on the above-mentioned contrast-learned model.

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 association between multiple columns included in the table data, the importance of each of the multiple columns is calculated, and Based on the importance calculated above, the table data is augmented to generate augmented table data, and The above plurality of segmented table data and the above augmented table data are each input into an encoder, and From the encoder, a plurality of embedding vectors corresponding to each of the plurality of segmented table data are obtained, and From the encoder above, an embedding vector corresponding to the augmented table data is obtained, and A data augmentation method for learning tabular data, characterized by training a target model using the plurality of embedding vectors corresponding to each of the plurality of segmented table data and the embedding vector corresponding to the augmented table data.

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, Table data configured to include multiple different columns and at least one record having a value corresponding to each of the multiple columns, and Based on the association between the plurality of columns included in the table data above, the importance of each of the plurality of columns is calculated, and Based on the importance calculated above, augment the table data to generate augmented table data, and The above table data and the above augmented table data are each input into the encoder, and From the encoder above, a first embedding vector corresponding to the table data is obtained, and From the encoder, a second embedding vector corresponding to the augmented 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 specifying table data configured to include multiple different columns and at least one record having a value corresponding to each of the multiple columns; A step of calculating the importance of each of the plurality of columns based on the association between the plurality of columns included in the table data; A step of generating augmented table data by augmenting the table data based on the importance calculated above; A step of inputting the table data and the augmented table data into an encoder, respectively; A step of obtaining a first embedding vector corresponding to the table data from the encoder; A step of obtaining a second embedding vector corresponding to the augmented table data from the 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