Method and system for providing tabular data analysis-based service using artificial intelligence model pre-trained by means of tabular augmented data

Abstract

An embodiment provides a method performed by a computer, the method comprising the steps of: receiving tabular data for analysis; generating, by at least one processor, analysis result information about the tabular data for analysis by using at least one artificial intelligence model, wherein the at least one artificial intelligence model is pre-trained on the basis of a training dataset comprising tabular augmented data generated on the basis of recognition of a user labeling pattern; and inputting the analysis result information into at least one subsequent processing component.

Description

Method and system for providing tabular data analysis-based services using an artificial intelligence model pre-trained on tabular augmented data

[0001] The present disclosure relates to a method and system for providing a service based on tabular data analysis using an artificial intelligence model pre-trained with tabular augmented data, and more specifically, to a method and system for providing a service based on tabular data analysis based on data augmentation for tabular data that enables efficient prediction model training by utilizing a small number of labeled tabular data.

[0002] Machine learning enables artificial intelligence models to learn patterns from large volumes of data to perform predictions or classifications. In particular, tabular data is the most commonly encountered data format in various fields such as healthcare, finance, and business. However, unlike images or text, tabular data presents a fundamental problem in that obtaining labels requires significant cost and manual labor by experts. Consequently, in situations where labels are extremely scarce—that is, in few-shot learning environments—existing supervised learning methodologies face limitations where performance degrades significantly.

[0003] To overcome these limitations, the need for learning methodologies that effectively utilize large amounts of unlabeled data has emerged. Previously, three approaches were attempted to address the problem of insufficient labeling.

[0004] The first is semi-supervised learning, a method in which a model is first trained using a small amount of labeled data, and then predictions are generated for unlabeled data to be utilized as if they were labeled. However, this approach has the drawback that if inaccurate predictions are generated early on, it can actually hinder model training.

[0005] The second is self-supervised learning, a method that utilizes latent relationships inherent in the data itself to create and solve pretext tasks autonomously without labels. However, tabular data lacks structural characteristics, making it difficult to design meaningful data augmentation methods, which can lead to uncertainty in the quality of the augmentation.

[0006] The third method is rule-based labeling, which involves manually defining explicit rules from specific domain experts and applying them to unlabeled data. However, this approach has limitations in that manual rules cannot encompass all data, as expert knowledge is often difficult to express as clear rules and involves complex and subtle implicit judgment patterns.

[0007] As such, the performance of AI models based on existing semi-supervised learning (SSL) methods can be hindered by initial prediction errors; self-supervised and contrastive learning methods suffer from uncertainty in data augmentation quality due to the lack of structural characteristics in tabular data; and expert knowledge-based automatic labeling has limitations in failing to capture the complex and subtle implicit judgment patterns of experts.

[0008] Accordingly, there is a need for research on new learning methodologies that generate high-quality tabular augmented data by overcoming the error accumulation problem of existing semi-supervised learning, the uncertainty of augmentation quality in self-supervised learning, and the limitations of implicit pattern capture in rule-based labeling, while minimizing the loss of potential information inherent in large amounts of tabular unlabeled data.

[0009] According to various embodiments of the present disclosure, an artificial intelligence model intelligently recognizes and learns complex and subtle expert labeling patterns inherent in a small number of user-labeled data, performs automatic labeling on unlabeled data to generate tabular augmented data, and aims to provide a method and system for providing a tabular data analysis-based service using a pre-trained artificial intelligence model based thereon.

[0010] According to various embodiments of the present disclosure, a method for training a prediction model based on data augmentation for tabular data, a prediction method based on tabular data analysis, and a system are provided that enable learning semantically robust representations using only unlabeled data by establishing a new inductive bias that samples located close to the data distribution based on the distribution of tabular data will have high semantic similarity, and performing data augmentation based on this.

[0011] According to various embodiments of the present disclosure, the present invention aims to provide a prediction model training method based on data augmentation for tabular data, a prediction method based on tabular data analysis, and a system that can learn the intrinsic characteristics of data by performing intervals on augmented data based on the distribution of tabular data, thereby reducing data complexity, leaving only meaningful information, removing unnecessary noise, and learning the intrinsic characteristics of the data.

[0012] However, the technical problems that the various embodiments of the present disclosure aim to solve are not limited to the technical problems described above, and other technical problems may exist.

[0013] One embodiment is,

[0014] A method performed by a computer comprises the steps of: receiving tabular data for analysis and storing it in at least one memory; at least one processor loading the tabular data for analysis from the at least one memory; at least one processor generating analysis result information for the tabular data for analysis using at least one artificial intelligence model, wherein the at least one artificial intelligence model is pre-trained based on a training dataset including tabular augmented data generated based on recognition of user labeling patterns; and inputting the analysis result information to at least one subsequent processing component.

[0015] In another aspect, the method may further include the step of the at least one subsequent processing component manifesting the analysis result information through at least one user interface.

[0016] In another aspect, the analysis result information may include at least one of classification information, regression prediction information, clustering information, or anomaly detection information for the tabular data for analysis.

[0017] In another aspect, the method may further include a step in which the subsequent processing component performs at least one decision-making based on the analysis result information.

[0018] In another aspect, the above at least one decision may include at least one of automated recommendation, risk assessment, resource allocation, and real-time control.

[0019] In another aspect, the at least one artificial intelligence model may be pre-trained by a learning method comprising the steps of: receiving tabular labeling data containing user labeling information and storing it in at least one memory; training the at least one artificial intelligence model to recognize the user labeling pattern based on the tabular labeling data; receiving tabular unlabeled data and storing it in the at least one memory; generating tabular augmented data by performing automatic labeling on the tabular unlabeled data using the at least one artificial intelligence model; and fine-tuning the at least one artificial intelligence model based on a learning dataset including the tabular labeling data and the tabular augmented data.

[0020] In another aspect, the step of generating the tabular augmented data by performing the automatic labeling may include the step of estimating the label distribution of the tabular user-labeled data having characteristics similar to the tabular unlabeled data, and the step of performing automatic labeling on the tabular unlabeled data based on the estimated label distribution.

[0021] In another aspect, the above-mentioned at least one artificial intelligence model may include at least one of a transformer model that learns the correlation between column information of tabular data and a decision tree algorithm-based model.

[0022] In another aspect, the method may further include the steps of: generating additional tabular augmented data by having at least one processor perform limited range augmentation on numerical variable values ​​of tabular data included in the training dataset; training at least one artificial intelligence model to recognize patterns of the training dataset and the additional tabular augmented data; loading tabular labeling data stored in at least one memory; and using the tabular labeling data, fine-tuning the at least one artificial intelligence model so as to output a result value for the tabular data.

[0023] In another aspect, the step of generating the additional tabular augmented data may include the step of determining a plurality of ranges for each of the plurality of numerical variable values ​​based on the distribution of a plurality of numerical variable values ​​corresponding to each of the plurality of numerical variables included in the training dataset, and the step of generating the additional tabular augmented data by changing at least some of the plurality of numerical variable values ​​included in the training dataset to other values ​​within the corresponding range.

[0024] In another aspect, the step of generating the additional tabular augmented data may include at least one of the steps of exchanging at least some of the plurality of numeric variable values ​​with other numeric variable values ​​within the same range as the range, and changing at least some of the plurality of numeric variable values ​​to arbitrary values ​​within the range.

[0025] In another aspect, the method further comprises the step of obtaining segmented data by performing segmentation on the training dataset and the additional tabular augmented data; and the step of training at least one artificial intelligence model may include the step of training the at least one artificial intelligence model to output the segmented data by taking the training dataset and the additional tabular augmented data as inputs.

[0026] One embodiment is,

[0027] The present invention provides a system comprising at least one memory and at least one processor that reads at least one instruction stored in the at least one memory and executes a method for providing a service based on tabular data analysis, wherein the at least one instruction includes the step of receiving tabular data for analysis and storing it in at least one memory; the step of the at least one processor loading the tabular data for analysis from the at least one memory; the step of the at least one processor generating analysis result information for the tabular data for analysis using at least one artificial intelligence model, wherein the at least one artificial intelligence model is pre-trained based on a training dataset including tabular augmented data generated based on recognition of user labeling patterns, and the step of inputting the analysis result information to at least one subsequent processing component.

[0028] In another aspect, the system may further include a Field Programmable Gate Array (FPGA) implementation for a predetermined artificial neural network comprising a plurality of neurons arranged in an array including at least one register, at least one programmable logic, and at least one input interface, a plurality of synapse circuits storing synapse weights that regulate the connection strength between the plurality of neurons, and at least one routing network that controls the data flow between the plurality of neurons, wherein each of the plurality of neurons is connected to at least one other neuron through the routing network to establish a transmission path for the weights.

[0029] In another aspect, the system may further include a plurality of neurons organized into an array comprising at least one register, at least one microprocessor, and at least one input, and a plurality of synapse circuits storing synapse weights that regulate the connection strength between the plurality of neurons, wherein each of the plurality of neurons may further include an Application Specific Integrated Circuit (ASIC) for a predetermined artificial neural network connected to at least one other neuron through any one of the plurality of synapse circuits.

[0030] According to various embodiments of the present disclosure, an artificial intelligence model intelligently recognizes and learns a small number of user labeling patterns, generates large-scale tabular augmented data based on this, and trains the model based on this data, thereby enabling the acquisition of large-scale data with the same reliability as the professional judgment criteria of users. By utilizing such quantitatively and qualitatively improved data, the prediction accuracy and generalization performance of tabular data analysis models can be dramatically improved.

[0031] According to various embodiments of the present disclosure, unlike conventional semi-supervised learning that uses simple hard labels, automatic labeling is performed by recognizing and learning a user-labeled distribution and assigning labels to tabular unlabeled data in proportion to the probability of this distribution. Through this, the diversity and robustness of the data can be preserved by reflecting the complex and subtle ambiguities and uncertainty criteria inherent in user-labeling into the augmented data, and higher quality tabular augmented data can be generated by resolving the error accumulation problem of existing methodologies. By using such tabular augmented data, overfitting of tabular data analysis models can be prevented and prediction accuracy can be improved more stably.

[0032] According to various embodiments of the present disclosure, analysis of tabular data with enhanced accuracy and reliability can be performed using a model trained on the high-quality tabular data generated as described above. Accordingly, by providing analysis result information with a low probability of error regarding the tabular data to a subsequent system, intelligent decision-making, such as automated recommendations, risk assessment, and resource allocation, can be performed quickly and accurately in various industrial fields such as finance, healthcare, and manufacturing.

[0033] In addition, according to various embodiments of the present disclosure, by providing a data augmentation method optimized for the characteristics of tabular data that is difficult to label, prediction performance in a few-shot learning environment can be dramatically improved.

[0034] Conventional contrastive learning methods have found it difficult to define meaningful data augmentation methods due to the ambiguous structural characteristics of tabular data. However, various embodiments of the present disclosure limit the augmentation range based on the data distribution, thereby maintaining a strong assumption that the original data and the augmented data are semantically similar. Through this, the prediction model can learn the intrinsic characteristics of the data without confusion, and consequently, can derive prediction results of high accuracy even in situations where labels are scarce.

[0035] According to various embodiments of the present disclosure, by allowing a model to learn rich representations of data on its own through self-supervised learning, the time and cost required for data labeling can be significantly reduced.

[0036] By pre-training encoders using large amounts of unlabeled data and then fine-tuning only the prediction head with a small amount of labeled data, a much more efficient training pipeline can be constructed compared to the traditional method of training complex models from scratch. This is particularly useful in industrial environments where data is continuously updated and labeling is difficult.

[0037] According to various embodiments of the present disclosure, by performing segmentation on augmented tabular data, the complexity of the data is reduced and only meaningful information is retained, thereby improving the generalization ability and reliability of the prediction model.

[0038] In this process, the prediction model becomes robust against noise, can learn the semantic invariance of the data, and identifies hidden relationships that were previously difficult to learn. This robust learning approach enhances the model's generalization ability, enabling accurate predictions even on new data it has not previously observed, which can increase the reliability of the final prediction results.

[0039] However, the effects obtainable through the various embodiments of the present disclosure are not limited to those mentioned above, and other unmentioned effects can be clearly understood from the description below.

[0040] FIG. 1 illustrates an example of a block diagram of a computing system implementing a table-type data analysis-based service according to one embodiment.

[0041] FIG. 2 briefly illustrates the structure of a neuromorphic circuit that may be included in a processor according to one embodiment.

[0042] FIG. 3 is a block diagram of a computing device implementing a table-type data analysis-based service according to one embodiment.

[0043] FIG. 4 is a block diagram of a computing device implementing a table-type data analysis-based service according to another embodiment.

[0044] FIG. 5 is a block diagram of a computing device implementing a tabular data analysis-based service according to another embodiment.

[0045] FIG. 6 is intended to explain the structure of table-type data according to one embodiment.

[0046] FIG. 7 is intended to explain how an artificial intelligence model performs automatic labeling on tabular unlabeled data according to a user labeling pattern according to one embodiment.

[0047] FIG. 8 is a block diagram of a computing device implementing a tabular data analysis-based service according to another embodiment.

[0048] FIG. 9 is intended to explain a method of augmenting multiple ranges for multiple numeric variable values ​​for each of multiple numeric variables for tabular data according to one embodiment.

[0049] FIG. 10 is intended to illustrate a method for performing limited range augmentation on tabular data according to one embodiment.

[0050] FIG. 11 is intended to illustrate a contrast learning method for original tabular data and augmented tabular data according to one embodiment.

[0051] FIG. 12 is intended to explain the fine-tuning process for a prediction head according to one embodiment.

[0052] FIGS. 13 and 14 are intended to illustrate the process of segmentation for tabular data according to one embodiment.

[0053] FIG. 15 is a flowchart of a method for providing a service based on tabular data analysis using an artificial intelligence model pre-trained by tabular augmented data according to one embodiment.

[0054] FIG. 16 is a flowchart of a method for pre-training an artificial intelligence model using tabular augmented data according to one embodiment.

[0055] FIG. 17 is a flowchart of a fine-tuning method for an artificial intelligence model based on limited range augmentation for tabular data according to one embodiment.

[0056] FIG. 18 is a flowchart illustrating steps that may be included in the step of training an artificial intelligence model to recognize patterns in the tabular data of FIG. 17.

[0057] As various modifications can be made to the various embodiments of the present disclosure, specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the various embodiments of the present disclosure, and the methods for achieving them, will become clear by referring to the embodiments described in detail below together with the drawings. However, the various embodiments of the present disclosure are not limited to the embodiments disclosed below but can be implemented in various forms. In the following embodiments, terms such as "first," "second," etc., are used not in a limiting sense but for the purpose of distinguishing one component from another. Also, singular expressions include plural expressions unless the context clearly indicates otherwise. Furthermore, terms such as "include" or "have" mean that the features or components described in the specification exist, and do not preclude the possibility that one or more other features or components may be added. Additionally, in the drawings, the size of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are arbitrarily depicted for convenience of explanation, so the various embodiments of the present disclosure are not necessarily limited to those depicted.

[0058] Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same reference numerals, and redundant descriptions thereof will be omitted.

[0059]

[0060] - Tabular data analysis-based service provision system (1000)

[0061] A system (1000) according to one embodiment can generate high-quality tabular augmented data by receiving a small amount of tabular user-labeled data, intelligently recognizing and learning complex and subtle expert labeling patterns embedded therein, and, based on this, intelligently assigning labels to a large amount of unlabeled tabular data by replicating the statistical ratio of the label distribution. This augmented data can overcome the problem of error accumulation in existing semi-supervised learning or the limitations of rule-based labeling, and contribute to securing the accuracy and robustness of model learning.

[0062] The system (1000) can generate accurate analysis result information (e.g., classification, prediction, clustering, etc.) for tabular data using an artificial intelligence analysis model that is pre-trained based on the generated tabular augmented data. Based on this analysis result information for tabular data, the system (1000) determines intelligent action policies such as automated recommendations, risk assessment, and resource allocation, and thereby can provide a highly reliable service based on tabular data analysis to the user.

[0063] Meanwhile, a system (1000) according to one embodiment can perform training of a tabular data analysis model to innovatively improve the prediction performance of tabular data that is difficult to label. While existing machine learning models required a sufficient amount of labeled data for prediction, the system (1000) can overcome this limitation by utilizing self-supervised learning. For example, through a methodology that limits the augmentation range based on the data distribution, the prediction model can effectively learn the intrinsic patterns of the data using only unlabeled data.

[0064] The system (1000) can generate augmented data semantically similar to the original data from a large amount of unlabeled data, allowing the encoder to learn the invariance inherent in the tabular data. Subsequently, the structure connecting the pre-trained encoder and prediction model based on a small amount of labeled data can be fine-tuned to achieve final prediction capability.

[0065] Through this, the system (1000) can provide an efficient solution to the problem of requiring massive cost and time for data labeling, and can have high prediction accuracy and generalization ability even in a few-shot learning environment with extremely few labels.

[0066] Furthermore, by achieving performance that surpasses complex foundation models utilizing large-scale datasets while using a simple model structure, it is possible to provide practical and highly reliable prediction results.

[0067] FIG. 1 illustrates an example of a block diagram of a computing system (1000) implementing a table-type data analysis-based service according to one embodiment.

[0068] Referring to FIG. 1, a computing system (1000) implementing a table-type data analysis-based service according to one embodiment includes a user computing device (110), a server computing system (130), and a training computing system (150), and the devices can communicate through a network (170).

[0069] A method for providing a table-type data analysis-based service according to one embodiment may be implemented and provided locally by a user computing device (110), implemented and provided in the form of a web service by a server computing system (130) communicating with the user computing device (110), or implemented and provided by the user computing device (110) and the server computing system (130) in conjunction with each other.

[0070] In this embodiment, the user computing device (110) and / or the server computing system (130) can train a machine learning model (120 and / or 140) through interaction with a training computing system (150) that is communicatedly connected via a network (170). The training computing system (150) may be separate from the server computing system (130) or may be part of the server computing system (130).

[0071] And at this time, the artificial intelligence model can be 1) trained directly locally by a user computing device (110), 2) trained by the server computing system (130) and the user computing device (110) interacting with each other through a network (170), and 3) trained by a separate training computing system (150) using various training and learning techniques. It may also be implemented by transmitting the artificial intelligence model trained by the training computing system (150) to the user computing device (110) and / or the server computing system (130) through the network (170) to provide / update it.

[0072] In some embodiments, the training computing system (150) may be part of the server computing system (130) or part of the user computing device (110).

[0073]

[0074] - User Computing Device (110: User Computing Device)

[0075] The user computing device (110) may include all other types of computing devices, such as a smartphone, a mobile phone, a digital broadcasting device, a PDA (personal digital assistants), a PMP (portable multimedia player), a desktop, a wearable device, an embedded computing device, a tablet PC, an augmented reality (VR) device, and / or a virtual reality (AR) device.

[0076] The user computing device (110) may include at least one processor (111) and memory (112). Here, the processor (111) may be composed of at least one of a central processing unit (CPU), a graphics processing unit (GPU), ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions, or a plurality of electrically connected processors.

[0077] In particular, according to the embodiment, this processor (111) may be configured based on a Field Programmable Gate Array (FPGA) implementation and / or an Application Specific Integrated Circuit (ASIC), which is a hardware technology for implementing a certain digital circuit.

[0078] Here, a field programmable gate array (FPGA) can refer to a flexible digital circuit that is programmable according to user needs.

[0079] As an example, a field programmable gate array implementation may include a register that temporarily stores data and controls the flow and timing of signals to maintain intermediate results or state information of operations to support synchronized operation of the FPGA, programmable logic that programs operations within the FPGA to perform specific functions or operations as logic circuits configurable according to user needs, and an input interface that receives signals from external devices or sensors and transmits them to internal circuits as a channel for receiving data from outside the FPGA.

[0080] Through the combination of the above components, a field-programmable gate array implementation can provide flexible and various types of digital circuits.

[0081] Meanwhile, an Application-Specific Integrated Circuit (ASIC) can refer to a custom integrated circuit that is fixedly designed to perform a specific use or function.

[0082] As an example, the application-dedicated integrated circuit may include a register, which is a small memory device for temporarily storing and managing data and supports the rapid processing of ASIC operations by storing intermediate calculation results or state information; a microprocessor, which is a central processing unit that performs control and operations within the ASIC and coordinates the operation of the entire system by performing various operations or generating control signals when necessary; and an input block, which is an interface for receiving data from the outside, which receives data to be processed by the ASIC and transmits it internally, and receives various input data through connections with sensors or external devices.

[0083] Through the combination of the components mentioned above, an application-specific integrated circuit can perform specific purpose tasks in an optimized manner.

[0084] For example, ASICs can have a structure of a neuromorphic circuit in the form of an array containing multiple neuron circuits.

[0085] FIG. 2 briefly illustrates the structure of a neuromorphic circuit (300) that may be included in a processor (111, 131, 151) according to one embodiment.

[0086] Referring to FIG. 2, for example, a neuromorphic circuit (300) may include a plurality of presynaptic neuron circuits (310), a plurality of presynaptic lines (311) extending laterally from the plurality of presynaptic neuron circuits (310), a plurality of postsynaptic neuron circuits (320), a plurality of postsynaptic lines (321) extending longitudinally from the plurality of postsynaptic neuron circuits (320), and a plurality of synaptic circuits (330) provided at the intersection of the plurality of presynaptic lines (311) and the plurality of postsynaptic lines (321).

[0087] A plurality of free synaptic neuron circuits (310) can transmit signals input from the outside in the form of electrical signals to a plurality of synaptic circuits (330) through a plurality of free synaptic lines (311).

[0088] Additionally, a plurality of post-synaptic neuron circuits (320) can receive electrical signals from a plurality of synaptic circuits (330) through a plurality of post-synaptic lines (321).

[0089] Furthermore, multiple post-synaptic neuron circuits (320) may transmit electrical signals to multiple synaptic circuits (330) through multiple post-synaptic lines (321).

[0090] A plurality of synapse circuits (330) can store weights included in layers constituting a neural network system implemented by a neuromorphic circuit (300) and perform a predetermined operation based on the weights and input data.

[0091] For example, each of the plurality of synaptic circuits (330) may include a resistive memory cell having a variable resistance. In this case, the resistance value of the plurality of synaptic circuits (330) changes by a voltage applied through the plurality of presynaptic neuron circuits (310) or the plurality of postsynaptic neuron circuits (320), and can store weight data according to this resistance change.

[0092] The neuromorphic circuit (300) is formed by mimicking the structure of neurons and synapses, which are essential elements of the human brain. When a deep neural network (DNN) is realized using the neuromorphic circuit (300), the data processing speed can be improved and power consumption can be reduced compared to when the existing von Neumann structure is utilized.

[0093] The memory (112) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, and combinations thereof, and may include web storage of a server that performs memory storage functions on the internet. This memory (112) may store data (113) and instructions (114) necessary for the at least one processor (111) to perform functional operations such as training an artificial intelligence model or executing tabular data analysis through an artificial intelligence model.

[0094] In one embodiment, the user computing device (110) can store at least one machine learning model (120).

[0095] For example, the machine learning model (120) may be various machine learning models, such as multiple neural networks (e.g., deep neural networks) for performing a tabular data analysis-based service provision method, or other types of machine learning models including non-linear models and / or linear models, and may be composed of a combination thereof.

[0096] For example, machine learning models may include linear regression, decision trees, random forests, gradient-boosting pre-trained language models or / and deep learning models. And neural networks may include at least one of feed-forward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks or / and other forms of neural networks.

[0097] Additionally, according to various embodiments, the user computing device (110) may store a model to be used in each process and a prompt template that serves as the basis for input to the model in order to perform at least part of the process for a tabular data analysis-based service provision method through a large language model (LLM).

[0098] In one embodiment, a user computing device (110) can receive at least one machine learning model (120) from a server computing system (130) through a network (170), store it in memory (112), and then execute the stored machine learning model (120) through a processor (111) to perform operations for tabular data analysis.

[0099] In another embodiment, the server computing system (130) includes at least one machine learning model (140) and performs operations through the machine learning model (140), and can provide a table-type data analysis-based service to the user by communicating with the user computing device (110) and related data.

[0100] For example, a user computing device (110) can perform a method of providing a tabular data analysis-based service by having a server computing system (130) provide an output for the user's input using a machine learning model (140) via the web.

[0101] Additionally, the artificial intelligence model can be implemented in such a way that at least some of the machine learning models (120 and / or 140) are executed on a user computing device (110) and the rest are executed on a server computing system (130).

[0102] Additionally, the user computing device (110) may include at least one input component (121) for detecting user input. For example, the user input component (121) may include a touch sensor (e.g., a touch screen and / or a touch pad, etc.) for detecting a touch of a user input medium (e.g., a finger or a stylus), an image sensor for detecting user motion input, a microphone for detecting user voice input, a button, a mouse and / or a keyboard, etc. Additionally, the user input component (121) may include an interface and an external controller when receiving input to an external controller (e.g., a mouse and / or a keyboard, etc.) through an interface.

[0103]

[0104] -Server Computing System (130: Server Computing System)

[0105] The server computing system (130) can perform a series of processes to provide a tabular data analysis-based service.

[0106] In detail, in an embodiment, the server computing system (130) can provide a table-type data analysis-based service by exchanging data necessary to enable a table-type data analysis-based service provision process to be executed on an external device such as a user computing device (110).

[0107] More specifically, in an embodiment, the server computing system (130) can provide an environment in which an application for providing a tabular data analysis-based service on a user computing device (110) can operate.

[0108] To this end, the server computing system (130) may include an application program, data and / or instructions, etc. for the application to operate, and may transmit and receive various data based thereon with the external device.

[0109] A server computing system (130) may include at least one processor (131) and memory (132). Here, the processor (131) may be composed of at least one of a central processing unit (CPU), a graphics processing unit (GPU), ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and / or electrical units for performing other functions, or a plurality of electrically connected processors.

[0110] For example, ASICs may have a structure of a neuromorphic circuit in the form of an array containing multiple neuron circuits (see Fig. 2).

[0111] And the memory (132) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory device, magnetic disk, etc. and combinations thereof. This memory (132) may store data (133) and instructions (134) necessary for the processor (131) to perform functional operations, such as training an artificial intelligence model or executing a method for providing a tabular data analysis-based service through the artificial intelligence model.

[0112] In one embodiment, the server computing system (130) may be implemented to include at least one computing device. For example, the server computing system (130) may be implemented to operate a plurality of computing devices according to a sequential computing architecture, a parallel computing architecture, or a combination thereof. Additionally, the server computing system (130) may include a plurality of computing devices connected to a network (170).

[0113] Additionally, the server computing system (130) may store at least one machine learning model (140). For example, the server computing system (130) may include a neural network and / or other multi-layer non-linear model as the machine learning model (140). Exemplary neural networks may include a feed-forward neural network, a deep neural network, a recurrent neural network, and a convolutional neural network.

[0114] In an embodiment, the server computing system (130) may further include a data store computing system (hereinafter, data store) which is a storage for continuously storing and managing raw data that forms the basis of a tabular data analysis-based service.

[0115] Such data stores may include various forms of data storage, ranging from file systems to cloud storage. For example, a data store may include at least one database among a relational database that uses a structured query language (SQL) to define and manipulate data, a NoSQL database designed for flexibility and scalability to process unstructured and semi-structured data, a data warehouse optimized for querying and analysis by centralizing large volumes of data from multiple sources as a system used for reporting and data analysis, a data warehouse that stores large volumes of raw data in basic formats such as structured data, semi-structured data, and unstructured data, and a local storage device or Network Attached Storage (NAS) that stores data in files in a format generally accessible by a computer operating system.

[0116]

[0117] - Training Computing System (150: Training Computing System)

[0118] The training computing system (150) may include at least one processor (151) and memory (152). Here, the processor (151) may be composed of at least one of a central processing unit (CPU), a graphics processing unit (GPU), ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions, or a plurality of electrically connected processors.

[0119] For example, ASICs may have a structure of a neuromorphic circuit in the form of an array containing multiple neuron circuits (see Fig. 2).

[0120] And the memory (152) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory device, magnetic disk, etc. and combinations thereof. This memory (152) may store data (153) and instructions (154) necessary for the processor (151) to perform learning of an artificial intelligence model, etc.

[0121] For example, the training computing system (150) may include a model trainer (160) that trains a machine learning model (120 and / or 140) stored in a user computing device (110) and / or a server computing system (130) using various training or learning techniques, such as back propagation of error.

[0122] For example, such a model trainer (160) can perform updates to one or more parameters of a machine learning model (120 and / or 140) for a goal-oriented conversational service in a backpropagation manner based on a defined loss function.

[0123] In some embodiments, performing backpropagation of the error may include performing truncated backpropagation through time. The model trainer (160) may perform a number of generalization techniques (e.g., weight devaluation, dropout and / or knowledge distillation, etc.) to improve the generalization ability of the machine learning model (120 and / or 140) being trained.

[0124] For example, a model trainer (160) can train a machine learning model (120 and / or 140) based on a series of training data (161). Here, the training data (161) may include data of different forms, such as, for example, images, audio samples and / or text.

[0125] Examples of image types that can be used may include video frames, LiDAR point clouds, X-ray images, computed tomography scans, hyperspectral images, and / or various other forms of images.

[0126] These training data (161) may be provided by a user computing device (110) and / or a server computing system (130). When the training computing device trains a machine learning model (120 and / or 140) on specific data of the user computing device (110), the machine learning model (120 and / or 140) may be characterized as a personalized model.

[0127] And the model trainer (160) includes computer logic that is utilized to provide the desired function.

[0128] Additionally, the model trainer (160) may be implemented as hardware, firmware, and / or software that controls a general-purpose processor. In one embodiment, the model trainer (160) may include a program file stored in a storage device, be loaded into memory (152), and be executed by one or more processors (151). In another embodiment, the model trainer (160) includes one or more sets of computer-executable data (153) and instructions (154) stored in a tangible computer-readable storage medium, such as a RAM hard disk or an optical or magnetic medium.

[0129] Network (170) includes, but is not limited to, 3GPP (3rd Generation Partnership Project) network, LTE (Long Term Evolution) network, WIMAX (World Interoperability for Microwave Access) network, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), Bluetooth network, satellite broadcasting network, analog broadcasting network and / or DMB (Digital Multimedia Broadcasting) network.

[0130] Generally, communication through the network (170) can be performed using any type of wired and / or wireless connection through various communication protocols (e.g., TCP / IP, HTTP, SMTP and / or FTP, etc.), encodings or formats (e.g., HTML and / or XML, etc.), and / or protection schemes (e.g., VPN, Secure HTTP and / or SSL, etc.).

[0131] FIG. 3 is a block diagram of a computing device (100) implementing a table-type data analysis-based service according to one embodiment.

[0132] Referring to FIG. 3, the computing device (100) included in the user computing device (110), server computing system (130), and training computing system (150) includes a plurality of applications (e.g., applications 1 to N). Each application may include a machine learning library and one or more machine learning models. For example, the applications may include an image processing application (e.g., detection, classification, and / or segmentation, etc.), a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, and / or a chat-bot application.

[0133] In an embodiment, the computing device (100) may include a model trainer (160) for training an artificial intelligence model, and by storing and operating the trained artificial intelligence model, it may provide output data according to a predetermined input data.

[0134] Each application of the computing device (100) can communicate with a number of other components of the computing device (100), such as, for example, at least one sensor, a context manager, a device state component, and / or additional components. In one embodiment, each application can communicate with each device component using an API (e.g., a public API). In one embodiment, the API used by each application may be specific to that application.

[0135] FIG. 4 is a block diagram of a computing device (200) implementing a table-type data analysis-based service according to another embodiment.

[0136] Referring to FIG. 4, the computing device (200) includes a plurality of applications (e.g., Application 1 to Application N). Each application can communicate with a central intelligence layer. For example, applications may include an image processing application, a text messaging application, an email application, a dictation application, a virtual keyboard application and / or a browser application. In one embodiment, each application can communicate with the central intelligence layer (and a model stored therein) using an API (e.g., a common API across all applications).

[0137] The central intelligence layer may include a number of machine learning models. For example, as illustrated in FIG. 4, at least some of the machine learning models may be provided for each application and managed by the central intelligence layer. In other embodiments, two or more applications may share a single machine learning model. For example, in some embodiments, the central intelligence layer may provide a single model for all applications. In some embodiments, the central intelligence layer may be included within the operating system of the computing device (200) or otherwise implemented.

[0138] The central intelligence layer can communicate with the central device data layer. The central device data layer may be a centralized data store for the computing device (200). As illustrated in FIG. 4, the central device data layer can communicate with a number of other components of the computing device (200), such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In some embodiments, the central device data layer can communicate with each device component using an API (e.g., a private API).

[0139] The technology described herein may refer to servers, databases, software applications, and other computer-based systems, as well as actions taken and information transmitted to or from said systems. It will be recognized that the inherent flexibility of computer-based systems allows for a wide range of possible configurations, combinations, division of tasks, and functionality between and from components. For example, the processes described herein may be implemented using a single device or component or multiple devices or components operating in combination. Databases and applications may be implemented in a single system or in a distributed system across multiple systems. Distributed components may operate sequentially or in parallel.

[0140] FIG. 5 is a block diagram of a computing device (400) implementing a tabular data analysis-based service according to another embodiment. FIG. 6 is a block diagram of a computing device (401) implementing a tabular data analysis-based service according to another embodiment. FIG. 7 is intended to explain a method in which an artificial intelligence model performs automatic labeling on tabular unlabeled data according to a user labeling pattern according to an embodiment. FIG. 8 is intended to explain the structure of tabular data according to an embodiment. FIG. 9 is intended to explain a method of augmenting multiple ranges for multiple numerical variable values ​​for each of multiple numerical variables for tabular data according to an embodiment. FIG. 10 is intended to explain a method of performing limited range augmentation on tabular data according to an embodiment. FIG. 11 is intended to explain a contrastive learning method for original tabular data and augmented tabular data according to an embodiment. FIG. 12 is intended to explain a fine-tuning process for a prediction head according to an embodiment. FIG. 13 and FIG. 14 are intended to explain a process of segmentation for tabular data according to an embodiment.

[0141] Referring to FIG. 5, the computing device (400) included in the user computing device (110), server computing system (130) and training computing system (150) may include a data augmentation module (10), a tabular data analysis model learning module (20), and a decision module (30).

[0142] The data augmentation module (10) can overcome the quantitative and qualitative limitations of data for training artificial intelligence models by generating a large amount of high-quality tabular augmented data based on a small number of user labeling patterns.

[0143] First, tabular data can be structured in a table form that includes, for example, row information representing information on various multiple heterogeneous features of individual samples and column information representing information on one feature of multiple samples.

[0144] For example, referring to FIG. 6, the tabular data may have a table form consisting of multiple rows (Ro1~Ro11) representing multiple heterogeneous feature information including age (A~K), height (Height), weight (Weight), and gender (Gender) for each of the multiple samples (A~K), and multiple columns (Cn1~Cn4) representing information on one type of the multiple heterogeneous feature information of the multiple samples (A~K).

[0145] Here, the heterogeneous features of a tabular dataset may include categorical variables, such as gender and country, which are difficult to express as numbers and can be represented as words, nouns, or combinations of specific words, and numerical variables, such as height, weight, and age, which can be expressed as continuous numbers.

[0146] For example, tabular 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 variables such as the product's brand name and color.

[0147] In addition, for example, tabular data may be 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 variables such as information on the manufacturing process line and information on the manufacturing process supervisor.

[0148] As such, tabular data may include at least one of numerical variables and categorical variables, and may be data of information structured in the form of a table containing various types of information regarding multiple samples. Additionally, tabular data may include information regarding various types of samples related to various industrial fields.

[0149] A computing device (400) according to one embodiment may receive and obtain table-type data from an external source. In this case, the computing device (400) may receive table-type data directly from an external source, or may extract table-type data from data of documents (e.g., papers, books, patent documents, reports, etc.) received from an external source.

[0150] For example, the computing device (400) can extract tabular data structured in a table format from input document data based on a document understanding model included in the machine learning model (120 and / or 140).

[0151] The data augmentation module (10) may include a user labeling pattern recognition module (11) and an automatic labeling module (12).

[0152] The user labeling pattern recognition module (11) can be configured to receive a small number of tabular labeling data (TLD) labeled by the user and to recognize and learn complex and subtle labeling patterns (tacit knowledge) of experts.

[0153] For example, the user labeling pattern recognition module (11) may enable at least one artificial intelligence model, including at least one of a decision tree algorithm-based model or an attention mechanism-based transformer model, to learn the user's labeling patterns included in tabular labeling data (TLD). Here, the transformer model can learn complex and non-linear interrelationships between multiple column information included in the tabular data based on an attention mechanism.

[0154] In this process, the user labeling pattern recognition module (11) can train at least one artificial intelligence model to perform automatic labeling that reflects the reliability of expert judgment by assigning weights to labeling patterns, taking into account the consistency and frequency of user labeling. Through this, the user labeling pattern recognition module (11) can enable at least one artificial intelligence model to learn the complex and subtle judgment criteria and statistical label distribution inherent in the user labeling patterns.

[0155] For example, it can be assumed that in a financial service environment, a skilled underwriter assigns 'high risk' and 'low risk' labels to the data (annual salary, debt ratio, credit rating, etc.) of 100 loan applicants. In this case, the user consistently assigns a 'high risk' label to applicants with an annual salary of less than 30 million won and a debt ratio of 40% or more, but an ambiguous pattern may be observed where only some (30%) of applicants with an annual salary of 50 million won or more are assigned a 'medium risk' label, while the rest (70%) are assigned a 'low risk' label.

[0156] The user labeling pattern recognition module (11) can enable at least one artificial intelligence model to receive this small amount of labeling data as input and to recognize and learn a probabilistic labeling distribution as an implicit pattern of experts, such as '70% are judged as low risk and 30% as medium risk even within the group with an annual salary of 50 million won or more,' rather than a simple rule. This learned pattern can subsequently be used as a basis for intelligent automatic labeling of a large amount of unlabeled data.

[0157] The automatic labeling module (12) can perform intelligent automatic labeling that reflects expert labeling patterns for a large amount of tabular unlabeled data using at least one artificial intelligence model learned from the user labeling pattern recognition module (11).

[0158] For example, referring to FIG. 7, the tabular labeling data (TLD) shown in the table on the left includes labels (L1 to L5) assigned by a user to a small number of samples (A to E) according to professional criteria, and the automatic labeling module (12) can perform automatic labeling on the tabular non-labeling data (F to N) shown in the table on the right based on this user labeling pattern learned from the pattern recognition module (11).

[0159] In this process, the module (12) can probabilistically assign labels (Lx1~Lxn) based on the statistical distribution of the user labeling pattern if the features (X1, X2, X3) of each sample (F, G, H, ..., N) of the unlabeled data are similar to the features (X1, X2, X3) of each sample (A~E) of the user labeling data.

[0160] For example, if 70% of samples with similar features are labeled L1 and 30% are labeled L2, then the same distribution can be applied to samples with similar features to the samples in the tabular labeling data for tabular unlabeled data. That is, the automatic labeling module (12) can generate high-quality tabular augmented data that preserves the ambiguity and diversity inherent in the data by probabilistically replicating the statistical ratio of the label distribution inherent in the learned user labeling pattern to the unlabeled data.

[0161] In this case, the automatic labeling module (12) can estimate the label distribution of user-labeled data having characteristics similar to tabular non-labeled data and perform automatic labeling on tabular non-labeled data based on this estimated label distribution. In this case, for example, labels can be randomly assigned to tabular non-labeled data in proportion to the probability of the label distribution of user-labeled data.

[0162] For example, if a label distribution of ‘risk 70%, safety 30%’ is estimated for multiple samples having similar characteristics for user-labeled data, the automatic labeling module (12) can assign a ‘risk’ label to approximately 70% of the data and a ‘safe’ label to 30% of the data according to the label ratio according to the estimated distribution for multiple samples having similar characteristics to the above features included in tabular non-labeled data, and accordingly, statistical ambiguity of the user-labeled pattern can be preserved.

[0163] The high-quality tabular augmented data generated through this process can resolve the error accumulation problem of existing semi-supervised learning and dramatically improve the robustness of the final tabular data analysis model.

[0164] The tabular data analysis model training module (20) can train the tabular data analysis model by utilizing high-quality tabular augmented data generated from the data augmentation module (10).

[0165] The tabular data analysis model learning module (20) can train an artificial intelligence model based on an expanded learning dataset that integrates a small number of tabular labeled data (TLD) labeled by the user and tabular augmented data generated by the automatic labeling module (12).

[0166] At least one artificial intelligence model used for learning may include at least one of an attention mechanism-based transformer model that learns interactions between column information by considering the complexity of tabular data, or an ensemble model that combines multiple decision trees (e.g., XGBoost, LightGBM).

[0167] By using such high-quality training datasets, the tabular data analysis model training module (20) generates accurate analysis result information (e.g., classification result, regression prediction value, clustering result, or anomaly detection result) for the new input tabular data, thereby improving the prediction accuracy and generalization performance of the tabular data analysis model. The generated analysis result information can be transmitted to the decision module (30) for decision-making in the subsequent system.

[0168] The decision module (30) can be configured to receive accurate and reliable analysis result information (classification results, regression prediction values, etc.) generated from the tabular data analysis model learning module (20) and to determine an automated course of action based thereon.

[0169] The decision module (30) can perform at least one of the following decisions: automated recommendation, risk assessment, resource allocation, or real-time control based on the enhanced accuracy of at least one artificial intelligence model.

[0170] For example, the decision module (30) can provide the user with a table-type data analysis-based service, such as automatically determining loan approval / rejection based on risk prediction results received in a financial environment, or automatically optimizing the distribution of maintenance personnel and spare parts based on the probability of equipment failure in a manufacturing environment.

[0171] However, it is not limited to this, and for example, the decision module (30) can perform recommendation and marketing strategy decisions such as automatically providing a customized coupon to a customer when it is predicted that a specific customer has a high probability of churn based on the results of tabular data analysis, and automatically excluding customers who have already confirmed to churn from the advertisement list. In addition, it may be used to determine various action plans throughout management and operations, such as production planning decisions based on demand forecasting, inventory management planning decisions, or marketing strategy decisions based on customer churn prediction.

[0172] In this way, the decision module (30) can perform analysis of tabular data with high accuracy and maximize operational efficiency and business value through intelligent action policy decisions using the data, thereby providing highly reliable services across various fields.

[0173] Meanwhile, referring to FIG. 8, a computing device (401) included in a user computing device (110), a server computing system (130), and a training computing system (150) according to another embodiment may include a data augmentation module (50), an encoder learning module (60), a prediction head learning module (70), and a segmentation module (80). Here, the data augmentation module (50) is a module that augments input original tabular data (OTD) and may include a range determination module (51) and a variable value change module (52).

[0174] The data augmentation module (50) can go beyond simply augmenting data and can process the original tabular data (OTD) to generate meaningful augmented data suitable for self-supervised learning.

[0175] The data augmentation module (50) can effectively augment tabular data that lacks clear structural characteristics, unlike image or text data. The data augmentation module (50) may include a range determination module (51) and a variable value substitution module (52).

[0176] The range determination module (51) can analyze the data distribution of each numerical feature included in the original tabular data (OTD). Based on the distribution of multiple numerical feature values ​​corresponding to each of the multiple numerical features included in the original tabular data (OTD), the range determination module (51) can determine multiple ranges for the multiple numerical feature values ​​for each of the multiple numerical features.

[0177] In this case, the range determination module (51) can determine multiple ranges based on the overall distribution of multiple numeric variable values ​​so that the number of data samples belonging to each range is substantially equal.

[0178] For example, referring to FIG. 9, the range determination module (51) can determine multiple ranges by using a specific point as a boundary so that the number of numeric variable values ​​belonging to each range is substantially equal based on the overall distribution of tabular data represented by a histogram.

[0179] For example, multiple ranges can be determined using multiple reference values ​​(Q1, Q2, Q3) as boundaries for the entire distribution of numerical variable values. In this case, four ranges (Range1 to Range4) with substantially equal numbers of numerical variable values ​​can be determined, and through this, numerical variable values ​​that are semantically similar in the data distribution can be induced to belong to one range group.

[0180] For example, referring to FIG. 10, multiple ranges for numeric variable values ​​can be determined so that the number of numeric variable values ​​of age, which is one of the numeric variables, is substantially evenly distributed among each range. For example, a first range (15–19) including 15 and 18, a second range (20–34) including 23, 26, and 33, a third range (35–49) including 35, 42, and 48, and a fourth range (50–66) including 50, 60, and 65 can be determined.

[0181] As such, multiple ranges for multiple numeric variable values ​​determined by the range determination module (51) are determined to include substantially the same number of variable values, and depending on the case, the size of each range may differ from one another. That is, the multiple ranges are not determined based on equal size, but rather can be determined so that the number of numeric variable values ​​belonging to each range is equal or as equal as possible.

[0182] Through this range determination process, artificial intelligence models (e.g., encoders) can be guided to focus on the intrinsic characteristics of tabular data instead of unnecessary noise during learning.

[0183] The variable value change module (52) can generate actual augmented data by utilizing multiple ranges determined by the range determination module (51). The variable value replacement module (52) can perform augmentation by changing some of the numeric variable values ​​of the original tabular data (OTD) to other values ​​within the corresponding range.

[0184] For example, referring to FIG. 10, 18 and 15 belonging to the first range (15-19) may be changed to 19 and 16, respectively, which are values ​​within the first range (15-19); 23, 26, and 33 belonging to the second range (20-34) may be changed to 33, 30, and 34, respectively, which are values ​​within the second range (20-34); 35, 42, and 48 belonging to the third range (35-49) may be changed to 40, 36, and 43, respectively, which are values ​​within the third range (35-49); and 50, 60, and 65 belonging to the fourth range (50-66) may be changed to 60, 63, and 58, respectively, which are values ​​within the fourth range.

[0185] In this case, the change can be made through either shuffling or sampling.

[0186] Shuffling may include a method of rearranging existing numeric variable values ​​within the same range. In this case, the variable value change module (52) may perform shuffling by exchanging at least some of the multiple numeric variable values ​​with other numeric variable values ​​within the same range as the corresponding range. For example, the variable value change module (52) may change the numeric variable values ​​for each sample by rearranging 23, 26, and 33 belonging to the second range (20 to 34) by exchanging them with each other.

[0187] Sampling may include a method of changing numeric variable values ​​to new values ​​within the same range. In this case, the variable value change module (52) may perform sampling by changing at least some of the multiple numeric variable values ​​to arbitrary values ​​within the corresponding range. For example, the variable value change module (12) may change 23, 26, and 33 belonging to the second range (20 to 34) to arbitrary values ​​33, 30, and 34 within the second range (20 to 34).

[0188] The augmented tabular data generated through this process has semantic similarity to the original tabular data (OTD) but with some values ​​modified, which can form positive pairs essential for the model to learn the semantic invariance of the data.

[0189] Meanwhile, the data augmentation module (50) can change multiple categorical variable values ​​corresponding to multiple categorical variables included in the original tabular data (OTD) into numerical values.

[0190] For example, referring to Fig. 6, tabular data may include a categorical variable such as gender in the fourth column (Cn4), and gender may include a first categorical variable value 'Male' and a second categorical variable value 'Female'.

[0191] The data augmentation module (50) can change these multiple categorical variable values ​​into numeric values. For example, a first numeric value can be assigned to the first categorical variable value 'Male', and a second numeric value different from the first numeric value can be assigned to the second categorical variable value 'Female'.

[0192] In this case, the variable value determination module (52) can change at least some of the numeric variable values ​​corresponding to the numeric variable to other values ​​within the corresponding range while keeping the numeric values ​​corresponding to the numeric variable values ​​fixed.

[0193] For example, the variable value determination module (52) can change a plurality of numeric variable values ​​included in the first to third columns (Cn1 to Cn3) of the tabular data of FIG. 6 to appropriate numeric values ​​based on a determined range. In this case, the variable value determination module (52) can fix the numeric values ​​for categorical variable values ​​included in the fourth column (Cn4) without changing them to other values.

[0194] Accordingly, the semantic consistency of categorical variables is maintained, and confusion during model training can be prevented. Categorical variables, such as 'male' or 'female' or 'purchased' or 'not purchased,' carry inherent meanings based on their values; randomly altering them can compromise the data's essential label information. Therefore, by fixing categorical variables, the model can clearly grasp the core characteristics of the data without unnecessary noise, thereby improving predictive performance.

[0195] The encoder learning module (60) can learn the semantic features of unlabeled tabular data to generate an effective representation.

[0196] For example, the encoder learning module (60) can perform self-supervised learning on the encoder (53) which performs the function of converting input tabular data into meaningful low-dimensional embedding vectors.

[0197] The encoder (53) may include, for example, a fully connected neural network. However, it is not limited thereto, and the encoder may include at least one neural network model of various forms, such as a convolutional neural network (CNN) or a recursive neural network (RNN).

[0198] For example, referring to FIG. 11, the encoder learning module (60) can train the encoder (53) based on contrast learning of the original tabular data and the augmented tabular data.

[0199] In this case, the encoder learning module (60) can determine a positive pair of any one of the multiple numeric variable values ​​included in the original tabular data and a change value for the said numeric variable value.

[0200] For example, referring to FIG. 11, the encoder learning module (60) takes a plurality of numeric variable values ​​(a1, b1) included in the original tabular data as anchor values ​​(I a Set to ), and change values ​​(a1', b1') for the corresponding multiple numeric variable values ​​(a1, b1) to positive values ​​(I + You can set it to ) and determine multiple positive pairs by matching two corresponding values.

[0201] Additionally, the encoder learning module (60) can determine at least one change value belonging to a different range from any one numeric variable value included in the original tabular data and any one numeric variable value identical to it as a negative pair.

[0202] For example, referring to FIG. 11, the encoder learning module (60) has an anchor value (I aMultiple numeric variable values ​​(a3', b3') belonging to a different range from the multiple numeric variable values ​​(a1, b1) included in the original tabular data set to ) are negative values ​​(I - By setting it to ), multiple negative pairs can be determined by matching multiple numeric variable values ​​(a3', b3') corresponding to negative values ​​with the same multiple numeric variable values ​​(a1, b1).

[0203] In this way, the encoder learning module (60) receives original tabular data and corresponding augmented tabular data from the data augmentation module (50), and by processing the two data according to the range to which the numerical variable value included in each data belongs, it can determine positive pairs that are considered semantically identical and negative pairs that are semantically unrelated.

[0204] The encoder (53) processes both the original tabular data and the augmented tabular data to produce multiple embedding vectors (f 0 , f + , f - Can generate ).

[0205] The encoder training module (60) can train the encoder (53) according to a loss function (e.g., InfoNCE Loss) that causes the distance between positive pairs to become closer and the distance between multiple negative pairs to become farther.

[0206] For example, the encoder learning module (60) can train the encoder (53) by minimizing the loss function (L) according to the following equation (1).

[0207]

[0208] Equation (1):

[0209] (step, : Embedding vector of anchor value, : Embedding vector of numeric variable values ​​modified and augmented from anchor values, Similarity calculation function, : Temperature parameter, : Embedding vector of all positive values ​​belonging to the same range as the anchor, : Embedding vector of all negative values ​​belonging to a range different from the anchor)

[0210]

[0211] In the process of minimizing the loss function (L) as described above, positive pairs (f a , f + The distance between ) is minimized, and the negative pair (f a , f - The distance between them can be maximized. Through this process, the encoder (21) can learn on its own the rule that data with the same meaning should be close to each other, and data with different meanings should be far apart. As this learning is repeated, the encoder (21) can more firmly grasp the essential patterns of the tabular data.

[0212] In this way, the encoder learning module (60) can effectively perform the role of extracting rich features necessary for prediction using only a large amount of unlabeled data.

[0213] The prediction head learning module (70) can perform a final prediction by utilizing a powerful representation of the tabular data of the pre-trained encoder (53). If the encoder learning module (60) acts as a learner that understands the intrinsic patterns of the data, the prediction head learning module (70) can act as an expert that outputs a prediction for a specific task based on that knowledge.

[0214] The prediction head learning module (70) can perform fine-tuning to update the parameters of the prediction model that performs the prediction while keeping the parameters of the pre-trained encoder (53) frozen.

[0215] For example, when labeled tabular data (LTD) with predicted values ​​(predicted labels) for tabular data is input into the prediction head learning module (70), the tabular data can first pass through a fixed encoder (53) to be converted into a meaningful compressed embedding vector (Z).

[0216] Since this embedding vector (Z) contains rich features of the data, the prediction head can perform predictions efficiently without a complex structure.

[0217] The prediction head learning module (70) can input a compressed embedding vector (Z) into a predetermined prediction head, and the prediction head can be trained to output a prediction value for the compressed embedding vector (Z). In this case, the parameters of the encoder (53) can be frozen.

[0218] In this process, the prediction head calculates a loss function using a small number of labeled tabular data (LTD) and updates only its own parameters in a direction that minimizes this loss. In this case, the parameters of the encoder (53) are not updated, so general knowledge obtained through prior learning can be preserved. Thanks to this two-stage separate learning method, accurate and reliable prediction results can be obtained even in environments with very few labels.

[0219] The computing device (401) can accurately perform predictions on newly input tabular data by using a structure in which an encoder and a prediction model learned in this way are connected.

[0220] For example, referring to FIG. 12, the tabular data (TTD) to be analyzed according to user input is input into the encoder (53) and converted into a compressed embedding vector (Z), and the prediction head (71) receives this compressed embedding vector (Z) and outputs result data (RD), which is the final prediction value (Prediction).

[0221] The segmentation module (80) can perform segmentation on the original tabular data and the augmented tabular data.

[0222] Here, sizing refers to assigning a single representative value to values ​​that fall within a predetermined numerical range among the various numerical variable values ​​contained in tabular data. In this context, the representative value may be referred to as a bin value.

[0223] For example, referring to FIG. 6 and FIG. 13, a segmentation module (80) according to one embodiment may assign empty values ​​to each of the multiple age variable values ​​of the multiple samples (A to K). In this case, for example, empty values ​​of 8 may be assigned for ages 1 to 15, 23 for ages 16 to 30, 38 for ages 31 to 45, 53 for ages 46 to 60, and 68 for ages 61 to 76.

[0224] Meanwhile, the number of multiple empty values ​​(T) assigned to multiple numeric variable values ​​included in tabular data can be set in various ways.

[0225] For example, the number of multiple empty values ​​(T) may be from 2 to 100. However, it is not limited thereto, and the number of multiple empty values ​​(T) may be set to be less than 2 or more than 100.

[0226] For example, the number of multiple empty values ​​(T) may be 5 to 100. When the number of multiple empty values ​​(T) is set to 5 to 100, the performance of downstream tasks may be improved by using an autoencoder that is trained based on the interval data generated by performing intervals on tabular data.

[0227] Since such interval data is generated by replacing the original values ​​of numerical variables in existing tabular data with empty values, interval data may be data to which inductive bias has been applied to the tabular data.

[0228] Afterwards, the encoder learning module (60) can train the autoencoder including the encoder (53) to output segmented data using the original tabular data and augmented tabular data as inputs.

[0229] For example, referring to FIG. 14, the autoencoder may include an encoder and a decoder. The encoder can generate a compressed embedding vector (Z) based on input tabular data. In this case, the encoder can generate a compressed embedding vector (Z) with reduced dimensions of the vector for the tabular data input. Additionally, the decoder can generate a predetermined output value by expanding the dimensions of the compressed embedding vector (Z) from the encoder.

[0230] The encoder and decoder may include, for example, a fully connected neural network. However, they are not limited thereto, and the encoder and decoder may include at least one neural network model of various forms, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). The size of the neural network layers of the encoder may be configured to gradually decrease, and the size of the neural network layers of the decoder may be configured to gradually increase.

[0231] The compressed embedding vector (Z) generated by the encoder can be input to the decoder, and the decoder can be trained to generate target data based on the compressed embedding vector (Z). In this case, the target data can be set to segmented data generated through segmentation, i.e., data with multiple empty values. Accordingly, the decoder can be trained to output segmented data based on the compressed embedding vector (Z).

[0232] For example, a first empty value (b1) may be assigned to numeric data in the first interval (-2.91 ≤ x < -0.85), a second empty value (b2) may be assigned to numeric data in the second interval (-0.85 ≤ x < -0.26), a third empty value (b3) may be assigned to numeric data in the third interval (-0.26 ≤ x < 0.26), and a fourth empty value (b4) may be assigned to numeric data in the fourth interval (0.26 ≤ x < 0.85).

[0233] The encoder can generate a compressed embedding vector (Z) corresponding to a tabular data input, and the decoder can be trained to output segmented data generated by performing segmentation on the tabular data input based on the compressed embedding vector (Z).

[0234] For example, the autoencoder can be trained to output the first empty value (b1) for first numeric data to which the first empty value (b1) is assigned, and to output the second empty value (b2) for second numeric data to which the second empty value (b2) different from the first empty value (b1) is assigned.

[0235] In this case, the loss function (L according to the following equation (2) BinRecon The autoencoder can be trained by optimizing ).

[0236]

[0237] Equation (2):

[0238] (where t i is the i-th empty value, is the decoder output value, z i is the i-th latent variable)

[0239]

[0240] In addition, for example, intervals can be performed on numeric values ​​corresponding to multiple categorical variable values ​​included in tabular data input, so that multiple empty values ​​can be set.

[0241] A segmentation module (80) according to one embodiment can generate an embedding vector for a numeric value of a categorical variable, and can train an autoencoder to output segmentation data for a categorical variable based on the embedding vector for a numeric value of a categorical variable.

[0242] In this case, the loss function (L according to the following equation (3) BinXent The autoencoder can be trained by optimizing ).

[0243]

[0244] Equation (3):

[0245] (step, is the i-th empty value for the j-th feature One-hot vector for, is the decoder output value, is the i-th latent variable)

[0246]

[0247] In this way, the autoencoder can be trained to output segmented data from tabular data by the encoder learning module (60) according to one embodiment, and accordingly, the finally trained autoencoder can effectively extract features including irregularities of the tabular data.

[0248] In addition, the autoencoder is trained based on segmented data generated by performing segmentation on tabular data; since this trains the autoencoder based on data to which inductive bias has been appropriately applied to the tabular data, irregular functions applicable to the autoencoder can be effectively learned during this process.

[0249]

[0250] - Method for providing services based on tabular data analysis (S100)

[0251] A method (S100) for providing a service based on tabular data analysis using an artificial intelligence model pre-trained by tabular augmented data according to one embodiment can overcome the limitations of existing data through intelligent recognition of user labeling patterns, secure high-quality training data, and utilize it to improve the reliability of tabular data analysis and provide a service based thereon.

[0252] Specifically, the method (S100) enables an artificial intelligence model to learn complex and subtle labeling patterns of experts inherent in a small number of user labeling data, and thereby the artificial intelligence model can recognize the ambiguity (label distribution) inherent in tabular labeling data as a statistical pattern along with the implicit judgment criteria of experts.

[0253] Furthermore, the method (S100) can perform intelligent automatic labeling on a large amount of unlabeled tabular data based on learned user labeling patterns. In this automatic labeling process, the artificial intelligence model assigns labels to the tabular unlabeled data according to the statistical ratio of the user labeling patterns, thereby overcoming the limitations of semi-supervised learning, which has the problem of accumulating errors, and generating tabular augmented data with minimized potential information loss.

[0254] Finally, by utilizing an artificial intelligence model trained on high-quality tabular augmented data, it is possible to generate analysis result information with improved accuracy and provide users with a tabular data analysis-based service that performs automated decision-making (e.g., recommendations, risk assessment, resource allocation, etc.) based on this information.

[0255] FIG. 15 is a flowchart of a method for providing a service based on tabular data analysis using an artificial intelligence model pre-trained by tabular augmented data according to one embodiment (S100). FIG. 16 is a flowchart of a method for pre-training an artificial intelligence model by tabular augmented data according to one embodiment.

[0256] Referring to FIG. 15, a method (S100) according to one embodiment may include the steps of receiving tabular data for analysis (S101), generating analysis result information for tabular data for analysis using at least one artificial intelligence model (S103), and inputting the analysis result information to at least one subsequent processing component (S105).

[0257] In one embodiment, the method (S100) may be performed by a processor (131) included in a server computing system (130). However, it is not limited thereto, and at least a part of the method (S100) may be performed by a processor (111) of a user computing device (110) or a processor (151) of a training computing system (150), and another part may be performed by a processor (131) included in a server computing system (130).

[0258] For convenience of explanation, the following description describes a processor (131) included in a server computing system (130) performing the method (S100).

[0259] In step (S101), the processor (131) may receive tabular data for analysis. The tabular data for analysis may include a large amount of unlabeled tabular data to be used for prediction and classification by the model. The processor (131) may store the received tabular data for analysis in at least one memory.

[0260] For example, it includes structured features across various fields, such as customer loan application information in a financial environment, sensor data of equipment in a manufacturing environment, or user behavior log data in a marketing environment, and this data can be used as input data for analysis using an artificial intelligence model in a subsequent step (S103).

[0261] In step (S103), the processor (131) can generate analysis result information for the received tabular data for analysis using at least one artificial intelligence model.

[0262] Here, at least one artificial intelligence model may include a model pre-trained based on a training dataset containing tabular augmented data generated based on recognition of user labeling patterns. This is a high-performance model that resolves the error accumulation problem of existing semi-supervised learning and dramatically improves the prediction accuracy and reliability of the model.

[0263] At least one such artificial intelligence model may include at least one of an attention mechanism-based transformer model that learns correlations between column information of tabular data, and an ensemble model that learns by combining multiple decision trees.

[0264] The processor (131) can generate analysis result information including at least one of a classification result, a regression prediction value, a clustering result, or an anomaly detection result for the input tabular data for analysis using at least one artificial intelligence model.

[0265] For example, referring to FIG. 16, at least one artificial intelligence model may be pre-trained by a learning method comprising the steps of: receiving tabular labeling data containing user labeling information (S1031); training at least one artificial intelligence model to recognize user labeling patterns based on tabular labeling data (S1033); receiving tabular unlabeled data (S1035); generating tabular augmented data by performing automatic labeling on tabular unlabeled data using at least one artificial intelligence model (S1037); and fine-tuning at least one artificial intelligence model based on a learning dataset including tabular labeling data and tabular augmented data (S1039).

[0266] In step (S1031), the processor (131) can receive table-type labeling data (TLD) containing user labeling information and can store it in at least one memory.

[0267] Typical labeled data (TLD) includes tabular data directly labeled by a small number of experts (users), and although the amount is small within the entire dataset, it can serve as reliable ground truth data. This data can be used in a subsequent step (S1033) for an artificial intelligence model to initially recognize and learn the experts' implicit labeling patterns and judgment criteria.

[0268] In step (S1033), the processor (131) can train at least one artificial intelligence model to recognize user labeling patterns based on the received tabular labeling data (TLD).

[0269] This step (S1033) is a process of establishing an intelligent foundation for a model to perform large-scale automatic labeling, and the processor (131) can train an artificial intelligence model (e.g., a transformer model or a decision tree ensemble model) to recognize the labeling patterns of a small number of experts included in tabular labeling data (TLD).

[0270] Through the above learning, the artificial intelligence model can internalize complex and subtle judgment criteria for labeling combinations of specific column information by the user. In particular, the model can acquire intelligent reasoning capabilities for automatic labeling by recognizing patterns beyond simple rules, such as statistical label distributions like '70% of this data group is labeled A, and 30% is labeled B'. This learned model can be used to replicate expert judgment on a large amount of unlabeled data in a subsequent step (S1037).

[0271] In step (S1035), the processor (131) can receive tabular unlabeled data and store it in at least one memory.

[0272] Tabular unlabeled data may include large tabular datasets that are not labeled but can be utilized to dramatically improve the training performance of tabular data analysis models.

[0273] The tabular unlabeled data becomes the target for intelligent automatic labeling by an artificial intelligence model that has learned user labeling patterns in step (S1033) by replicating the judgment of an expert, and serves as the basis for generating high-quality tabular augmented data.

[0274] Effectively utilizing large amounts of tabular unlabeled data in this way is essential to overcome the limitations of tabular data learning methodologies, such as the problem of error accumulation in existing semi-supervised learning.

[0275] In step (S1037), the processor (131) can generate tabular augmented data by performing automatic labeling on tabular unlabeled data using at least one artificial intelligence model.

[0276] At least the artificial intelligence model is a model that has learned user labeling patterns in step (S1033), and through this, can intelligently apply the complex and subtle tacit knowledge of experts to a large amount of unlabeled data.

[0277] In particular, this automatic labeling can be performed not simply by assigning fixed labels, but by estimating the statistical label distribution of user labeling patterns for unlabeled data and intelligently assigning it as a probabilistic label (Soft Label).

[0278] For example, if 70% of samples with similar features are labeled L1 and 30% are labeled L2 by a user, the processor (131) can also label samples with similar features to the user-labeled data samples in the same distribution for tabular unlabeled data.

[0279] That is, the processor (131) can generate high-quality tabular augmented data that preserves the ambiguity and potential information inherent in the data by probabilistically replicating the statistical ratio of the label distribution inherent in the learned user labeling pattern to the unlabeled data.

[0280] In step (S1039), the processor (131) can fine-tun at least one artificial intelligence model based on a training dataset including tabular labeling data (TLD) and tabular augmented data.

[0281] Step (S1039) is a process of completing a tabular data analysis model by finally utilizing high-quality data that has been quantitatively and qualitatively expanded through all processes from user pattern recognition to automatic labeling. The processor (131) can train the model with an expanded dataset that combines a small amount of tabular labeled data (TLD) with a large amount of tabular augmented data. Through this pre-training, at least one artificial intelligence model can be exposed to a large amount of data to dramatically improve its generalization performance, and learn the soft label ambiguity of the label distribution inherent in the tabular augmented data to prevent overfitting and secure stable and robust prediction capabilities suitable for actual service environments.

[0282] In step (S105), the processor (131) can input the analysis result information generated in step (S103) into at least one subsequent processing component.

[0283] This step (S105) may correspond to a core process of converting analysis result information into an automated action plan in an actual service environment. For example, the subsequent processing component may include a decision module (30), and this module (30) may perform intelligent decision-making based on the received analysis result information.

[0284] This decision-making can be implemented in various forms, such as automated recommendations (e.g., providing coupons to customers with a high probability of churn), risk assessment (e.g., loan approval / rejection decisions), and resource allocation (e.g., prioritizing personnel allocation to equipment with a high probability of failure).

[0285] Since this decision is based on the improved prediction reliability of the tabular data analysis model through learning user labeling patterns, the processor (131) can quickly and accurately determine a course of action based on analysis results with a low probability of error. This can serve as a basis for maximizing business efficiency and increasing the reliability of the service compared to existing systems.

[0286] Additionally, the subsequent processing component receiving the analysis result information may include a terminal such as a server or user computing device (110) separate from the server computing system (130). This structure ensures system flexibility to efficiently perform data analysis and final decision logic in a distributed environment, and provides the advantage of being able to expand and deploy modules to suit various service environments.

[0287] Additionally, the method (S100) may further include the step of at least one subsequent processing component manifesting the analysis result information through at least one user interface.

[0288] This enables users to intuitively check analysis results of tabular data and utilize them for final decision-making.

[0289] For example, the follow-up processing component can visually display predicted loan risk scores in a financial environment on a user interface (e.g., a mobile app or web dashboard) or graph equipment failure probabilities in a manufacturing environment. Through this, experts can quickly grasp the basis for the model's judgment, perform additional review (Interactive Decision-Making), or approve final actions, thereby maximizing the reliability and user accessibility of services utilizing various embodiments of the present disclosure.

[0290] Meanwhile, according to another embodiment, the method (S100) can build an efficient and reliable prediction model in an environment where labels for tabular data are insufficient by performing limited range augmentation for numerical variable values, and can perform high-accuracy predictions by analyzing tabular data using the same.

[0291] Existing supervised learning methods have limitations in that prediction performance is significantly degraded when there is insufficient label data for tabular data, but the method (S100) according to another embodiment solves this problem through self-supervised learning. By utilizing a large amount of unlabeled tabular data to allow the model to learn the intrinsic features and patterns of the data on its own, high accuracy can be achieved with only a small amount of labeled data.

[0292] The method (S100) may include two additional training steps. The first is a pre-training step, in which the encoder is trained based on data augmentation of the original tabular data. During this process, the model acquires the ability to identify semantic invariants of the data and extract rich features necessary for prediction. The second is a fine-tuning step, in which the prediction head connected to the pre-trained encoder is trained using a small amount of labeled data to provide final prediction capability.

[0293] A model that has acquired general knowledge of the data through prior training can quickly become a prediction expert for a specific task with only a small amount of labels during the fine-tuning phase. This offers an effective solution to the real-world problem where data labeling is costly and time-consuming, and consequently improves the model's generalization ability and reliability.

[0294] FIG. 17 is a flowchart of a fine-tuning method for an artificial intelligence model based on limited range augmentation for tabular data according to one embodiment. FIG. 18 is a flowchart for explaining steps that may be included in the step of training an artificial intelligence model to recognize patterns in the tabular data of FIG. 17.

[0295] Referring to FIG. 17, a method (S100) according to one embodiment may further include the step of generating additional tabular augmented data by performing limited range augmentation on numerical variable values ​​of tabular data (S201), the step of training at least one artificial intelligence model to recognize patterns of a training dataset and additional tabular augmented data (S203), and the step of fine-tuning at least one artificial intelligence model to output result values ​​for tabular data using tabular labeling data (S205).

[0296] In step (S201), the processor (131) can analyze the distribution of values ​​corresponding to each numerical feature included in the original tabular data and, based on this, perform limited range augmentation on the original tabular data to generate additional tabular augmented data.

[0297] For example, the processor (131) can sort the variable values ​​of the numeric variable 'age' in ascending order and then determine multiple ranges such that the number of variable values ​​included in each range is substantially equal. This enables the realization of the assumption of the invention that variable values ​​located close to each other in the data distribution will have high semantic similarity.

[0298] Subsequently, the processor (131) can generate tabular augmented data using the determined range. The processor (131) can change at least some of the multiple numerical variable values ​​included in the original tabular data to other values ​​within the range.

[0299] Shuffling or sampling methods can be used for modification. Shuffling is a method of rearranging existing variable values ​​within the same range by exchanging them with each other, while sampling is a method of changing a numeric variable value to any other numeric value within the range that includes the numeric variable value.

[0300] Through this process, it is possible to obtain augmented data that is semantically identical to the original data but has modified values, and this data can be used as positive pairs in self-supervised learning.

[0301] In step (S203), the processor (131) can train at least one artificial intelligence model (e.g., encoder) using the original tabular data and the tabular augmented data generated in step (S201).

[0302] Here, learning is based on self-supervised learning performed using only unlabeled data, and a contrastive learning loss function can be used.

[0303] This loss function can induce minimizing the distance between the embeddings of the original data (anchor) and the augmented data (positive), and maximizing the distance between the embeddings of the anchor and all other data (negative).

[0304] For example, referring to FIG. 18, step (S203) may include a step (S2031) of determining a positive pair of one of a plurality of numeric variable values ​​and a change value for said numeric variable value, a step (S2031) of determining a negative pair of at least one change value belonging to a different range from said numeric variable value and said same numeric variable value, and a step (S2035) of training an artificial intelligence model according to a loss function that causes the distance between positive pairs to become closer and the distance between a plurality of negative pairs to become farther apart.

[0305] The processor (131) can determine a positive pair of one of the multiple numeric variable values ​​included in the original tabular data and a change value for the said numeric variable value (S2031).

[0306] Additionally, the processor (131) can determine at least one change value belonging to a different range from any one numeric variable value included in the original tabular data and any one numeric variable value that is the same as the negative pair (S2033).

[0307] In this way, the processor (131) receives original tabular data and corresponding augmented tabular data, and by processing the two data according to the range to which the numerical variable values ​​included in each data belong, it can determine positive pairs that are considered semantically identical and negative pairs that are semantically unrelated.

[0308] The processor (131) can train the encoder according to a loss function (e.g., InfoNCE Loss) that causes the distance between positive pairs to become closer and the distance between multiple negative pairs to become farther apart (S2035).

[0309] In the process of minimizing the loss function described above, the distance between positive pairs can be minimized, while the distance between negative pairs can be maximized. Through this process, the encoder can learn on its own the rule that data with the same meaning should be close to each other, and data with different meanings should be far apart. As this learning is repeated, the encoder can more robustly grasp the inherent patterns of tabular data.

[0310] Through this process, the encoder can effectively identify the intrinsic characteristics of the data and acquire the ability to generate a robust representation of tabular data without relying on labeling.

[0311] In step (S205), the processor (131) can fine-tune at least one artificial intelligence model (e.g., a prediction model connected to a pre-trained encoder) to provide prediction capabilities.

[0312] The processor (131) can train the prediction model to output a correct prediction value using a small number of labeled tabular data. Here, the small number of labeled tabular data is stored in at least one memory and can be loaded by the processor (131).

[0313] In this case, the encoder parameters remain fixed, and only the parameters of the prediction head can be updated. Through this fine-tuning, the general knowledge of the pre-trained encoder is preserved, and the prediction head can be trained as a prediction expert for a specific task based on label data.

[0314] Various embodiments of the present disclosure described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the computer-readable recording medium may be those specifically designed and configured for the various embodiments of the present disclosure, or may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. Hardware devices may be modified into one or more software modules to perform processing according to the various embodiments of the present disclosure, and vice versa.

[0315] The specific embodiments described in this disclosure are exemplary and do not limit the scope of the various embodiments of this disclosure in any way. For the sake of brevity of the specification, descriptions of conventional electronic configurations, control systems, software, and other functional aspects of said systems may be omitted. Additionally, the connections of lines or connecting members between components shown in the drawings are exemplary representations of functional connections and / or physical or circuit connections, and may be replaced or additionally represented as various functional connections, physical connections, or circuit connections in actual devices. Furthermore, unless specifically stated as “essential,” “importantly,” etc., a component may not be strictly necessary for the application of the various embodiments of this disclosure.

[0316] Furthermore, although the detailed description of the present disclosure has been described with reference to preferred embodiments of the present disclosure, those skilled in the art or those with ordinary knowledge in the art will understand that various modifications and changes can be made to the various embodiments of the present disclosure without departing from the spirit and technical scope of the various embodiments of the present disclosure as set forth in the claims below. Accordingly, the technical scope of the various embodiments of the present disclosure should not be limited to the contents described in the detailed description of the specification but should be determined by the claims.

[0317] According to various embodiments of the present disclosure, it is possible to obtain large-scale tabular augmented data with the same reliability as the professional judgment criteria of users obtained based on the recognition of a small number of user labeling patterns, and by utilizing this, the prediction accuracy and generalization performance of tabular data analysis models can be dramatically improved, which makes it industrially applicable.

Claims

1. As a method performed by a computer, A step of receiving tabular data for analysis and storing it in at least one memory; A step in which at least one processor loads the tabular data for analysis from the at least one memory; A step in which at least one processor generates analysis result information for the above-mentioned tabular data for analysis using at least one artificial intelligence model; wherein the at least one artificial intelligence model is pre-trained based on a training dataset including tabular augmented data generated based on recognition of user labeling patterns, and A method comprising the step of inputting the above analysis result information into at least one subsequent processing component.

2. In Paragraph 1, A method further comprising the step of the above at least one subsequent processing component manifesting the analysis result information through at least one user interface.

3. In Paragraph 1, A method comprising at least one of classification information, regression prediction information, clustering information, or anomaly detection information for the tabular data for analysis.

4. In Paragraph 1, A method comprising further including the step of the subsequent processing component performing at least one decision-making based on the analysis result information.

5. In Paragraph 4, The above method, wherein at least one decision is made by including at least one of automated recommendation, risk assessment, resource allocation, and real-time control.

6. In Paragraph 1, The above-mentioned at least one artificial intelligence model is, A step of receiving tabular user labeling data containing user labeling information and storing it in at least one memory; A step of training at least one artificial intelligence model to recognize the user labeling pattern based on the above table-type labeling data; A step of receiving tabular non-labeled data and storing it in at least one memory; A step of generating tabular augmented data by performing automatic labeling on the tabular unlabeled data using the above at least one artificial intelligence model; and A method that is pre-trained by a learning method comprising the step of fine-tuning at least one artificial intelligence model based on the learning dataset including the tabular labeling data and the tabular augmentation data.

7. In Paragraph 6, The step of generating the tabular augmented data by performing the above automatic labeling is: A step of estimating the label distribution of the tabular user-labeled data having characteristics similar to the tabular non-labeled data; and A method comprising the step of performing automatic labeling on the tabular non-labeled data based on the estimated label distribution above.

8. In Paragraph 6, The above-mentioned at least one artificial intelligence model comprises at least one of a transformer model that learns the correlation between column information of tabular data and a decision tree algorithm-based model.

9. In Paragraph 1, A step in which at least one processor performs limited range augmentation on numeric variable values ​​of tabular data included in the training dataset to generate additional tabular augmented data; A step of training at least one artificial intelligence model to recognize patterns in the above-mentioned training dataset and the above-mentioned additional tabular augmented data; A step of loading tabular labeling data stored in at least one memory, and A method further comprising the step of fine-tuning at least one artificial intelligence model using the above tabular labeling data so that it outputs a result value for the tabular data.

10. In Paragraph 9, The step of generating the above additional tabular augmented data is, A step of determining a plurality of ranges for the plurality of numeric variable values ​​for each of the plurality of numeric variable values ​​based on the distribution of the plurality of numeric variable values ​​corresponding to each of the plurality of numeric variable values ​​included in the above training dataset; and A method comprising the step of generating additional tabular augmented data by changing each of at least some of the values ​​of a plurality of numeric variables included in the training dataset to another value within the corresponding range.

11. In Paragraph 10, The step of generating the above additional tabular augmented data is, A step of exchanging at least some of the above-mentioned plurality of numeric variable values ​​with other some of the numeric variable values ​​within the same range as the corresponding range; and A method comprising at least one of the step of changing at least some of the above-mentioned numeric variable values ​​to arbitrary values ​​within a corresponding range.

12. In Paragraph 9, The method further includes the step of obtaining segmented data by performing segmentation on the above training dataset and the above additional tabular augmented data; The step of training at least one artificial intelligence model is, A method comprising the step of training at least one artificial intelligence model to output segmented data by taking the training dataset and the additional tabular augmented data as inputs.

13. At least one memory; and At least one processor that reads at least one instruction stored in the above at least one memory and executes a method for providing a table-type data analysis-based service; comprising The above at least one instruction is, A step of receiving tabular data for analysis and storing it in at least one memory; The step of the above at least one processor loading the analysis table-type data from the above at least one memory; The step of generating analysis result information for the tabular data for analysis using at least one artificial intelligence model by the at least one processor; wherein the at least one artificial intelligence model is pre-trained based on a training dataset including tabular augmented data generated based on recognition of user labeling patterns, and A system comprising a command that performs the step of ingesting the above analysis result information into at least one subsequent processing component.

14. In Paragraph 13, A plurality of neurons comprising an array including at least one register, at least one programmable logic, and at least one input interface; a plurality of synapse circuits storing synapse weights that regulate the connection strength between the plurality of neurons; and at least one routing network that controls the data flow between the plurality of neurons; comprising A system comprising a Field Programmable Gate Array (FPGA) implementation for a predetermined artificial neural network, wherein each of the plurality of neurons is connected to at least one other neuron through the routing network to establish a transmission path for the weights.

15. In Paragraph 13, A plurality of neurons organized into an array comprising at least one register, at least one microprocessor, and at least one input; and a plurality of synapse circuits storing synapse weights that regulate the connection strength between the plurality of neurons; comprising A system comprising an Application Specific Integrated Circuit (ASIC) for a predetermined artificial neural network, wherein each of the plurality of neurons is connected to at least one other neuron through any one of the plurality of synaptic circuits.

Images